* [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver
@ 2025-09-26 3:34 Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 01/23] crypto: iaa - Reorganize the iaa_crypto driver code Kanchana P Sridhar
` (23 more replies)
0 siblings, 24 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
v12: zswap compression batching with optimized iaa_crypto driver
Based on Herbert's suggestions, the most significant change in v12 is, the
batching interfaces from zswap to crypto, from crypto to iaa_crypto, and
the batching implementation within iaa_crypto now use the folio directly as the
source (sg_page_iter for retrieving pages) and destination SG lists. A unit_size
has been added to struct acomp_req, with kernel users such as zswap using the
new acomp_request_set_unit_size() API to set the unit size to use while breaking
down the request's src/dst scatterlists. zswap sets the unit-size to PAGE_SIZE.
Following Andrew's suggestion, the next two paragraphs emphasize generality and
alignment with current kernel efforts.
Architectural considerations for the zswap batching framework:
==============================================================
We have designed the zswap batching framework to be
hardware-agnostic. It has no dependencies on Intel-specific features and
can be leveraged by any hardware accelerator or software-based
compressor. In other words, the framework is open and inclusive by
design.
Other ongoing work that can use batching:
=========================================
This patch-series demonstrates the performance benefits of compress
batching when used in zswap_store() of large folios. shrink_folio_list()
"reclaim batching" of any-order folios is the next major work that uses
this zswap compress batching framework: our testing of kernel_compilation
with writeback and the zswap shrinker indicates 10X fewer pages get
written back when we reclaim 32 folios as a batch, as compared to one
folio at a time: this is with deflate-iaa and with zstd. We expect to
submit a patch-series with this data and the resulting performance
improvements shortly. Reclaim batching relieves memory pressure faster
than reclaiming one folio at a time, hence alleviates the need to scan
slab memory for writeback.
Many thanks to Nhat for suggesting ideas on using batching with the
ongoing kcompressd work, as well as beneficially using decompression
batching & block IO batching to improve zswap writeback efficiency.
Experiments with kernel compilation benchmark (allmod config) that
combine zswap compress batching, reclaim batching, swapin_readahead()
decompression batching of prefetched pages, and writeback batching show
that 0 pages are written back to disk with deflate-iaa and zstd. For
comparison, the baselines for these compressors see 200K-800K pages
written to disk.
To summarize, these are future clients of the batching framework:
- shrink_folio_list() reclaim batching of multiple folios:
Implemented, will submit patch-series.
- zswap writeback with decompress batching:
Implemented, will submit patch-series.
- zram:
Implemented, will submit patch-series.
- kcompressd:
Not yet implemented.
- file systems:
Not yet implemented.
- swapin_readahead() decompression batching of prefetched pages:
Implemented, will submit patch-series.
iaa_crypto Driver Rearchitecting and Optimizations:
===================================================
The most significant highlight of v11 is a new, lightweight and highly
optimized iaa_crypto driver, resulting directly in the latency and
throughput improvements noted later in this cover letter.
1) Better stability, more functionally versatile to support zswap and
zram with better performance on different Intel platforms.
a) Patches 0002, 0005 and 0010 together resolve a race condition in
mainline v6.15, reported from internal validation, when IAA
wqs/devices are disabled while workloads are using IAA.
b) Patch 0002 introduces a new architecture for mapping cores to
IAAs based on packages instead of NUMA nodes, and generalizing
how WQs are used: as package level shared resources for all
same-package cores (default for compress WQs), or dedicated to
mapped cores (default for decompress WQs). Further, users are
able to configure multiple WQs and specify how many of those are
for compress jobs only vs. decompress jobs only. sysfs iaa_crypto
driver parameters can be used to change the default settings for
performance tuning.
c) idxd descriptor allocation moved from blocking to non-blocking
with retry limits and mitigations if limits are exceeded.
d) Code cleanup for readability and clearer code flow.
e) Fixes IAA re-registration errors upon disabling/enabling IAA wqs
and devices that exists in the mainline v6.15.
f) Rearchitecting iaa_crypto to be independent of crypto_acomp to
enable a zram/zcomp backend_deflate_iaa.c, while fully supporting
the crypto_acomp interfaces for zswap. A new
include/linux/iaa_comp.h is added.
g) New Dynamic compression mode for Granite Rapids to get better
compression ratio by echo-ing 'deflate-iaa-dynamic' as the zswap
compressor.
h) New crypto_acomp API crypto_acomp_batch_size() that will return
the driver's max batch size if the driver has registered a batch_size
that's greater than 1; or 1 if there is no driver specific definition of
batch_size.
Accordingly, iaa_crypto sets the acomp_alg batch_size to its internal
IAA_CRYPTO_MAX_BATCH_SIZE for fixed and dynamic modes.
i) A versatile set of interfaces independent of crypto_acomp for use
in developing a zram zcomp backend for iaa_crypto.
2) Performance optimizations (please refer to the latency data per
optimization in the commit logs):
a) Distributing [de]compress jobs in round-robin manner to available
IAAs on package.
b) Replacing the compute-intensive iaa_wq_get()/iaa_wq_put() with a
percpu_ref in struct iaa_wq, thereby eliminating acquiring a
spinlock in the fast path, while using a combination of the
iaa_crypto_enabled atomic with spinlocks in the slow path to
ensure the compress/decompress code sees a consistent state of the
wq tables.
c) Directly call movdir64b for non-irq use cases, i.e., the most
common usage. Avoid the overhead of irq-specific computes in
idxd_submit_desc() to gain latency.
d) Batching of compressions/decompressions using async submit-poll
mechanism to derive the benefits of hardware parallelism.
e) Batching compressors need to manage their own "requests"
abstraction, and remove this driver-specific aspect from being
managed by kernel users such as zswap. iaa_crypto maintains
per-CPU "struct iaa_req **reqs" to submit multiple jobs to the
hardware accelerator to run in parallel.
f) Modifies the iaa_crypto batching API and their implementation to expect a
src SG list that contains the batch's pages and a dst SG list that has
multiple scatterlists for the batch's output buffers.
g) Submit the two largest data buffers first for decompression
batching, so that the longest running jobs get a head start,
reducing latency for the batch.
Main Changes in Zswap Compression Batching:
===========================================
Note to zswap maintainers:
--------------------------
Patches 20 and 21 can be reviewed and improved/merged independently
of this series, since they are zswap centric. These 2 patches help
batching but the crypto_acomp_batch_size() from the iaa_crypto commits
in this series is not a requirement, unlike patches 22-23.
1) v12 preserves the pool acomp_ctx resources creation/deletion
simplification of v11, namely, lasting from pool creation-deletion,
persisting through CPU hot[un]plug operations. Further, zswap no
longer needs to create multiple "struct acomp_req" in the per-CPU
acomp_ctx. zswap only needs to manage multiple "u8 **buffers".
2) We store the compressor's batch-size (@pool->compr_batch_size) and
the batch-size to use during compression batching
(@pool->store_batch_size) directly in struct zswap_pool for quick
retrieval in the zswap_store() fast path.
3) Optimizations to not cause regressions in software compressors with
the introduction of the new unified zswap_compress() framework that
implements compression batching for all compressors. These optimizations
help recover the performance for non-batching compressors:
a) kmem_cache_alloc_bulk(), kmem_cache_free_bulk() to allocate/free
batch zswap_entry-s. These kmem_cache API allow allocator
optimizations with internal locks for multiple allocations.
b) Writes to the zswap_entry right after it is allocated without
modifying the publishing order. This avoids different code blocks
in zswap_store_pages() having to bring the zswap_entries to the
cache for writing, potentially evicting other working set
structures, impacting performance.
c) ZSWAP_MAX_BATCH_SIZE is used as the batch-size for software
compressors, since this gives the best performance with zstd.
d) More likely()/unlikely() annotations to try and minimize branch
mis-predicts.
4) During pool creation, these key additions are allocated as part of the
per-CPU acomp_ctx so as to recover performance with the new, generalized SG
lists based zswap_compress() batching interface:
a) An sg_table "acomp_ctx->sg_outputs" is allocated to contain the
compressor's batch-size number of SG lists that will contain the
destination buffers/lengths after batch compression.
b) The per-CPU destination buffers are mapped to the per-CPU SG lists: this
needs to be done only once, and optimizes performance.
c) Only for batching compressors, extra memory is allocated in the
acomp_req, so as to avail of struct acomp_req's "void *__ctx[]" to store
pool->compr_batch_size number of int pointers to the
acomp_ctx->sg_outputs-sgl's component scatterlists. This is also added
to optimize/recover performance by avoiding traversing the sg_outputs per
batch compress call. Without this optimization, it is difficult to
realize the benefits of batching within large folios.
5) Finally, zswap_compress() has been re-written to incorporate Herbert's
suggestions to use source folios and output SG lists for batching. The new
zswap_compress() code has been made as generic to software and batching
compressors as possible, so that it is easy to read and maintain. The
recent changes related to PAGE_SIZE dst buffers, zsmalloc and incompressible
pages have been incorporated into the batched zswap_compress() as well! To
resolve regressions with zstd, I took the liberty of not explicitly checking
for dlen == 0 and dlen > PAGE_SIZE (as is assumed in the mainline); instead,
expecting that a negative err value will be returned by the software
compressor in such cases.
Compression Batching:
=====================
This patch-series introduces batch compression of pages in large folios to
improve zswap swapout latency. It preserves the existing zswap protocols
for non-batching software compressors by calling crypto_acomp sequentially
per page in the batch. Additionally, in support of hardware accelerators
that can process a batch as an integral unit, the patch-series allows
zswap to call crypto_acomp without API changes, for compressors
that intrinsically support batching. The zswap_compress() code has very minimal
special casing for software/batching compressors.
The patch series provides a proof point by using the Intel Analytics
Accelerator (IAA) for implementing the compress/decompress batching API
using hardware parallelism in the iaa_crypto driver and another proof point
with a sequential software compressor, zstd.
SUMMARY:
========
The first proof point is to test with IAA using a sequential call (fully
synchronous, compress one page at a time) vs. a batching call (fully
asynchronous, submit a batch to IAA for parallel compression, then poll for
completion statuses).
The performance testing data with 30 usemem processes/64K folios
shows 49% throughput gains and 25% elapsed/sys time reductions with
deflate-iaa; and 1% sys time reduction with zstd for a small
throughput increase. For PMD folios, a 55% throughput gain and 15%
elapsed/sys time reduction is seen.
Kernel compilation test with 64K folios using 32 threads and the
zswap shrinker_enabled set to "N", demonstrates similar
improvements: zswap_store() large folios using IAA compress batching
improves the workload performance by 6.1% and reduces sys time by
1% as compared to IAA sequential. For zstd, compress batching
improves workload performance by 7.8% and reduces sys time by
2.8% as compared to sequentially calling zswap_compress() per page
in a folio.
The main takeaway from usemem, a workload that is mostly compression
dominated (very few swapins) is that the higher the number of batches,
such as, with larger folios, the more the benefit of batching cost
amortization, as shown by the PMD usemem data. This aligns well with the
future direction for batching.
The second proof point is to make sure that software algorithms such as
zstd do not regress. The data indicates that for sequential software
algorithms a performance gain is achieved.
With the performance optimizations implemented in patches 22-23 of v12:
* zstd usemem metrics with 64K folios are within range of variation
with a slight sys time improvement. zstd usemem30 throughput with PMD
folios increases by 1%, workload performance improves by 6% and sys time
reduces by 8%.
* With kernel compilation, I used zstd without the zswap shrinker to enable
more direct comparisons with the changes in this series. Subsequent patch
series I expect to submit in collaboration with Nhat, will enable the
zswap shrinker to quantify the benefits of decompression batching during
writeback. With this series' compression batching within large folios, we
get a 1%-2.8% reduction in sys time, a 6.1%-7.8% improvement in workload
performance with 64K folios for deflate-iaa/zstd respectively.
These optimizations pertain to ensuring common code paths, removing
redundant branches/computes, selectively annotating branches with
likely()/unlikely() compiler directives to minimize branch
mis-prediction penalty. Additionally, using the batching code for
non-batching compressors to sequentially compress/store batches of up
to ZSWAP_MAX_BATCH_SIZE pages seems to help, most likely due to
cache locality of working set structures such as the array of
zswap_entry-s for the batch.
Our internal validation of zstd with the batching interface vs. IAA with
the batching interface on Emerald Rapids has shown that IAA
compress/decompress batching gives 21.3% more memory savings as compared
to zstd, for 5% performance loss as compared to the baseline without any
memory pressure. IAA batching demonstrates more than 2X the memory
savings obtained by zstd at this 95% performance KPI.
The compression ratio with IAA is 2.23, and with zstd 2.96. Even with
this compression ratio deficit for IAA, batching is extremely
beneficial. As we improve the compression ratio of the IAA accelerator,
we expect to see even better memory savings with IAA as compared to
software compressors.
Batching Roadmap:
=================
1) Compression batching within large folios (this series).
2) zswap writeback decompression batching:
This is being co-developed with Nhat Pham, and shows promising
results. We plan to submit an RFC shortly.
3) Reclaim batching of hybrid folios:
We can expect to see even more significant performance and throughput
improvements if we use the parallelism offered by IAA to do reclaim
batching of 4K/large folios (really any-order folios), and using the
zswap_store() high throughput compression pipeline to batch-compress
pages comprising these folios, not just batching within large
folios. This is the reclaim batching patch 13 in v1, which we expect
to submit in a separate patch-series. As mentioned earlier, reclaim
batching reduces the # of writeback pages by 10X for zstd and
deflate-iaa.
4) swapin_readahead() decompression batching:
We have developed a zswap load batching interface to be used
for parallel decompression batching, using swapin_readahead().
These capabilities are architected so as to be useful to zswap and
zram. We have integrated these components with zram and expect to submit an
RFC soon.
v12 Performance Summary:
========================
This is a performance testing summary of results with usemem30
(30 usemem processes running in a cgroup limited at 150G, each trying to
allocate 10G).
usemem30 with 64K folios:
=========================
zswap shrinker_enabled = N.
-----------------------------------------------------------------------
mm-unstable-9-18-2025 v12
-----------------------------------------------------------------------
zswap compressor deflate-iaa deflate-iaa IAA Batching
vs.
IAA Sequential
-----------------------------------------------------------------------
Total throughput (KB/s) 7,191,949 10,702,115 49%
Average throughput (KB/s) 239,731 356,737 49%
elapsed time (sec) 93.21 69.98 -25%
sys time (sec) 2,190.52 1,651.51 -25%
-----------------------------------------------------------------------
-----------------------------------------------------------------------
mm-unstable-9-18-2025 v12
-----------------------------------------------------------------------
zswap compressor zstd zstd v12 zstd
improvement
-----------------------------------------------------------------------
Total throughput (KB/s) 6,258,312 6,269,511 0.2%
Average throughput (KB/s) 208,610 208,983 0.2%
elapsed time (sec) 100.01 100.50 0.5%
sys time (sec) 2,505.14 2,490.00 -1%
-----------------------------------------------------------------------
usemem30 with 2M folios:
========================
-----------------------------------------------------------------------
mm-unstable-9-18-2025 v12
-----------------------------------------------------------------------
zswap compressor deflate-iaa deflate-iaa IAA Batching
vs.
IAA Sequential
-----------------------------------------------------------------------
Total throughput (KB/s) 7,237,676 11,228,928 55%
Average throughput (KB/s) 241,255 374,297 55%
elapsed time (sec) 82.26 69.30 -16%
sys time (sec) 1,901.90 1,634.78 -14%
-----------------------------------------------------------------------
-----------------------------------------------------------------------
mm-unstable-9-18-2025 v12
-----------------------------------------------------------------------
zswap compressor zstd zstd v12 zstd
improvement
-----------------------------------------------------------------------
Total throughput (KB/s) 6,796,376 6,865,781 1.0%
Average throughput (KB/s) 226,545 228,859 1.0%
elapsed time (sec) 94.07 88.80 -6%
sys time (sec) 2,261.67 2,082.91 -8%
-----------------------------------------------------------------------
This is a performance testing summary of results with
kernel_compilation test (allmod config, 32 cores, cgroup limited to 2G).
zswap shrinker_enabled = N.
kernel_compilation with 64K folios:
===================================
--------------------------------------------------------------------------
mm-unstable-9-18-2025 v12
--------------------------------------------------------------------------
zswap compressor deflate-iaa deflate-iaa IAA Batching
vs.
IAA Sequential
--------------------------------------------------------------------------
real_sec 874.74 821.59 -6.1%
sys_sec 3,834.35 3,791.12 -1%
--------------------------------------------------------------------------
--------------------------------------------------------------------------
mm-unstable-9-18-2025 v12
--------------------------------------------------------------------------
zswap compressor zstd zstd Improvement
--------------------------------------------------------------------------
real_sec 925.08 853.14 -7.8%
sys_sec 5,318.65 5,172.23 -2.8%
--------------------------------------------------------------------------
kernel_compilation with PMD folios:
===================================
--------------------------------------------------------------------------
mm-unstable-9-18-2025 v12
--------------------------------------------------------------------------
zswap compressor deflate-iaa deflate-iaa IAA Batching
vs.
IAA Sequential
--------------------------------------------------------------------------
real_sec 808.10 794.85 -1.6%
sys_sec 4,351.01 4,266.95 -2%
--------------------------------------------------------------------------
--------------------------------------------------------------------------
mm-unstable-9-18-2025 v12
--------------------------------------------------------------------------
zswap compressor zstd zstd Improvement
--------------------------------------------------------------------------
real_sec 848.06 845.42 -0.3%
sys_sec 5,898.58 5,741.31 -2.7%
--------------------------------------------------------------------------
DETAILS:
========
(A) From zswap's perspective, the most significant changes are:
===============================================================
1) A unified zswap_compress() API is added to compress multiple
pages:
- If the compressor has multiple acomp requests, i.e., internally
supports batching, crypto_acomp_batch_compress() is called. If all
pages are successfully compressed, or, if folio writeback is enabled, the
batch is stored in zsmalloc, preserving latest mm-stable behavior.
- If the compressor can only compress one page at a time, each page
is compressed and stored sequentially.
Many thanks to Yosry for this suggestion, because it is an essential
component of unifying common code paths between sequential/batching
compressions.
Overall, the likely()/unlikely() annotations prevent regressions with
software compressors like zstd, and generally improve non-batching
compressors' performance with the batching code by 1% to 8%.
2) A new zswap_store_pages() is added, that stores multiple pages in a
folio in a range of indices. This is an extension of the earlier
zswap_store_page(), except it operates on a batch of pages.
3) zswap_store() is modified to store the folio's pages in batches
by calling zswap_store_pages(). If the compressor supports batching,
the folio will be compressed in batches of
"pool->compr_batch_size". If the compressor does not support
batching, the folio will be compressed in batches of
ZSWAP_MAX_BATCH_SIZE pages, where each page in the batch is
compressed sequentially. We see better performance by processing
the folio in batches of ZSWAP_MAX_BATCH_SIZE, due to cache locality
of working set structures such as the array of zswap_entry-s for the
batch.
Many thanks to Yosry and Johannes for steering towards a common
design and code paths for sequential and batched compressions (i.e.,
for software compressors and hardware accelerators such as IAA). As per
Yosry's suggestion in v8, the "batch_size" is an attribute of the
compressor/pool, and hence is stored in struct zswap_pool instead of
in struct crypto_acomp_ctx.
4) Simplifications to the acomp_ctx resources allocation/deletion
vis-a-vis CPU hot[un]plug. This further improves upon v8 of this
patch-series based on the discussion with Yosry, and formalizes the
lifetime of these resources from pool creation to pool
deletion. zswap does not register a CPU hotplug teardown
callback. The acomp_ctx resources will persist through CPU
online/offline transitions. The main changes made to avoid UAF/race
conditions, and correctly handle process migration, are:
a) No acomp_ctx mutex locking in zswap_cpu_comp_prepare().
b) No CPU hotplug teardown callback, no acomp_ctx resources deleted.
c) New acomp_ctx_dealloc() procedure that cleans up the acomp_ctx
resources, and is shared by
zswap_cpu_comp_prepare()/zswap_pool_create() error handling and
zswap_pool_destroy().
d) The zswap_pool node list instance is removed right after the node
list add function in zswap_pool_create().
e) We directly call mutex_[un]lock(&acomp_ctx->mutex) in
zswap_[de]compress(). acomp_ctx_get_cpu_lock()/acomp_ctx_put_unlock()
are deleted.
The commit log of patch 0020 has a more detailed analysis.
(B) Main changes in crypto_acomp and iaa_crypto:
================================================
1) A new architecture is introduced for IAA device WQs' usage as:
- compress only
- decompress only
- generic, i.e., both compress/decompress.
Further, IAA devices/wqs are assigned to cores based on packages
instead of NUMA nodes.
The WQ rebalancing algorithm that is invoked as WQs are
discovered/deleted has been made very general and flexible so that
the user can control exactly how IAA WQs are used. In addition to the
user being able to specify a WQ type as comp/decomp/generic, the user
can also configure if WQs need to be shared among all same-package
cores, or, whether the cores should be divided up amongst the
available IAA devices.
If distribute_[de]comps is enabled, from a given core's perspective,
the iaa_crypto driver will distribute comp/decomp jobs among all
devices' WQs in round-robin manner. This improves batching latency
and can improve compression/decompression throughput for workloads
that see a lot of swap activity.
The commit log of patch 0002 provides more details on new iaa_crypto
driver parameters added, along with recommended settings (defaults
are optimal settings).
2) Compress/decompress batching are implemented using
crypto_acomp_[de]compress() with SG lists used as the batching interface.
(C) The patch-series is organized as follows:
=============================================
1) crypto acomp & iaa_crypto driver enablers for batching: Relevant
patches are tagged with "crypto:" in the subject:
Patch 1) Reorganizes the iaa_crypto driver code into logically related
sections and avoids forward declarations, in order to facilitate
subsequent iaa_crypto patches. This patch makes no
functional changes.
Patch 2) Makes an infrastructure change in the iaa_crypto driver
to map IAA devices/work-queues to cores based on packages
instead of NUMA nodes. This doesn't impact performance on
the Sapphire Rapids system used for performance
testing. However, this change fixes functional problems we
found on Granite Rapids during internal validation, where the
number of NUMA nodes is greater than the number of packages,
which was resulting in over-utilization of some IAA devices
and non-usage of other IAA devices as per the current NUMA
based mapping infrastructure.
This patch also develops a new architecture that
generalizes how IAA device WQs are used. It enables
designating IAA device WQs as either compress-only or
decompress-only or generic. Once IAA device WQ types are
thus defined, it also allows the configuration of whether
device WQs will be shared by all cores on the package, or
used only by "mapped cores" obtained by a simple allocation
of available IAAs to cores on the package.
As a result of the overhaul of wq_table definition,
allocation and rebalancing, this patch eliminates
duplication of device WQs in per-CPU wq_tables, thereby
saving 140MiB on a 384 cores dual socket Granite Rapids server
with 8 IAAs.
Regardless of how the user has configured the WQs' usage,
the next WQ to use is obtained through a direct look-up in
per-CPU "cpu_comp_wqs" and "cpu_decomp_wqs" structures so
as to minimize latency in the critical path driver compress
and decompress routines.
Patch 3) Code cleanup, consistency of function parameters.
Patch 4) Makes a change to iaa_crypto driver's descriptor allocation,
from blocking to non-blocking with retries/timeouts and
mitigations in case of timeouts during compress/decompress
ops. This prevents tasks getting blocked indefinitely, which
was observed when testing 30 cores running workloads, with
only 1 IAA enabled on Sapphire Rapids (out of 4). These
timeouts are typically only encountered, and associated
mitigations exercised, only in configurations with 1 IAA
device shared by 30+ cores.
Patch 5) Optimize iaa_wq refcounts using a percpu_ref instead of
spinlocks and "int refcount".
Patch 6) Code simplification and restructuring for understandability
in core iaa_compress() and iaa_decompress() routines.
Patch 7) Refactor hardware descriptor setup to their own procedures
to reduce code clutter.
Patch 8) Simplify and optimize (i.e. reduce computes) job submission
for the most commonly used non-irq async mode by directly
calling movdir64b.
Patch 9) Deprecate exporting symbols for adding IAA compression
modes.
Patch 10) All dma_map_sg() calls will pass in 1 for the nents instead of
sg_nents(), for these main reasons: performance; no existing
iaa_crypto use cases that allow multiple SG lists to be mapped for
dma at once; facilitates new SG lists batching interface through
crypto.
Patch 11) Rearchitect iaa_crypto to be agnostic of crypto_acomp for
it be usable in both zswap and zram. crypto_acomp interfaces are
maintained as earlier, for use in zswap.
Patch 12) Descriptor submit and polling mechanisms, enablers for batching.
Patch 13) Define a unit_size in struct acomp_req to enable batching, and
provides acomp_request_set_unit_size() for use by kernel modules.
Patch 14) Implement IAA batching of compressions and decompressions
for deriving hardware parallelism.
Patch 15) Enables the "async" mode, sets it as the default.
Patch 16) Disables verify_compress by default.
Patch 17) Decompress batching optimization: Find the two largest
buffers in the batch and submit them first.
Patch 18) Add a new Dynamic compression mode that can be used on
Granite Rapids.
Patch 19) Add a batch_size data member to structs acomp_alg/crypto_acomp and
a crypto_acomp_batch_size() API that returns the compressor's
batch-size, if it has defined one; 1 otherwise. Add the definition
of batch_size to iaa_crypto acomp_algs.
2) zswap modifications to enable compress batching in zswap_store()
of large folios (including pmd-mappable folios):
Patch 20) Simplifies the zswap_pool's per-CPU acomp_ctx resource
management and lifetime to be from pool creation to pool
deletion.
Patch 21) Uses IS_ERR_OR_NULL() in zswap_cpu_comp_prepare() to check for
valid acomp/req, thereby making it consistent with the resource
de-allocation code.
Patch 22) Defines a zswap-specific ZSWAP_MAX_BATCH_SIZE (currently set
as 8U) to denote the maximum number of acomp_ctx batching
resources to allocate, thus limiting the amount of extra
memory used for batching. Further, the "struct
crypto_acomp_ctx" is modified to contain multiple buffers.
New "u8 compr_batch_size" and "u8 store_batch_size" members are
added to "struct zswap_pool" to track the number of dst
buffers associated with the compressor (more than 1 if
the compressor supports batching) and the unit for storing
large folios using compression batching respectively.
Modifies zswap_store() to store the folio in batches of
pool->store_batch_size by calling a new zswap_store_pages() that
takes a range of indices in the folio to be stored.
zswap_store_pages() pre-allocates zswap entries for the batch,
calls zswap_compress() for each page in this range, and stores
the entries in xarray/LRU.
Patch 23) Introduces a new unified batching implementation of
zswap_compress() for compressors that do and do not support
batching. This eliminates code duplication and facilitates
maintainability of the code with the introduction of compress
batching. Further, there are many optimizations to this common
code that result in workload throughput and performance
improvements with software compressors and hardware accelerators
such as IAA.
zstd performance is better or on par with mm-unstable. We
see impressive throughput/performance improvements with
IAA and zstd batching vs. no-batching.
With v12 of this patch series, the IAA compress batching feature will be
enabled seamlessly on Intel platforms that have IAA by selecting
'deflate-iaa' as the zswap compressor, and using the iaa_crypto 'async'
sync_mode driver attribute (the default).
System setup for testing:
=========================
Testing of this patch-series was done with mm-unstable as of 9-18-2025,
commit 1f98191f08b4, without and with this patch-series. Data was
gathered on an Intel Sapphire Rapids (SPR) server, dual-socket 56 cores
per socket, 4 IAA devices per socket, each IAA has total 128 WQ entries,
503 GiB RAM and 525G SSD disk partition swap. Core frequency was fixed
at 2500MHz.
Other kernel configuration parameters:
zswap compressor : zstd, deflate-iaa
zswap allocator : zsmalloc
vm.page-cluster : 0
IAA "compression verification" is disabled and IAA is run in the async
mode (the defaults with this series).
I ran experiments with these workloads:
1) usemem 30 processes with zswap shrinker_enabled=N. Two sets of
experiments, one with 64K folios, another with PMD folios.
2) Kernel compilation allmodconfig with 2G max memory, 32 threads, with
zswap shrinker_enabled=N to test batching performance impact in
isolation. Two sets of experiments, one with 64K folios, another with PMD
folios.
IAA configuration is done by a CLI: script is included at the end of the
cover letter.
Performance testing (usemem30):
===============================
The vm-scalability "usemem" test was run in a cgroup whose memory.high
was fixed at 150G. There is no swap limit set for the cgroup. 30 usemem
processes were run, each allocating and writing 10G of memory, and
sleeping for 10 sec before exiting:
usemem --init-time -w -O -b 1 -s 10 -n 30 10g
echo 0 > /sys/module/zswap/parameters/shrinker_enabled
IAA WQ Configuration (script is iincluded at the end of the cover
letter):
./enable_iaa.sh -d 4 -q 1
This enables all 4 IAAs on the socket, and configures 1 WQ per IAA
device, each containing 128 entries. The driver distributes compress
jobs from each core to wqX.0 of all same-package IAAs in a
round-robin manner. Decompress jobs are send to the wqX.0 of the
mapped IAA device.
Since usemem has significantly more swapouts than swapins, this
configuration is the most optimal.
64K folios: usemem30: deflate-iaa:
==================================
-------------------------------------------------------------------------------
mm-unstable-9-18-2025 v12
-------------------------------------------------------------------------------
zswap compressor deflate-iaa deflate-iaa IAA Batching
vs.
IAA Sequential
-------------------------------------------------------------------------------
Total throughput (KB/s) 7,191,949 10,702,115 49%
Avg throughput (KB/s) 239,731 356,737 49%
elapsed time (sec) 93.21 69.98 -25%
sys time (sec) 2,190.52 1,651.51 -25%
-------------------------------------------------------------------------------
memcg_high 1,056,505 1,320,579
memcg_swap_fail 3,966 321
64kB_swpout_fallback 3,965 321
zswpout 61,577,440 71,448,317
zswpin 488 141
pswpout 0 0
pswpin 0 0
ZSWPOUT-64kB 3,844,593 4,465,129
SWPOUT-64kB 0 0
pgmajfault 2,786 2,454
zswap_reject_compress_fail 0 0
zswap_reject_reclaim_fail 0 0
IAA incompressible pages 0 0
-------------------------------------------------------------------------------
2M folios: usemem30: deflate-iaa:
=================================
-------------------------------------------------------------------------------
mm-unstable-9-18-2025 v12
-------------------------------------------------------------------------------
zswap compressor deflate-iaa deflate-iaa IAA Batching
vs.
IAA Sequential
-------------------------------------------------------------------------------
Total throughput (KB/s) 7,237,676 11,228,928 55%
Avg throughput (KB/s) 241,255 374,297 55%
elapsed time (sec) 82.26 69.30 -16%
sys time (sec) 1,901.90 1,634.78 -14%
-------------------------------------------------------------------------------
memcg_high 100,141 119,644
memcg_swap_fail 356 12
thp_swpout_fallback 356 12
zswpout 63,470,927 71,773,688
zswpin 455 169
pswpout 0 0
pswpin 0 0
ZSWPOUT-2048kB 123,610 140,169
thp_swpout 0 0
pgmajfault 2,743 2,471
zswap_reject_compress_fail 0 0
zswap_reject_reclaim_fail 0 0
IAA incompressible pages 0 0
-------------------------------------------------------------------------------
64K folios: usemem30: zstd:
===========================
-------------------------------------------------------------------------------
mm-unstable-9-18-2025 v12
-------------------------------------------------------------------------------
zswap compressor zstd zstd v12 zstd
improvement
-------------------------------------------------------------------------------
Total throughput (KB/s) 6,258,312 6,269,511 0.2%
Avg throughput (KB/s) 208,610 208,983 0.2%
elapsed time (sec) 100.01 100.50 0.5%
sys time (sec) 2,505.14 2,490.00 -1%
-------------------------------------------------------------------------------
memcg_high 1,071,250 1,043,886
memcg_swap_fail 2,251 73
64kB_swpout_fallback 2,251 73
zswpout 49,491,137 49,693,220
zswpin 455 179
pswpout 0 0
pswpin 0 0
ZSWPOUT-64kB 3,090,904 3,105,690
SWPOUT-64kB 0 0
pgmajfault 2,780 2,456
zswap_reject_compress_fail 0 0
zswap_reject_reclaim_fail 0 0
-------------------------------------------------------------------------------
2M folios: usemem30: zstd:
==========================
-------------------------------------------------------------------------------
mm-unstable-9-18-2025 v12
-------------------------------------------------------------------------------
zswap compressor zstd zstd v12 zstd
improvement
-------------------------------------------------------------------------------
Total throughput (KB/s) 6,796,376 6,865,781 1.0%
Avg throughput (KB/s) 226,545 228,859 1.0%
elapsed time (sec) 94.07 88.80 -6%
sys time (sec) 2,261.67 2,082.91 -8%
-------------------------------------------------------------------------------
memcg_high 89,582 88,103
memcg_swap_fail 14 56
thp_swpout_fallback 14 56
zswpout 48,846,660 48,000,106
zswpin 58 266
pswpout 0 0
pswpin 0 0
ZSWPOUT-2048kB 95,388 93,692
thp_swpout 0 0
pgmajfault 2,348 2,557
zswap_reject_compress_fail 0 0
zswap_reject_reclaim_fail 0 0
-------------------------------------------------------------------------------
Performance testing (Kernel compilation, allmodconfig):
=======================================================
The experiments with kernel compilation test use 32 threads and build
the "allmodconfig" that takes ~14 minutes, and has considerable
swapout/swapin activity. The cgroup's memory.max is set to 2G. zswap
writeback is not enabled so as to isolate the performance impact of only large
folio batch compression.
echo 0 > /sys/module/zswap/parameters/shrinker_enabled
IAA WQ Configuration (script is at the end of the cover letter):
./enable_iaa.sh -d 4 -q 2
This enables all 4 IAAs on the socket, and configures 2 WQs per IAA,
each containing 64 entries. The driver sends decompresses to wqX.0 of
the mapped IAA device, and distributes compresses to wqX.1 of all
same-package IAAs in a round-robin manner.
64K folios: Kernel compilation/allmodconfig: deflate-iaa:
=========================================================
-------------------------------------------------------------------------------
mm-unstable-9-18-2025 v12
-------------------------------------------------------------------------------
zswap compressor deflate-iaa deflate-iaa IAA Batching
vs.
IAA Sequential
-------------------------------------------------------------------------------
real_sec 874.74 821.59 -6.1%
user_sec 15,701.49 15,718.29
sys_sec 3,834.35 3,791.12 -1%
-------------------------------------------------------------------------------
Max_Res_Set_Size_KB 1,871,264 1,872,780
-------------------------------------------------------------------------------
memcg_high 0 0
memcg_swap_fail 0 0
64kB_swpout_fallback 0 0
zswpout 91,614,674 103,057,250
zswpin 27,028,688 30,946,488
pswpout 0 0
pswpin 0 0
ZSWPOUT-64kB 3,007,705 3,337,682
SWPOUT-64kB 0 0
pgmajfault 28,697,315 32,778,092
zswap_reject_compress_fail 0 0
zswap_reject_reclaim_fail 0 0
IAA incompressible pages 440 239
-------------------------------------------------------------------------------
2M folios: Kernel compilation/allmodconfig: deflate-iaa:
========================================================
-------------------------------------------------------------------------------
mm-unstable-9-18-2025 v12
-------------------------------------------------------------------------------
zswap compressor deflate-iaa deflate-iaa IAA Batching
vs.
IAA Sequential
-------------------------------------------------------------------------------
real_sec 808.10 794.85 -1.6%
user_sec 15,807.59 15,794.49
sys_sec 4,351.01 4,266.95 -2%
-------------------------------------------------------------------------------
Max_Res_Set_Size_KB 1,871,100 1,874,684
-------------------------------------------------------------------------------
memcg_high 0 0
memcg_swap_fail 0 0
thp_swpout_fallback 0 0
zswpout 109,563,983 111,586,692
zswpin 38,396,482 39,089,000
pswpout 0 0
pswpin 0 0
ZSWPOUT-2048kB 16,661 17,020
thp_swpout 0 0
pgmajfault 39,727,295 40,433,559
zswap_reject_compress_fail 0 0
zswap_reject_reclaim_fail 0 0
IAA incompressible pages 531 317
-------------------------------------------------------------------------------
With the iaa_crypto driver changes for non-blocking descriptor allocations,
no timeouts-with-mitigations were seen in compress/decompress jobs, for all
of the above experiments.
64K folios: Kernel compilation/allmodconfig: zstd:
==================================================
-------------------------------------------------------------------------------
mm-unstable-9-18-2025 v12
-------------------------------------------------------------------------------
zswap compressor zstd zstd Improvement
-------------------------------------------------------------------------------
real_sec 925.08 853.14 -7.8%
user_sec 15,708.03 15,703.87
sys_sec 5,318.65 5,172.23 -2.8%
-------------------------------------------------------------------------------
Max_Res_Set_Size_KB 1,874,060 1,874,276
-------------------------------------------------------------------------------
memcg_high 0 0
memcg_swap_fail 0 0
64kB_swpout_fallback 0 0
zswpout 80,735,528 78,500,218
zswpin 24,155,464 22,898,395
pswpout 0 0
pswpin 0 0
ZSWPOUT-64kB 2,580,722 2,489,796
SWPOUT-64kB 0 0
pgmajfault 25,564,063 24,223,218
zswap_reject_compress_fail 0 0
zswap_reject_reclaim_fail 0 0
-------------------------------------------------------------------------------
2M folios: Kernel compilation/allmodconfig: zstd:
=================================================
-------------------------------------------------------------------------------
mm-unstable-9-18-2025 v12
-------------------------------------------------------------------------------
zswap compressor zstd zstd Improvement
-------------------------------------------------------------------------------
real_sec 848.06 845.42 -0.3%
user_sec 15,822.19 15,826.68
sys_sec 5,898.58 5,741.31 -2.7%
-------------------------------------------------------------------------------
Max_Res_Set_Size_KB 1,872,888 1,874,684
-------------------------------------------------------------------------------
memcg_high 0 0
memcg_swap_fail 0 0
thp_swpout_fallback 0 0
zswpout 91,399,677 87,856,850
zswpin 29,871,078 28,452,891
pswpout 0 0
pswpin 0 0
ZSWPOUT-2048kB 12,130 11,968
thp_swpout 0 0
pgmajfault 30,723,070 29,308,735
zswap_reject_compress_fail 0 0
zswap_reject_reclaim_fail 0 0
-------------------------------------------------------------------------------
IAA configuration script "enable_iaa.sh":
=========================================
Acknowledgements: Binuraj Ravindran and Rakib Al-Fahad.
Usage:
------
./enable_iaa.sh -d <num_IAAs> -q <num_WQs_per_IAA>
#---------------------------------<cut here>----------------------------------
#!/usr/bin/env bash
#SPDX-License-Identifier: BSD-3-Clause
#Copyright (c) 2025, Intel Corporation
#Description: Configure IAA devices
VERIFY_COMPRESS_PATH="/sys/bus/dsa/drivers/crypto/verify_compress"
iax_dev_id="0cfe"
num_iaa=$(lspci -d:${iax_dev_id} | wc -l)
sockets=$(lscpu | grep Socket | awk '{print $2}')
echo "Found ${num_iaa} instances in ${sockets} sockets(s)"
# The same number of devices will be configured in each socket, if there
# are more than one socket.
# Normalize with respect to the number of sockets.
device_num_per_socket=$(( num_iaa/sockets ))
num_iaa_per_socket=$(( num_iaa / sockets ))
iaa_wqs=2
verbose=0
iaa_engines=8
mode="dedicated"
wq_type="kernel"
iaa_crypto_mode="async"
verify_compress=0
# Function to handle errors
handle_error() {
echo "Error: $1"
exit 1
}
# Process arguments
while getopts "d:hm:q:vD" opt; do
case $opt in
d)
device_num_per_socket=$OPTARG
;;
m)
iaa_crypto_mode=$OPTARG
;;
q)
iaa_wqs=$OPTARG
;;
D)
verbose=1
;;
v)
verify_compress=1
;;
h)
echo "Usage: $0 [-d <device_count>][-q <wq_per_device>][-v]"
echo " -d - number of devices"
echo " -q - number of WQs per device"
echo " -v - verbose mode"
echo " -h - help"
exit
;;
\?)
echo "Invalid option: -$OPTARG" >&2
exit
;;
esac
done
LOG="configure_iaa.log"
# Update wq_size based on number of wqs
wq_size=$(( 128 / iaa_wqs ))
# Take care of the enumeration, if DSA is enabled.
dsa=`lspci | grep -c 0b25`
# set first,step counters to correctly enumerate iax devices based on
# whether running on guest or host with or without dsa
first=0
step=1
[[ $dsa -gt 0 && -d /sys/bus/dsa/devices/dsa0 ]] && first=1 && step=2
echo "first index: ${first}, step: ${step}"
#
# Switch to software compressors and disable IAAs to have a clean start
#
COMPRESSOR=/sys/module/zswap/parameters/compressor
last_comp=`cat ${COMPRESSOR}`
echo lzo > ${COMPRESSOR}
echo "Disable IAA devices before configuring"
for ((i = ${first}; i < ${step} * ${num_iaa}; i += ${step})); do
for ((j = 0; j < ${iaa_wqs}; j += 1)); do
cmd="accel-config disable-wq iax${i}/wq${i}.${j} >& /dev/null"
[[ $verbose == 1 ]] && echo $cmd; eval $cmd
done
cmd="accel-config disable-device iax${i} >& /dev/null"
[[ $verbose == 1 ]] && echo $cmd; eval $cmd
done
rmmod iaa_crypto
modprobe iaa_crypto
# apply crypto parameters
echo $verify_compress > ${VERIFY_COMPRESS_PATH} || handle_error "did not change verify_compress"
# Note: This is a temporary solution for during the kernel transition.
if [ -f /sys/bus/dsa/drivers/crypto/g_comp_wqs_per_iaa ];then
echo 1 > /sys/bus/dsa/drivers/crypto/g_comp_wqs_per_iaa || handle_error "did not set g_comp_wqs_per_iaa"
elif [ -f /sys/bus/dsa/drivers/crypto/g_wqs_per_iaa ];then
echo 1 > /sys/bus/dsa/drivers/crypto/g_wqs_per_iaa || handle_error "did not set g_wqs_per_iaa"
fi
if [ -f /sys/bus/dsa/drivers/crypto/g_consec_descs_per_gwq ];then
echo 1 > /sys/bus/dsa/drivers/crypto/g_consec_descs_per_gwq || handle_error "did not set g_consec_descs_per_gwq"
fi
echo ${iaa_crypto_mode} > /sys/bus/dsa/drivers/crypto/sync_mode || handle_error "could not set sync_mode"
echo "Configuring ${device_num_per_socket} device(s) out of $num_iaa_per_socket per socket"
if [ "${device_num_per_socket}" -le "${num_iaa_per_socket}" ]; then
echo "Configuring all devices"
start=${first}
end=$(( ${step} * ${device_num_per_socket} ))
else
echo "ERROR: Not enough devices"
exit
fi
#
# enable all iax devices and wqs
#
for (( socket = 0; socket < ${sockets}; socket += 1 )); do
for ((i = ${start}; i < ${end}; i += ${step})); do
echo "Configuring iaa$i on socket ${socket}"
for ((j = 0; j < ${iaa_engines}; j += 1)); do
cmd="accel-config config-engine iax${i}/engine${i}.${j} --group-id=0"
[[ $verbose == 1 ]] && echo $cmd; eval $cmd
done
# Config WQs
for ((j = 0; j < ${iaa_wqs}; j += 1)); do
# Config WQ: group 0, priority=10, mode=shared, type = kernel name=kernel, driver_name=crypto
cmd="accel-config config-wq iax${i}/wq${i}.${j} -g 0 -s ${wq_size} -p 10 -m ${mode} -y ${wq_type} -n iaa_crypto${i}${j} -d crypto"
[[ $verbose == 1 ]] && echo $cmd; eval $cmd
done
# Enable Device and WQs
cmd="accel-config enable-device iax${i}"
[[ $verbose == 1 ]] && echo $cmd; eval $cmd
for ((j = 0; j < ${iaa_wqs}; j += 1)); do
cmd="accel-config enable-wq iax${i}/wq${i}.${j}"
[[ $verbose == 1 ]] && echo $cmd; eval $cmd
done
done
start=$(( start + ${step} * ${num_iaa_per_socket} ))
end=$(( start + (${step} * ${device_num_per_socket}) ))
done
# Restore the last compressor
echo "$last_comp" > ${COMPRESSOR}
# Check if the configuration is correct
echo "Configured IAA devices:"
accel-config list | grep iax
#---------------------------------<cut here>----------------------------------
Changes since v11:
==================
1) Rebased to mm-unstable as of 9-18-2025, commit 1f98191f08b4.
2) Incorporated Herbert's suggestions on submitting the folio as the source and
SG lists for the destination to create the compress batching interface from
zswap to crypto.
3) As per Herbert's suggestion, added a new unit_size member to struct
acomp_req, along with a acomp_request_set_unit_size() API for kernel modules
to set the unit size to use while breaking down the request's src/dst
scatterlists.
4) Implemented iaa_crypto batching using the new SG lists based architecture and
crypto interfaces.
5) To make the SG lists based approach functional and performant for IAA, I have
changed all the calls to dma_map_sg() to use nents of 1. This should not be a
concern, since it eliminates redundant computes to scan an SG list with only
one scatterlist for existing kernel users, i.e. zswap with the
zswap_compress() modifications in this series. This will continue to hold
true with the zram IAA batching support I am developing. There are no kernel
use cases for the iaa_crypto driver that will break this assumption.
6) Addressed Herbert's comment about batch_size being a statically defined data
member in struct acomp_alg and struct crypto_acomp.
7) Addressed Nhat's comment about VM_WARN_ON_ONCE(nr_pages >
ZSWAP_MAX_BATCH_SIZE) in zswap_store_pages().
8) Nhat's comment about deleting struct swap_batch_decomp_data is automatically
addressed by the SG lists based rewrite of the crypto batching interface.
9) Addressed Barry's comment about renaming pool->batch_size to
pool->store_batch_size.
10) Incorporated Barry's suggestion to merge patches that introduce data members
to structures and/or API and their usage.
11) Added performance data to patch 0023's commit log, as suggested by Barry.
Changes since v10:
==================
1) Rebased to mm-unstable as of 7-30-2025, commit 01da54f10fdd.
2) Added change logging in patch 0024 on there being no Intel-specific
dependencies in the batching framework, as suggested by
Andrew Morton. Thanks Andrew!
3) Added change logging in patch 0024 on other ongoing work that can use
batching, as per Andrew's suggestion. Thanks Andrew!
4) Added the IAA configuration script in the cover letter, as suggested
by Nhat Pham. Thanks Nhat!
5) As suggested by Nhat, dropped patch 0020 from v10, that moves CPU
hotplug procedures to pool functions.
6) Gathered kernel_compilation 'allmod' config performance data with
writeback and zswap shrinker_enabled=Y.
7) Changed the pool->batch_size for software compressors to be
ZSWAP_MAX_BATCH_SIZE since this gives better performance with the zswap
shrinker enabled.
8) Was unable to replicate in v11 the issue seen in v10 with higher
memcg_swap_fail than in the baseline, with usemem30/zstd.
Changes since v9:
=================
1) Rebased to mm-unstable as of 6-24-2025, commit 23b9c0472ea3.
2) iaa_crypto rearchitecting, mainline race condition fix, performance
optimizations, code cleanup.
3) Addressed Herbert's comments in v9 patch 10, that an array based
crypto_acomp interface is not acceptable.
4) Optimized the implementation of the batching zswap_compress() and
zswap_store_pages() added in v9, to recover performance when
integrated with the changes in commit 56e5a103a721 ("zsmalloc: prefer
the the original page's node for compressed data").
Changes since v8:
=================
1) Rebased to mm-unstable as of 4-21-2025, commit 2c01d9f3c611.
2) Backported commits for reverting request chaining, since these are
in cryptodev-2.6 but not yet in mm-unstable: without these backports,
deflate-iaa is non-functional in mm-unstable:
commit 64929fe8c0a4 ("crypto: acomp - Remove request chaining")
commit 5976fe19e240 ("Revert "crypto: testmgr - Add multibuffer acomp
testing"")
Backported this hotfix as well:
commit 002ba346e3d7 ("crypto: scomp - Fix off-by-one bug when
calculating last page").
3) crypto_acomp_[de]compress() restored to non-request chained
implementations since request chaining has been removed from acomp in
commit 64929fe8c0a4 ("crypto: acomp - Remove request chaining").
4) New IAA WQ architecture to denote WQ type and whether or not a WQ
should be shared among all package cores, or only to the "mapped"
ones from an even cores-to-IAA distribution scheme.
5) Compress/decompress batching are implemented in iaa_crypto using new
crypto_acomp_batch_compress()/crypto_acomp_batch_decompress() API.
6) Defines a "void *data" in struct acomp_req, based on Herbert advising
against using req->base.data in the driver. This is needed for async
submit-poll to work.
7) In zswap.c, moved the CPU hotplug callbacks to reside in "pool
functions", per Yosry's suggestion to move procedures in a distinct
patch before refactoring patches.
8) A new "u8 nr_reqs" member is added to "struct zswap_pool" to track
the number of requests/buffers associated with the per-cpu acomp_ctx,
as per Yosry's suggestion.
9) Simplifications to the acomp_ctx resources allocation, deletion,
locking, and for these to exist from pool creation to pool deletion,
based on v8 code review discussions with Yosry.
10) Use IS_ERR_OR_NULL() consistently in zswap_cpu_comp_prepare() and
acomp_ctx_dealloc(), as per Yosry's v8 comment.
11) zswap_store_folio() is deleted, and instead, the loop over
zswap_store_pages() is moved inline in zswap_store(), per Yosry's
suggestion.
12) Better structure in zswap_compress(), unified procedure that
compresses/stores a batch of pages for both, non-batching and
batching compressors. Renamed from zswap_batch_compress() to
zswap_compress(): Thanks Yosry for these suggestions.
Changes since v7:
=================
1) Rebased to mm-unstable as of 3-3-2025, commit 5f089a9aa987.
2) Changed the acomp_ctx->nr_reqs to be u8 since ZSWAP_MAX_BATCH_SIZE is
defined as 8U, for saving memory in this per-cpu structure.
3) Fixed a typo in code comments in acomp_ctx_get_cpu_lock():
acomp_ctx->initialized to acomp_ctx->__online.
4) Incorporated suggestions from Yosry, Chengming, Nhat and Johannes,
thanks to all!
a) zswap_batch_compress() replaces zswap_compress(). Thanks Yosry
for this suggestion!
b) Process the folio in sub-batches of ZSWAP_MAX_BATCH_SIZE, regardless
of whether or not the compressor supports batching. This gets rid of
the kmalloc(entries), and allows us to allocate an array of
ZSWAP_MAX_BATCH_SIZE entries on the stack. This is implemented in
zswap_store_pages().
c) Use of a common structure and code paths for compressing a folio in
batches, either as a request chain (in parallel in IAA hardware) or
sequentially. No code duplication since zswap_compress() has been
replaced with zswap_batch_compress(), simplifying maintainability.
5) A key difference between compressors that support batching and
those that do not, is that for the latter, the acomp_ctx mutex is
locked/unlocked per ZSWAP_MAX_BATCH_SIZE batch, so that decompressions
to handle page-faults can make progress. This fixes the zstd kernel
compilation regression seen in v7. For compressors that support
batching, for e.g. IAA, the mutex is locked/released once for storing
the folio.
6) Used likely/unlikely compiler directives and prefetchw to restore
performance with the common code paths.
Changes since v6:
=================
1) Rebased to mm-unstable as of 2-27-2025, commit d58172d128ac.
2) Deleted crypto_acomp_batch_compress() and
crypto_acomp_batch_decompress() interfaces, as per Herbert's
suggestion. Batching is instead enabled by chaining the requests. For
non-batching compressors, there is no request chaining involved. Both,
batching and non-batching compressions are accomplished by zswap by
calling:
crypto_wait_req(crypto_acomp_compress(acomp_ctx->reqs[0]), &acomp_ctx->wait);
3) iaa_crypto implementation of batch compressions/decompressions using
request chaining, as per Herbert's suggestions.
4) Simplification of the acomp_ctx resource allocation/deletion with
respect to CPU hot[un]plug, to address Yosry's suggestions to explore the
mutex options in zswap_cpu_comp_prepare(). Yosry, please let me know if
the per-cpu memory cost of this proposed change is acceptable (IAA:
64.8KB, Software compressors: 8.2KB). On the positive side, I believe
restarting reclaim on a CPU after it has been through an offline-online
transition, will be much faster by not deleting the acomp_ctx resources
when the CPU gets offlined.
5) Use of lockdep assertions rather than comments for internal locking
rules, as per Yosry's suggestion.
6) No specific references to IAA in zswap.c, as suggested by Yosry.
7) Explored various solutions other than the v6 zswap_store_folio()
implementation, to fix the zstd regression seen in v5, to attempt to
unify common code paths, and to allocate smaller arrays for the zswap
entries on the stack. All these options were found to cause usemem30
latency regression with zstd. The v6 version of zswap_store_folio() is
the only implementation that does not cause zstd regression, confirmed
by 10 consecutive runs, each giving quite consistent latency
numbers. Hence, the v6 implementation is carried forward to v7, with
changes for branching for batching vs. sequential compression API
calls.
Changes since v5:
=================
1) Rebased to mm-unstable as of 2-1-2025, commit 7de6fd8ab650.
Several improvements, regression fixes and bug fixes, based on Yosry's
v5 comments (Thanks Yosry!):
2) Fix for zstd performance regression in v5.
3) Performance debug and fix for marginal improvements with IAA batching
vs. sequential.
4) Performance testing data compares IAA with and without batching, instead
of IAA batching against zstd.
5) Commit logs/zswap comments not mentioning crypto_acomp implementation
details.
6) Delete the pr_info_once() when batching resources are allocated in
zswap_cpu_comp_prepare().
7) Use kcalloc_node() for the multiple acomp_ctx buffers/reqs in
zswap_cpu_comp_prepare().
8) Simplify and consolidate error handling cleanup code in
zswap_cpu_comp_prepare().
9) Introduce zswap_compress_folio() in a separate patch.
10) Bug fix in zswap_store_folio() when xa_store() failure can cause all
compressed objects and entries to be freed, and UAF when zswap_store()
tries to free the entries that were already added to the xarray prior
to the failure.
11) Deleting compressed_bytes/bytes. zswap_store_folio() also comprehends
the recent fixes in commit bf5eaaaf7941 ("mm/zswap: fix inconsistency
when zswap_store_page() fails") by Hyeonggon Yoo.
iaa_crypto improvements/fixes/changes:
12) Enables asynchronous mode and makes it the default. With commit
4ebd9a5ca478 ("crypto: iaa - Fix IAA disabling that occurs when
sync_mode is set to 'async'"), async mode was previously just sync. We
now have true async support.
13) Change idxd descriptor allocations from blocking to non-blocking with
timeouts, and mitigations for compress/decompress ops that fail to
obtain a descriptor. This is a fix for tasks blocked errors seen in
configurations where 30+ cores are running workloads under high memory
pressure, and sending comps/decomps to 1 IAA device.
14) Fixes a bug with unprotected access of "deflate_generic_tfm" in
deflate_generic_decompress(), which can cause data corruption and
zswap_decompress() kernel crash.
15) zswap uses crypto_acomp_batch_compress() with async polling instead of
request chaining for slightly better latency. However, the request
chaining framework itself is unchanged, preserved from v5.
Changes since v4:
=================
1) Rebased to mm-unstable as of 12-20-2024, commit 5555a83c82d6.
2) Added acomp request chaining, as suggested by Herbert. Thanks Herbert!
3) Implemented IAA compress batching using request chaining.
4) zswap_store() batching simplifications suggested by Chengming, Yosry and
Nhat, thanks to all!
- New zswap_compress_folio() that is called by zswap_store().
- Move the loop over folio's pages out of zswap_store() and into a
zswap_store_folio() that stores all pages.
- Allocate all zswap entries for the folio upfront.
- Added zswap_batch_compress().
- Branch to call zswap_compress() or zswap_batch_compress() inside
zswap_compress_folio().
- All iterations over pages kept in same function level.
- No helpers other than the newly added zswap_store_folio() and
zswap_compress_folio().
Changes since v3:
=================
1) Rebased to mm-unstable as of 11-18-2024, commit 5a7056135bb6.
2) Major re-write of iaa_crypto driver's mapping of IAA devices to cores,
based on packages instead of NUMA nodes.
3) Added acomp_has_async_batching() API to crypto acomp, that allows
zswap/zram to query if a crypto_acomp has registered batch_compress and
batch_decompress interfaces.
4) Clear the poll bits on the acomp_reqs passed to
iaa_comp_a[de]compress_batch() so that a module like zswap can be
confident about the acomp_reqs[0] not having the poll bit set before
calling the fully synchronous API crypto_acomp_[de]compress().
Herbert, I would appreciate it if you can review changes 2-4; in patches
1-8 in v4. I did not want to introduce too many iaa_crypto changes in
v4, given that patch 7 is already making a major change. I plan to work
on incorporating the request chaining using the ahash interface in v5
(I need to understand the basic crypto ahash better). Thanks Herbert!
5) Incorporated Johannes' suggestion to not have a sysctl to enable
compress batching.
6) Incorporated Yosry's suggestion to allocate batching resources in the
cpu hotplug onlining code, since there is no longer a sysctl to control
batching. Thanks Yosry!
7) Incorporated Johannes' suggestions related to making the overall
sequence of events between zswap_store() and zswap_batch_store() similar
as much as possible for readability and control flow, better naming of
procedures, avoiding forward declarations, not inlining error path
procedures, deleting zswap internal details from zswap.h, etc. Thanks
Johannes, really appreciate the direction!
I have tried to explain the minimal future-proofing in terms of the
zswap_batch_store() signature and the definition of "struct
zswap_batch_store_sub_batch" in the comments for this struct. I hope the
new code explains the control flow a bit better.
Changes since v2:
=================
1) Rebased to mm-unstable as of 11-5-2024, commit 7994b7ea6ac8.
2) Fixed an issue in zswap_create_acomp_ctx() with checking for NULL
returned by kmalloc_node() for acomp_ctx->buffers and for
acomp_ctx->reqs.
3) Fixed a bug in zswap_pool_can_batch() for returning true if
pool->can_batch_comp is found to be equal to BATCH_COMP_ENABLED, and if
the per-cpu acomp_batch_ctx tests true for batching resources having
been allocated on this cpu. Also, changed from per_cpu_ptr() to
raw_cpu_ptr().
4) Incorporated the zswap_store_propagate_errors() compilation warning fix
suggested by Dan Carpenter. Thanks Dan!
5) Replaced the references to SWAP_CRYPTO_SUB_BATCH_SIZE in comments in
zswap.h, with SWAP_CRYPTO_BATCH_SIZE.
Changes since v1:
=================
1) Rebased to mm-unstable as of 11-1-2024, commit 5c4cf96cd702.
2) Incorporated Herbert's suggestions to use an acomp_req flag to indicate
async/poll mode, and to encapsulate the polling functionality in the
iaa_crypto driver. Thanks Herbert!
3) Incorporated Herbert's and Yosry's suggestions to implement the batching
API in iaa_crypto and to make its use seamless from zswap's
perspective. Thanks Herbert and Yosry!
4) Incorporated Yosry's suggestion to make it more convenient for the user
to enable compress batching, while minimizing the memory footprint
cost. Thanks Yosry!
5) Incorporated Yosry's suggestion to de-couple the shrink_folio_list()
reclaim batching patch from this series, since it requires a broader
discussion.
I would greatly appreciate code review comments for the iaa_crypto driver
and mm patches included in this series!
Thanks,
Kanchana
Kanchana P Sridhar (23):
crypto: iaa - Reorganize the iaa_crypto driver code.
crypto: iaa - New architecture for IAA device WQ comp/decomp usage &
core mapping.
crypto: iaa - Simplify, consistency of function parameters, minor
stats bug fix.
crypto: iaa - Descriptor allocation timeouts with mitigations.
crypto: iaa - iaa_wq uses percpu_refs for get/put reference counting.
crypto: iaa - Simplify the code flow in iaa_compress() and
iaa_decompress().
crypto: iaa - Refactor hardware descriptor setup into separate
procedures.
crypto: iaa - Simplified, efficient job submissions for non-irq mode.
crypto: iaa - Deprecate exporting add/remove IAA compression modes.
crypto: iaa - Expect a single scatterlist for a [de]compress request's
src/dst.
crypto: iaa - Rearchitect the iaa_crypto driver to be usable by zswap
and zram.
crypto: iaa - Enablers for submitting descriptors then polling for
completion.
crypto: acomp - Define a unit_size in struct acomp_req to enable
batching.
crypto: iaa - IAA Batching for parallel compressions/decompressions.
crypto: iaa - Enable async mode and make it the default.
crypto: iaa - Disable iaa_verify_compress by default.
crypto: iaa - Submit the two largest source buffers first in
decompress batching.
crypto: iaa - Add deflate-iaa-dynamic compression mode.
crypto: acomp - Add crypto_acomp_batch_size() to get an algorithm's
batch-size.
mm: zswap: Per-CPU acomp_ctx resources exist from pool creation to
deletion.
mm: zswap: Consistently use IS_ERR_OR_NULL() to check acomp_ctx
resources.
mm: zswap: zswap_store() will process a large folio in batches.
mm: zswap: Batched zswap_compress() with compress batching of large
folios.
.../driver-api/crypto/iaa/iaa-crypto.rst | 168 +-
crypto/acompress.c | 1 +
crypto/testmgr.c | 10 +
crypto/testmgr.h | 74 +
drivers/crypto/intel/iaa/Makefile | 4 +-
drivers/crypto/intel/iaa/iaa_crypto.h | 60 +-
.../intel/iaa/iaa_crypto_comp_dynamic.c | 22 +
drivers/crypto/intel/iaa/iaa_crypto_main.c | 2936 ++++++++++++-----
drivers/crypto/intel/iaa/iaa_crypto_stats.c | 8 +
drivers/crypto/intel/iaa/iaa_crypto_stats.h | 2 +
include/crypto/acompress.h | 63 +
include/crypto/internal/acompress.h | 3 +
include/linux/iaa_comp.h | 122 +
mm/zswap.c | 766 +++--
14 files changed, 3181 insertions(+), 1058 deletions(-)
create mode 100644 drivers/crypto/intel/iaa/iaa_crypto_comp_dynamic.c
create mode 100644 include/linux/iaa_comp.h
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 01/23] crypto: iaa - Reorganize the iaa_crypto driver code.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 02/23] crypto: iaa - New architecture for IAA device WQ comp/decomp usage & core mapping Kanchana P Sridhar
` (22 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch merely reorganizes the code in iaa_crypto_main.c, so that
the functions are consolidated into logically related sub-sections of
code, without requiring forward declarations.
This is expected to make the code more maintainable and for it to be
easier to replace functional layers and/or add new features.
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
drivers/crypto/intel/iaa/iaa_crypto_main.c | 677 +++++++++++----------
1 file changed, 350 insertions(+), 327 deletions(-)
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index 23f585219fb4..760997eee8fe 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -24,6 +24,10 @@
#define IAA_ALG_PRIORITY 300
+/**************************************
+ * Driver internal global variables.
+ **************************************/
+
/* number of iaa instances probed */
static unsigned int nr_iaa;
static unsigned int nr_cpus;
@@ -36,54 +40,6 @@ static unsigned int cpus_per_iaa;
/* Per-cpu lookup table for balanced wqs */
static struct wq_table_entry __percpu *wq_table;
-static struct idxd_wq *wq_table_next_wq(int cpu)
-{
- struct wq_table_entry *entry = per_cpu_ptr(wq_table, cpu);
-
- if (++entry->cur_wq >= entry->n_wqs)
- entry->cur_wq = 0;
-
- if (!entry->wqs[entry->cur_wq])
- return NULL;
-
- pr_debug("%s: returning wq at idx %d (iaa wq %d.%d) from cpu %d\n", __func__,
- entry->cur_wq, entry->wqs[entry->cur_wq]->idxd->id,
- entry->wqs[entry->cur_wq]->id, cpu);
-
- return entry->wqs[entry->cur_wq];
-}
-
-static void wq_table_add(int cpu, struct idxd_wq *wq)
-{
- struct wq_table_entry *entry = per_cpu_ptr(wq_table, cpu);
-
- if (WARN_ON(entry->n_wqs == entry->max_wqs))
- return;
-
- entry->wqs[entry->n_wqs++] = wq;
-
- pr_debug("%s: added iaa wq %d.%d to idx %d of cpu %d\n", __func__,
- entry->wqs[entry->n_wqs - 1]->idxd->id,
- entry->wqs[entry->n_wqs - 1]->id, entry->n_wqs - 1, cpu);
-}
-
-static void wq_table_free_entry(int cpu)
-{
- struct wq_table_entry *entry = per_cpu_ptr(wq_table, cpu);
-
- kfree(entry->wqs);
- memset(entry, 0, sizeof(*entry));
-}
-
-static void wq_table_clear_entry(int cpu)
-{
- struct wq_table_entry *entry = per_cpu_ptr(wq_table, cpu);
-
- entry->n_wqs = 0;
- entry->cur_wq = 0;
- memset(entry->wqs, 0, entry->max_wqs * sizeof(struct idxd_wq *));
-}
-
LIST_HEAD(iaa_devices);
DEFINE_MUTEX(iaa_devices_lock);
@@ -91,36 +47,11 @@ DEFINE_MUTEX(iaa_devices_lock);
static bool iaa_crypto_enabled;
static bool iaa_crypto_registered;
+static struct iaa_compression_mode *iaa_compression_modes[IAA_COMP_MODES_MAX];
+
/* Verify results of IAA compress or not */
static bool iaa_verify_compress = true;
-static ssize_t verify_compress_show(struct device_driver *driver, char *buf)
-{
- return sprintf(buf, "%d\n", iaa_verify_compress);
-}
-
-static ssize_t verify_compress_store(struct device_driver *driver,
- const char *buf, size_t count)
-{
- int ret = -EBUSY;
-
- mutex_lock(&iaa_devices_lock);
-
- if (iaa_crypto_enabled)
- goto out;
-
- ret = kstrtobool(buf, &iaa_verify_compress);
- if (ret)
- goto out;
-
- ret = count;
-out:
- mutex_unlock(&iaa_devices_lock);
-
- return ret;
-}
-static DRIVER_ATTR_RW(verify_compress);
-
/*
* The iaa crypto driver supports three 'sync' methods determining how
* compressions and decompressions are performed:
@@ -155,6 +86,37 @@ static bool async_mode;
/* Use interrupts */
static bool use_irq;
+/**************************************************
+ * Driver attributes along with get/set functions.
+ **************************************************/
+
+static ssize_t verify_compress_show(struct device_driver *driver, char *buf)
+{
+ return sprintf(buf, "%d\n", iaa_verify_compress);
+}
+
+static ssize_t verify_compress_store(struct device_driver *driver,
+ const char *buf, size_t count)
+{
+ int ret = -EBUSY;
+
+ mutex_lock(&iaa_devices_lock);
+
+ if (iaa_crypto_enabled)
+ goto out;
+
+ ret = kstrtobool(buf, &iaa_verify_compress);
+ if (ret)
+ goto out;
+
+ ret = count;
+out:
+ mutex_unlock(&iaa_devices_lock);
+
+ return ret;
+}
+static DRIVER_ATTR_RW(verify_compress);
+
/**
* set_iaa_sync_mode - Set IAA sync mode
* @name: The name of the sync mode
@@ -217,7 +179,9 @@ static ssize_t sync_mode_store(struct device_driver *driver,
}
static DRIVER_ATTR_RW(sync_mode);
-static struct iaa_compression_mode *iaa_compression_modes[IAA_COMP_MODES_MAX];
+/****************************
+ * Driver compression modes.
+ ****************************/
static int find_empty_iaa_compression_mode(void)
{
@@ -409,11 +373,6 @@ static void free_device_compression_mode(struct iaa_device *iaa_device,
IDXD_OP_FLAG_WR_SRC2_AECS_COMP | \
IDXD_OP_FLAG_AECS_RW_TGLS)
-static int check_completion(struct device *dev,
- struct iax_completion_record *comp,
- bool compress,
- bool only_once);
-
static int init_device_compression_mode(struct iaa_device *iaa_device,
struct iaa_compression_mode *mode,
int idx, struct idxd_wq *wq)
@@ -500,6 +459,11 @@ static void remove_device_compression_modes(struct iaa_device *iaa_device)
}
}
+/***********************************************************
+ * Functions for use in crypto probe and remove interfaces:
+ * allocate/init/query/deallocate devices/wqs.
+ ***********************************************************/
+
static struct iaa_device *iaa_device_alloc(void)
{
struct iaa_device *iaa_device;
@@ -513,18 +477,6 @@ static struct iaa_device *iaa_device_alloc(void)
return iaa_device;
}
-static bool iaa_has_wq(struct iaa_device *iaa_device, struct idxd_wq *wq)
-{
- struct iaa_wq *iaa_wq;
-
- list_for_each_entry(iaa_wq, &iaa_device->wqs, list) {
- if (iaa_wq->wq == wq)
- return true;
- }
-
- return false;
-}
-
static struct iaa_device *add_iaa_device(struct idxd_device *idxd)
{
struct iaa_device *iaa_device;
@@ -560,6 +512,27 @@ static void del_iaa_device(struct iaa_device *iaa_device)
nr_iaa--;
}
+static void free_iaa_device(struct iaa_device *iaa_device)
+{
+ if (!iaa_device)
+ return;
+
+ remove_device_compression_modes(iaa_device);
+ kfree(iaa_device);
+}
+
+static bool iaa_has_wq(struct iaa_device *iaa_device, struct idxd_wq *wq)
+{
+ struct iaa_wq *iaa_wq;
+
+ list_for_each_entry(iaa_wq, &iaa_device->wqs, list) {
+ if (iaa_wq->wq == wq)
+ return true;
+ }
+
+ return false;
+}
+
static int add_iaa_wq(struct iaa_device *iaa_device, struct idxd_wq *wq,
struct iaa_wq **new_wq)
{
@@ -612,23 +585,23 @@ static void del_iaa_wq(struct iaa_device *iaa_device, struct idxd_wq *wq)
}
}
-static void clear_wq_table(void)
+static void remove_iaa_wq(struct idxd_wq *wq)
{
- int cpu;
-
- for (cpu = 0; cpu < nr_cpus; cpu++)
- wq_table_clear_entry(cpu);
-
- pr_debug("cleared wq table\n");
-}
+ struct iaa_device *iaa_device;
-static void free_iaa_device(struct iaa_device *iaa_device)
-{
- if (!iaa_device)
- return;
+ list_for_each_entry(iaa_device, &iaa_devices, list) {
+ if (iaa_has_wq(iaa_device, wq)) {
+ del_iaa_wq(iaa_device, wq);
+ break;
+ }
+ }
- remove_device_compression_modes(iaa_device);
- kfree(iaa_device);
+ if (nr_iaa) {
+ cpus_per_iaa = (nr_nodes * nr_cpus_per_node) / nr_iaa;
+ if (!cpus_per_iaa)
+ cpus_per_iaa = 1;
+ } else
+ cpus_per_iaa = 1;
}
static void __free_iaa_wq(struct iaa_wq *iaa_wq)
@@ -655,6 +628,75 @@ static void free_iaa_wq(struct iaa_wq *iaa_wq)
idxd_wq_set_private(wq, NULL);
}
+static int save_iaa_wq(struct idxd_wq *wq)
+{
+ struct iaa_device *iaa_device, *found = NULL;
+ struct idxd_device *idxd;
+ struct pci_dev *pdev;
+ struct device *dev;
+ int ret = 0;
+
+ list_for_each_entry(iaa_device, &iaa_devices, list) {
+ if (iaa_device->idxd == wq->idxd) {
+ idxd = iaa_device->idxd;
+ pdev = idxd->pdev;
+ dev = &pdev->dev;
+ /*
+ * Check to see that we don't already have this wq.
+ * Shouldn't happen but we don't control probing.
+ */
+ if (iaa_has_wq(iaa_device, wq)) {
+ dev_dbg(dev, "same wq probed multiple times for iaa_device %p\n",
+ iaa_device);
+ goto out;
+ }
+
+ found = iaa_device;
+
+ ret = add_iaa_wq(iaa_device, wq, NULL);
+ if (ret)
+ goto out;
+
+ break;
+ }
+ }
+
+ if (!found) {
+ struct iaa_device *new_device;
+ struct iaa_wq *new_wq;
+
+ new_device = add_iaa_device(wq->idxd);
+ if (!new_device) {
+ ret = -ENOMEM;
+ goto out;
+ }
+
+ ret = add_iaa_wq(new_device, wq, &new_wq);
+ if (ret) {
+ del_iaa_device(new_device);
+ free_iaa_device(new_device);
+ goto out;
+ }
+
+ ret = init_iaa_device(new_device, new_wq);
+ if (ret) {
+ del_iaa_wq(new_device, new_wq->wq);
+ del_iaa_device(new_device);
+ free_iaa_wq(new_wq);
+ goto out;
+ }
+ }
+
+ if (WARN_ON(nr_iaa == 0))
+ return -EINVAL;
+
+ cpus_per_iaa = (nr_nodes * nr_cpus_per_node) / nr_iaa;
+ if (!cpus_per_iaa)
+ cpus_per_iaa = 1;
+out:
+ return 0;
+}
+
static int iaa_wq_get(struct idxd_wq *wq)
{
struct idxd_device *idxd = wq->idxd;
@@ -702,6 +744,37 @@ static int iaa_wq_put(struct idxd_wq *wq)
return ret;
}
+/***************************************************************
+ * Mapping IAA devices and wqs to cores with per-cpu wq_tables.
+ ***************************************************************/
+
+static void wq_table_free_entry(int cpu)
+{
+ struct wq_table_entry *entry = per_cpu_ptr(wq_table, cpu);
+
+ kfree(entry->wqs);
+ memset(entry, 0, sizeof(*entry));
+}
+
+static void wq_table_clear_entry(int cpu)
+{
+ struct wq_table_entry *entry = per_cpu_ptr(wq_table, cpu);
+
+ entry->n_wqs = 0;
+ entry->cur_wq = 0;
+ memset(entry->wqs, 0, entry->max_wqs * sizeof(struct idxd_wq *));
+}
+
+static void clear_wq_table(void)
+{
+ int cpu;
+
+ for (cpu = 0; cpu < nr_cpus; cpu++)
+ wq_table_clear_entry(cpu);
+
+ pr_debug("cleared wq table\n");
+}
+
static void free_wq_table(void)
{
int cpu;
@@ -739,92 +812,18 @@ static int alloc_wq_table(int max_wqs)
return 0;
}
-static int save_iaa_wq(struct idxd_wq *wq)
+static void wq_table_add(int cpu, struct idxd_wq *wq)
{
- struct iaa_device *iaa_device, *found = NULL;
- struct idxd_device *idxd;
- struct pci_dev *pdev;
- struct device *dev;
- int ret = 0;
-
- list_for_each_entry(iaa_device, &iaa_devices, list) {
- if (iaa_device->idxd == wq->idxd) {
- idxd = iaa_device->idxd;
- pdev = idxd->pdev;
- dev = &pdev->dev;
- /*
- * Check to see that we don't already have this wq.
- * Shouldn't happen but we don't control probing.
- */
- if (iaa_has_wq(iaa_device, wq)) {
- dev_dbg(dev, "same wq probed multiple times for iaa_device %p\n",
- iaa_device);
- goto out;
- }
-
- found = iaa_device;
-
- ret = add_iaa_wq(iaa_device, wq, NULL);
- if (ret)
- goto out;
-
- break;
- }
- }
-
- if (!found) {
- struct iaa_device *new_device;
- struct iaa_wq *new_wq;
-
- new_device = add_iaa_device(wq->idxd);
- if (!new_device) {
- ret = -ENOMEM;
- goto out;
- }
-
- ret = add_iaa_wq(new_device, wq, &new_wq);
- if (ret) {
- del_iaa_device(new_device);
- free_iaa_device(new_device);
- goto out;
- }
-
- ret = init_iaa_device(new_device, new_wq);
- if (ret) {
- del_iaa_wq(new_device, new_wq->wq);
- del_iaa_device(new_device);
- free_iaa_wq(new_wq);
- goto out;
- }
- }
-
- if (WARN_ON(nr_iaa == 0))
- return -EINVAL;
-
- cpus_per_iaa = (nr_nodes * nr_cpus_per_node) / nr_iaa;
- if (!cpus_per_iaa)
- cpus_per_iaa = 1;
-out:
- return 0;
-}
+ struct wq_table_entry *entry = per_cpu_ptr(wq_table, cpu);
-static void remove_iaa_wq(struct idxd_wq *wq)
-{
- struct iaa_device *iaa_device;
+ if (WARN_ON(entry->n_wqs == entry->max_wqs))
+ return;
- list_for_each_entry(iaa_device, &iaa_devices, list) {
- if (iaa_has_wq(iaa_device, wq)) {
- del_iaa_wq(iaa_device, wq);
- break;
- }
- }
+ entry->wqs[entry->n_wqs++] = wq;
- if (nr_iaa) {
- cpus_per_iaa = (nr_nodes * nr_cpus_per_node) / nr_iaa;
- if (!cpus_per_iaa)
- cpus_per_iaa = 1;
- } else
- cpus_per_iaa = 1;
+ pr_debug("%s: added iaa wq %d.%d to idx %d of cpu %d\n", __func__,
+ entry->wqs[entry->n_wqs - 1]->idxd->id,
+ entry->wqs[entry->n_wqs - 1]->id, entry->n_wqs - 1, cpu);
}
static int wq_table_add_wqs(int iaa, int cpu)
@@ -930,6 +929,44 @@ static void rebalance_wq_table(void)
pr_debug("could not add any wqs for iaa %d to cpu %d!\n", iaa, cpu);
}
+/***************************************************************
+ * Assign work-queues for driver ops using per-cpu wq_tables.
+ ***************************************************************/
+
+static struct idxd_wq *wq_table_next_wq(int cpu)
+{
+ struct wq_table_entry *entry = per_cpu_ptr(wq_table, cpu);
+
+ if (++entry->cur_wq >= entry->n_wqs)
+ entry->cur_wq = 0;
+
+ if (!entry->wqs[entry->cur_wq])
+ return NULL;
+
+ pr_debug("%s: returning wq at idx %d (iaa wq %d.%d) from cpu %d\n", __func__,
+ entry->cur_wq, entry->wqs[entry->cur_wq]->idxd->id,
+ entry->wqs[entry->cur_wq]->id, cpu);
+
+ return entry->wqs[entry->cur_wq];
+}
+
+/*************************************************
+ * Core iaa_crypto compress/decompress functions.
+ *************************************************/
+
+static int deflate_generic_decompress(struct acomp_req *req)
+{
+ ACOMP_FBREQ_ON_STACK(fbreq, req);
+ int ret;
+
+ ret = crypto_acomp_decompress(fbreq);
+ req->dlen = fbreq->dlen;
+
+ update_total_sw_decomp_calls();
+
+ return ret;
+}
+
static inline int check_completion(struct device *dev,
struct iax_completion_record *comp,
bool compress,
@@ -990,27 +1027,132 @@ static inline int check_completion(struct device *dev,
return ret;
}
-static int deflate_generic_decompress(struct acomp_req *req)
+static int iaa_remap_for_verify(struct device *dev, struct iaa_wq *iaa_wq,
+ struct acomp_req *req,
+ dma_addr_t *src_addr, dma_addr_t *dst_addr)
{
- ACOMP_FBREQ_ON_STACK(fbreq, req);
- int ret;
+ int ret = 0;
+ int nr_sgs;
- ret = crypto_acomp_decompress(fbreq);
- req->dlen = fbreq->dlen;
+ dma_unmap_sg(dev, req->dst, sg_nents(req->dst), DMA_FROM_DEVICE);
+ dma_unmap_sg(dev, req->src, sg_nents(req->src), DMA_TO_DEVICE);
- update_total_sw_decomp_calls();
+ nr_sgs = dma_map_sg(dev, req->src, sg_nents(req->src), DMA_FROM_DEVICE);
+ if (nr_sgs <= 0 || nr_sgs > 1) {
+ dev_dbg(dev, "verify: couldn't map src sg for iaa device %d,"
+ " wq %d: ret=%d\n", iaa_wq->iaa_device->idxd->id,
+ iaa_wq->wq->id, ret);
+ ret = -EIO;
+ goto out;
+ }
+ *src_addr = sg_dma_address(req->src);
+ dev_dbg(dev, "verify: dma_map_sg, src_addr %llx, nr_sgs %d, req->src %p,"
+ " req->slen %d, sg_dma_len(sg) %d\n", *src_addr, nr_sgs,
+ req->src, req->slen, sg_dma_len(req->src));
+ nr_sgs = dma_map_sg(dev, req->dst, sg_nents(req->dst), DMA_TO_DEVICE);
+ if (nr_sgs <= 0 || nr_sgs > 1) {
+ dev_dbg(dev, "verify: couldn't map dst sg for iaa device %d,"
+ " wq %d: ret=%d\n", iaa_wq->iaa_device->idxd->id,
+ iaa_wq->wq->id, ret);
+ ret = -EIO;
+ dma_unmap_sg(dev, req->src, sg_nents(req->src), DMA_FROM_DEVICE);
+ goto out;
+ }
+ *dst_addr = sg_dma_address(req->dst);
+ dev_dbg(dev, "verify: dma_map_sg, dst_addr %llx, nr_sgs %d, req->dst %p,"
+ " req->dlen %d, sg_dma_len(sg) %d\n", *dst_addr, nr_sgs,
+ req->dst, req->dlen, sg_dma_len(req->dst));
+out:
return ret;
}
-static int iaa_remap_for_verify(struct device *dev, struct iaa_wq *iaa_wq,
- struct acomp_req *req,
- dma_addr_t *src_addr, dma_addr_t *dst_addr);
-
static int iaa_compress_verify(struct crypto_tfm *tfm, struct acomp_req *req,
struct idxd_wq *wq,
dma_addr_t src_addr, unsigned int slen,
- dma_addr_t dst_addr, unsigned int *dlen);
+ dma_addr_t dst_addr, unsigned int *dlen)
+{
+ struct iaa_device_compression_mode *active_compression_mode;
+ struct iaa_compression_ctx *ctx = crypto_tfm_ctx(tfm);
+ u32 *compression_crc = acomp_request_ctx(req);
+ struct iaa_device *iaa_device;
+ struct idxd_desc *idxd_desc;
+ struct iax_hw_desc *desc;
+ struct idxd_device *idxd;
+ struct iaa_wq *iaa_wq;
+ struct pci_dev *pdev;
+ struct device *dev;
+ int ret = 0;
+
+ iaa_wq = idxd_wq_get_private(wq);
+ iaa_device = iaa_wq->iaa_device;
+ idxd = iaa_device->idxd;
+ pdev = idxd->pdev;
+ dev = &pdev->dev;
+
+ active_compression_mode = get_iaa_device_compression_mode(iaa_device, ctx->mode);
+
+ idxd_desc = idxd_alloc_desc(wq, IDXD_OP_BLOCK);
+ if (IS_ERR(idxd_desc)) {
+ dev_dbg(dev, "idxd descriptor allocation failed\n");
+ dev_dbg(dev, "iaa compress failed: ret=%ld\n",
+ PTR_ERR(idxd_desc));
+ return PTR_ERR(idxd_desc);
+ }
+ desc = idxd_desc->iax_hw;
+
+ /* Verify (optional) - decompress and check crc, suppress dest write */
+
+ desc->flags = IDXD_OP_FLAG_CRAV | IDXD_OP_FLAG_RCR | IDXD_OP_FLAG_CC;
+ desc->opcode = IAX_OPCODE_DECOMPRESS;
+ desc->decompr_flags = IAA_DECOMP_FLAGS | IAA_DECOMP_SUPPRESS_OUTPUT;
+ desc->priv = 0;
+
+ desc->src1_addr = (u64)dst_addr;
+ desc->src1_size = *dlen;
+ desc->dst_addr = (u64)src_addr;
+ desc->max_dst_size = slen;
+ desc->completion_addr = idxd_desc->compl_dma;
+
+ dev_dbg(dev, "(verify) compression mode %s,"
+ " desc->src1_addr %llx, desc->src1_size %d,"
+ " desc->dst_addr %llx, desc->max_dst_size %d,"
+ " desc->src2_addr %llx, desc->src2_size %d\n",
+ active_compression_mode->name,
+ desc->src1_addr, desc->src1_size, desc->dst_addr,
+ desc->max_dst_size, desc->src2_addr, desc->src2_size);
+
+ ret = idxd_submit_desc(wq, idxd_desc);
+ if (ret) {
+ dev_dbg(dev, "submit_desc (verify) failed ret=%d\n", ret);
+ goto err;
+ }
+
+ ret = check_completion(dev, idxd_desc->iax_completion, false, false);
+ if (ret) {
+ dev_dbg(dev, "(verify) check_completion failed ret=%d\n", ret);
+ goto err;
+ }
+
+ if (*compression_crc != idxd_desc->iax_completion->crc) {
+ ret = -EINVAL;
+ dev_dbg(dev, "(verify) iaa comp/decomp crc mismatch:"
+ " comp=0x%x, decomp=0x%x\n", *compression_crc,
+ idxd_desc->iax_completion->crc);
+ print_hex_dump(KERN_INFO, "cmp-rec: ", DUMP_PREFIX_OFFSET,
+ 8, 1, idxd_desc->iax_completion, 64, 0);
+ goto err;
+ }
+
+ idxd_free_desc(wq, idxd_desc);
+out:
+ return ret;
+err:
+ idxd_free_desc(wq, idxd_desc);
+ dev_dbg(dev, "iaa compress failed: ret=%d\n", ret);
+
+ goto out;
+}
static void iaa_desc_complete(struct idxd_desc *idxd_desc,
enum idxd_complete_type comp_type,
@@ -1226,133 +1368,6 @@ static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
goto out;
}
-static int iaa_remap_for_verify(struct device *dev, struct iaa_wq *iaa_wq,
- struct acomp_req *req,
- dma_addr_t *src_addr, dma_addr_t *dst_addr)
-{
- int ret = 0;
- int nr_sgs;
-
- dma_unmap_sg(dev, req->dst, sg_nents(req->dst), DMA_FROM_DEVICE);
- dma_unmap_sg(dev, req->src, sg_nents(req->src), DMA_TO_DEVICE);
-
- nr_sgs = dma_map_sg(dev, req->src, sg_nents(req->src), DMA_FROM_DEVICE);
- if (nr_sgs <= 0 || nr_sgs > 1) {
- dev_dbg(dev, "verify: couldn't map src sg for iaa device %d,"
- " wq %d: ret=%d\n", iaa_wq->iaa_device->idxd->id,
- iaa_wq->wq->id, ret);
- ret = -EIO;
- goto out;
- }
- *src_addr = sg_dma_address(req->src);
- dev_dbg(dev, "verify: dma_map_sg, src_addr %llx, nr_sgs %d, req->src %p,"
- " req->slen %d, sg_dma_len(sg) %d\n", *src_addr, nr_sgs,
- req->src, req->slen, sg_dma_len(req->src));
-
- nr_sgs = dma_map_sg(dev, req->dst, sg_nents(req->dst), DMA_TO_DEVICE);
- if (nr_sgs <= 0 || nr_sgs > 1) {
- dev_dbg(dev, "verify: couldn't map dst sg for iaa device %d,"
- " wq %d: ret=%d\n", iaa_wq->iaa_device->idxd->id,
- iaa_wq->wq->id, ret);
- ret = -EIO;
- dma_unmap_sg(dev, req->src, sg_nents(req->src), DMA_FROM_DEVICE);
- goto out;
- }
- *dst_addr = sg_dma_address(req->dst);
- dev_dbg(dev, "verify: dma_map_sg, dst_addr %llx, nr_sgs %d, req->dst %p,"
- " req->dlen %d, sg_dma_len(sg) %d\n", *dst_addr, nr_sgs,
- req->dst, req->dlen, sg_dma_len(req->dst));
-out:
- return ret;
-}
-
-static int iaa_compress_verify(struct crypto_tfm *tfm, struct acomp_req *req,
- struct idxd_wq *wq,
- dma_addr_t src_addr, unsigned int slen,
- dma_addr_t dst_addr, unsigned int *dlen)
-{
- struct iaa_device_compression_mode *active_compression_mode;
- struct iaa_compression_ctx *ctx = crypto_tfm_ctx(tfm);
- u32 *compression_crc = acomp_request_ctx(req);
- struct iaa_device *iaa_device;
- struct idxd_desc *idxd_desc;
- struct iax_hw_desc *desc;
- struct idxd_device *idxd;
- struct iaa_wq *iaa_wq;
- struct pci_dev *pdev;
- struct device *dev;
- int ret = 0;
-
- iaa_wq = idxd_wq_get_private(wq);
- iaa_device = iaa_wq->iaa_device;
- idxd = iaa_device->idxd;
- pdev = idxd->pdev;
- dev = &pdev->dev;
-
- active_compression_mode = get_iaa_device_compression_mode(iaa_device, ctx->mode);
-
- idxd_desc = idxd_alloc_desc(wq, IDXD_OP_BLOCK);
- if (IS_ERR(idxd_desc)) {
- dev_dbg(dev, "idxd descriptor allocation failed\n");
- dev_dbg(dev, "iaa compress failed: ret=%ld\n",
- PTR_ERR(idxd_desc));
- return PTR_ERR(idxd_desc);
- }
- desc = idxd_desc->iax_hw;
-
- /* Verify (optional) - decompress and check crc, suppress dest write */
-
- desc->flags = IDXD_OP_FLAG_CRAV | IDXD_OP_FLAG_RCR | IDXD_OP_FLAG_CC;
- desc->opcode = IAX_OPCODE_DECOMPRESS;
- desc->decompr_flags = IAA_DECOMP_FLAGS | IAA_DECOMP_SUPPRESS_OUTPUT;
- desc->priv = 0;
-
- desc->src1_addr = (u64)dst_addr;
- desc->src1_size = *dlen;
- desc->dst_addr = (u64)src_addr;
- desc->max_dst_size = slen;
- desc->completion_addr = idxd_desc->compl_dma;
-
- dev_dbg(dev, "(verify) compression mode %s,"
- " desc->src1_addr %llx, desc->src1_size %d,"
- " desc->dst_addr %llx, desc->max_dst_size %d,"
- " desc->src2_addr %llx, desc->src2_size %d\n",
- active_compression_mode->name,
- desc->src1_addr, desc->src1_size, desc->dst_addr,
- desc->max_dst_size, desc->src2_addr, desc->src2_size);
-
- ret = idxd_submit_desc(wq, idxd_desc);
- if (ret) {
- dev_dbg(dev, "submit_desc (verify) failed ret=%d\n", ret);
- goto err;
- }
-
- ret = check_completion(dev, idxd_desc->iax_completion, false, false);
- if (ret) {
- dev_dbg(dev, "(verify) check_completion failed ret=%d\n", ret);
- goto err;
- }
-
- if (*compression_crc != idxd_desc->iax_completion->crc) {
- ret = -EINVAL;
- dev_dbg(dev, "(verify) iaa comp/decomp crc mismatch:"
- " comp=0x%x, decomp=0x%x\n", *compression_crc,
- idxd_desc->iax_completion->crc);
- print_hex_dump(KERN_INFO, "cmp-rec: ", DUMP_PREFIX_OFFSET,
- 8, 1, idxd_desc->iax_completion, 64, 0);
- goto err;
- }
-
- idxd_free_desc(wq, idxd_desc);
-out:
- return ret;
-err:
- idxd_free_desc(wq, idxd_desc);
- dev_dbg(dev, "iaa compress failed: ret=%d\n", ret);
-
- goto out;
-}
-
static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
struct idxd_wq *wq,
dma_addr_t src_addr, unsigned int slen,
@@ -1662,6 +1677,10 @@ static void compression_ctx_init(struct iaa_compression_ctx *ctx)
ctx->use_irq = use_irq;
}
+/*********************************************
+ * Interfaces to crypto_alg and crypto_acomp.
+ *********************************************/
+
static int iaa_comp_init_fixed(struct crypto_acomp *acomp_tfm)
{
struct crypto_tfm *tfm = crypto_acomp_tfm(acomp_tfm);
@@ -1864,6 +1883,10 @@ static struct idxd_device_driver iaa_crypto_driver = {
.desc_complete = iaa_desc_complete,
};
+/********************
+ * Module init/exit.
+ ********************/
+
static int __init iaa_crypto_init_module(void)
{
int ret = 0;
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 02/23] crypto: iaa - New architecture for IAA device WQ comp/decomp usage & core mapping.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 01/23] crypto: iaa - Reorganize the iaa_crypto driver code Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 03/23] crypto: iaa - Simplify, consistency of function parameters, minor stats bug fix Kanchana P Sridhar
` (21 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch re-architects the iaa_crypto driver in three main aspects, to
make it more robust, stable, generic and functionally versatile to
support zswap users on platforms with different number of cores/IAAs
running workloads with different swap characteristics, and most
importantly, better performance.
Summary of latency improvement for large folio compression:
===========================================================
When measured in zswap using a simple madvise workload, where 64K
Folios are stored using IAA batch compressions, this is how the
per-page compress latency changes just by setting the
"distribute_comps" driver parameter to "1":
--------------------------------------------------------------
zswap compressor: deflate-iaa
64K Folios: zswap_store() latency normalized to per-page
--------------------------------------------------------------
p50 (ns) p99 (ns)
--------------------------------------------------------------
Sequential store 3,503 3,695
Batch compress, distribute_comps=0 1,356 1,384
Batch compress, distribute_comps=1 706 763
--------------------------------------------------------------
The rearchitecting aspects are:
A) Map IAA devices/wqs to cores based on packages instead of NUMA.
B) The WQ rebalancing algorithm that is invoked as WQs are
discovered/deleted has been made very general and flexible so that
the user can control exactly how IAA WQs are used, for optimizing
performance.
C) Additionally, the "iaa_crypto_enabled" driver global has been
modified to be an atomic, and used for synchronization between
dynamic/asynchronous WQ discovery/deletion and the fundamental
routines comp_wq_table_next_wq() and decomp_wq_table_next_wq() that
are queried by compress/decompress job submissions.
Description/motivation for (A):
===============================
This patch modifies the algorithm for mapping available IAA devices and
WQs to cores based on packages instead of NUMA nodes. This leads to a
more realistic mapping of IAA devices as compression/decompression
resources for a package, rather than for a NUMA node. This also resolves
problems that were observed during internal validation on Intel Granite
Rapids platforms with many more NUMA nodes than packages: for such
cases, the earlier NUMA based allocation caused some IAAs to be
over-subscribed and some to not be utilized at all.
As a result of this change from NUMA to packages, some of the core
functions used by the iaa_crypto driver's "probe" and "remove" API
have been re-written. The new infrastructure maintains a static mapping
of wqs per IAA device, in the "struct iaa_device" itself. The earlier
implementation would allocate memory per-cpu for this data, which never
changes once the IAA devices/wqs have been initialized.
Two main outcomes from this new iaa_crypto driver infrastructure are:
1) Resolves "task blocked for more than x seconds" errors observed during
internal validation on Intel systems with the earlier NUMA node based
mappings, which was root-caused to the non-optimal IAA-to-core mappings
described earlier.
2) Results in a NUM_THREADS factor reduction in memory footprint cost of
initializing IAA devices/wqs, due to eliminating the per-cpu copies of
each IAA device's wqs. On a 384 cores Intel Granite Rapids server with
8 IAA devices, this saves 140MiB.
An auxiliary change included in this patch is that the driver's "nr_iaa",
"nr_iaa_per_package" and "cpus_per_iaa" global variables are made
atomic, because iaa_crypto_probe() and iaa_crypto_remove() change the
values of these variables asynchronously and concurrently as wqs get
added/deleted and rebalance_wq_table() is called. This change allows the
rebalance_wq_table() code to see consistent values of the number of IAA
devices.
Description/motivation for (B):
===============================
This builds upon the package-based driver infrastructure, to provide
more flexibility in using particular WQs for compress-only or
decompress-only jobs. It also introduces the notion of using all the IAA
devices on a package as resources that are shared by all cores on the
package: this significantly improves batching (to be added in subsequent
patches) latency and compress/decompress throughput. sysfs driver
paramters provide configurability of these features.
Two main concepts are introduced as part of the rebalancing changes:
1) An IAA WQ can be used for specific ops, that determines a WQ "type"
for the iaa_crypto driver to submit compress/decompress jobs:
- compress only
- decompress only
- generic, i.e, for both compresses and decompresses
The WQ type is decided based on the number of WQs configured for a
given IAA device, and the new "g_comp_wqs_per_iaa" driver parameter.
2) An IAA WQ can be mapped to cores using either of the following
balancing techniques:
a) Shared by all cores on a package. The iaa_crypto driver will
dispatch compress/decompress jobs to all WQs of the same type,
across all IAA devices on the package:
- IAA compress jobs will be distributed to all same-package IAA
compress-only/generic WQs.
- IAA decompress jobs will be distributed to all same-package IAA
decompress-only/generic WQs.
b) Handles compress/decompress jobs only from "mapped cores", i.e.,
the cores derived by evenly dividing the number of IAAs among the
number of cores, per package.
Server setups that are moderately to highly contended can benefit from
(2.a). When the mix of workloads running on a system need high compress
throughput, and have relatively lower decompress activity, (2.b) might
be more optimal.
These approaches can be accomplished with the following new iaa_crypto
driver parameters. These parameters are global settings and will apply
to all IAAs on a package, interpreted in the context of the number of
WQs configured per IAA device.
g_comp_wqs_per_iaa:
===================
Number of compress-only WQs. The default is 1, but is applicable only
if the device has more than 1 WQ. If the device has exactly 1 WQ
configured, "g_comp_wqs_per_iaa" is a don't care.
If the IAA device has more than "g_comp_wqs_per_iaa" WQs configured,
the last "g_comp_wqs_per_iaa" number of WQs will be considered as
"compress only". The remaining WQs will be considered as
"decompress only".
If the device has less than or equal to "g_comp_wqs_per_iaa" WQs, all
the device's WQs will be considered "generic", i.e., the driver will
submit compress and decompress jobs to all the WQs configured for the
device.
For e.g., if an IAA "X" has 2 WQs, this will set up 1 decompress WQ and
1 compress WQ:
echo 1 > /sys/bus/dsa/drivers/crypto/g_comp_wqs_per_iaa
wqX.0: decompress jobs only.
wqX.1: compress jobs only.
This setting would typically benefit workloads that see a high
level of compress and decompress activity.
If an IAA has 1 WQ, that WQ will be considered "generic": the driver
will submit compress and decompress jobs to the same WQ (this is
independent of the "g_comp_wqs_per_iaa" setting):
wqX.0: compress and decompress jobs.
This would typically benefit workloads that see significant cold
memory being reclaimed, and consequently, high swapout and low swapin
activity.
distribute_comps:
=================
Distribute compressions to all IAAs on package (default is Y).
Assuming the WQ type has been established as
compress-only/decompress-only/generic, this setting will determine if
the driver will distribute compress jobs to all IAAs on a package
(default behavior) or not.
If this is turned off, the driver will dispatch compress jobs to a
given IAA "compression enabled" WQ only from cores that are mapped to
that IAA using an algorithm that evenly distributes IAAs per package
to cores per package. For e.g., on a Sapphire Rapids server with
56-physical-cores and 4 IAAs per package, with Hyperthreading, 28
logical cores will be assigned to each IAA. With the
"distribute_comps" driver parameter turned off, the driver will send
compress jobs only to it's assigned IAA device.
Enabling "distribute_comps" would typically benefit workloads in
terms of batch compress latency and throughput.
distribute_decomps:
===================
Distribute decompressions to all IAAs on package (default is N).
Assuming the WQ type has been established as
compress-only/decompress-only/generic, this setting will determine if
the driver will distribute decompress jobs to all IAAs on a package
(default behavior) or not.
We recommend leaving this parameter at its default setting of "N".
Enabling "distribute_decomps = Y" can be evaluated for workloads that
are sensitive to p99 decompress latency, and see a high level of
compress and decompress activity (for e.g. warm memory reclaim/swapin).
Recommended settings for best compress/decompress latency, throughput
and hence memory savings for a moderately contended server, are:
2 WQs per IAA
g_comp_wqs_per_iaa = 1 (separate WQ for comps/decomps per IAA)
distribute_decomps = N
distribute_comps = Y
For systems that have one IAA device, the distribute_[de]comps settings
will be a no-op. Even for such systems, as long as considerable swapout
and swapin activity is expected, we recommend setting up 2 WQs
for the IAA, one each for compressions/decompressions. If swapouts are
significantly more than swapins, 1 WQ would be a better configuration,
as mentioned earlier.
Examples:
=========
For a Sapphire Rapids server with 2 packages, 56 cores and 4 IAAs per
package, each IAA has 2 WQs, and these settings are in effect:
echo 1 > /sys/bus/dsa/drivers/crypto/g_comp_wqs_per_iaa
echo 1 > /sys/bus/dsa/drivers/crypto/distribute_comps
echo 0 > /sys/bus/dsa/drivers/crypto/distribute_decomps
wqX.0: decompress jobs only.
wqX.1: compress jobs only.
Compress jobs from all cores on package-0 will be distributed in
round-robin manner to [iax1, iax3, iax5, iax7]'s wqX.1, to maximize
compression throughput/latency/memory savings:
wq1.1
wq3.1
wq5.1
wq7.1
Likewise, compress jobs from all cores on package-1 will be
distributed in round-robin manner to [iax9, iax11, iax13, iax15]'s
wqX.1, to maximize compression throughput/latency/memory savings for
workloads running on package-1:
wq9.1
wq11.1
wq13.1
wq15.1
Decompress jobs will be submitted from mapped logical cores only, as
follows:
package-0:
CPU 0-13,112-125 14-27,126-139 28-41,140-153 42-55,154-167
IAA: iax1 iax3 iax5 iax7
WQ: wq1.0 wq3.0 wq5.0 wq7.0
package-1:
CPU 56-69,168-181 70-83,182-195 84-97,196-209 98-111,210-223
IAA: iax9 iax11 iax13 iax15
WQ: wq9.0 wq11.0 wq13.0 wq15.0
IAA WQs can be configured using higher level scripts as described in
Documentation/driver-api/crypto/iaa/iaa-crypto.rst. This documentation
has been updated for the above new parameters.
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
.../driver-api/crypto/iaa/iaa-crypto.rst | 136 +++
drivers/crypto/intel/iaa/iaa_crypto.h | 18 +-
drivers/crypto/intel/iaa/iaa_crypto_main.c | 889 ++++++++++++++----
3 files changed, 872 insertions(+), 171 deletions(-)
diff --git a/Documentation/driver-api/crypto/iaa/iaa-crypto.rst b/Documentation/driver-api/crypto/iaa/iaa-crypto.rst
index f815d4fd8372..0ff4ec603b43 100644
--- a/Documentation/driver-api/crypto/iaa/iaa-crypto.rst
+++ b/Documentation/driver-api/crypto/iaa/iaa-crypto.rst
@@ -290,6 +290,142 @@ The available attributes are:
'sync' mode. This is to ensure correct iaa_crypto behavior until true
async polling without interrupts is enabled in iaa_crypto.
+ - g_comp_wqs_per_iaa
+
+ Number of compress-only WQs. The default is 1, but is applicable only
+ if the device has more than 1 WQ. If the device has exactly 1 WQ
+ configured, "g_comp_wqs_per_iaa" is a don't care.
+
+ If the IAA device has more than "g_comp_wqs_per_iaa" WQs configured,
+ the last "g_comp_wqs_per_iaa" number of WQs will be considered as
+ "compress only". The remaining WQs will be considered as "decomp only".
+
+ If the device has less than or equal to "g_comp_wqs_per_iaa" WQs, all
+ the device's WQs will be considered "generic", i.e., the driver will
+ submit compress and decompress jobs to all the WQs configured for the
+ device.
+
+ For e.g., if an IAA "X" has 2 WQs, this will set up 1 decompress WQ and
+ 1 compress WQ::
+
+ echo 1 > /sys/bus/dsa/drivers/crypto/g_comp_wqs_per_iaa
+
+ wqX.0: decompress jobs only.
+ wqX.1: compress jobs only.
+
+ This setting would typically benefit workloads that see a high
+ level of compress and decompress activity.
+
+ If an IAA has 1 WQ, that WQ will be considered "generic": the driver
+ will submit compress and decompress jobs to the same WQ (this is
+ independent of the "g_comp_wqs_per_iaa" setting):
+
+ wqX.0: compress and decompress jobs.
+
+ This would typically benefit workloads that see significant cold
+ memory being reclaimed, and consequently, high swapout and low swapin
+ activity.
+
+ - distribute_comps
+
+ Distribute compressions to all IAAs on package (default is Y).
+
+ Assuming the WQ type has been established as
+ compress-only/decompress-only/generic, this setting will determine if
+ the driver will distribute compress jobs to all IAAs on a package
+ (default behavior) or not.
+
+ If this is turned off, the driver will dispatch compress jobs to a
+ given IAA "compression enabled" WQ only from cores that are mapped to
+ that IAA using an algorithm that evenly distributes IAAs per package
+ to cores per package. For e.g., on a Sapphire Rapids server with
+ 56-physical-cores and 4 IAAs per package, with Hyperthreading, 28
+ logical cores will be assigned to each IAA. With the
+ "distribute_comps" driver parameter turned off, the driver will send
+ compress jobs only to it's assigned IAA device.
+
+ Enabling "distribute_comps" would typically benefit workloads in
+ terms of batch compress latency and throughput.
+
+ - distribute_decomps
+
+ Distribute decompressions to all IAAs on package (default is Y).
+
+ Assuming the WQ type has been established as
+ compress-only/decompress-only/generic, this setting will determine if
+ the driver will distribute decompress jobs to all IAAs on a package
+ (default behavior) or not.
+
+ Enabling "distribute_decomps" would typically benefit workloads that
+ see a high level of compress and decompress activity, especially
+ p99 decompress latency.
+
+ Recommended settings for best compress/decompress latency, throughput
+ and hence memory savings for a moderately contended server that
+ has more than 1 IAA device enabled on a given package:
+
+ 2 WQs per IAA
+ g_comp_wqs_per_iaa = 1 (separate WQ for comps/decomps per IAA)
+ distribute_decomps = Y
+ distribute_comps = Y
+
+ For a system that has only 1 IAA device enabled on a given package,
+ the recommended settings are:
+
+ 1 WQ per IAA
+ g_comp_wqs_per_iaa = 0 (same WQ for comps/decomps)
+ distribute_decomps = N
+ distribute_comps = N
+
+ Examples:
+
+ For a Sapphire Rapids server with 2 packages, 56 cores and 4 IAAs per
+ package, each IAA has 2 WQs, and these settings are in effect::
+
+ echo 1 > /sys/bus/dsa/drivers/crypto/g_comp_wqs_per_iaa
+ echo 1 > /sys/bus/dsa/drivers/crypto/distribute_comps
+ echo 0 > /sys/bus/dsa/drivers/crypto/distribute_decomps
+
+ This enables the following behavior:
+
+ wqX.0: decompress jobs only.
+ wqX.1: compress jobs only.
+
+ Compress jobs from all cores on package-0 will be distributed in
+ round-robin manner to [iax1, iax3, iax5, iax7]'s wqX.1, to maximize
+ compression throughput/latency/memory savings:
+
+ wq1.1
+ wq3.1
+ wq5.1
+ wq7.1
+
+ Likewise, compress jobs from all cores on package-1 will be
+ distributed in round-robin manner to [iax9, iax11, iax13, iax15]'s
+ wqX.1, to maximize compression throughput/latency/memory savings for
+ workloads running on package-1:
+
+ wq9.1
+ wq11.1
+ wq13.1
+ wq15.1
+
+ Decompress jobs will be submitted from mapped logical cores only, as
+ follows:
+
+ package-0:
+
+ CPU 0-13,112-125 14-27,126-139 28-41,140-153 42-55,154-167
+ IAA: iax1 iax3 iax5 iax7
+ WQ: wq1.0 wq3.0 wq5.0 wq7.0
+
+ package-1:
+
+ CPU 56-69,168-181 70-83,182-195 84-97,196-209 98-111,210-223
+ IAA: iax9 iax11 iax13 iax15
+ WQ: wq9.0 wq11.0 wq13.0 wq15.0
+
+
.. _iaa_default_config:
IAA Default Configuration
diff --git a/drivers/crypto/intel/iaa/iaa_crypto.h b/drivers/crypto/intel/iaa/iaa_crypto.h
index 56985e395263..549ac98a9366 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto.h
+++ b/drivers/crypto/intel/iaa/iaa_crypto.h
@@ -46,6 +46,7 @@ struct iaa_wq {
struct idxd_wq *wq;
int ref;
bool remove;
+ bool mapped;
struct iaa_device *iaa_device;
@@ -63,6 +64,13 @@ struct iaa_device_compression_mode {
dma_addr_t aecs_comp_table_dma_addr;
};
+struct wq_table_entry {
+ struct idxd_wq **wqs;
+ unsigned int max_wqs;
+ unsigned int n_wqs;
+ unsigned int cur_wq;
+};
+
/* Representation of IAA device with wqs, populated by probe */
struct iaa_device {
struct list_head list;
@@ -73,19 +81,15 @@ struct iaa_device {
int n_wq;
struct list_head wqs;
+ struct wq_table_entry *generic_wq_table;
+ struct wq_table_entry *comp_wq_table;
+
atomic64_t comp_calls;
atomic64_t comp_bytes;
atomic64_t decomp_calls;
atomic64_t decomp_bytes;
};
-struct wq_table_entry {
- struct idxd_wq **wqs;
- int max_wqs;
- int n_wqs;
- int cur_wq;
-};
-
#define IAA_AECS_ALIGN 32
/*
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index 760997eee8fe..c6db721eaa79 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -23,32 +23,86 @@
#define pr_fmt(fmt) "idxd: " IDXD_SUBDRIVER_NAME ": " fmt
#define IAA_ALG_PRIORITY 300
+#define MAX_PKG_IAA 8
+#define MAX_IAA_WQ 8
/**************************************
* Driver internal global variables.
**************************************/
/* number of iaa instances probed */
-static unsigned int nr_iaa;
+static atomic_t nr_iaa = ATOMIC_INIT(0);
static unsigned int nr_cpus;
-static unsigned int nr_nodes;
-static unsigned int nr_cpus_per_node;
+static unsigned int nr_packages;
+static unsigned int nr_cpus_per_package;
+static atomic_t nr_iaa_per_package = ATOMIC_INIT(0);
/* Number of physical cpus sharing each iaa instance */
-static unsigned int cpus_per_iaa;
+static atomic_t cpus_per_iaa = ATOMIC_INIT(0);
-/* Per-cpu lookup table for balanced wqs */
-static struct wq_table_entry __percpu *wq_table;
+/* Per-cpu lookup table for decomp wqs. */
+static struct wq_table_entry __percpu *cpu_decomp_wqs;
+
+/* Per-cpu lookup table for comp wqs. */
+static struct wq_table_entry __percpu *cpu_comp_wqs;
+
+/* All decomp wqs from IAAs on a package. */
+static struct wq_table_entry **pkg_global_decomp_wqs;
+/* All comp wqs from IAAs on a package. */
+static struct wq_table_entry **pkg_global_comp_wqs;
LIST_HEAD(iaa_devices);
DEFINE_MUTEX(iaa_devices_lock);
-/* If enabled, IAA hw crypto algos are registered, unavailable otherwise */
-static bool iaa_crypto_enabled;
+/*
+ * If enabled, IAA hw crypto algos are registered, unavailable otherwise:
+ *
+ * We use the atomic @iaa_crypto_enabled to know if the per-CPU
+ * compress/decompress wq tables have been setup successfully.
+ * Since @iaa_crypto_enabled is atomic, the core functions that
+ * return a wq for compression/decompression, namely,
+ * comp_wq_table_next_wq() and decomp_wq_table_next_wq() will
+ * test this atomic before proceeding to query the per-cpu wq tables.
+ *
+ * These events will set @iaa_crypto_enabled to 1:
+ * - Successful rebalance_wq_table() after individual wq addition/removal.
+ *
+ * These events will set @iaa_crypto_enabled to 0:
+ * - Error during rebalance_wq_table() after individual wq addition/removal.
+ * - check_completion() timeouts.
+ * - @nr_iaa is 0.
+ * - module cleanup.
+ */
+static atomic_t iaa_crypto_enabled = ATOMIC_INIT(0);
+
+/*
+ * First wq probed, to use until @iaa_crypto_enabled is 1:
+ *
+ * The first wq probed will be entered in the per-CPU comp/decomp wq tables
+ * until the IAA compression modes are registered. This is done to facilitate
+ * the compress/decompress calls from the crypto testmgr resulting from
+ * calling crypto_register_acomp().
+ *
+ * With the new dynamic package-level rebalancing of WQs being
+ * discovered asynchronously and concurrently with tests
+ * triggered from device registration, this is needed to
+ * determine when it is safe for the rebalancing of decomp/comp
+ * WQs to de-allocate the per-package WQs and re-allocate them
+ * based on the latest number of IAA devices and WQs.
+ */
+static struct idxd_wq *first_wq_found;
+DEFINE_MUTEX(first_wq_found_lock);
+
static bool iaa_crypto_registered;
static struct iaa_compression_mode *iaa_compression_modes[IAA_COMP_MODES_MAX];
+/* Distribute decompressions across all IAAs on the package. */
+static bool iaa_distribute_decomps;
+
+/* Distribute compressions across all IAAs on the package. */
+static bool iaa_distribute_comps = true;
+
/* Verify results of IAA compress or not */
static bool iaa_verify_compress = true;
@@ -86,6 +140,9 @@ static bool async_mode;
/* Use interrupts */
static bool use_irq;
+/* Number of compress-only wqs per iaa*/
+static unsigned int g_comp_wqs_per_iaa = 1;
+
/**************************************************
* Driver attributes along with get/set functions.
**************************************************/
@@ -102,7 +159,7 @@ static ssize_t verify_compress_store(struct device_driver *driver,
mutex_lock(&iaa_devices_lock);
- if (iaa_crypto_enabled)
+ if (atomic_read(&iaa_crypto_enabled))
goto out;
ret = kstrtobool(buf, &iaa_verify_compress);
@@ -166,7 +223,7 @@ static ssize_t sync_mode_store(struct device_driver *driver,
mutex_lock(&iaa_devices_lock);
- if (iaa_crypto_enabled)
+ if (atomic_read(&iaa_crypto_enabled))
goto out;
ret = set_iaa_sync_mode(buf);
@@ -179,6 +236,87 @@ static ssize_t sync_mode_store(struct device_driver *driver,
}
static DRIVER_ATTR_RW(sync_mode);
+static ssize_t g_comp_wqs_per_iaa_show(struct device_driver *driver, char *buf)
+{
+ return sprintf(buf, "%u\n", g_comp_wqs_per_iaa);
+}
+
+static ssize_t g_comp_wqs_per_iaa_store(struct device_driver *driver,
+ const char *buf, size_t count)
+{
+ int ret = -EBUSY;
+
+ mutex_lock(&iaa_devices_lock);
+
+ if (atomic_read(&iaa_crypto_enabled))
+ goto out;
+
+ ret = kstrtouint(buf, 10, &g_comp_wqs_per_iaa);
+ if (ret)
+ goto out;
+
+ ret = count;
+out:
+ mutex_unlock(&iaa_devices_lock);
+
+ return ret;
+}
+static DRIVER_ATTR_RW(g_comp_wqs_per_iaa);
+
+static ssize_t distribute_decomps_show(struct device_driver *driver, char *buf)
+{
+ return sprintf(buf, "%d\n", iaa_distribute_decomps);
+}
+
+static ssize_t distribute_decomps_store(struct device_driver *driver,
+ const char *buf, size_t count)
+{
+ int ret = -EBUSY;
+
+ mutex_lock(&iaa_devices_lock);
+
+ if (atomic_read(&iaa_crypto_enabled))
+ goto out;
+
+ ret = kstrtobool(buf, &iaa_distribute_decomps);
+ if (ret)
+ goto out;
+
+ ret = count;
+out:
+ mutex_unlock(&iaa_devices_lock);
+
+ return ret;
+}
+static DRIVER_ATTR_RW(distribute_decomps);
+
+static ssize_t distribute_comps_show(struct device_driver *driver, char *buf)
+{
+ return sprintf(buf, "%d\n", iaa_distribute_comps);
+}
+
+static ssize_t distribute_comps_store(struct device_driver *driver,
+ const char *buf, size_t count)
+{
+ int ret = -EBUSY;
+
+ mutex_lock(&iaa_devices_lock);
+
+ if (atomic_read(&iaa_crypto_enabled))
+ goto out;
+
+ ret = kstrtobool(buf, &iaa_distribute_comps);
+ if (ret)
+ goto out;
+
+ ret = count;
+out:
+ mutex_unlock(&iaa_devices_lock);
+
+ return ret;
+}
+static DRIVER_ATTR_RW(distribute_comps);
+
/****************************
* Driver compression modes.
****************************/
@@ -464,32 +602,81 @@ static void remove_device_compression_modes(struct iaa_device *iaa_device)
* allocate/init/query/deallocate devices/wqs.
***********************************************************/
-static struct iaa_device *iaa_device_alloc(void)
+static struct iaa_device *iaa_device_alloc(struct idxd_device *idxd)
{
struct iaa_device *iaa_device;
+ struct wq_table_entry *wqt;
iaa_device = kzalloc(sizeof(*iaa_device), GFP_KERNEL);
if (!iaa_device)
- return NULL;
+ goto err;
+
+ iaa_device->idxd = idxd;
+
+ /* IAA device's generic/decomp wqs. */
+ iaa_device->generic_wq_table = kzalloc(sizeof(struct wq_table_entry), GFP_KERNEL);
+ if (!iaa_device->generic_wq_table)
+ goto err;
+
+ wqt = iaa_device->generic_wq_table;
+
+ wqt->wqs = kcalloc(iaa_device->idxd->max_wqs, sizeof(struct idxd_wq *), GFP_KERNEL);
+ if (!wqt->wqs)
+ goto err;
+
+ wqt->max_wqs = iaa_device->idxd->max_wqs;
+ wqt->n_wqs = 0;
+
+ /*
+ * IAA device's comp wqs (optional). If the device has more than
+ * "g_comp_wqs_per_iaa" WQs configured, the last "g_comp_wqs_per_iaa"
+ * number of WQs will be considered as "comp only". The remaining
+ * WQs will be considered as "decomp only".
+ * If the device has <= "g_comp_wqs_per_iaa" WQs, all the
+ * device's WQs will be considered "generic", i.e., cores can submit
+ * comp and decomp jobs to all the WQs configured for the device.
+ */
+ iaa_device->comp_wq_table = kzalloc(sizeof(struct wq_table_entry), GFP_KERNEL);
+ if (!iaa_device->comp_wq_table)
+ goto err;
+
+ wqt = iaa_device->comp_wq_table;
+
+ wqt->wqs = kcalloc(iaa_device->idxd->max_wqs, sizeof(struct idxd_wq *), GFP_KERNEL);
+ if (!wqt->wqs)
+ goto err;
+
+ wqt->max_wqs = iaa_device->idxd->max_wqs;
+ wqt->n_wqs = 0;
INIT_LIST_HEAD(&iaa_device->wqs);
return iaa_device;
+
+err:
+ if (iaa_device) {
+ if (iaa_device->generic_wq_table) {
+ kfree(iaa_device->generic_wq_table->wqs);
+ kfree(iaa_device->generic_wq_table);
+ }
+ kfree(iaa_device->comp_wq_table);
+ kfree(iaa_device);
+ }
+
+ return NULL;
}
static struct iaa_device *add_iaa_device(struct idxd_device *idxd)
{
struct iaa_device *iaa_device;
- iaa_device = iaa_device_alloc();
+ iaa_device = iaa_device_alloc(idxd);
if (!iaa_device)
return NULL;
- iaa_device->idxd = idxd;
-
list_add_tail(&iaa_device->list, &iaa_devices);
- nr_iaa++;
+ atomic_inc(&nr_iaa);
return iaa_device;
}
@@ -509,7 +696,7 @@ static void del_iaa_device(struct iaa_device *iaa_device)
{
list_del(&iaa_device->list);
- nr_iaa--;
+ atomic_dec(&nr_iaa);
}
static void free_iaa_device(struct iaa_device *iaa_device)
@@ -518,6 +705,17 @@ static void free_iaa_device(struct iaa_device *iaa_device)
return;
remove_device_compression_modes(iaa_device);
+
+ if (iaa_device->generic_wq_table) {
+ kfree(iaa_device->generic_wq_table->wqs);
+ kfree(iaa_device->generic_wq_table);
+ }
+
+ if (iaa_device->comp_wq_table) {
+ kfree(iaa_device->comp_wq_table->wqs);
+ kfree(iaa_device->comp_wq_table);
+ }
+
kfree(iaa_device);
}
@@ -576,7 +774,7 @@ static void del_iaa_wq(struct iaa_device *iaa_device, struct idxd_wq *wq)
dev_dbg(dev, "removed wq %d from iaa_device %d, n_wq %d, nr_iaa %d\n",
wq->id, iaa_device->idxd->id,
- iaa_device->n_wq, nr_iaa);
+ iaa_device->n_wq, atomic_read(&nr_iaa));
if (iaa_device->n_wq == 0)
del_iaa_device(iaa_device);
@@ -588,6 +786,7 @@ static void del_iaa_wq(struct iaa_device *iaa_device, struct idxd_wq *wq)
static void remove_iaa_wq(struct idxd_wq *wq)
{
struct iaa_device *iaa_device;
+ unsigned int num_pkg_iaa = 0;
list_for_each_entry(iaa_device, &iaa_devices, list) {
if (iaa_has_wq(iaa_device, wq)) {
@@ -596,12 +795,20 @@ static void remove_iaa_wq(struct idxd_wq *wq)
}
}
- if (nr_iaa) {
- cpus_per_iaa = (nr_nodes * nr_cpus_per_node) / nr_iaa;
- if (!cpus_per_iaa)
- cpus_per_iaa = 1;
- } else
- cpus_per_iaa = 1;
+ if (atomic_read(&nr_iaa)) {
+ atomic_set(&cpus_per_iaa, (nr_packages * nr_cpus_per_package) / atomic_read(&nr_iaa));
+ if (!atomic_read(&cpus_per_iaa))
+ atomic_set(&cpus_per_iaa, 1);
+
+ num_pkg_iaa = atomic_read(&nr_iaa) / nr_packages;
+ if (!num_pkg_iaa)
+ num_pkg_iaa = 1;
+ } else {
+ atomic_set(&cpus_per_iaa, 1);
+ num_pkg_iaa = 1;
+ }
+
+ atomic_set(&nr_iaa_per_package, num_pkg_iaa);
}
static void __free_iaa_wq(struct iaa_wq *iaa_wq)
@@ -635,6 +842,7 @@ static int save_iaa_wq(struct idxd_wq *wq)
struct pci_dev *pdev;
struct device *dev;
int ret = 0;
+ unsigned int num_pkg_iaa = 0;
list_for_each_entry(iaa_device, &iaa_devices, list) {
if (iaa_device->idxd == wq->idxd) {
@@ -687,12 +895,19 @@ static int save_iaa_wq(struct idxd_wq *wq)
}
}
- if (WARN_ON(nr_iaa == 0))
+ if (WARN_ON(atomic_read(&nr_iaa) == 0))
return -EINVAL;
- cpus_per_iaa = (nr_nodes * nr_cpus_per_node) / nr_iaa;
- if (!cpus_per_iaa)
- cpus_per_iaa = 1;
+ atomic_set(&cpus_per_iaa, (nr_packages * nr_cpus_per_package) / atomic_read(&nr_iaa));
+ if (!atomic_read(&cpus_per_iaa))
+ atomic_set(&cpus_per_iaa, 1);
+
+ num_pkg_iaa = atomic_read(&nr_iaa) / nr_packages;
+ if (!num_pkg_iaa)
+ num_pkg_iaa = 1;
+
+ atomic_set(&nr_iaa_per_package, num_pkg_iaa);
+
out:
return 0;
}
@@ -748,105 +963,284 @@ static int iaa_wq_put(struct idxd_wq *wq)
* Mapping IAA devices and wqs to cores with per-cpu wq_tables.
***************************************************************/
-static void wq_table_free_entry(int cpu)
+/*
+ * Given a cpu, find the closest IAA instance.
+ */
+static inline int cpu_to_iaa(int cpu)
{
- struct wq_table_entry *entry = per_cpu_ptr(wq_table, cpu);
+ int package_id, base_iaa, iaa = 0;
+
+ if (!nr_packages || !atomic_read(&nr_iaa_per_package) || !atomic_read(&nr_iaa))
+ return -1;
+
+ package_id = topology_logical_package_id(cpu);
+ base_iaa = package_id * atomic_read(&nr_iaa_per_package);
+ iaa = base_iaa + ((cpu % nr_cpus_per_package) / atomic_read(&cpus_per_iaa));
- kfree(entry->wqs);
- memset(entry, 0, sizeof(*entry));
+ pr_debug("cpu = %d, package_id = %d, base_iaa = %d, iaa = %d",
+ cpu, package_id, base_iaa, iaa);
+
+ if (iaa >= 0 && iaa < atomic_read(&nr_iaa))
+ return iaa;
+
+ return (atomic_read(&nr_iaa) - 1);
}
-static void wq_table_clear_entry(int cpu)
+static void free_wq_tables(void)
{
- struct wq_table_entry *entry = per_cpu_ptr(wq_table, cpu);
+ if (cpu_decomp_wqs) {
+ free_percpu(cpu_decomp_wqs);
+ cpu_decomp_wqs = NULL;
+ }
- entry->n_wqs = 0;
- entry->cur_wq = 0;
- memset(entry->wqs, 0, entry->max_wqs * sizeof(struct idxd_wq *));
+ if (cpu_comp_wqs) {
+ free_percpu(cpu_comp_wqs);
+ cpu_comp_wqs = NULL;
+ }
+
+ pr_debug("freed comp/decomp wq tables\n");
}
-static void clear_wq_table(void)
+static void pkg_global_wqs_dealloc(void)
{
- int cpu;
+ int i;
- for (cpu = 0; cpu < nr_cpus; cpu++)
- wq_table_clear_entry(cpu);
+ if (pkg_global_decomp_wqs) {
+ for (i = 0; i < nr_packages; ++i) {
+ kfree(pkg_global_decomp_wqs[i]->wqs);
+ kfree(pkg_global_decomp_wqs[i]);
+ }
+ kfree(pkg_global_decomp_wqs);
+ pkg_global_decomp_wqs = NULL;
+ }
- pr_debug("cleared wq table\n");
+ if (pkg_global_comp_wqs) {
+ for (i = 0; i < nr_packages; ++i) {
+ kfree(pkg_global_comp_wqs[i]->wqs);
+ kfree(pkg_global_comp_wqs[i]);
+ }
+ kfree(pkg_global_comp_wqs);
+ pkg_global_comp_wqs = NULL;
+ }
}
-static void free_wq_table(void)
+static bool pkg_global_wqs_alloc(void)
{
- int cpu;
+ int i;
+
+ pkg_global_decomp_wqs = kcalloc(nr_packages, sizeof(*pkg_global_decomp_wqs), GFP_KERNEL);
+ if (!pkg_global_decomp_wqs)
+ return false;
+
+ for (i = 0; i < nr_packages; ++i) {
+ pkg_global_decomp_wqs[i] = kzalloc(sizeof(struct wq_table_entry), GFP_KERNEL);
+ if (!pkg_global_decomp_wqs[i])
+ goto err;
+
+ pkg_global_decomp_wqs[i]->wqs = kcalloc(MAX_PKG_IAA * MAX_IAA_WQ, sizeof(struct idxd_wq *), GFP_KERNEL);
+ if (!pkg_global_decomp_wqs[i]->wqs)
+ goto err;
+
+ pkg_global_decomp_wqs[i]->max_wqs = MAX_PKG_IAA * MAX_IAA_WQ;
+ }
+
+ pkg_global_comp_wqs = kcalloc(nr_packages, sizeof(*pkg_global_comp_wqs), GFP_KERNEL);
+ if (!pkg_global_comp_wqs)
+ goto err;
+
+ for (i = 0; i < nr_packages; ++i) {
+ pkg_global_comp_wqs[i] = kzalloc(sizeof(struct wq_table_entry), GFP_KERNEL);
+ if (!pkg_global_comp_wqs[i])
+ goto err;
- for (cpu = 0; cpu < nr_cpus; cpu++)
- wq_table_free_entry(cpu);
+ pkg_global_comp_wqs[i]->wqs = kcalloc(MAX_PKG_IAA * MAX_IAA_WQ, sizeof(struct idxd_wq *), GFP_KERNEL);
+ if (!pkg_global_comp_wqs[i]->wqs)
+ goto err;
+
+ pkg_global_comp_wqs[i]->max_wqs = MAX_PKG_IAA * MAX_IAA_WQ;
+ }
- free_percpu(wq_table);
+ return true;
- pr_debug("freed wq table\n");
+err:
+ pkg_global_wqs_dealloc();
+ return false;
}
static int alloc_wq_table(int max_wqs)
{
- struct wq_table_entry *entry;
- int cpu;
-
- wq_table = alloc_percpu(struct wq_table_entry);
- if (!wq_table)
+ cpu_decomp_wqs = alloc_percpu_gfp(struct wq_table_entry, GFP_KERNEL | __GFP_ZERO);
+ if (!cpu_decomp_wqs)
return -ENOMEM;
- for (cpu = 0; cpu < nr_cpus; cpu++) {
- entry = per_cpu_ptr(wq_table, cpu);
- entry->wqs = kcalloc(max_wqs, sizeof(*entry->wqs), GFP_KERNEL);
- if (!entry->wqs) {
- free_wq_table();
- return -ENOMEM;
- }
+ cpu_comp_wqs = alloc_percpu_gfp(struct wq_table_entry, GFP_KERNEL | __GFP_ZERO);
+ if (!cpu_comp_wqs)
+ goto err;
- entry->max_wqs = max_wqs;
- }
+ if (!pkg_global_wqs_alloc())
+ goto err;
pr_debug("initialized wq table\n");
return 0;
+
+err:
+ free_wq_tables();
+ return -ENOMEM;
+}
+
+/*
+ * The caller should have established that device_iaa_wqs is not empty,
+ * i.e., every IAA device in "iaa_devices" has at least one WQ.
+ */
+static void add_device_wqs_to_wq_table(struct wq_table_entry *dst_wq_table,
+ struct wq_table_entry *device_wq_table)
+{
+ int i;
+
+ for (i = 0; i < device_wq_table->n_wqs; ++i)
+ dst_wq_table->wqs[dst_wq_table->n_wqs++] = device_wq_table->wqs[i];
+}
+
+static bool reinit_pkg_global_wqs(bool comp)
+{
+ int cur_iaa = 0, pkg = 0;
+ struct iaa_device *iaa_device;
+ struct wq_table_entry **pkg_wqs = comp ? pkg_global_comp_wqs : pkg_global_decomp_wqs;
+
+ for (pkg = 0; pkg < nr_packages; ++pkg)
+ pkg_wqs[pkg]->n_wqs = 0;
+
+ pkg = 0;
+
+one_iaa_special_case:
+ /* Re-initialize per-package wqs. */
+ list_for_each_entry(iaa_device, &iaa_devices, list) {
+ struct wq_table_entry *device_wq_table = comp ?
+ ((iaa_device->comp_wq_table->n_wqs > 0) ?
+ iaa_device->comp_wq_table : iaa_device->generic_wq_table) :
+ iaa_device->generic_wq_table;
+
+ if (pkg_wqs[pkg]->n_wqs + device_wq_table->n_wqs > pkg_wqs[pkg]->max_wqs) {
+ pkg_wqs[pkg]->wqs = krealloc(pkg_wqs[pkg]->wqs,
+ ksize(pkg_wqs[pkg]->wqs) +
+ max((MAX_PKG_IAA * MAX_IAA_WQ), iaa_device->n_wq) * sizeof(struct idxd_wq *),
+ GFP_KERNEL | __GFP_ZERO);
+ if (!pkg_wqs[pkg]->wqs)
+ return false;
+
+ pkg_wqs[pkg]->max_wqs = ksize(pkg_wqs[pkg]->wqs)/sizeof(struct idxd_wq *);
+ }
+
+ add_device_wqs_to_wq_table(pkg_wqs[pkg], device_wq_table);
+
+ pr_debug("pkg_global_%s_wqs[%d] has %u n_wqs %u max_wqs",
+ (comp ? "comp" : "decomp"), pkg, pkg_wqs[pkg]->n_wqs, pkg_wqs[pkg]->max_wqs);
+
+ if (++cur_iaa == atomic_read(&nr_iaa_per_package)) {
+ if (++pkg == nr_packages)
+ break;
+ cur_iaa = 0;
+ if (atomic_read(&nr_iaa) == 1)
+ goto one_iaa_special_case;
+ }
+ }
+
+ return true;
}
-static void wq_table_add(int cpu, struct idxd_wq *wq)
+static void create_cpu_wq_table(int cpu, struct wq_table_entry *wq_table, bool comp)
{
- struct wq_table_entry *entry = per_cpu_ptr(wq_table, cpu);
+ struct wq_table_entry *entry = comp ?
+ per_cpu_ptr(cpu_comp_wqs, cpu) :
+ per_cpu_ptr(cpu_decomp_wqs, cpu);
+
+ if (!atomic_read(&iaa_crypto_enabled)) {
+ mutex_lock(&first_wq_found_lock);
+
+ BUG_ON(!first_wq_found && !wq_table->n_wqs);
+
+ if (!first_wq_found)
+ first_wq_found = wq_table->wqs[0];
+
+ mutex_unlock(&first_wq_found_lock);
- if (WARN_ON(entry->n_wqs == entry->max_wqs))
+ entry->wqs = &first_wq_found;
+ entry->max_wqs = 1;
+ entry->n_wqs = 1;
+ entry->cur_wq = 0;
+ pr_debug("%s: cpu %d: added %u first_wq_found for %s wqs up to wq %d.%d\n", __func__,
+ cpu, entry->n_wqs, comp ? "comp":"decomp",
+ entry->wqs[entry->n_wqs - 1]->idxd->id,
+ entry->wqs[entry->n_wqs - 1]->id);
return;
+ }
+
+ entry->wqs = wq_table->wqs;
+ entry->max_wqs = wq_table->max_wqs;
+ entry->n_wqs = wq_table->n_wqs;
+ entry->cur_wq = 0;
+
+ if (entry->n_wqs)
+ pr_debug("%s: cpu %d: added %u iaa %s wqs up to wq %d.%d: entry->max_wqs = %u\n", __func__,
+ cpu, entry->n_wqs, comp ? "comp":"decomp",
+ entry->wqs[entry->n_wqs - 1]->idxd->id, entry->wqs[entry->n_wqs - 1]->id,
+ entry->max_wqs);
+}
+
+static void set_cpu_wq_table_start_wq(int cpu, bool comp)
+{
+ struct wq_table_entry *entry = comp ?
+ per_cpu_ptr(cpu_comp_wqs, cpu) :
+ per_cpu_ptr(cpu_decomp_wqs, cpu);
+ unsigned int num_pkg_iaa = atomic_read(&nr_iaa_per_package);
+
+ int start_wq = (entry->n_wqs / num_pkg_iaa) * (cpu_to_iaa(cpu) % num_pkg_iaa);
+
+ if ((start_wq >= 0) && (start_wq < entry->n_wqs))
+ entry->cur_wq = start_wq;
+}
- entry->wqs[entry->n_wqs++] = wq;
+static void create_cpu_wq_table_from_pkg_wqs(bool comp)
+{
+ int cpu;
- pr_debug("%s: added iaa wq %d.%d to idx %d of cpu %d\n", __func__,
- entry->wqs[entry->n_wqs - 1]->idxd->id,
- entry->wqs[entry->n_wqs - 1]->id, entry->n_wqs - 1, cpu);
+ /*
+ * All CPU on the same package share the same "package global"
+ * [de]comp_wqs.
+ */
+ for (cpu = 0; cpu < nr_cpus; cpu += nr_cpus_per_package) {
+ int package_id = topology_logical_package_id(cpu);
+ struct wq_table_entry *pkg_wq_table = comp ?
+ ((pkg_global_comp_wqs[package_id]->n_wqs > 0) ?
+ pkg_global_comp_wqs[package_id] : pkg_global_decomp_wqs[package_id])
+ : pkg_global_decomp_wqs[package_id];
+ int pkg_cpu;
+
+ for (pkg_cpu = cpu; pkg_cpu < cpu + nr_cpus_per_package; ++pkg_cpu) {
+ /* Initialize decomp/comp wq_table for CPU. */
+ create_cpu_wq_table(pkg_cpu, pkg_wq_table, comp);
+ /* Stagger the starting WQ in the package WQ table, for each CPU. */
+ set_cpu_wq_table_start_wq(pkg_cpu, comp);
+ }
+ }
}
-static int wq_table_add_wqs(int iaa, int cpu)
+static int add_mapped_device_wq_table_for_cpu(int iaa, int cpu, bool comp)
{
struct iaa_device *iaa_device, *found_device = NULL;
- int ret = 0, cur_iaa = 0, n_wqs_added = 0;
- struct idxd_device *idxd;
- struct iaa_wq *iaa_wq;
- struct pci_dev *pdev;
- struct device *dev;
+ struct wq_table_entry *device_wq_table;
+ int ret = 0, cur_iaa = 0;
list_for_each_entry(iaa_device, &iaa_devices, list) {
- idxd = iaa_device->idxd;
- pdev = idxd->pdev;
- dev = &pdev->dev;
-
if (cur_iaa != iaa) {
cur_iaa++;
continue;
}
found_device = iaa_device;
- dev_dbg(dev, "getting wq from iaa_device %d, cur_iaa %d\n",
+ dev_dbg(&found_device->idxd->pdev->dev,
+ "getting wq from iaa_device %d, cur_iaa %d\n",
found_device->idxd->id, cur_iaa);
break;
}
@@ -861,93 +1255,219 @@ static int wq_table_add_wqs(int iaa, int cpu)
}
cur_iaa = 0;
- idxd = found_device->idxd;
- pdev = idxd->pdev;
- dev = &pdev->dev;
- dev_dbg(dev, "getting wq from only iaa_device %d, cur_iaa %d\n",
+ dev_dbg(&found_device->idxd->pdev->dev,
+ "getting wq from only iaa_device %d, cur_iaa %d\n",
found_device->idxd->id, cur_iaa);
}
- list_for_each_entry(iaa_wq, &found_device->wqs, list) {
- wq_table_add(cpu, iaa_wq->wq);
- pr_debug("rebalance: added wq for cpu=%d: iaa wq %d.%d\n",
- cpu, iaa_wq->wq->idxd->id, iaa_wq->wq->id);
- n_wqs_added++;
+ device_wq_table = comp ?
+ ((found_device->comp_wq_table->n_wqs > 0) ?
+ found_device->comp_wq_table : found_device->generic_wq_table) :
+ found_device->generic_wq_table;
+
+ create_cpu_wq_table(cpu, device_wq_table, comp);
+
+out:
+ return ret;
+}
+
+static void create_cpu_wq_table_from_mapped_device(bool comp)
+{
+ int cpu, iaa;
+
+ for (cpu = 0; cpu < nr_cpus; cpu++) {
+ iaa = cpu_to_iaa(cpu);
+ pr_debug("rebalance: cpu=%d iaa=%d\n", cpu, iaa);
+
+ if (WARN_ON(iaa == -1)) {
+ pr_debug("rebalance (cpu_to_iaa(%d)) failed!\n", cpu);
+ return;
+ }
+
+ if (WARN_ON(add_mapped_device_wq_table_for_cpu(iaa, cpu, comp))) {
+ pr_debug("could not add any wqs of iaa %d to cpu %d!\n", iaa, cpu);
+ return;
+ }
+ }
+}
+
+static int map_iaa_device_wqs(struct iaa_device *iaa_device)
+{
+ struct wq_table_entry *generic, *for_comps;
+ int ret = 0, n_wqs_added = 0;
+ struct iaa_wq *iaa_wq;
+
+ generic = iaa_device->generic_wq_table;
+ for_comps = iaa_device->comp_wq_table;
+
+ list_for_each_entry(iaa_wq, &iaa_device->wqs, list) {
+ if (iaa_wq->mapped && ++n_wqs_added)
+ continue;
+
+ pr_debug("iaa_device %p: processing wq %d.%d\n", iaa_device, iaa_device->idxd->id, iaa_wq->wq->id);
+
+ if ((!n_wqs_added || ((n_wqs_added + g_comp_wqs_per_iaa) < iaa_device->n_wq)) &&
+ (generic->n_wqs < generic->max_wqs)) {
+
+ generic->wqs[generic->n_wqs++] = iaa_wq->wq;
+ pr_debug("iaa_device %p: added decomp wq %d.%d\n", iaa_device, iaa_device->idxd->id, iaa_wq->wq->id);
+ } else {
+ if (WARN_ON(for_comps->n_wqs == for_comps->max_wqs))
+ break;
+
+ for_comps->wqs[for_comps->n_wqs++] = iaa_wq->wq;
+ pr_debug("iaa_device %p: added comp wq %d.%d\n", iaa_device, iaa_device->idxd->id, iaa_wq->wq->id);
+ }
+
+ iaa_wq->mapped = true;
+ ++n_wqs_added;
}
- if (!n_wqs_added) {
- pr_debug("couldn't find any iaa wqs!\n");
+ if (!n_wqs_added && !iaa_device->n_wq) {
+ pr_debug("iaa_device %d: couldn't find any iaa wqs!\n", iaa_device->idxd->id);
ret = -EINVAL;
- goto out;
}
-out:
+
return ret;
}
+static void map_iaa_devices(void)
+{
+ struct iaa_device *iaa_device;
+
+ list_for_each_entry(iaa_device, &iaa_devices, list) {
+ BUG_ON(map_iaa_device_wqs(iaa_device));
+ }
+}
+
/*
- * Rebalance the wq table so that given a cpu, it's easy to find the
- * closest IAA instance. The idea is to try to choose the most
- * appropriate IAA instance for a caller and spread available
- * workqueues around to clients.
+ * Rebalance the per-cpu wq table based on available IAA devices/WQs.
+ * Three driver parameters control how this algorithm works:
+ *
+ * - g_comp_wqs_per_iaa:
+ *
+ * If multiple WQs are configured for a given device, this setting determines
+ * the number of WQs to be used as "compress only" WQs. The remaining WQs will
+ * be used as "decompress only WQs".
+ * Note that the comp WQ can be the same as the decomp WQ, for e.g., if
+ * g_comp_wqs_per_iaa is 0 (regardless of the # of available WQs per device), or,
+ * if there is only 1 WQ configured for a device (regardless of
+ * g_comp_wqs_per_iaa).
+ *
+ * - distribute_decomps, distribute_comps:
+ *
+ * If this is enabled, all [de]comp WQs found from the IAA devices on a
+ * package, will be aggregated into pkg_global_[de]comp_wqs, then assigned to
+ * each CPU on the package.
+ *
+ * Note:
+ * -----
+ * rebalance_wq_table() will return true if it was able to successfully
+ * configure comp/decomp wqs for all CPUs, without changing the
+ * @iaa_crypto_enabled atomic. The caller can re-enable the use of the wq
+ * tables after rebalance_wq_table() returns true, by setting the
+ * @iaa_crypto_enabled atomic to 1.
+ * In case of any errors, the @iaa_crypto_enabled atomic will be set to 0,
+ * and rebalance_wq_table() will return false.
*/
-static void rebalance_wq_table(void)
+static bool rebalance_wq_table(void)
{
- const struct cpumask *node_cpus;
- int node_cpu, node, cpu, iaa = 0;
+ int cpu;
- if (nr_iaa == 0)
- return;
+ if (atomic_read(&nr_iaa) == 0)
+ goto err;
- pr_debug("rebalance: nr_nodes=%d, nr_cpus %d, nr_iaa %d, cpus_per_iaa %d\n",
- nr_nodes, nr_cpus, nr_iaa, cpus_per_iaa);
+ map_iaa_devices();
- clear_wq_table();
+ pr_info("rebalance: nr_packages=%d, nr_cpus %d, nr_iaa %d, nr_iaa_per_package %d, cpus_per_iaa %d\n",
+ nr_packages, nr_cpus, atomic_read(&nr_iaa),
+ atomic_read(&nr_iaa_per_package), atomic_read(&cpus_per_iaa));
- if (nr_iaa == 1) {
- for_each_possible_cpu(cpu) {
- if (WARN_ON(wq_table_add_wqs(0, cpu)))
- goto err;
- }
+ if (iaa_distribute_decomps) {
+ /* Each CPU uses all IAA devices on package for decomps. */
+ if (!reinit_pkg_global_wqs(false))
+ goto err;
+ create_cpu_wq_table_from_pkg_wqs(false);
+ } else {
+ /*
+ * Each CPU uses the decomp WQ on the mapped IAA device using
+ * a balanced mapping of cores to IAA.
+ */
+ create_cpu_wq_table_from_mapped_device(false);
+ }
- return;
+ if (iaa_distribute_comps) {
+ /* Each CPU uses all IAA devices on package for comps. */
+ if (!reinit_pkg_global_wqs(true))
+ goto err;
+ create_cpu_wq_table_from_pkg_wqs(true);
+ } else {
+ /*
+ * Each CPU uses the comp WQ on the mapped IAA device using
+ * a balanced mapping of cores to IAA.
+ */
+ create_cpu_wq_table_from_mapped_device(true);
}
- for_each_node_with_cpus(node) {
- cpu = 0;
- node_cpus = cpumask_of_node(node);
+ /* Verify that each cpu has comp and decomp wqs.*/
+ for (cpu = 0; cpu < nr_cpus; cpu++) {
+ struct wq_table_entry *entry = per_cpu_ptr(cpu_decomp_wqs, cpu);
- for_each_cpu(node_cpu, node_cpus) {
- iaa = cpu / cpus_per_iaa;
- if (WARN_ON(wq_table_add_wqs(iaa, node_cpu)))
- goto err;
- cpu++;
+ if (!entry->wqs || !entry->n_wqs) {
+ pr_err("%s: cpu %d does not have decomp_wqs", __func__, cpu);
+ goto err;
+ }
+
+ entry = per_cpu_ptr(cpu_comp_wqs, cpu);
+ if (!entry->wqs || !entry->n_wqs) {
+ pr_err("%s: cpu %d does not have comp_wqs", __func__, cpu);
+ goto err;
}
}
- return;
+ pr_debug("Finished rebalance decomp/comp wqs.");
+ return true;
+
err:
- pr_debug("could not add any wqs for iaa %d to cpu %d!\n", iaa, cpu);
+ atomic_set(&iaa_crypto_enabled, 0);
+ pr_debug("Error during rebalance decomp/comp wqs.");
+ return false;
}
/***************************************************************
* Assign work-queues for driver ops using per-cpu wq_tables.
***************************************************************/
-static struct idxd_wq *wq_table_next_wq(int cpu)
+static struct idxd_wq *decomp_wq_table_next_wq(int cpu)
{
- struct wq_table_entry *entry = per_cpu_ptr(wq_table, cpu);
+ struct wq_table_entry *entry = per_cpu_ptr(cpu_decomp_wqs, cpu);
+ struct idxd_wq *wq;
+
+ if (!atomic_read(&iaa_crypto_enabled))
+ return NULL;
+
+ wq = entry->wqs[entry->cur_wq];
- if (++entry->cur_wq >= entry->n_wqs)
+ if (++entry->cur_wq == entry->n_wqs)
entry->cur_wq = 0;
- if (!entry->wqs[entry->cur_wq])
+ return wq;
+}
+
+static struct idxd_wq *comp_wq_table_next_wq(int cpu)
+{
+ struct wq_table_entry *entry = per_cpu_ptr(cpu_comp_wqs, cpu);
+ struct idxd_wq *wq;
+
+ if (!atomic_read(&iaa_crypto_enabled))
return NULL;
- pr_debug("%s: returning wq at idx %d (iaa wq %d.%d) from cpu %d\n", __func__,
- entry->cur_wq, entry->wqs[entry->cur_wq]->idxd->id,
- entry->wqs[entry->cur_wq]->id, cpu);
+ wq = entry->wqs[entry->cur_wq];
- return entry->wqs[entry->cur_wq];
+ if (++entry->cur_wq == entry->n_wqs)
+ entry->cur_wq = 0;
+
+ return wq;
}
/*************************************************
@@ -985,7 +1505,7 @@ static inline int check_completion(struct device *dev,
dev_err(dev, "%s completion timed out - "
"assuming broken hw, iaa_crypto now DISABLED\n",
op_str);
- iaa_crypto_enabled = false;
+ atomic_set(&iaa_crypto_enabled, 0);
ret = -ETIMEDOUT;
goto out;
}
@@ -1501,18 +2021,13 @@ static int iaa_comp_acompress(struct acomp_req *req)
compression_ctx = crypto_tfm_ctx(tfm);
- if (!iaa_crypto_enabled) {
- pr_debug("iaa_crypto disabled, not compressing\n");
- return -ENODEV;
- }
-
if (!req->src || !req->slen) {
pr_debug("invalid src, not compressing\n");
return -EINVAL;
}
cpu = get_cpu();
- wq = wq_table_next_wq(cpu);
+ wq = comp_wq_table_next_wq(cpu);
put_cpu();
if (!wq) {
pr_debug("no wq configured for cpu=%d\n", cpu);
@@ -1599,18 +2114,13 @@ static int iaa_comp_adecompress(struct acomp_req *req)
struct device *dev;
struct idxd_wq *wq;
- if (!iaa_crypto_enabled) {
- pr_debug("iaa_crypto disabled, not decompressing\n");
- return -ENODEV;
- }
-
if (!req->src || !req->slen) {
pr_debug("invalid src, not decompressing\n");
return -EINVAL;
}
cpu = get_cpu();
- wq = wq_table_next_wq(cpu);
+ wq = decomp_wq_table_next_wq(cpu);
put_cpu();
if (!wq) {
pr_debug("no wq configured for cpu=%d\n", cpu);
@@ -1725,6 +2235,8 @@ static int iaa_register_compression_device(void)
static int iaa_unregister_compression_device(void)
{
+ atomic_set(&iaa_crypto_enabled, 0);
+
if (iaa_crypto_registered)
crypto_unregister_acomp(&iaa_acomp_fixed_deflate);
@@ -1746,10 +2258,13 @@ static int iaa_crypto_probe(struct idxd_dev *idxd_dev)
if (data->type != IDXD_TYPE_IAX)
return -ENODEV;
+ mutex_lock(&iaa_devices_lock);
+
mutex_lock(&wq->wq_lock);
if (idxd_wq_get_private(wq)) {
mutex_unlock(&wq->wq_lock);
+ mutex_unlock(&iaa_devices_lock);
return -EBUSY;
}
@@ -1771,8 +2286,6 @@ static int iaa_crypto_probe(struct idxd_dev *idxd_dev)
goto err;
}
- mutex_lock(&iaa_devices_lock);
-
if (list_empty(&iaa_devices)) {
ret = alloc_wq_table(wq->idxd->max_wqs);
if (ret)
@@ -1784,24 +2297,33 @@ static int iaa_crypto_probe(struct idxd_dev *idxd_dev)
if (ret)
goto err_save;
- rebalance_wq_table();
+ if (!rebalance_wq_table()) {
+ dev_dbg(dev, "%s: IAA rebalancing device wq tables failed\n", __func__);
+ goto err_register;
+ }
+ atomic_set(&iaa_crypto_enabled, 1);
if (first_wq) {
- iaa_crypto_enabled = true;
ret = iaa_register_compression_device();
if (ret != 0) {
- iaa_crypto_enabled = false;
dev_dbg(dev, "IAA compression device registration failed\n");
goto err_register;
}
+
+ if (!rebalance_wq_table()) {
+ dev_dbg(dev, "%s: Rerun after registration: IAA rebalancing device wq tables failed\n", __func__);
+ goto err_register;
+ }
+ atomic_set(&iaa_crypto_enabled, 1);
+
try_module_get(THIS_MODULE);
pr_info("iaa_crypto now ENABLED\n");
}
- mutex_unlock(&iaa_devices_lock);
out:
mutex_unlock(&wq->wq_lock);
+ mutex_unlock(&iaa_devices_lock);
return ret;
@@ -1810,9 +2332,8 @@ static int iaa_crypto_probe(struct idxd_dev *idxd_dev)
free_iaa_wq(idxd_wq_get_private(wq));
err_save:
if (first_wq)
- free_wq_table();
+ free_wq_tables();
err_alloc:
- mutex_unlock(&iaa_devices_lock);
idxd_drv_disable_wq(wq);
err:
wq->type = IDXD_WQT_NONE;
@@ -1827,13 +2348,17 @@ static void iaa_crypto_remove(struct idxd_dev *idxd_dev)
struct iaa_wq *iaa_wq;
bool free = false;
+ atomic_set(&iaa_crypto_enabled, 0);
idxd_wq_quiesce(wq);
- mutex_lock(&wq->wq_lock);
mutex_lock(&iaa_devices_lock);
+ mutex_lock(&wq->wq_lock);
remove_iaa_wq(wq);
+ if (!rebalance_wq_table())
+ pr_debug("%s: IAA rebalancing device wq tables failed\n", __func__);
+
spin_lock(&idxd->dev_lock);
iaa_wq = idxd_wq_get_private(wq);
if (!iaa_wq) {
@@ -1856,18 +2381,22 @@ static void iaa_crypto_remove(struct idxd_dev *idxd_dev)
}
idxd_drv_disable_wq(wq);
- rebalance_wq_table();
- if (nr_iaa == 0) {
- iaa_crypto_enabled = false;
- free_wq_table();
+ if (atomic_read(&nr_iaa) == 0) {
+ atomic_set(&iaa_crypto_enabled, 0);
+ pkg_global_wqs_dealloc();
+ free_wq_tables();
+ BUG_ON(!list_empty(&iaa_devices));
+ INIT_LIST_HEAD(&iaa_devices);
module_put(THIS_MODULE);
pr_info("iaa_crypto now DISABLED\n");
+ } else {
+ atomic_set(&iaa_crypto_enabled, 1);
}
out:
- mutex_unlock(&iaa_devices_lock);
mutex_unlock(&wq->wq_lock);
+ mutex_unlock(&iaa_devices_lock);
}
static enum idxd_dev_type dev_types[] = {
@@ -1890,16 +2419,12 @@ static struct idxd_device_driver iaa_crypto_driver = {
static int __init iaa_crypto_init_module(void)
{
int ret = 0;
- int node;
+
+ INIT_LIST_HEAD(&iaa_devices);
nr_cpus = num_possible_cpus();
- for_each_node_with_cpus(node)
- nr_nodes++;
- if (!nr_nodes) {
- pr_err("IAA couldn't find any nodes with cpus\n");
- return -ENODEV;
- }
- nr_cpus_per_node = nr_cpus / nr_nodes;
+ nr_cpus_per_package = topology_num_cores_per_package();
+ nr_packages = topology_max_packages();
ret = iaa_aecs_init_fixed();
if (ret < 0) {
@@ -1913,6 +2438,27 @@ static int __init iaa_crypto_init_module(void)
goto err_driver_reg;
}
+ ret = driver_create_file(&iaa_crypto_driver.drv,
+ &driver_attr_g_comp_wqs_per_iaa);
+ if (ret) {
+ pr_debug("IAA g_comp_wqs_per_iaa attr creation failed\n");
+ goto err_g_comp_wqs_per_iaa_attr_create;
+ }
+
+ ret = driver_create_file(&iaa_crypto_driver.drv,
+ &driver_attr_distribute_decomps);
+ if (ret) {
+ pr_debug("IAA distribute_decomps attr creation failed\n");
+ goto err_distribute_decomps_attr_create;
+ }
+
+ ret = driver_create_file(&iaa_crypto_driver.drv,
+ &driver_attr_distribute_comps);
+ if (ret) {
+ pr_debug("IAA distribute_comps attr creation failed\n");
+ goto err_distribute_comps_attr_create;
+ }
+
ret = driver_create_file(&iaa_crypto_driver.drv,
&driver_attr_verify_compress);
if (ret) {
@@ -1938,6 +2484,15 @@ static int __init iaa_crypto_init_module(void)
driver_remove_file(&iaa_crypto_driver.drv,
&driver_attr_verify_compress);
err_verify_attr_create:
+ driver_remove_file(&iaa_crypto_driver.drv,
+ &driver_attr_distribute_comps);
+err_distribute_comps_attr_create:
+ driver_remove_file(&iaa_crypto_driver.drv,
+ &driver_attr_distribute_decomps);
+err_distribute_decomps_attr_create:
+ driver_remove_file(&iaa_crypto_driver.drv,
+ &driver_attr_g_comp_wqs_per_iaa);
+err_g_comp_wqs_per_iaa_attr_create:
idxd_driver_unregister(&iaa_crypto_driver);
err_driver_reg:
iaa_aecs_cleanup_fixed();
@@ -1956,6 +2511,12 @@ static void __exit iaa_crypto_cleanup_module(void)
&driver_attr_sync_mode);
driver_remove_file(&iaa_crypto_driver.drv,
&driver_attr_verify_compress);
+ driver_remove_file(&iaa_crypto_driver.drv,
+ &driver_attr_distribute_comps);
+ driver_remove_file(&iaa_crypto_driver.drv,
+ &driver_attr_distribute_decomps);
+ driver_remove_file(&iaa_crypto_driver.drv,
+ &driver_attr_g_comp_wqs_per_iaa);
idxd_driver_unregister(&iaa_crypto_driver);
iaa_aecs_cleanup_fixed();
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 03/23] crypto: iaa - Simplify, consistency of function parameters, minor stats bug fix.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 01/23] crypto: iaa - Reorganize the iaa_crypto driver code Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 02/23] crypto: iaa - New architecture for IAA device WQ comp/decomp usage & core mapping Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 04/23] crypto: iaa - Descriptor allocation timeouts with mitigations Kanchana P Sridhar
` (20 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch further simplifies the code in some places and makes it more
consistent and readable:
1) Change iaa_compress_verify() @dlen parameter to be a value instead of
a pointer, because @dlen's value is only read, not modified by this
procedure.
2) Simplify the success/error return paths in iaa_compress(),
iaa_decompress() and iaa_compress_verify().
3) Delete dev_dbg() statements to make the code more readable.
4) Change return value from descriptor allocation failures to be
-ENODEV, for better maintainability.
5) Fix a minor statistics bug in iaa_decompress(), with the
decomp_bytes getting updated in case of errors.
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
drivers/crypto/intel/iaa/iaa_crypto_main.c | 107 +++++----------------
1 file changed, 22 insertions(+), 85 deletions(-)
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index c6db721eaa79..ed3325bb3291 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -1590,7 +1590,7 @@ static int iaa_remap_for_verify(struct device *dev, struct iaa_wq *iaa_wq,
static int iaa_compress_verify(struct crypto_tfm *tfm, struct acomp_req *req,
struct idxd_wq *wq,
dma_addr_t src_addr, unsigned int slen,
- dma_addr_t dst_addr, unsigned int *dlen)
+ dma_addr_t dst_addr, unsigned int dlen)
{
struct iaa_device_compression_mode *active_compression_mode;
struct iaa_compression_ctx *ctx = crypto_tfm_ctx(tfm);
@@ -1614,10 +1614,8 @@ static int iaa_compress_verify(struct crypto_tfm *tfm, struct acomp_req *req,
idxd_desc = idxd_alloc_desc(wq, IDXD_OP_BLOCK);
if (IS_ERR(idxd_desc)) {
- dev_dbg(dev, "idxd descriptor allocation failed\n");
- dev_dbg(dev, "iaa compress failed: ret=%ld\n",
- PTR_ERR(idxd_desc));
- return PTR_ERR(idxd_desc);
+ dev_dbg(dev, "iaa compress_verify failed: idxd descriptor allocation failure: ret=%ld\n", PTR_ERR(idxd_desc));
+ return -ENODEV;
}
desc = idxd_desc->iax_hw;
@@ -1629,19 +1627,11 @@ static int iaa_compress_verify(struct crypto_tfm *tfm, struct acomp_req *req,
desc->priv = 0;
desc->src1_addr = (u64)dst_addr;
- desc->src1_size = *dlen;
+ desc->src1_size = dlen;
desc->dst_addr = (u64)src_addr;
desc->max_dst_size = slen;
desc->completion_addr = idxd_desc->compl_dma;
- dev_dbg(dev, "(verify) compression mode %s,"
- " desc->src1_addr %llx, desc->src1_size %d,"
- " desc->dst_addr %llx, desc->max_dst_size %d,"
- " desc->src2_addr %llx, desc->src2_size %d\n",
- active_compression_mode->name,
- desc->src1_addr, desc->src1_size, desc->dst_addr,
- desc->max_dst_size, desc->src2_addr, desc->src2_size);
-
ret = idxd_submit_desc(wq, idxd_desc);
if (ret) {
dev_dbg(dev, "submit_desc (verify) failed ret=%d\n", ret);
@@ -1664,14 +1654,10 @@ static int iaa_compress_verify(struct crypto_tfm *tfm, struct acomp_req *req,
goto err;
}
- idxd_free_desc(wq, idxd_desc);
-out:
- return ret;
err:
idxd_free_desc(wq, idxd_desc);
- dev_dbg(dev, "iaa compress failed: ret=%d\n", ret);
- goto out;
+ return ret;
}
static void iaa_desc_complete(struct idxd_desc *idxd_desc,
@@ -1751,7 +1737,7 @@ static void iaa_desc_complete(struct idxd_desc *idxd_desc,
}
ret = iaa_compress_verify(ctx->tfm, ctx->req, iaa_wq->wq, src_addr,
- ctx->req->slen, dst_addr, &ctx->req->dlen);
+ ctx->req->slen, dst_addr, ctx->req->dlen);
if (ret) {
dev_dbg(dev, "%s: compress verify failed ret=%d\n", __func__, ret);
err = -EIO;
@@ -1777,7 +1763,7 @@ static void iaa_desc_complete(struct idxd_desc *idxd_desc,
iaa_wq_put(idxd_desc->wq);
}
-static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
+static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
struct idxd_wq *wq,
dma_addr_t src_addr, unsigned int slen,
dma_addr_t dst_addr, unsigned int *dlen)
@@ -1804,9 +1790,9 @@ static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
idxd_desc = idxd_alloc_desc(wq, IDXD_OP_BLOCK);
if (IS_ERR(idxd_desc)) {
- dev_dbg(dev, "idxd descriptor allocation failed\n");
- dev_dbg(dev, "iaa compress failed: ret=%ld\n", PTR_ERR(idxd_desc));
- return PTR_ERR(idxd_desc);
+ dev_dbg(dev, "iaa compress failed: idxd descriptor allocation failure: ret=%ld\n",
+ PTR_ERR(idxd_desc));
+ return -ENODEV;
}
desc = idxd_desc->iax_hw;
@@ -1832,21 +1818,8 @@ static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
idxd_desc->crypto.src_addr = src_addr;
idxd_desc->crypto.dst_addr = dst_addr;
idxd_desc->crypto.compress = true;
-
- dev_dbg(dev, "%s use_async_irq: compression mode %s,"
- " src_addr %llx, dst_addr %llx\n", __func__,
- active_compression_mode->name,
- src_addr, dst_addr);
}
- dev_dbg(dev, "%s: compression mode %s,"
- " desc->src1_addr %llx, desc->src1_size %d,"
- " desc->dst_addr %llx, desc->max_dst_size %d,"
- " desc->src2_addr %llx, desc->src2_size %d\n", __func__,
- active_compression_mode->name,
- desc->src1_addr, desc->src1_size, desc->dst_addr,
- desc->max_dst_size, desc->src2_addr, desc->src2_size);
-
ret = idxd_submit_desc(wq, idxd_desc);
if (ret) {
dev_dbg(dev, "submit_desc failed ret=%d\n", ret);
@@ -1859,7 +1832,6 @@ static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
if (ctx->async_mode) {
ret = -EINPROGRESS;
- dev_dbg(dev, "%s: returning -EINPROGRESS\n", __func__);
goto out;
}
@@ -1877,15 +1849,10 @@ static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
*compression_crc = idxd_desc->iax_completion->crc;
- if (!ctx->async_mode)
- idxd_free_desc(wq, idxd_desc);
-out:
- return ret;
err:
idxd_free_desc(wq, idxd_desc);
- dev_dbg(dev, "iaa compress failed: ret=%d\n", ret);
-
- goto out;
+out:
+ return ret;
}
static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
@@ -1914,10 +1881,10 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
idxd_desc = idxd_alloc_desc(wq, IDXD_OP_BLOCK);
if (IS_ERR(idxd_desc)) {
- dev_dbg(dev, "idxd descriptor allocation failed\n");
- dev_dbg(dev, "iaa decompress failed: ret=%ld\n",
+ ret = -ENODEV;
+ dev_dbg(dev, "%s: idxd descriptor allocation failed: ret=%ld\n", __func__,
PTR_ERR(idxd_desc));
- return PTR_ERR(idxd_desc);
+ return ret;
}
desc = idxd_desc->iax_hw;
@@ -1941,21 +1908,8 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
idxd_desc->crypto.src_addr = src_addr;
idxd_desc->crypto.dst_addr = dst_addr;
idxd_desc->crypto.compress = false;
-
- dev_dbg(dev, "%s: use_async_irq compression mode %s,"
- " src_addr %llx, dst_addr %llx\n", __func__,
- active_compression_mode->name,
- src_addr, dst_addr);
}
- dev_dbg(dev, "%s: decompression mode %s,"
- " desc->src1_addr %llx, desc->src1_size %d,"
- " desc->dst_addr %llx, desc->max_dst_size %d,"
- " desc->src2_addr %llx, desc->src2_size %d\n", __func__,
- active_compression_mode->name,
- desc->src1_addr, desc->src1_size, desc->dst_addr,
- desc->max_dst_size, desc->src2_addr, desc->src2_size);
-
ret = idxd_submit_desc(wq, idxd_desc);
if (ret) {
dev_dbg(dev, "submit_desc failed ret=%d\n", ret);
@@ -1968,7 +1922,6 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
if (ctx->async_mode) {
ret = -EINPROGRESS;
- dev_dbg(dev, "%s: returning -EINPROGRESS\n", __func__);
goto out;
}
@@ -1990,23 +1943,19 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
}
} else {
req->dlen = idxd_desc->iax_completion->output_size;
+
+ /* Update stats */
+ update_total_decomp_bytes_in(slen);
+ update_wq_decomp_bytes(wq, slen);
}
*dlen = req->dlen;
- if (!ctx->async_mode)
+err:
+ if (idxd_desc)
idxd_free_desc(wq, idxd_desc);
-
- /* Update stats */
- update_total_decomp_bytes_in(slen);
- update_wq_decomp_bytes(wq, slen);
out:
return ret;
-err:
- idxd_free_desc(wq, idxd_desc);
- dev_dbg(dev, "iaa decompress failed: ret=%d\n", ret);
-
- goto out;
}
static int iaa_comp_acompress(struct acomp_req *req)
@@ -2053,9 +2002,6 @@ static int iaa_comp_acompress(struct acomp_req *req)
goto out;
}
src_addr = sg_dma_address(req->src);
- dev_dbg(dev, "dma_map_sg, src_addr %llx, nr_sgs %d, req->src %p,"
- " req->slen %d, sg_dma_len(sg) %d\n", src_addr, nr_sgs,
- req->src, req->slen, sg_dma_len(req->src));
nr_sgs = dma_map_sg(dev, req->dst, sg_nents(req->dst), DMA_FROM_DEVICE);
if (nr_sgs <= 0 || nr_sgs > 1) {
@@ -2066,9 +2012,6 @@ static int iaa_comp_acompress(struct acomp_req *req)
goto err_map_dst;
}
dst_addr = sg_dma_address(req->dst);
- dev_dbg(dev, "dma_map_sg, dst_addr %llx, nr_sgs %d, req->dst %p,"
- " req->dlen %d, sg_dma_len(sg) %d\n", dst_addr, nr_sgs,
- req->dst, req->dlen, sg_dma_len(req->dst));
ret = iaa_compress(tfm, req, wq, src_addr, req->slen, dst_addr,
&req->dlen);
@@ -2083,7 +2026,7 @@ static int iaa_comp_acompress(struct acomp_req *req)
}
ret = iaa_compress_verify(tfm, req, wq, src_addr, req->slen,
- dst_addr, &req->dlen);
+ dst_addr, req->dlen);
if (ret)
dev_dbg(dev, "asynchronous compress verification failed ret=%d\n", ret);
@@ -2146,9 +2089,6 @@ static int iaa_comp_adecompress(struct acomp_req *req)
goto out;
}
src_addr = sg_dma_address(req->src);
- dev_dbg(dev, "dma_map_sg, src_addr %llx, nr_sgs %d, req->src %p,"
- " req->slen %d, sg_dma_len(sg) %d\n", src_addr, nr_sgs,
- req->src, req->slen, sg_dma_len(req->src));
nr_sgs = dma_map_sg(dev, req->dst, sg_nents(req->dst), DMA_FROM_DEVICE);
if (nr_sgs <= 0 || nr_sgs > 1) {
@@ -2159,9 +2099,6 @@ static int iaa_comp_adecompress(struct acomp_req *req)
goto err_map_dst;
}
dst_addr = sg_dma_address(req->dst);
- dev_dbg(dev, "dma_map_sg, dst_addr %llx, nr_sgs %d, req->dst %p,"
- " req->dlen %d, sg_dma_len(sg) %d\n", dst_addr, nr_sgs,
- req->dst, req->dlen, sg_dma_len(req->dst));
ret = iaa_decompress(tfm, req, wq, src_addr, req->slen,
dst_addr, &req->dlen);
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 04/23] crypto: iaa - Descriptor allocation timeouts with mitigations.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (2 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 03/23] crypto: iaa - Simplify, consistency of function parameters, minor stats bug fix Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 05/23] crypto: iaa - iaa_wq uses percpu_refs for get/put reference counting Kanchana P Sridhar
` (19 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch modifies the descriptor allocation from blocking to
non-blocking with bounded retries or "timeouts".
This is necessary to prevent task blocked errors in high contention
scenarios, for instance, when the platform has only 1 IAA device
enabled. With 1 IAA device enabled per package on a dual-package
Sapphire Rapids with 56 cores/package, there are 112 logical cores
mapped to this single IAA device. In this scenario, the task blocked
errors can occur because idxd_alloc_desc() is called with
IDXD_OP_BLOCK. With batching, multiple descriptors will need to be
allocated per batch. Any process that is able to do so, can cause
contention for allocating descriptors for all other processes that share
the use of the same sbitmap_queue. Under IDXD_OP_BLOCK, this causes
compress/decompress jobs to stall in stress test scenarios
(e.g. zswap_store() of 2M folios).
In order to make the iaa_crypto driver be more fail-safe, this commit
implements the following:
1) Change compress/decompress descriptor allocations to be non-blocking
with retries ("timeouts").
2) Return compress error to zswap if descriptor allocation with timeouts
fails during compress ops. zswap_store() will return an error and the
folio gets stored in the backing swap device.
3) Fallback to software decompress if descriptor allocation with timeouts
fails during decompress ops.
With these fixes, there are no task blocked errors seen under stress
testing conditions, and no performance degradation observed.
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
drivers/crypto/intel/iaa/iaa_crypto.h | 5 ++
drivers/crypto/intel/iaa/iaa_crypto_main.c | 58 +++++++++++++++-------
2 files changed, 44 insertions(+), 19 deletions(-)
diff --git a/drivers/crypto/intel/iaa/iaa_crypto.h b/drivers/crypto/intel/iaa/iaa_crypto.h
index 549ac98a9366..cc76a047b54a 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto.h
+++ b/drivers/crypto/intel/iaa/iaa_crypto.h
@@ -21,6 +21,9 @@
#define IAA_COMPLETION_TIMEOUT 1000000
+#define IAA_ALLOC_DESC_COMP_TIMEOUT 1000
+#define IAA_ALLOC_DESC_DECOMP_TIMEOUT 500
+
#define IAA_ANALYTICS_ERROR 0x0a
#define IAA_ERROR_DECOMP_BUF_OVERFLOW 0x0b
#define IAA_ERROR_COMP_BUF_OVERFLOW 0x19
@@ -141,6 +144,8 @@ enum iaa_mode {
struct iaa_compression_ctx {
enum iaa_mode mode;
+ u16 alloc_comp_desc_timeout;
+ u16 alloc_decomp_desc_timeout;
bool verify_compress;
bool async_mode;
bool use_irq;
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index ed3325bb3291..1169cd44c8e7 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -1596,7 +1596,8 @@ static int iaa_compress_verify(struct crypto_tfm *tfm, struct acomp_req *req,
struct iaa_compression_ctx *ctx = crypto_tfm_ctx(tfm);
u32 *compression_crc = acomp_request_ctx(req);
struct iaa_device *iaa_device;
- struct idxd_desc *idxd_desc;
+ struct idxd_desc *idxd_desc = ERR_PTR(-EAGAIN);
+ u16 alloc_desc_retries = 0;
struct iax_hw_desc *desc;
struct idxd_device *idxd;
struct iaa_wq *iaa_wq;
@@ -1612,7 +1613,11 @@ static int iaa_compress_verify(struct crypto_tfm *tfm, struct acomp_req *req,
active_compression_mode = get_iaa_device_compression_mode(iaa_device, ctx->mode);
- idxd_desc = idxd_alloc_desc(wq, IDXD_OP_BLOCK);
+ while ((idxd_desc == ERR_PTR(-EAGAIN)) && (alloc_desc_retries++ < ctx->alloc_decomp_desc_timeout)) {
+ idxd_desc = idxd_alloc_desc(wq, IDXD_OP_NONBLOCK);
+ cpu_relax();
+ }
+
if (IS_ERR(idxd_desc)) {
dev_dbg(dev, "iaa compress_verify failed: idxd descriptor allocation failure: ret=%ld\n", PTR_ERR(idxd_desc));
return -ENODEV;
@@ -1772,7 +1777,8 @@ static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
struct iaa_compression_ctx *ctx = crypto_tfm_ctx(tfm);
u32 *compression_crc = acomp_request_ctx(req);
struct iaa_device *iaa_device;
- struct idxd_desc *idxd_desc;
+ struct idxd_desc *idxd_desc = ERR_PTR(-EAGAIN);
+ u16 alloc_desc_retries = 0;
struct iax_hw_desc *desc;
struct idxd_device *idxd;
struct iaa_wq *iaa_wq;
@@ -1788,7 +1794,11 @@ static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
active_compression_mode = get_iaa_device_compression_mode(iaa_device, ctx->mode);
- idxd_desc = idxd_alloc_desc(wq, IDXD_OP_BLOCK);
+ while ((idxd_desc == ERR_PTR(-EAGAIN)) && (alloc_desc_retries++ < ctx->alloc_comp_desc_timeout)) {
+ idxd_desc = idxd_alloc_desc(wq, IDXD_OP_NONBLOCK);
+ cpu_relax();
+ }
+
if (IS_ERR(idxd_desc)) {
dev_dbg(dev, "iaa compress failed: idxd descriptor allocation failure: ret=%ld\n",
PTR_ERR(idxd_desc));
@@ -1863,7 +1873,8 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
struct iaa_device_compression_mode *active_compression_mode;
struct iaa_compression_ctx *ctx = crypto_tfm_ctx(tfm);
struct iaa_device *iaa_device;
- struct idxd_desc *idxd_desc;
+ struct idxd_desc *idxd_desc = ERR_PTR(-EAGAIN);
+ u16 alloc_desc_retries = 0;
struct iax_hw_desc *desc;
struct idxd_device *idxd;
struct iaa_wq *iaa_wq;
@@ -1879,12 +1890,17 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
active_compression_mode = get_iaa_device_compression_mode(iaa_device, ctx->mode);
- idxd_desc = idxd_alloc_desc(wq, IDXD_OP_BLOCK);
+ while ((idxd_desc == ERR_PTR(-EAGAIN)) && (alloc_desc_retries++ < ctx->alloc_decomp_desc_timeout)) {
+ idxd_desc = idxd_alloc_desc(wq, IDXD_OP_NONBLOCK);
+ cpu_relax();
+ }
+
if (IS_ERR(idxd_desc)) {
ret = -ENODEV;
dev_dbg(dev, "%s: idxd descriptor allocation failed: ret=%ld\n", __func__,
PTR_ERR(idxd_desc));
- return ret;
+ idxd_desc = NULL;
+ goto fallback_software_decomp;
}
desc = idxd_desc->iax_hw;
@@ -1913,7 +1929,7 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
ret = idxd_submit_desc(wq, idxd_desc);
if (ret) {
dev_dbg(dev, "submit_desc failed ret=%d\n", ret);
- goto err;
+ goto fallback_software_decomp;
}
/* Update stats */
@@ -1926,19 +1942,21 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
}
ret = check_completion(dev, idxd_desc->iax_completion, false, false);
+
+fallback_software_decomp:
if (ret) {
- dev_dbg(dev, "%s: check_completion failed ret=%d\n", __func__, ret);
- if (idxd_desc->iax_completion->status == IAA_ANALYTICS_ERROR) {
+ dev_dbg(dev, "%s: desc allocation/submission/check_completion failed ret=%d\n", __func__, ret);
+ if (idxd_desc && idxd_desc->iax_completion->status == IAA_ANALYTICS_ERROR) {
pr_warn("%s: falling back to deflate-generic decompress, "
"analytics error code %x\n", __func__,
idxd_desc->iax_completion->error_code);
- ret = deflate_generic_decompress(req);
- if (ret) {
- dev_dbg(dev, "%s: deflate-generic failed ret=%d\n",
- __func__, ret);
- goto err;
- }
- } else {
+ }
+
+ ret = deflate_generic_decompress(req);
+
+ if (ret) {
+ pr_err("%s: iaa decompress failed: deflate-generic fallback error ret=%d\n",
+ __func__, ret);
goto err;
}
} else {
@@ -2119,6 +2137,8 @@ static int iaa_comp_adecompress(struct acomp_req *req)
static void compression_ctx_init(struct iaa_compression_ctx *ctx)
{
+ ctx->alloc_comp_desc_timeout = IAA_ALLOC_DESC_COMP_TIMEOUT;
+ ctx->alloc_decomp_desc_timeout = IAA_ALLOC_DESC_DECOMP_TIMEOUT;
ctx->verify_compress = iaa_verify_compress;
ctx->async_mode = async_mode;
ctx->use_irq = use_irq;
@@ -2133,10 +2153,10 @@ static int iaa_comp_init_fixed(struct crypto_acomp *acomp_tfm)
struct crypto_tfm *tfm = crypto_acomp_tfm(acomp_tfm);
struct iaa_compression_ctx *ctx = crypto_tfm_ctx(tfm);
- compression_ctx_init(ctx);
-
ctx->mode = IAA_MODE_FIXED;
+ compression_ctx_init(ctx);
+
return 0;
}
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 05/23] crypto: iaa - iaa_wq uses percpu_refs for get/put reference counting.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (3 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 04/23] crypto: iaa - Descriptor allocation timeouts with mitigations Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 06/23] crypto: iaa - Simplify the code flow in iaa_compress() and iaa_decompress() Kanchana P Sridhar
` (18 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch modifies the reference counting on "struct iaa_wq" to be a
percpu_ref in atomic mode, instead of an "int refcount" combined with
the "idxd->dev_lock" spin_lock currently used as a synchronization
mechanism to achieve get/put semantics.
This enables a more light-weight, cleaner and effective refcount
implementation for the iaa_wq, significantly reducing latency per
compress/decompress job submitted to the IAA accelerator:
p50: -136 ns
p99: -880 ns
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
drivers/crypto/intel/iaa/iaa_crypto.h | 4 +-
drivers/crypto/intel/iaa/iaa_crypto_main.c | 119 +++++++--------------
2 files changed, 41 insertions(+), 82 deletions(-)
diff --git a/drivers/crypto/intel/iaa/iaa_crypto.h b/drivers/crypto/intel/iaa/iaa_crypto.h
index cc76a047b54a..9611f2518f42 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto.h
+++ b/drivers/crypto/intel/iaa/iaa_crypto.h
@@ -47,8 +47,8 @@ struct iaa_wq {
struct list_head list;
struct idxd_wq *wq;
- int ref;
- bool remove;
+ struct percpu_ref ref;
+ bool free;
bool mapped;
struct iaa_device *iaa_device;
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index 1169cd44c8e7..5cb7c930158e 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -701,7 +701,7 @@ static void del_iaa_device(struct iaa_device *iaa_device)
static void free_iaa_device(struct iaa_device *iaa_device)
{
- if (!iaa_device)
+ if (!iaa_device || iaa_device->n_wq)
return;
remove_device_compression_modes(iaa_device);
@@ -731,6 +731,13 @@ static bool iaa_has_wq(struct iaa_device *iaa_device, struct idxd_wq *wq)
return false;
}
+static void __iaa_wq_release(struct percpu_ref *ref)
+{
+ struct iaa_wq *iaa_wq = container_of(ref, typeof(*iaa_wq), ref);
+
+ iaa_wq->free = true;
+}
+
static int add_iaa_wq(struct iaa_device *iaa_device, struct idxd_wq *wq,
struct iaa_wq **new_wq)
{
@@ -738,11 +745,20 @@ static int add_iaa_wq(struct iaa_device *iaa_device, struct idxd_wq *wq,
struct pci_dev *pdev = idxd->pdev;
struct device *dev = &pdev->dev;
struct iaa_wq *iaa_wq;
+ int ret;
iaa_wq = kzalloc(sizeof(*iaa_wq), GFP_KERNEL);
if (!iaa_wq)
return -ENOMEM;
+ ret = percpu_ref_init(&iaa_wq->ref, __iaa_wq_release,
+ PERCPU_REF_INIT_ATOMIC, GFP_KERNEL);
+
+ if (ret) {
+ kfree(iaa_wq);
+ return -ENOMEM;
+ }
+
iaa_wq->wq = wq;
iaa_wq->iaa_device = iaa_device;
idxd_wq_set_private(wq, iaa_wq);
@@ -818,6 +834,9 @@ static void __free_iaa_wq(struct iaa_wq *iaa_wq)
if (!iaa_wq)
return;
+ WARN_ON(!percpu_ref_is_zero(&iaa_wq->ref));
+ percpu_ref_exit(&iaa_wq->ref);
+
iaa_device = iaa_wq->iaa_device;
if (iaa_device->n_wq == 0)
free_iaa_device(iaa_wq->iaa_device);
@@ -912,53 +931,6 @@ static int save_iaa_wq(struct idxd_wq *wq)
return 0;
}
-static int iaa_wq_get(struct idxd_wq *wq)
-{
- struct idxd_device *idxd = wq->idxd;
- struct iaa_wq *iaa_wq;
- int ret = 0;
-
- spin_lock(&idxd->dev_lock);
- iaa_wq = idxd_wq_get_private(wq);
- if (iaa_wq && !iaa_wq->remove) {
- iaa_wq->ref++;
- idxd_wq_get(wq);
- } else {
- ret = -ENODEV;
- }
- spin_unlock(&idxd->dev_lock);
-
- return ret;
-}
-
-static int iaa_wq_put(struct idxd_wq *wq)
-{
- struct idxd_device *idxd = wq->idxd;
- struct iaa_wq *iaa_wq;
- bool free = false;
- int ret = 0;
-
- spin_lock(&idxd->dev_lock);
- iaa_wq = idxd_wq_get_private(wq);
- if (iaa_wq) {
- iaa_wq->ref--;
- if (iaa_wq->ref == 0 && iaa_wq->remove) {
- idxd_wq_set_private(wq, NULL);
- free = true;
- }
- idxd_wq_put(wq);
- } else {
- ret = -ENODEV;
- }
- spin_unlock(&idxd->dev_lock);
- if (free) {
- __free_iaa_wq(iaa_wq);
- kfree(iaa_wq);
- }
-
- return ret;
-}
-
/***************************************************************
* Mapping IAA devices and wqs to cores with per-cpu wq_tables.
***************************************************************/
@@ -1765,7 +1737,7 @@ static void iaa_desc_complete(struct idxd_desc *idxd_desc,
if (free_desc)
idxd_free_desc(idxd_desc->wq, idxd_desc);
- iaa_wq_put(idxd_desc->wq);
+ percpu_ref_put(&iaa_wq->ref);
}
static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
@@ -1996,19 +1968,13 @@ static int iaa_comp_acompress(struct acomp_req *req)
cpu = get_cpu();
wq = comp_wq_table_next_wq(cpu);
put_cpu();
- if (!wq) {
- pr_debug("no wq configured for cpu=%d\n", cpu);
- return -ENODEV;
- }
- ret = iaa_wq_get(wq);
- if (ret) {
+ iaa_wq = wq ? idxd_wq_get_private(wq) : NULL;
+ if (unlikely(!iaa_wq || !percpu_ref_tryget(&iaa_wq->ref))) {
pr_debug("no wq available for cpu=%d\n", cpu);
return -ENODEV;
}
- iaa_wq = idxd_wq_get_private(wq);
-
dev = &wq->idxd->pdev->dev;
nr_sgs = dma_map_sg(dev, req->src, sg_nents(req->src), DMA_TO_DEVICE);
@@ -2061,7 +2027,7 @@ static int iaa_comp_acompress(struct acomp_req *req)
err_map_dst:
dma_unmap_sg(dev, req->src, sg_nents(req->src), DMA_TO_DEVICE);
out:
- iaa_wq_put(wq);
+ percpu_ref_put(&iaa_wq->ref);
return ret;
}
@@ -2083,19 +2049,13 @@ static int iaa_comp_adecompress(struct acomp_req *req)
cpu = get_cpu();
wq = decomp_wq_table_next_wq(cpu);
put_cpu();
- if (!wq) {
- pr_debug("no wq configured for cpu=%d\n", cpu);
- return -ENODEV;
- }
- ret = iaa_wq_get(wq);
- if (ret) {
+ iaa_wq = wq ? idxd_wq_get_private(wq) : NULL;
+ if (unlikely(!iaa_wq || !percpu_ref_tryget(&iaa_wq->ref))) {
pr_debug("no wq available for cpu=%d\n", cpu);
- return -ENODEV;
+ return deflate_generic_decompress(req);
}
- iaa_wq = idxd_wq_get_private(wq);
-
dev = &wq->idxd->pdev->dev;
nr_sgs = dma_map_sg(dev, req->src, sg_nents(req->src), DMA_TO_DEVICE);
@@ -2130,7 +2090,7 @@ static int iaa_comp_adecompress(struct acomp_req *req)
err_map_dst:
dma_unmap_sg(dev, req->src, sg_nents(req->src), DMA_TO_DEVICE);
out:
- iaa_wq_put(wq);
+ percpu_ref_put(&iaa_wq->ref);
return ret;
}
@@ -2303,7 +2263,6 @@ static void iaa_crypto_remove(struct idxd_dev *idxd_dev)
struct idxd_wq *wq = idxd_dev_to_wq(idxd_dev);
struct idxd_device *idxd = wq->idxd;
struct iaa_wq *iaa_wq;
- bool free = false;
atomic_set(&iaa_crypto_enabled, 0);
idxd_wq_quiesce(wq);
@@ -2324,18 +2283,18 @@ static void iaa_crypto_remove(struct idxd_dev *idxd_dev)
goto out;
}
- if (iaa_wq->ref) {
- iaa_wq->remove = true;
- } else {
- wq = iaa_wq->wq;
- idxd_wq_set_private(wq, NULL);
- free = true;
- }
+ /* Drop the initial reference. */
+ percpu_ref_kill(&iaa_wq->ref);
+
+ while (!iaa_wq->free)
+ cpu_relax();
+
+ __free_iaa_wq(iaa_wq);
+
+ idxd_wq_set_private(wq, NULL);
spin_unlock(&idxd->dev_lock);
- if (free) {
- __free_iaa_wq(iaa_wq);
- kfree(iaa_wq);
- }
+
+ kfree(iaa_wq);
idxd_drv_disable_wq(wq);
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 06/23] crypto: iaa - Simplify the code flow in iaa_compress() and iaa_decompress().
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (4 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 05/23] crypto: iaa - iaa_wq uses percpu_refs for get/put reference counting Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 07/23] crypto: iaa - Refactor hardware descriptor setup into separate procedures Kanchana P Sridhar
` (17 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This commit simplifies and streamlines the logic in the core
iaa_compress() and iaa_decompress() routines, eliminates branches, etc.
This makes it easier to add improvements such as polling for job
completions, essential to accomplish batching with hardware
parallelism.
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
drivers/crypto/intel/iaa/iaa_crypto_main.c | 114 ++++++++++++---------
1 file changed, 67 insertions(+), 47 deletions(-)
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index 5cb7c930158e..38b4be0c10b0 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -1792,7 +1792,34 @@ static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
desc->src2_size = sizeof(struct aecs_comp_table_record);
desc->completion_addr = idxd_desc->compl_dma;
- if (ctx->use_irq) {
+ if (likely(!ctx->use_irq)) {
+ ret = idxd_submit_desc(wq, idxd_desc);
+ if (ret) {
+ dev_dbg(dev, "submit_desc failed ret=%d\n", ret);
+ goto out;
+ }
+
+ /* Update stats */
+ update_total_comp_calls();
+ update_wq_comp_calls(wq);
+
+ if (ctx->async_mode)
+ return -EINPROGRESS;
+
+ ret = check_completion(dev, idxd_desc->iax_completion, true, false);
+ if (ret) {
+ dev_dbg(dev, "check_completion failed ret=%d\n", ret);
+ goto out;
+ }
+
+ *dlen = idxd_desc->iax_completion->output_size;
+
+ /* Update stats */
+ update_total_comp_bytes_out(*dlen);
+ update_wq_comp_bytes(wq, *dlen);
+
+ *compression_crc = idxd_desc->iax_completion->crc;
+ } else {
desc->flags |= IDXD_OP_FLAG_RCI;
idxd_desc->crypto.req = req;
@@ -1800,40 +1827,23 @@ static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
idxd_desc->crypto.src_addr = src_addr;
idxd_desc->crypto.dst_addr = dst_addr;
idxd_desc->crypto.compress = true;
- }
-
- ret = idxd_submit_desc(wq, idxd_desc);
- if (ret) {
- dev_dbg(dev, "submit_desc failed ret=%d\n", ret);
- goto err;
- }
- /* Update stats */
- update_total_comp_calls();
- update_wq_comp_calls(wq);
+ ret = idxd_submit_desc(wq, idxd_desc);
+ if (ret) {
+ dev_dbg(dev, "submit_desc failed ret=%d\n", ret);
+ goto out;
+ }
- if (ctx->async_mode) {
- ret = -EINPROGRESS;
- goto out;
- }
+ /* Update stats */
+ update_total_comp_calls();
+ update_wq_comp_calls(wq);
- ret = check_completion(dev, idxd_desc->iax_completion, true, false);
- if (ret) {
- dev_dbg(dev, "check_completion failed ret=%d\n", ret);
- goto err;
+ return -EINPROGRESS;
}
- *dlen = idxd_desc->iax_completion->output_size;
-
- /* Update stats */
- update_total_comp_bytes_out(*dlen);
- update_wq_comp_bytes(wq, *dlen);
-
- *compression_crc = idxd_desc->iax_completion->crc;
-
-err:
- idxd_free_desc(wq, idxd_desc);
out:
+ idxd_free_desc(wq, idxd_desc);
+
return ret;
}
@@ -1888,7 +1898,22 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
desc->src1_size = slen;
desc->completion_addr = idxd_desc->compl_dma;
- if (ctx->use_irq) {
+ if (likely(!ctx->use_irq)) {
+ ret = idxd_submit_desc(wq, idxd_desc);
+ if (ret) {
+ dev_dbg(dev, "submit_desc failed ret=%d\n", ret);
+ goto fallback_software_decomp;
+ }
+
+ /* Update stats */
+ update_total_decomp_calls();
+ update_wq_decomp_calls(wq);
+
+ if (ctx->async_mode)
+ return -EINPROGRESS;
+
+ ret = check_completion(dev, idxd_desc->iax_completion, false, false);
+ } else {
desc->flags |= IDXD_OP_FLAG_RCI;
idxd_desc->crypto.req = req;
@@ -1896,25 +1921,20 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
idxd_desc->crypto.src_addr = src_addr;
idxd_desc->crypto.dst_addr = dst_addr;
idxd_desc->crypto.compress = false;
- }
- ret = idxd_submit_desc(wq, idxd_desc);
- if (ret) {
- dev_dbg(dev, "submit_desc failed ret=%d\n", ret);
- goto fallback_software_decomp;
- }
+ ret = idxd_submit_desc(wq, idxd_desc);
+ if (ret) {
+ dev_dbg(dev, "submit_desc failed ret=%d\n", ret);
+ goto fallback_software_decomp;
+ }
- /* Update stats */
- update_total_decomp_calls();
- update_wq_decomp_calls(wq);
+ /* Update stats */
+ update_total_decomp_calls();
+ update_wq_decomp_calls(wq);
- if (ctx->async_mode) {
- ret = -EINPROGRESS;
- goto out;
+ return -EINPROGRESS;
}
- ret = check_completion(dev, idxd_desc->iax_completion, false, false);
-
fallback_software_decomp:
if (ret) {
dev_dbg(dev, "%s: desc allocation/submission/check_completion failed ret=%d\n", __func__, ret);
@@ -1929,7 +1949,7 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
if (ret) {
pr_err("%s: iaa decompress failed: deflate-generic fallback error ret=%d\n",
__func__, ret);
- goto err;
+ goto out;
}
} else {
req->dlen = idxd_desc->iax_completion->output_size;
@@ -1941,10 +1961,10 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
*dlen = req->dlen;
-err:
+out:
if (idxd_desc)
idxd_free_desc(wq, idxd_desc);
-out:
+
return ret;
}
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 07/23] crypto: iaa - Refactor hardware descriptor setup into separate procedures.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (5 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 06/23] crypto: iaa - Simplify the code flow in iaa_compress() and iaa_decompress() Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 08/23] crypto: iaa - Simplified, efficient job submissions for non-irq mode Kanchana P Sridhar
` (16 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch refactors the code that sets up the "struct iax_hw_desc" for
compress/decompress ops, into distinct procedures to make the code more
readable.
Also, get_iaa_device_compression_mode() is deleted and the compression
mode directly accessed from the iaa_device in the calling procedures.
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
drivers/crypto/intel/iaa/iaa_crypto_main.c | 99 ++++++++++++----------
1 file changed, 56 insertions(+), 43 deletions(-)
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index 38b4be0c10b0..c94e7abd3909 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -483,12 +483,6 @@ int add_iaa_compression_mode(const char *name,
}
EXPORT_SYMBOL_GPL(add_iaa_compression_mode);
-static struct iaa_device_compression_mode *
-get_iaa_device_compression_mode(struct iaa_device *iaa_device, int idx)
-{
- return iaa_device->compression_modes[idx];
-}
-
static void free_device_compression_mode(struct iaa_device *iaa_device,
struct iaa_device_compression_mode *device_mode)
{
@@ -1564,7 +1558,6 @@ static int iaa_compress_verify(struct crypto_tfm *tfm, struct acomp_req *req,
dma_addr_t src_addr, unsigned int slen,
dma_addr_t dst_addr, unsigned int dlen)
{
- struct iaa_device_compression_mode *active_compression_mode;
struct iaa_compression_ctx *ctx = crypto_tfm_ctx(tfm);
u32 *compression_crc = acomp_request_ctx(req);
struct iaa_device *iaa_device;
@@ -1583,8 +1576,6 @@ static int iaa_compress_verify(struct crypto_tfm *tfm, struct acomp_req *req,
pdev = idxd->pdev;
dev = &pdev->dev;
- active_compression_mode = get_iaa_device_compression_mode(iaa_device, ctx->mode);
-
while ((idxd_desc == ERR_PTR(-EAGAIN)) && (alloc_desc_retries++ < ctx->alloc_decomp_desc_timeout)) {
idxd_desc = idxd_alloc_desc(wq, IDXD_OP_NONBLOCK);
cpu_relax();
@@ -1660,8 +1651,7 @@ static void iaa_desc_complete(struct idxd_desc *idxd_desc,
pdev = idxd->pdev;
dev = &pdev->dev;
- active_compression_mode = get_iaa_device_compression_mode(iaa_device,
- compression_ctx->mode);
+ active_compression_mode = iaa_device->compression_modes[compression_ctx->mode];
dev_dbg(dev, "%s: compression mode %s,"
" ctx->src_addr %llx, ctx->dst_addr %llx\n", __func__,
active_compression_mode->name,
@@ -1740,12 +1730,63 @@ static void iaa_desc_complete(struct idxd_desc *idxd_desc,
percpu_ref_put(&iaa_wq->ref);
}
+static __always_inline struct iax_hw_desc *
+iaa_setup_compress_hw_desc(struct idxd_desc *idxd_desc,
+ dma_addr_t src_addr,
+ unsigned int slen,
+ dma_addr_t dst_addr,
+ unsigned int dlen,
+ enum iaa_mode mode,
+ struct iaa_device_compression_mode *active_compression_mode)
+{
+ struct iax_hw_desc *desc = idxd_desc->iax_hw;
+
+ desc->flags = IDXD_OP_FLAG_CRAV | IDXD_OP_FLAG_RCR | IDXD_OP_FLAG_CC;
+ desc->opcode = IAX_OPCODE_COMPRESS;
+ desc->compr_flags = IAA_COMP_FLAGS;
+ desc->priv = 0;
+
+ desc->src1_addr = (u64)src_addr;
+ desc->src1_size = slen;
+ desc->dst_addr = (u64)dst_addr;
+ desc->max_dst_size = dlen;
+ desc->flags |= IDXD_OP_FLAG_RD_SRC2_AECS;
+ desc->src2_addr = active_compression_mode->aecs_comp_table_dma_addr;
+ desc->src2_size = sizeof(struct aecs_comp_table_record);
+ desc->completion_addr = idxd_desc->compl_dma;
+
+ return desc;
+}
+
+static __always_inline struct iax_hw_desc *
+iaa_setup_decompress_hw_desc(struct idxd_desc *idxd_desc,
+ dma_addr_t src_addr,
+ unsigned int slen,
+ dma_addr_t dst_addr,
+ unsigned int dlen)
+{
+ struct iax_hw_desc *desc = idxd_desc->iax_hw;
+
+ desc->flags = IDXD_OP_FLAG_CRAV | IDXD_OP_FLAG_RCR | IDXD_OP_FLAG_CC;
+ desc->opcode = IAX_OPCODE_DECOMPRESS;
+ desc->max_dst_size = PAGE_SIZE;
+ desc->decompr_flags = IAA_DECOMP_FLAGS;
+ desc->priv = 0;
+
+ desc->src1_addr = (u64)src_addr;
+ desc->dst_addr = (u64)dst_addr;
+ desc->max_dst_size = dlen;
+ desc->src1_size = slen;
+ desc->completion_addr = idxd_desc->compl_dma;
+
+ return desc;
+}
+
static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
struct idxd_wq *wq,
dma_addr_t src_addr, unsigned int slen,
dma_addr_t dst_addr, unsigned int *dlen)
{
- struct iaa_device_compression_mode *active_compression_mode;
struct iaa_compression_ctx *ctx = crypto_tfm_ctx(tfm);
u32 *compression_crc = acomp_request_ctx(req);
struct iaa_device *iaa_device;
@@ -1764,8 +1805,6 @@ static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
pdev = idxd->pdev;
dev = &pdev->dev;
- active_compression_mode = get_iaa_device_compression_mode(iaa_device, ctx->mode);
-
while ((idxd_desc == ERR_PTR(-EAGAIN)) && (alloc_desc_retries++ < ctx->alloc_comp_desc_timeout)) {
idxd_desc = idxd_alloc_desc(wq, IDXD_OP_NONBLOCK);
cpu_relax();
@@ -1776,21 +1815,9 @@ static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
PTR_ERR(idxd_desc));
return -ENODEV;
}
- desc = idxd_desc->iax_hw;
- desc->flags = IDXD_OP_FLAG_CRAV | IDXD_OP_FLAG_RCR |
- IDXD_OP_FLAG_RD_SRC2_AECS | IDXD_OP_FLAG_CC;
- desc->opcode = IAX_OPCODE_COMPRESS;
- desc->compr_flags = IAA_COMP_FLAGS;
- desc->priv = 0;
-
- desc->src1_addr = (u64)src_addr;
- desc->src1_size = slen;
- desc->dst_addr = (u64)dst_addr;
- desc->max_dst_size = *dlen;
- desc->src2_addr = active_compression_mode->aecs_comp_table_dma_addr;
- desc->src2_size = sizeof(struct aecs_comp_table_record);
- desc->completion_addr = idxd_desc->compl_dma;
+ desc = iaa_setup_compress_hw_desc(idxd_desc, src_addr, slen, dst_addr, *dlen,
+ ctx->mode, iaa_device->compression_modes[ctx->mode]);
if (likely(!ctx->use_irq)) {
ret = idxd_submit_desc(wq, idxd_desc);
@@ -1852,7 +1879,6 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
dma_addr_t src_addr, unsigned int slen,
dma_addr_t dst_addr, unsigned int *dlen)
{
- struct iaa_device_compression_mode *active_compression_mode;
struct iaa_compression_ctx *ctx = crypto_tfm_ctx(tfm);
struct iaa_device *iaa_device;
struct idxd_desc *idxd_desc = ERR_PTR(-EAGAIN);
@@ -1870,8 +1896,6 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
pdev = idxd->pdev;
dev = &pdev->dev;
- active_compression_mode = get_iaa_device_compression_mode(iaa_device, ctx->mode);
-
while ((idxd_desc == ERR_PTR(-EAGAIN)) && (alloc_desc_retries++ < ctx->alloc_decomp_desc_timeout)) {
idxd_desc = idxd_alloc_desc(wq, IDXD_OP_NONBLOCK);
cpu_relax();
@@ -1884,19 +1908,8 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
idxd_desc = NULL;
goto fallback_software_decomp;
}
- desc = idxd_desc->iax_hw;
- desc->flags = IDXD_OP_FLAG_CRAV | IDXD_OP_FLAG_RCR | IDXD_OP_FLAG_CC;
- desc->opcode = IAX_OPCODE_DECOMPRESS;
- desc->max_dst_size = PAGE_SIZE;
- desc->decompr_flags = IAA_DECOMP_FLAGS;
- desc->priv = 0;
-
- desc->src1_addr = (u64)src_addr;
- desc->dst_addr = (u64)dst_addr;
- desc->max_dst_size = *dlen;
- desc->src1_size = slen;
- desc->completion_addr = idxd_desc->compl_dma;
+ desc = iaa_setup_decompress_hw_desc(idxd_desc, src_addr, slen, dst_addr, *dlen);
if (likely(!ctx->use_irq)) {
ret = idxd_submit_desc(wq, idxd_desc);
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 08/23] crypto: iaa - Simplified, efficient job submissions for non-irq mode.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (6 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 07/23] crypto: iaa - Refactor hardware descriptor setup into separate procedures Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 09/23] crypto: iaa - Deprecate exporting add/remove IAA compression modes Kanchana P Sridhar
` (15 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch adds a new procedure, iaa_submit_desc_movdir64b(), that
directly calls movdir64b. The core iaa_crypto routines that submit
compress and decompress jobs now invoke iaa_submit_desc_movdir64b() in
non-irq driver modes, instead of idxd_submit_desc().
idxd_submit_desc() is called only in irq mode.
This improves latency for the most commonly used iaa_crypto usage
(i.e., async non-irq) in zswap/zram by eliminating redundant computes
that would otherwise be incurred in idxd_submit_desc():
p50: -32 ns
p99: -1,048 ns
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
drivers/crypto/intel/iaa/iaa_crypto_main.c | 30 ++++++++++++++--------
1 file changed, 20 insertions(+), 10 deletions(-)
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index c94e7abd3909..cac39b418cf0 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -1782,6 +1782,24 @@ iaa_setup_decompress_hw_desc(struct idxd_desc *idxd_desc,
return desc;
}
+/*
+ * Call this for non-irq, non-enqcmds job submissions.
+ */
+static __always_inline void iaa_submit_desc_movdir64b(struct idxd_wq *wq,
+ struct idxd_desc *desc)
+{
+ void __iomem *portal = idxd_wq_portal_addr(wq);
+
+ /*
+ * The wmb() flushes writes to coherent DMA data before
+ * possibly triggering a DMA read. The wmb() is necessary
+ * even on UP because the recipient is a device.
+ */
+ wmb();
+
+ iosubmit_cmds512(portal, desc->hw, 1);
+}
+
static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
struct idxd_wq *wq,
dma_addr_t src_addr, unsigned int slen,
@@ -1820,11 +1838,7 @@ static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
ctx->mode, iaa_device->compression_modes[ctx->mode]);
if (likely(!ctx->use_irq)) {
- ret = idxd_submit_desc(wq, idxd_desc);
- if (ret) {
- dev_dbg(dev, "submit_desc failed ret=%d\n", ret);
- goto out;
- }
+ iaa_submit_desc_movdir64b(wq, idxd_desc);
/* Update stats */
update_total_comp_calls();
@@ -1912,11 +1926,7 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
desc = iaa_setup_decompress_hw_desc(idxd_desc, src_addr, slen, dst_addr, *dlen);
if (likely(!ctx->use_irq)) {
- ret = idxd_submit_desc(wq, idxd_desc);
- if (ret) {
- dev_dbg(dev, "submit_desc failed ret=%d\n", ret);
- goto fallback_software_decomp;
- }
+ iaa_submit_desc_movdir64b(wq, idxd_desc);
/* Update stats */
update_total_decomp_calls();
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 09/23] crypto: iaa - Deprecate exporting add/remove IAA compression modes.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (7 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 08/23] crypto: iaa - Simplified, efficient job submissions for non-irq mode Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 10/23] crypto: iaa - Expect a single scatterlist for a [de]compress request's src/dst Kanchana P Sridhar
` (14 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
There is no use case right now for kernel users to dynamically
add/remove IAA compression modes; hence this commit deletes the symbol
exports of add_iaa_compression_mode() and remove_iaa_compression_mode().
The only supported usage model of IAA compression modes is for the code
to be statically linked during the iaa_crypto module build,
e.g. iaa_crypto_comp_fixed.c, and for available modes to be registered
when the first IAA device wq is probed.
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
drivers/crypto/intel/iaa/iaa_crypto_main.c | 6 ------
1 file changed, 6 deletions(-)
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index cac39b418cf0..dd7c4831e092 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -367,10 +367,6 @@ static void free_iaa_compression_mode(struct iaa_compression_mode *mode)
* These tables are typically generated and captured using statistics
* collected from running actual compress/decompress workloads.
*
- * A module or other kernel code can add and remove compression modes
- * with a given name using the exported @add_iaa_compression_mode()
- * and @remove_iaa_compression_mode functions.
- *
* When a new compression mode is added, the tables are saved in a
* global compression mode list. When IAA devices are added, a
* per-IAA device dma mapping is created for each IAA device, for each
@@ -404,7 +400,6 @@ void remove_iaa_compression_mode(const char *name)
out:
mutex_unlock(&iaa_devices_lock);
}
-EXPORT_SYMBOL_GPL(remove_iaa_compression_mode);
/**
* add_iaa_compression_mode - Add an IAA compression mode
@@ -481,7 +476,6 @@ int add_iaa_compression_mode(const char *name,
free_iaa_compression_mode(mode);
goto out;
}
-EXPORT_SYMBOL_GPL(add_iaa_compression_mode);
static void free_device_compression_mode(struct iaa_device *iaa_device,
struct iaa_device_compression_mode *device_mode)
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 10/23] crypto: iaa - Expect a single scatterlist for a [de]compress request's src/dst.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (8 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 09/23] crypto: iaa - Deprecate exporting add/remove IAA compression modes Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 11/23] crypto: iaa - Rearchitect the iaa_crypto driver to be usable by zswap and zram Kanchana P Sridhar
` (13 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
The calls to dma_map_sg() were passing sg_nents() for the @nents
parameter, then error-ing out if more than one @nr_sgs were
returned. Furthermore, there are no use-cases for iaa_crypto that allow
multiple SG lists to be mapped for dma at once.
Moreover, as per Herbert's direction in [1] for the batching API from
higher mm layers to interface with crypto using SG lists, batching
within iaa_crypto will rely on there being exactly one SG list per
"unit" of [de]compression in a batch, where the component SG lists are
obtained by breaking down the @req->src and @req->dst.
Given all of the above, this patch simplifies the design by expecting
only 1 @nents in req->src and req->dst, which aligns with current and
batching use cases that will be developed in subsequent patches.
This alleviates the latency penalty of calling sg_nents() per
[de]compress op submitted to the hardware.
Some unlikely() annotations are added to conditionals in the core
[de]compress routines to further improve latency per op.
[1]: https://lore.kernel.org/all/aJ7Fk6RpNc815Ivd@gondor.apana.org.au/T/#m99aea2ce3d284e6c5a3253061d97b08c4752a798
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
drivers/crypto/intel/iaa/iaa_crypto_main.c | 54 +++++++++++-----------
1 file changed, 27 insertions(+), 27 deletions(-)
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index dd7c4831e092..16b071058f2b 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -1514,11 +1514,11 @@ static int iaa_remap_for_verify(struct device *dev, struct iaa_wq *iaa_wq,
int ret = 0;
int nr_sgs;
- dma_unmap_sg(dev, req->dst, sg_nents(req->dst), DMA_FROM_DEVICE);
- dma_unmap_sg(dev, req->src, sg_nents(req->src), DMA_TO_DEVICE);
+ dma_unmap_sg(dev, req->dst, 1, DMA_FROM_DEVICE);
+ dma_unmap_sg(dev, req->src, 1, DMA_TO_DEVICE);
- nr_sgs = dma_map_sg(dev, req->src, sg_nents(req->src), DMA_FROM_DEVICE);
- if (nr_sgs <= 0 || nr_sgs > 1) {
+ nr_sgs = dma_map_sg(dev, req->src, 1, DMA_FROM_DEVICE);
+ if (unlikely(nr_sgs <= 0 || nr_sgs > 1)) {
dev_dbg(dev, "verify: couldn't map src sg for iaa device %d,"
" wq %d: ret=%d\n", iaa_wq->iaa_device->idxd->id,
iaa_wq->wq->id, ret);
@@ -1530,13 +1530,13 @@ static int iaa_remap_for_verify(struct device *dev, struct iaa_wq *iaa_wq,
" req->slen %d, sg_dma_len(sg) %d\n", *src_addr, nr_sgs,
req->src, req->slen, sg_dma_len(req->src));
- nr_sgs = dma_map_sg(dev, req->dst, sg_nents(req->dst), DMA_TO_DEVICE);
- if (nr_sgs <= 0 || nr_sgs > 1) {
+ nr_sgs = dma_map_sg(dev, req->dst, 1, DMA_TO_DEVICE);
+ if (unlikely(nr_sgs <= 0 || nr_sgs > 1)) {
dev_dbg(dev, "verify: couldn't map dst sg for iaa device %d,"
" wq %d: ret=%d\n", iaa_wq->iaa_device->idxd->id,
iaa_wq->wq->id, ret);
ret = -EIO;
- dma_unmap_sg(dev, req->src, sg_nents(req->src), DMA_FROM_DEVICE);
+ dma_unmap_sg(dev, req->src, 1, DMA_FROM_DEVICE);
goto out;
}
*dst_addr = sg_dma_address(req->dst);
@@ -1704,14 +1704,14 @@ static void iaa_desc_complete(struct idxd_desc *idxd_desc,
err = -EIO;
}
- dma_unmap_sg(dev, ctx->req->dst, sg_nents(ctx->req->dst), DMA_TO_DEVICE);
- dma_unmap_sg(dev, ctx->req->src, sg_nents(ctx->req->src), DMA_FROM_DEVICE);
+ dma_unmap_sg(dev, ctx->req->dst, 1, DMA_TO_DEVICE);
+ dma_unmap_sg(dev, ctx->req->src, 1, DMA_FROM_DEVICE);
goto out;
}
err:
- dma_unmap_sg(dev, ctx->req->dst, sg_nents(ctx->req->dst), DMA_FROM_DEVICE);
- dma_unmap_sg(dev, ctx->req->src, sg_nents(ctx->req->src), DMA_TO_DEVICE);
+ dma_unmap_sg(dev, ctx->req->dst, 1, DMA_FROM_DEVICE);
+ dma_unmap_sg(dev, ctx->req->src, 1, DMA_TO_DEVICE);
out:
if (ret != 0)
dev_dbg(dev, "asynchronous compress failed ret=%d\n", ret);
@@ -2014,8 +2014,8 @@ static int iaa_comp_acompress(struct acomp_req *req)
dev = &wq->idxd->pdev->dev;
- nr_sgs = dma_map_sg(dev, req->src, sg_nents(req->src), DMA_TO_DEVICE);
- if (nr_sgs <= 0 || nr_sgs > 1) {
+ nr_sgs = dma_map_sg(dev, req->src, 1, DMA_TO_DEVICE);
+ if (unlikely(nr_sgs <= 0 || nr_sgs > 1)) {
dev_dbg(dev, "couldn't map src sg for iaa device %d,"
" wq %d: ret=%d\n", iaa_wq->iaa_device->idxd->id,
iaa_wq->wq->id, ret);
@@ -2024,8 +2024,8 @@ static int iaa_comp_acompress(struct acomp_req *req)
}
src_addr = sg_dma_address(req->src);
- nr_sgs = dma_map_sg(dev, req->dst, sg_nents(req->dst), DMA_FROM_DEVICE);
- if (nr_sgs <= 0 || nr_sgs > 1) {
+ nr_sgs = dma_map_sg(dev, req->dst, 1, DMA_FROM_DEVICE);
+ if (unlikely(nr_sgs <= 0 || nr_sgs > 1)) {
dev_dbg(dev, "couldn't map dst sg for iaa device %d,"
" wq %d: ret=%d\n", iaa_wq->iaa_device->idxd->id,
iaa_wq->wq->id, ret);
@@ -2051,18 +2051,18 @@ static int iaa_comp_acompress(struct acomp_req *req)
if (ret)
dev_dbg(dev, "asynchronous compress verification failed ret=%d\n", ret);
- dma_unmap_sg(dev, req->dst, sg_nents(req->dst), DMA_TO_DEVICE);
- dma_unmap_sg(dev, req->src, sg_nents(req->src), DMA_FROM_DEVICE);
+ dma_unmap_sg(dev, req->dst, 1, DMA_TO_DEVICE);
+ dma_unmap_sg(dev, req->src, 1, DMA_FROM_DEVICE);
goto out;
}
- if (ret)
+ if (unlikely(ret))
dev_dbg(dev, "asynchronous compress failed ret=%d\n", ret);
- dma_unmap_sg(dev, req->dst, sg_nents(req->dst), DMA_FROM_DEVICE);
+ dma_unmap_sg(dev, req->dst, 1, DMA_FROM_DEVICE);
err_map_dst:
- dma_unmap_sg(dev, req->src, sg_nents(req->src), DMA_TO_DEVICE);
+ dma_unmap_sg(dev, req->src, 1, DMA_TO_DEVICE);
out:
percpu_ref_put(&iaa_wq->ref);
@@ -2095,8 +2095,8 @@ static int iaa_comp_adecompress(struct acomp_req *req)
dev = &wq->idxd->pdev->dev;
- nr_sgs = dma_map_sg(dev, req->src, sg_nents(req->src), DMA_TO_DEVICE);
- if (nr_sgs <= 0 || nr_sgs > 1) {
+ nr_sgs = dma_map_sg(dev, req->src, 1, DMA_TO_DEVICE);
+ if (unlikely(nr_sgs <= 0 || nr_sgs > 1)) {
dev_dbg(dev, "couldn't map src sg for iaa device %d,"
" wq %d: ret=%d\n", iaa_wq->iaa_device->idxd->id,
iaa_wq->wq->id, ret);
@@ -2105,8 +2105,8 @@ static int iaa_comp_adecompress(struct acomp_req *req)
}
src_addr = sg_dma_address(req->src);
- nr_sgs = dma_map_sg(dev, req->dst, sg_nents(req->dst), DMA_FROM_DEVICE);
- if (nr_sgs <= 0 || nr_sgs > 1) {
+ nr_sgs = dma_map_sg(dev, req->dst, 1, DMA_FROM_DEVICE);
+ if (unlikely(nr_sgs <= 0 || nr_sgs > 1)) {
dev_dbg(dev, "couldn't map dst sg for iaa device %d,"
" wq %d: ret=%d\n", iaa_wq->iaa_device->idxd->id,
iaa_wq->wq->id, ret);
@@ -2120,12 +2120,12 @@ static int iaa_comp_adecompress(struct acomp_req *req)
if (ret == -EINPROGRESS)
return ret;
- if (ret != 0)
+ if (unlikely(ret != 0))
dev_dbg(dev, "asynchronous decompress failed ret=%d\n", ret);
- dma_unmap_sg(dev, req->dst, sg_nents(req->dst), DMA_FROM_DEVICE);
+ dma_unmap_sg(dev, req->dst, 1, DMA_FROM_DEVICE);
err_map_dst:
- dma_unmap_sg(dev, req->src, sg_nents(req->src), DMA_TO_DEVICE);
+ dma_unmap_sg(dev, req->src, 1, DMA_TO_DEVICE);
out:
percpu_ref_put(&iaa_wq->ref);
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 11/23] crypto: iaa - Rearchitect the iaa_crypto driver to be usable by zswap and zram.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (9 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 10/23] crypto: iaa - Expect a single scatterlist for a [de]compress request's src/dst Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 12/23] crypto: iaa - Enablers for submitting descriptors then polling for completion Kanchana P Sridhar
` (12 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch rearchitects the iaa_crypto driver to be usable by
non-crypto_acomp kernel users such as zram. The crypto_acomp interface
is also preserved for use by zswap. The core driver code is moved under
a crypto_acomp-agnostic layer that relies only on idxd, dma and
scatterlist.
Additionally, this patch resolves a race condition triggered when
IAA wqs and devices are continuously disabled/enabled when workloads are
using IAA for compression/decompression. This commit, in combination
with patches 0002 ("crypto: iaa - New architecture for IAA device WQ
comp/decomp usage & core mapping.) and 0005 (crypto: iaa - iaa_wq uses
percpu_refs for get/put reference counting.) in this series fix the race
condition. This has been verified using bisecting.
The newly added include/linux/iaa_comp.h provides the data structures
and API for use by non-crypto_acomp kernel code such as zram.
This allows kernel users i.e., zswap and zram, to use IAA's hardware
acceleration for compression/decompression without/with crypto_acomp.
Towards this goal, most of the driver code has been made independent of
crypto_acomp, by introducing a new "struct iaa_req" data structure, and
light-weight internal translation routines to/from crypto_acomp, namely,
acomp_to_iaa() and iaa_to_acomp().
The exception is that the driver defines a "static struct crypto_acomp
*deflate_crypto_comp" for the software decompress fall-back
path. Hopefully this shouldn't be an issue for zram because it is
encapsulated within the iaa_crypto driver.
The acomp_alg .compress() and .decompress() interfaces call into
iaa_comp_acompress_main() and iaa_comp_adecompress_main(), which are
wrappers around the core crypto-independent driver functions.
A zram/zcomp backend for iaa_crypto will be submitted as a separate
patch series, using these interfaces from iaa_comp.h:
int iaa_comp_compress(enum iaa_mode mode, struct iaa_req *req);
int iaa_comp_decompress(enum iaa_mode mode, struct iaa_req *req);
These iaa_crypto interfaces will continue to be available through
crypto_acomp for use in zswap:
int crypto_acomp_compress(struct acomp_req *req);
int crypto_acomp_decompress(struct acomp_req *req);
Some other changes introduced by this commit are:
1) iaa_crypto symbol namespace is changed from "IDXD" to
"CRYPTO_DEV_IAA_CRYPTO".
2) Some constants and data structures are moved to
include/linux/iaa_comp.h so as to be usable in developing the zram
iaa_crypto backend.
Fixes: ea7a5cbb4369 ("crypto: iaa - Add Intel IAA Compression Accelerator crypto driver core")
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
drivers/crypto/intel/iaa/Makefile | 2 +-
drivers/crypto/intel/iaa/iaa_crypto.h | 7 +-
drivers/crypto/intel/iaa/iaa_crypto_main.c | 373 ++++++++++++++++++---
include/linux/iaa_comp.h | 86 +++++
4 files changed, 406 insertions(+), 62 deletions(-)
create mode 100644 include/linux/iaa_comp.h
diff --git a/drivers/crypto/intel/iaa/Makefile b/drivers/crypto/intel/iaa/Makefile
index 55bda7770fac..ebfa1a425f80 100644
--- a/drivers/crypto/intel/iaa/Makefile
+++ b/drivers/crypto/intel/iaa/Makefile
@@ -3,7 +3,7 @@
# Makefile for IAA crypto device drivers
#
-ccflags-y += -I $(srctree)/drivers/dma/idxd -DDEFAULT_SYMBOL_NAMESPACE='"IDXD"'
+ccflags-y += -I $(srctree)/drivers/dma/idxd -DDEFAULT_SYMBOL_NAMESPACE='"CRYPTO_DEV_IAA_CRYPTO"'
obj-$(CONFIG_CRYPTO_DEV_IAA_CRYPTO) := iaa_crypto.o
diff --git a/drivers/crypto/intel/iaa/iaa_crypto.h b/drivers/crypto/intel/iaa/iaa_crypto.h
index 9611f2518f42..190157967e3b 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto.h
+++ b/drivers/crypto/intel/iaa/iaa_crypto.h
@@ -6,6 +6,7 @@
#include <linux/crypto.h>
#include <linux/idxd.h>
+#include <linux/iaa_comp.h>
#include <uapi/linux/idxd.h>
#define IDXD_SUBDRIVER_NAME "crypto"
@@ -29,8 +30,6 @@
#define IAA_ERROR_COMP_BUF_OVERFLOW 0x19
#define IAA_ERROR_WATCHDOG_EXPIRED 0x24
-#define IAA_COMP_MODES_MAX 2
-
#define FIXED_HDR 0x2
#define FIXED_HDR_SIZE 3
@@ -138,10 +137,6 @@ int add_iaa_compression_mode(const char *name,
void remove_iaa_compression_mode(const char *name);
-enum iaa_mode {
- IAA_MODE_FIXED,
-};
-
struct iaa_compression_ctx {
enum iaa_mode mode;
u16 alloc_comp_desc_timeout;
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index 16b071058f2b..f5abad950371 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -11,6 +11,7 @@
#include <linux/highmem.h>
#include <linux/sched/smt.h>
#include <crypto/internal/acompress.h>
+#include <linux/iaa_comp.h>
#include "idxd.h"
#include "iaa_crypto.h"
@@ -51,6 +52,9 @@ static struct wq_table_entry **pkg_global_decomp_wqs;
/* All comp wqs from IAAs on a package. */
static struct wq_table_entry **pkg_global_comp_wqs;
+/* For software deflate fallback compress/decompress. */
+static struct crypto_acomp *deflate_crypto_acomp;
+
LIST_HEAD(iaa_devices);
DEFINE_MUTEX(iaa_devices_lock);
@@ -93,9 +97,18 @@ static atomic_t iaa_crypto_enabled = ATOMIC_INIT(0);
static struct idxd_wq *first_wq_found;
DEFINE_MUTEX(first_wq_found_lock);
-static bool iaa_crypto_registered;
+const char *iaa_compression_mode_names[IAA_COMP_MODES_MAX] = {
+ "fixed",
+};
+
+const char *iaa_compression_alg_names[IAA_COMP_MODES_MAX] = {
+ "deflate-iaa",
+};
static struct iaa_compression_mode *iaa_compression_modes[IAA_COMP_MODES_MAX];
+static struct iaa_compression_ctx *iaa_ctx[IAA_COMP_MODES_MAX];
+static bool iaa_mode_registered[IAA_COMP_MODES_MAX];
+static u8 num_iaa_modes_registered;
/* Distribute decompressions across all IAAs on the package. */
static bool iaa_distribute_decomps;
@@ -353,6 +366,20 @@ static struct iaa_compression_mode *find_iaa_compression_mode(const char *name,
return NULL;
}
+static bool iaa_alg_is_registered(const char *name, int *idx)
+{
+ int i;
+
+ for (i = 0; i < IAA_COMP_MODES_MAX; ++i) {
+ if (!strcmp(name, iaa_compression_alg_names[i]) && iaa_mode_registered[i]) {
+ *idx = i;
+ return true;
+ }
+ }
+
+ return false;
+}
+
static void free_iaa_compression_mode(struct iaa_compression_mode *mode)
{
kfree(mode->name);
@@ -466,6 +493,7 @@ int add_iaa_compression_mode(const char *name,
mode->name, idx);
iaa_compression_modes[idx] = mode;
+ ++num_iaa_modes_registered;
ret = 0;
out:
@@ -1434,11 +1462,15 @@ static struct idxd_wq *comp_wq_table_next_wq(int cpu)
* Core iaa_crypto compress/decompress functions.
*************************************************/
-static int deflate_generic_decompress(struct acomp_req *req)
+static int deflate_generic_decompress(struct iaa_req *req)
{
- ACOMP_FBREQ_ON_STACK(fbreq, req);
+ ACOMP_REQUEST_ON_STACK(fbreq, deflate_crypto_acomp);
int ret;
+ acomp_request_set_callback(fbreq, 0, NULL, NULL);
+ acomp_request_set_params(fbreq, req->src, req->dst, req->slen,
+ PAGE_SIZE);
+
ret = crypto_acomp_decompress(fbreq);
req->dlen = fbreq->dlen;
@@ -1447,6 +1479,25 @@ static int deflate_generic_decompress(struct acomp_req *req)
return ret;
}
+static __always_inline void acomp_to_iaa(struct acomp_req *areq,
+ struct iaa_req *req,
+ struct iaa_compression_ctx *ctx)
+{
+ req->src = areq->src;
+ req->dst = areq->dst;
+ req->slen = areq->slen;
+ req->dlen = areq->dlen;
+ req->flags = areq->base.flags;
+ if (unlikely(ctx->use_irq))
+ req->drv_data = areq;
+}
+
+static __always_inline void iaa_to_acomp(int dlen, struct acomp_req *areq)
+{
+ areq->dst->length = dlen;
+ areq->dlen = dlen;
+}
+
static inline int check_completion(struct device *dev,
struct iax_completion_record *comp,
bool compress,
@@ -1508,7 +1559,7 @@ static inline int check_completion(struct device *dev,
}
static int iaa_remap_for_verify(struct device *dev, struct iaa_wq *iaa_wq,
- struct acomp_req *req,
+ struct iaa_req *req,
dma_addr_t *src_addr, dma_addr_t *dst_addr)
{
int ret = 0;
@@ -1547,13 +1598,11 @@ static int iaa_remap_for_verify(struct device *dev, struct iaa_wq *iaa_wq,
return ret;
}
-static int iaa_compress_verify(struct crypto_tfm *tfm, struct acomp_req *req,
+static int iaa_compress_verify(struct iaa_compression_ctx *ctx, struct iaa_req *req,
struct idxd_wq *wq,
dma_addr_t src_addr, unsigned int slen,
dma_addr_t dst_addr, unsigned int dlen)
{
- struct iaa_compression_ctx *ctx = crypto_tfm_ctx(tfm);
- u32 *compression_crc = acomp_request_ctx(req);
struct iaa_device *iaa_device;
struct idxd_desc *idxd_desc = ERR_PTR(-EAGAIN);
u16 alloc_desc_retries = 0;
@@ -1606,10 +1655,10 @@ static int iaa_compress_verify(struct crypto_tfm *tfm, struct acomp_req *req,
goto err;
}
- if (*compression_crc != idxd_desc->iax_completion->crc) {
+ if (req->compression_crc != idxd_desc->iax_completion->crc) {
ret = -EINVAL;
dev_dbg(dev, "(verify) iaa comp/decomp crc mismatch:"
- " comp=0x%x, decomp=0x%x\n", *compression_crc,
+ " comp=0x%x, decomp=0x%x\n", req->compression_crc,
idxd_desc->iax_completion->crc);
print_hex_dump(KERN_INFO, "cmp-rec: ", DUMP_PREFIX_OFFSET,
8, 1, idxd_desc->iax_completion, 64, 0);
@@ -1635,6 +1684,7 @@ static void iaa_desc_complete(struct idxd_desc *idxd_desc,
struct iaa_wq *iaa_wq;
struct pci_dev *pdev;
struct device *dev;
+ struct iaa_req req;
int ret, err = 0;
compression_ctx = crypto_tfm_ctx(ctx->tfm);
@@ -1660,12 +1710,18 @@ static void iaa_desc_complete(struct idxd_desc *idxd_desc,
pr_warn("%s: falling back to deflate-generic decompress, "
"analytics error code %x\n", __func__,
idxd_desc->iax_completion->error_code);
- ret = deflate_generic_decompress(ctx->req);
+
+ acomp_to_iaa(ctx->req, &req, compression_ctx);
+ ret = deflate_generic_decompress(&req);
+ iaa_to_acomp(req.dlen, ctx->req);
+
if (ret) {
dev_dbg(dev, "%s: deflate-generic failed ret=%d\n",
__func__, ret);
err = -EIO;
goto err;
+ } else {
+ goto verify;
}
} else {
err = -EIO;
@@ -1684,21 +1740,26 @@ static void iaa_desc_complete(struct idxd_desc *idxd_desc,
update_wq_decomp_bytes(iaa_wq->wq, ctx->req->slen);
}
+verify:
if (ctx->compress && compression_ctx->verify_compress) {
- u32 *compression_crc = acomp_request_ctx(ctx->req);
dma_addr_t src_addr, dst_addr;
- *compression_crc = idxd_desc->iax_completion->crc;
+ acomp_to_iaa(ctx->req, &req, compression_ctx);
+ req.compression_crc = idxd_desc->iax_completion->crc;
+
+ ret = iaa_remap_for_verify(dev, iaa_wq, &req, &src_addr, &dst_addr);
+ iaa_to_acomp(req.dlen, ctx->req);
- ret = iaa_remap_for_verify(dev, iaa_wq, ctx->req, &src_addr, &dst_addr);
if (ret) {
dev_dbg(dev, "%s: compress verify remap failed ret=%d\n", __func__, ret);
err = -EIO;
goto out;
}
- ret = iaa_compress_verify(ctx->tfm, ctx->req, iaa_wq->wq, src_addr,
+ ret = iaa_compress_verify(compression_ctx, &req, iaa_wq->wq, src_addr,
ctx->req->slen, dst_addr, ctx->req->dlen);
+ iaa_to_acomp(req.dlen, ctx->req);
+
if (ret) {
dev_dbg(dev, "%s: compress verify failed ret=%d\n", __func__, ret);
err = -EIO;
@@ -1724,7 +1785,7 @@ static void iaa_desc_complete(struct idxd_desc *idxd_desc,
percpu_ref_put(&iaa_wq->ref);
}
-static __always_inline struct iax_hw_desc *
+static struct iax_hw_desc *
iaa_setup_compress_hw_desc(struct idxd_desc *idxd_desc,
dma_addr_t src_addr,
unsigned int slen,
@@ -1752,7 +1813,7 @@ iaa_setup_compress_hw_desc(struct idxd_desc *idxd_desc,
return desc;
}
-static __always_inline struct iax_hw_desc *
+static struct iax_hw_desc *
iaa_setup_decompress_hw_desc(struct idxd_desc *idxd_desc,
dma_addr_t src_addr,
unsigned int slen,
@@ -1794,13 +1855,11 @@ static __always_inline void iaa_submit_desc_movdir64b(struct idxd_wq *wq,
iosubmit_cmds512(portal, desc->hw, 1);
}
-static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
+static int iaa_compress(struct iaa_compression_ctx *ctx, struct iaa_req *req,
struct idxd_wq *wq,
dma_addr_t src_addr, unsigned int slen,
dma_addr_t dst_addr, unsigned int *dlen)
{
- struct iaa_compression_ctx *ctx = crypto_tfm_ctx(tfm);
- u32 *compression_crc = acomp_request_ctx(req);
struct iaa_device *iaa_device;
struct idxd_desc *idxd_desc = ERR_PTR(-EAGAIN);
u16 alloc_desc_retries = 0;
@@ -1848,17 +1907,18 @@ static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
}
*dlen = idxd_desc->iax_completion->output_size;
+ req->compression_crc = idxd_desc->iax_completion->crc;
/* Update stats */
update_total_comp_bytes_out(*dlen);
update_wq_comp_bytes(wq, *dlen);
-
- *compression_crc = idxd_desc->iax_completion->crc;
} else {
+ struct acomp_req *areq = req->drv_data;
+
desc->flags |= IDXD_OP_FLAG_RCI;
- idxd_desc->crypto.req = req;
- idxd_desc->crypto.tfm = tfm;
+ idxd_desc->crypto.req = areq;
+ idxd_desc->crypto.tfm = areq->base.tfm;
idxd_desc->crypto.src_addr = src_addr;
idxd_desc->crypto.dst_addr = dst_addr;
idxd_desc->crypto.compress = true;
@@ -1882,12 +1942,11 @@ static int iaa_compress(struct crypto_tfm *tfm, struct acomp_req *req,
return ret;
}
-static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
+static int iaa_decompress(struct iaa_compression_ctx *ctx, struct iaa_req *req,
struct idxd_wq *wq,
dma_addr_t src_addr, unsigned int slen,
dma_addr_t dst_addr, unsigned int *dlen)
{
- struct iaa_compression_ctx *ctx = crypto_tfm_ctx(tfm);
struct iaa_device *iaa_device;
struct idxd_desc *idxd_desc = ERR_PTR(-EAGAIN);
u16 alloc_desc_retries = 0;
@@ -1931,10 +1990,12 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
ret = check_completion(dev, idxd_desc->iax_completion, false, false);
} else {
+ struct acomp_req *areq = req->drv_data;
+
desc->flags |= IDXD_OP_FLAG_RCI;
- idxd_desc->crypto.req = req;
- idxd_desc->crypto.tfm = tfm;
+ idxd_desc->crypto.req = areq;
+ idxd_desc->crypto.tfm = areq->base.tfm;
idxd_desc->crypto.src_addr = src_addr;
idxd_desc->crypto.dst_addr = dst_addr;
idxd_desc->crypto.compress = false;
@@ -1985,20 +2046,16 @@ static int iaa_decompress(struct crypto_tfm *tfm, struct acomp_req *req,
return ret;
}
-static int iaa_comp_acompress(struct acomp_req *req)
+static int iaa_comp_acompress(struct iaa_compression_ctx *ctx, struct iaa_req *req)
{
- struct iaa_compression_ctx *compression_ctx;
- struct crypto_tfm *tfm = req->base.tfm;
dma_addr_t src_addr, dst_addr;
int nr_sgs, cpu, ret = 0;
struct iaa_wq *iaa_wq;
struct idxd_wq *wq;
struct device *dev;
- compression_ctx = crypto_tfm_ctx(tfm);
-
- if (!req->src || !req->slen) {
- pr_debug("invalid src, not compressing\n");
+ if (!req->src || !req->slen || !req->dst) {
+ pr_debug("invalid src/dst, not compressing\n");
return -EINVAL;
}
@@ -2034,19 +2091,19 @@ static int iaa_comp_acompress(struct acomp_req *req)
}
dst_addr = sg_dma_address(req->dst);
- ret = iaa_compress(tfm, req, wq, src_addr, req->slen, dst_addr,
+ ret = iaa_compress(ctx, req, wq, src_addr, req->slen, dst_addr,
&req->dlen);
if (ret == -EINPROGRESS)
return ret;
- if (!ret && compression_ctx->verify_compress) {
+ if (!ret && ctx->verify_compress) {
ret = iaa_remap_for_verify(dev, iaa_wq, req, &src_addr, &dst_addr);
if (ret) {
dev_dbg(dev, "%s: compress verify remap failed ret=%d\n", __func__, ret);
goto out;
}
- ret = iaa_compress_verify(tfm, req, wq, src_addr, req->slen,
+ ret = iaa_compress_verify(ctx, req, wq, src_addr, req->slen,
dst_addr, req->dlen);
if (ret)
dev_dbg(dev, "asynchronous compress verification failed ret=%d\n", ret);
@@ -2069,9 +2126,8 @@ static int iaa_comp_acompress(struct acomp_req *req)
return ret;
}
-static int iaa_comp_adecompress(struct acomp_req *req)
+static int iaa_comp_adecompress(struct iaa_compression_ctx *ctx, struct iaa_req *req)
{
- struct crypto_tfm *tfm = req->base.tfm;
dma_addr_t src_addr, dst_addr;
int nr_sgs, cpu, ret = 0;
struct iaa_wq *iaa_wq;
@@ -2115,7 +2171,7 @@ static int iaa_comp_adecompress(struct acomp_req *req)
}
dst_addr = sg_dma_address(req->dst);
- ret = iaa_decompress(tfm, req, wq, src_addr, req->slen,
+ ret = iaa_decompress(ctx, req, wq, src_addr, req->slen,
dst_addr, &req->dlen);
if (ret == -EINPROGRESS)
return ret;
@@ -2132,8 +2188,9 @@ static int iaa_comp_adecompress(struct acomp_req *req)
return ret;
}
-static void compression_ctx_init(struct iaa_compression_ctx *ctx)
+static void compression_ctx_init(struct iaa_compression_ctx *ctx, enum iaa_mode mode)
{
+ ctx->mode = mode;
ctx->alloc_comp_desc_timeout = IAA_ALLOC_DESC_COMP_TIMEOUT;
ctx->alloc_decomp_desc_timeout = IAA_ALLOC_DESC_DECOMP_TIMEOUT;
ctx->verify_compress = iaa_verify_compress;
@@ -2141,26 +2198,164 @@ static void compression_ctx_init(struct iaa_compression_ctx *ctx)
ctx->use_irq = use_irq;
}
+static __always_inline bool iaa_compressor_enabled(void)
+{
+ return (atomic_read(&iaa_crypto_enabled) && num_iaa_modes_registered);
+}
+
+static __always_inline enum iaa_mode iaa_compressor_is_registered(const char *compressor_name)
+{
+ u8 i;
+
+ if (!atomic_read(&iaa_crypto_enabled) || !num_iaa_modes_registered)
+ return IAA_MODE_NONE;
+
+ for (i = 0; i < IAA_COMP_MODES_MAX; ++i) {
+ if (iaa_mode_registered[i] &&
+ !strcmp(iaa_compression_alg_names[i], compressor_name))
+ return (enum iaa_mode)i;
+ }
+
+ return IAA_MODE_NONE;
+}
+
+/***********************************************************
+ * Interfaces for non-crypto_acomp kernel users, e.g. zram.
+ ***********************************************************/
+
+__always_inline bool iaa_comp_enabled(void)
+{
+ return iaa_compressor_enabled();
+}
+EXPORT_SYMBOL_GPL(iaa_comp_enabled);
+
+__always_inline enum iaa_mode iaa_comp_get_compressor_mode(const char *compressor_name)
+{
+ return iaa_compressor_is_registered(compressor_name);
+}
+EXPORT_SYMBOL_GPL(iaa_comp_get_compressor_mode);
+
+__always_inline bool iaa_comp_mode_is_registered(enum iaa_mode mode)
+{
+ return iaa_mode_registered[mode];
+}
+EXPORT_SYMBOL_GPL(iaa_comp_mode_is_registered);
+
+void iaa_comp_put_modes(char **iaa_mode_names, enum iaa_mode *iaa_modes, u8 nr_modes)
+{
+ u8 i;
+
+ if (iaa_mode_names) {
+ for (i = 0; i < nr_modes; ++i)
+ kfree(iaa_mode_names[i]);
+ kfree(iaa_mode_names);
+ }
+
+ kfree(iaa_modes);
+}
+EXPORT_SYMBOL_GPL(iaa_comp_put_modes);
+
+u8 iaa_comp_get_modes(char **iaa_mode_names, enum iaa_mode *iaa_modes)
+{
+ u8 i, nr_modes = 0;
+
+ if (!atomic_read(&iaa_crypto_enabled) || !num_iaa_modes_registered)
+ return 0;
+
+ iaa_mode_names = kcalloc(num_iaa_modes_registered, sizeof(char *), GFP_KERNEL);
+ if (!iaa_mode_names)
+ goto err;
+
+ iaa_modes = kcalloc(num_iaa_modes_registered, sizeof(enum iaa_mode), GFP_KERNEL);
+ if (!iaa_modes)
+ goto err;
+
+ for (i = 0; i < IAA_COMP_MODES_MAX; ++i) {
+ if (iaa_mode_registered[i]) {
+ iaa_mode_names[nr_modes] = kzalloc(sizeof(char) * 30, GFP_KERNEL);
+ if (!iaa_mode_names[nr_modes])
+ goto err;
+ strscpy(iaa_mode_names[nr_modes], iaa_compression_alg_names[i],
+ sizeof(iaa_mode_names[nr_modes]));
+ iaa_modes[nr_modes] = (enum iaa_mode)nr_modes;
+ ++nr_modes;
+ }
+ }
+
+ return nr_modes;
+
+err:
+ iaa_comp_put_modes(iaa_mode_names, iaa_modes, num_iaa_modes_registered);
+ return 0;
+}
+EXPORT_SYMBOL_GPL(iaa_comp_get_modes);
+
+__always_inline int iaa_comp_compress(enum iaa_mode mode, struct iaa_req *req)
+{
+ return iaa_comp_acompress(iaa_ctx[mode], req);
+}
+EXPORT_SYMBOL_GPL(iaa_comp_compress);
+
+__always_inline int iaa_comp_decompress(enum iaa_mode mode, struct iaa_req *req)
+{
+ return iaa_comp_adecompress(iaa_ctx[mode], req);
+}
+EXPORT_SYMBOL_GPL(iaa_comp_decompress);
+
/*********************************************
* Interfaces to crypto_alg and crypto_acomp.
*********************************************/
+static int iaa_comp_acompress_main(struct acomp_req *areq)
+{
+ struct crypto_tfm *tfm = areq->base.tfm;
+ struct iaa_compression_ctx *ctx;
+ struct iaa_req req;
+ int ret = -ENODEV, idx;
+
+ if (iaa_alg_is_registered(crypto_tfm_alg_driver_name(tfm), &idx)) {
+ ctx = iaa_ctx[idx];
+
+ acomp_to_iaa(areq, &req, ctx);
+ ret = iaa_comp_acompress(ctx, &req);
+ iaa_to_acomp(unlikely(ret) ? ret : req.dlen, areq);
+ }
+
+ return ret;
+}
+
+static int iaa_comp_adecompress_main(struct acomp_req *areq)
+{
+ struct crypto_tfm *tfm = areq->base.tfm;
+ struct iaa_compression_ctx *ctx;
+ struct iaa_req req;
+ int ret = -ENODEV, idx;
+
+ if (iaa_alg_is_registered(crypto_tfm_alg_driver_name(tfm), &idx)) {
+ ctx = iaa_ctx[idx];
+
+ acomp_to_iaa(areq, &req, ctx);
+ ret = iaa_comp_adecompress(ctx, &req);
+ iaa_to_acomp(unlikely(ret) ? ret : req.dlen, areq);
+ }
+
+ return ret;
+}
+
static int iaa_comp_init_fixed(struct crypto_acomp *acomp_tfm)
{
struct crypto_tfm *tfm = crypto_acomp_tfm(acomp_tfm);
struct iaa_compression_ctx *ctx = crypto_tfm_ctx(tfm);
- ctx->mode = IAA_MODE_FIXED;
-
- compression_ctx_init(ctx);
+ ctx = iaa_ctx[IAA_MODE_FIXED];
return 0;
}
static struct acomp_alg iaa_acomp_fixed_deflate = {
.init = iaa_comp_init_fixed,
- .compress = iaa_comp_acompress,
- .decompress = iaa_comp_adecompress,
+ .compress = iaa_comp_acompress_main,
+ .decompress = iaa_comp_adecompress_main,
.base = {
.cra_name = "deflate",
.cra_driver_name = "deflate-iaa",
@@ -2172,29 +2367,89 @@ static struct acomp_alg iaa_acomp_fixed_deflate = {
}
};
+/*******************************************
+ * Implement idxd_device_driver interfaces.
+ *******************************************/
+
+static void iaa_unregister_compression_device(void)
+{
+ unsigned int i;
+
+ atomic_set(&iaa_crypto_enabled, 0);
+
+ for (i = 0; i < IAA_COMP_MODES_MAX; ++i) {
+ iaa_mode_registered[i] = false;
+ kfree(iaa_ctx[i]);
+ iaa_ctx[i] = NULL;
+ }
+
+ num_iaa_modes_registered = 0;
+}
+
static int iaa_register_compression_device(void)
{
- int ret;
+ struct iaa_compression_mode *mode;
+ int i, idx;
+
+ for (i = 0; i < IAA_COMP_MODES_MAX; ++i) {
+ iaa_mode_registered[i] = false;
+ mode = find_iaa_compression_mode(iaa_compression_mode_names[i], &idx);
+ if (mode) {
+ iaa_ctx[i] = kmalloc(sizeof(struct iaa_compression_ctx), GFP_KERNEL);
+ if (!iaa_ctx[i])
+ goto err;
+
+ compression_ctx_init(iaa_ctx[i], (enum iaa_mode)i);
+ iaa_mode_registered[i] = true;
+ }
+ }
+
+ BUG_ON(!iaa_mode_registered[IAA_MODE_FIXED]);
+ return 0;
+
+err:
+ iaa_unregister_compression_device();
+ return -ENODEV;
+}
+
+static int iaa_register_acomp_compression_device(void)
+{
+ int ret = -ENOMEM;
+
+ deflate_crypto_acomp = crypto_alloc_acomp("deflate", 0, 0);
+ if (IS_ERR_OR_NULL(deflate_crypto_acomp))
+ goto err_deflate_acomp;
ret = crypto_register_acomp(&iaa_acomp_fixed_deflate);
if (ret) {
pr_err("deflate algorithm acomp fixed registration failed (%d)\n", ret);
- goto out;
+ goto err_fixed;
}
- iaa_crypto_registered = true;
-out:
+ return 0;
+
+err_fixed:
+ if (!IS_ERR_OR_NULL(deflate_crypto_acomp)) {
+ crypto_free_acomp(deflate_crypto_acomp);
+ deflate_crypto_acomp = NULL;
+ }
+
+err_deflate_acomp:
+ iaa_unregister_compression_device();
return ret;
}
-static int iaa_unregister_compression_device(void)
+static void iaa_unregister_acomp_compression_device(void)
{
atomic_set(&iaa_crypto_enabled, 0);
- if (iaa_crypto_registered)
+ if (iaa_mode_registered[IAA_MODE_FIXED])
crypto_unregister_acomp(&iaa_acomp_fixed_deflate);
- return 0;
+ if (!IS_ERR_OR_NULL(deflate_crypto_acomp)) {
+ crypto_free_acomp(deflate_crypto_acomp);
+ deflate_crypto_acomp = NULL;
+ }
}
static int iaa_crypto_probe(struct idxd_dev *idxd_dev)
@@ -2264,6 +2519,12 @@ static int iaa_crypto_probe(struct idxd_dev *idxd_dev)
goto err_register;
}
+ ret = iaa_register_acomp_compression_device();
+ if (ret != 0) {
+ dev_dbg(dev, "IAA compression device acomp registration failed\n");
+ goto err_register;
+ }
+
if (!rebalance_wq_table()) {
dev_dbg(dev, "%s: Rerun after registration: IAA rebalancing device wq tables failed\n", __func__);
goto err_register;
@@ -2340,6 +2601,8 @@ static void iaa_crypto_remove(struct idxd_dev *idxd_dev)
pkg_global_wqs_dealloc();
free_wq_tables();
BUG_ON(!list_empty(&iaa_devices));
+ iaa_unregister_acomp_compression_device();
+ iaa_unregister_compression_device();
INIT_LIST_HEAD(&iaa_devices);
module_put(THIS_MODULE);
@@ -2456,8 +2719,8 @@ static int __init iaa_crypto_init_module(void)
static void __exit iaa_crypto_cleanup_module(void)
{
- if (iaa_unregister_compression_device())
- pr_debug("IAA compression device unregister failed\n");
+ iaa_unregister_acomp_compression_device();
+ iaa_unregister_compression_device();
iaa_crypto_debugfs_cleanup();
driver_remove_file(&iaa_crypto_driver.drv,
diff --git a/include/linux/iaa_comp.h b/include/linux/iaa_comp.h
new file mode 100644
index 000000000000..ec061315f477
--- /dev/null
+++ b/include/linux/iaa_comp.h
@@ -0,0 +1,86 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/* Copyright(c) 2021 Intel Corporation. All rights rsvd. */
+
+#ifndef __IAA_COMP_H__
+#define __IAA_COMP_H__
+
+#if IS_ENABLED(CONFIG_CRYPTO_DEV_IAA_CRYPTO)
+
+#include <linux/scatterlist.h>
+
+#define IAA_COMP_MODES_MAX IAA_MODE_NONE
+
+enum iaa_mode {
+ IAA_MODE_FIXED = 0,
+ IAA_MODE_NONE = 1,
+};
+
+struct iaa_req {
+ struct scatterlist *src;
+ struct scatterlist *dst;
+ unsigned int slen;
+ unsigned int dlen;
+ u32 flags;
+ u32 compression_crc;
+ void *drv_data; /* for driver internal use */
+};
+
+extern bool iaa_comp_enabled(void);
+
+extern enum iaa_mode iaa_comp_get_compressor_mode(const char *compressor_name);
+
+extern bool iaa_comp_mode_is_registered(enum iaa_mode mode);
+
+extern u8 iaa_comp_get_modes(char **iaa_mode_names, enum iaa_mode *iaa_modes);
+
+extern void iaa_comp_put_modes(char **iaa_mode_names, enum iaa_mode *iaa_modes, u8 nr_modes);
+
+extern int iaa_comp_compress(enum iaa_mode mode, struct iaa_req *req);
+
+extern int iaa_comp_decompress(enum iaa_mode mode, struct iaa_req *req);
+
+#else /* CONFIG_CRYPTO_DEV_IAA_CRYPTO */
+
+enum iaa_mode {
+ IAA_MODE_NONE = 1,
+};
+
+struct iaa_req {};
+
+static inline bool iaa_comp_enabled(void)
+{
+ return false;
+}
+
+static inline enum iaa_mode iaa_comp_get_compressor_mode(const char *compressor_name)
+{
+ return IAA_MODE_NONE;
+}
+
+static inline bool iaa_comp_mode_is_registered(enum iaa_mode mode)
+{
+ return false;
+}
+
+static inline u8 iaa_comp_get_modes(char **iaa_mode_names, enum iaa_mode *iaa_modes)
+{
+ return 0;
+}
+
+static inline void iaa_comp_put_modes(char **iaa_mode_names, enum iaa_mode *iaa_modes, u8 nr_modes)
+{
+}
+
+static inline int iaa_comp_compress(enum iaa_mode mode, struct iaa_req *req)
+{
+ return -EINVAL;
+}
+
+static inline int iaa_comp_decompress(enum iaa_mode mode, struct iaa_req *req)
+{
+ return -EINVAL;
+}
+
+#endif /* CONFIG_CRYPTO_DEV_IAA_CRYPTO */
+
+#endif
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 12/23] crypto: iaa - Enablers for submitting descriptors then polling for completion.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (10 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 11/23] crypto: iaa - Rearchitect the iaa_crypto driver to be usable by zswap and zram Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 13/23] crypto: acomp - Define a unit_size in struct acomp_req to enable batching Kanchana P Sridhar
` (11 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch adds capabilities in the IAA driver for kernel users to avail
of the benefits of compressing/decompressing multiple jobs in parallel
using IAA hardware acceleration, without the use of interrupts. Instead,
this is accomplished using an async "submit-poll" mechanism.
To achieve this, we break down a compress/decompress job into two
separate activities if the driver is configured for non-irq async mode:
1) Submit a descriptor after caching the "idxd_desc" descriptor in the
req->drv_data, and return -EINPROGRESS.
2) Poll: Given a request, retrieve the descriptor and poll its completion
status for success/error.
This is enabled by the following additions in the driver:
1) The idxd_desc is cached in the "drv_data" member of "struct iaa_req".
2) IAA_REQ_POLL_FLAG: if set in the iaa_req's flags, this tells
the driver that it should submit the descriptor and return
-EINPROGRESS. If not set, the driver will proceed to call
check_completion() in fully synchronous mode, until the hardware
returns a completion status.
3) iaa_comp_poll() procedure: This routine is intended to be called
after submission returns -EINPROGRESS. It will check the completion
status once, and return -EAGAIN if the job has not completed. If the
job has completed, it will return the completion status.
The purpose of this commit is to allow kernel users of iaa_crypto, such
as zswap, to be able to invoke the crypto_acomp_compress() API in fully
synchronous mode for sequential/non-batching use cases (i.e. today's
status-quo), wherein zswap calls:
crypto_wait_req(crypto_acomp_compress(req), wait);
and to non-instrusively invoke the fully asynchronous batch
compress/decompress functionality that will be introduced in subsequent
patches. Both use cases need to reuse same code paths in the driver to
interface with hardware: the IAA_REQ_POLL_FLAG allows this
shared code to determine whether we need to process an iaa_req
synchronously/asynchronously. The idea is to simplify iaa_crypto's
sequential/batching interfaces for use by zswap and zram.
Thus, regardless of the iaa_crypto driver's 'sync_mode' setting, it
can still be forced to use synchronous mode by *not setting* the
IAA_REQ_POLL_FLAG in iaa_req->flags: this is the default to support
sequential use cases in zswap today.
When IAA batching functionality is introduced subsequently, it will set
the IAA_REQ_POLL_FLAG for the requests in a batch. We will submit the
descriptors for each request in the batch in iaa_[de]compress(), and
return -EINPROGRESS. The hardware will begin processing each request as
soon as it is submitted, essentially all compress/decompress jobs will
be parallelized. The polling function, "iaa_comp_poll()", will retrieve
the descriptor from each iaa_req->drv_data to check its completion
status. This enables the iaa_crypto driver to implement true async
"submit-polling" for parallel compressions and decompressions in the IAA
hardware accelerator.
To summarize, both these conditions need to be met for a request to be
processed in fully async submit-poll mode:
1) use_irq should be "false"
2) iaa_req->flags & IAA_REQ_POLL_FLAG should be "true"
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
drivers/crypto/intel/iaa/iaa_crypto.h | 6 ++
drivers/crypto/intel/iaa/iaa_crypto_main.c | 71 +++++++++++++++++++++-
2 files changed, 75 insertions(+), 2 deletions(-)
diff --git a/drivers/crypto/intel/iaa/iaa_crypto.h b/drivers/crypto/intel/iaa/iaa_crypto.h
index 190157967e3b..1cc383c94fb8 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto.h
+++ b/drivers/crypto/intel/iaa/iaa_crypto.h
@@ -41,6 +41,12 @@
IAA_DECOMP_CHECK_FOR_EOB | \
IAA_DECOMP_STOP_ON_EOB)
+/*
+ * If set, the driver must have a way to submit the req, then
+ * poll its completion status for success/error.
+ */
+#define IAA_REQ_POLL_FLAG 0x00000002
+
/* Representation of IAA workqueue */
struct iaa_wq {
struct list_head list;
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index f5abad950371..7395822430b1 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -1891,13 +1891,14 @@ static int iaa_compress(struct iaa_compression_ctx *ctx, struct iaa_req *req,
ctx->mode, iaa_device->compression_modes[ctx->mode]);
if (likely(!ctx->use_irq)) {
+ req->drv_data = idxd_desc;
iaa_submit_desc_movdir64b(wq, idxd_desc);
/* Update stats */
update_total_comp_calls();
update_wq_comp_calls(wq);
- if (ctx->async_mode)
+ if (req->flags & IAA_REQ_POLL_FLAG)
return -EINPROGRESS;
ret = check_completion(dev, idxd_desc->iax_completion, true, false);
@@ -1979,13 +1980,14 @@ static int iaa_decompress(struct iaa_compression_ctx *ctx, struct iaa_req *req,
desc = iaa_setup_decompress_hw_desc(idxd_desc, src_addr, slen, dst_addr, *dlen);
if (likely(!ctx->use_irq)) {
+ req->drv_data = idxd_desc;
iaa_submit_desc_movdir64b(wq, idxd_desc);
/* Update stats */
update_total_decomp_calls();
update_wq_decomp_calls(wq);
- if (ctx->async_mode)
+ if (req->flags & IAA_REQ_POLL_FLAG)
return -EINPROGRESS;
ret = check_completion(dev, idxd_desc->iax_completion, false, false);
@@ -2188,6 +2190,71 @@ static int iaa_comp_adecompress(struct iaa_compression_ctx *ctx, struct iaa_req
return ret;
}
+static int __maybe_unused iaa_comp_poll(struct iaa_compression_ctx *ctx, struct iaa_req *req)
+{
+ struct idxd_desc *idxd_desc;
+ struct idxd_device *idxd;
+ struct iaa_wq *iaa_wq;
+ struct pci_dev *pdev;
+ struct device *dev;
+ struct idxd_wq *wq;
+ bool compress_op;
+ int ret;
+
+ idxd_desc = req->drv_data;
+ if (!idxd_desc)
+ return -EAGAIN;
+
+ compress_op = (idxd_desc->iax_hw->opcode == IAX_OPCODE_COMPRESS);
+ wq = idxd_desc->wq;
+ iaa_wq = idxd_wq_get_private(wq);
+ idxd = iaa_wq->iaa_device->idxd;
+ pdev = idxd->pdev;
+ dev = &pdev->dev;
+
+ ret = check_completion(dev, idxd_desc->iax_completion, compress_op, true);
+ if (ret == -EAGAIN)
+ return ret;
+ if (ret)
+ goto out;
+
+ req->dlen = idxd_desc->iax_completion->output_size;
+
+ /* Update stats */
+ if (compress_op) {
+ update_total_comp_bytes_out(req->dlen);
+ update_wq_comp_bytes(wq, req->dlen);
+ } else {
+ update_total_decomp_bytes_in(req->slen);
+ update_wq_decomp_bytes(wq, req->slen);
+ }
+
+ if (compress_op && ctx->verify_compress) {
+ dma_addr_t src_addr, dst_addr;
+
+ req->compression_crc = idxd_desc->iax_completion->crc;
+
+ dma_sync_sg_for_device(dev, req->dst, 1, DMA_FROM_DEVICE);
+ dma_sync_sg_for_device(dev, req->src, 1, DMA_TO_DEVICE);
+
+ src_addr = sg_dma_address(req->src);
+ dst_addr = sg_dma_address(req->dst);
+
+ ret = iaa_compress_verify(ctx, req, wq, src_addr, req->slen,
+ dst_addr, req->dlen);
+ }
+
+out:
+ /* caller doesn't call crypto_wait_req, so no acomp_request_complete() */
+ dma_unmap_sg(dev, req->dst, 1, DMA_FROM_DEVICE);
+ dma_unmap_sg(dev, req->src, 1, DMA_TO_DEVICE);
+
+ idxd_free_desc(idxd_desc->wq, idxd_desc);
+ percpu_ref_put(&iaa_wq->ref);
+
+ return ret;
+}
+
static void compression_ctx_init(struct iaa_compression_ctx *ctx, enum iaa_mode mode)
{
ctx->mode = mode;
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 13/23] crypto: acomp - Define a unit_size in struct acomp_req to enable batching.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (11 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 12/23] crypto: iaa - Enablers for submitting descriptors then polling for completion Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 14/23] crypto: iaa - IAA Batching for parallel compressions/decompressions Kanchana P Sridhar
` (10 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
We add a new @unit_size data member to struct acomp_req along with a
helper function acomp_request_set_unit_size() for kernel modules to set
the unit size to use while breaking down the request's src/dst
scatterlists.
An acomp_alg can implement batching by using the @req->unit_size to
break down the SG lists passed in via @req->dst and/or @req->src, to
submit individual @req->slen/@req->unit_size compress jobs or
@req->dlen/@req->unit_size decompress jobs, for batch compression and
batch decompression respectively.
In case of batch compression, the folio's pages for the batch can be
retrieved from the @req->src scatterlist by using an struct sg_page_iter
after determining the number of pages as @req->slen/@req->unit_size.
Suggested-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
include/crypto/acompress.h | 36 ++++++++++++++++++++++++++++++++++++
1 file changed, 36 insertions(+)
diff --git a/include/crypto/acompress.h b/include/crypto/acompress.h
index 9eacb9fa375d..0f1334168f1b 100644
--- a/include/crypto/acompress.h
+++ b/include/crypto/acompress.h
@@ -79,6 +79,7 @@ struct acomp_req_chain {
* @dvirt: Destination virtual address
* @slen: Size of the input buffer
* @dlen: Size of the output buffer and number of bytes produced
+ * @unit_size: Unit size for the request for use in batching
* @chain: Private API code data, do not use
* @__ctx: Start of private context data
*/
@@ -94,6 +95,7 @@ struct acomp_req {
};
unsigned int slen;
unsigned int dlen;
+ unsigned int unit_size;
struct acomp_req_chain chain;
@@ -328,9 +330,43 @@ static inline void acomp_request_set_callback(struct acomp_req *req,
{
flgs &= ~CRYPTO_ACOMP_REQ_PRIVATE;
flgs |= req->base.flags & CRYPTO_ACOMP_REQ_PRIVATE;
+ req->unit_size = 0;
crypto_request_set_callback(&req->base, flgs, cmpl, data);
}
+/**
+ * acomp_request_set_unit_size() -- Sets the unit size for the request.
+ *
+ * As suggested by Herbert Xu, this is a new helper function that enables
+ * batching for zswap, IPComp, etc.
+ *
+ * Example usage model:
+ *
+ * A module like zswap that wants to use batch compression of @nr_pages with
+ * crypto_acomp must create an output SG table for the batch, initialized to
+ * contain @nr_pages SG lists. Each scatterlist is mapped to the nth
+ * destination buffer for the batch.
+ *
+ * An acomp_alg can implement batching by using the @req->unit_size to
+ * break down the SG lists passed in via @req->dst and/or @req->src, to
+ * submit individual @req->slen/@req->unit_size compress jobs or
+ * @req->dlen/@req->unit_size decompress jobs, for batch compression and
+ * batch decompression respectively.
+ *
+ * This API must be called after acomp_request_set_callback(),
+ * which sets @req->unit_size to 0.
+ *
+ * @du would be PAGE_SIZE for zswap, it could be the MTU for IPsec.
+ *
+ * @req: asynchronous compress request
+ * @du: data unit size of the input buffer scatterlist.
+ */
+static inline void acomp_request_set_unit_size(struct acomp_req *req,
+ unsigned int du)
+{
+ req->unit_size = du;
+}
+
/**
* acomp_request_set_params() -- Sets request parameters
*
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 14/23] crypto: iaa - IAA Batching for parallel compressions/decompressions.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (12 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 13/23] crypto: acomp - Define a unit_size in struct acomp_req to enable batching Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-10-17 1:09 ` Herbert Xu
2025-09-26 3:34 ` [PATCH v12 15/23] crypto: iaa - Enable async mode and make it the default Kanchana P Sridhar
` (9 subsequent siblings)
23 siblings, 1 reply; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch introduces batch compressions/decompressions in
iaa_crypto. Two new interfaces are provided for use in the kernel,
either directly, in the zram/zcomp backend, or by calling
crypto_acomp_[de]compress() in the case of zswap.
IAA Batching allows the kernel swap modules to compress/decompress
multiple pages/buffers in parallel in hardware, significantly improving
swapout/swapin latency and throughput.
The patch defines an iaa_crypto constant, IAA_CRYPTO_MAX_BATCH_SIZE
(set to 8U currently). This is the maximum batch-size for IAA, and
represents the maximum number of pages/buffers that can be
compressed/decompressed in parallel, respectively.
In order to support IAA batching, the iaa_crypto driver allocates
IAA_CRYPTO_MAX_BATCH_SIZE "struct iaa_req *reqs[]" per-CPU, upon
initialization. Notably, the task of allocating multiple requests to
submit to the hardware for parallel [de]compressions is taken over by
iaa_crypto, so that zswap/zram don't need to allocate the reqs.
Compress batching is expected to be called by kernel modules such as
zswap by passing the folio pages for the "source" SG list of the
acomp_req, and by constructing an SG table of SG lists for the output
buffers and setting the acomp_req's "dst" to the head of this list of
scatterlists. Thanks to Herbert Xu for suggesting this batching
architecture.
Within the iaa_crypto driver's compress batching function:
1) The per-CPU iaa_reqs are populated from the acomp_req's src/dst SG
lists.
2) All iaa_reqs are submitted to the hardware in async mode, using
movdir64b. This enables hardware parallelism, because we don't wait
for one compress/decompress job to finish before submitting the next
one.
3) The iaa_reqs submitted are polled for completion statuses in a
non-blocking manner in a while loop: each request that is still
pending is polled once, and this repeats, until all requests have
completed.
IAA's maximum batch-size can be queried with the following API:
unsigned int iaa_comp_get_max_batch_size(void);
This allows swap modules such as zram to allocate required batching
dst buffers and then invoke fully asynchronous batch parallel
compression/decompression of pages/buffers on systems with Intel IAA, by
invoking these batching API, respectively:
int iaa_comp_compress_batch(
enum iaa_mode mode,
struct iaa_req *parent_req,
unsigned int unit_size);
int iaa_comp_decompress_batch(
enum iaa_mode mode,
struct iaa_req *parent_req,
unsigned int unit_size);
The parameter @unit_size represents the unit size in bytes, for
dis-assembling the source/destination
@parent_req->slen/@parent_req->dlen and SG lists passed in through
@parent_req->src and @parent_req->dst.
A zram/zcomp backend_deflate_iaa.c will be submitted as a separate patch
series, and will enable single-page and batch IAA compress/decompress
ops.
The zswap interface to these batching API will be done by setting up the
acomp_req through these crypto API:
acomp_request_set_src_folio()
acomp_request_set_dst_sg()
acomp_request_set_unit_size()
before proceeding to invoke batch compression/decompression using the
existing crypto_acomp_compress()/crypto_acomp_decompress() interfaces.
The new crypto_acomp-agnostic iaa_comp_[de]compress_batch() API result
in impressive latency improvements for zswap batch [de]compression, as
compared to a crypto_acomp based batching interface, most likely because
we avoid the overhead of crypto_acomp: we observe 17.78 micro-seconds
p99 latency savings for a decompress batch of 8 with the new
iaa_comp_decompress_batch() API.
Suggested-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
drivers/crypto/intel/iaa/iaa_crypto.h | 15 +
drivers/crypto/intel/iaa/iaa_crypto_main.c | 346 ++++++++++++++++++++-
include/linux/iaa_comp.h | 35 +++
3 files changed, 388 insertions(+), 8 deletions(-)
diff --git a/drivers/crypto/intel/iaa/iaa_crypto.h b/drivers/crypto/intel/iaa/iaa_crypto.h
index 1cc383c94fb8..db1e50574662 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto.h
+++ b/drivers/crypto/intel/iaa/iaa_crypto.h
@@ -47,6 +47,21 @@
*/
#define IAA_REQ_POLL_FLAG 0x00000002
+/*
+ * The maximum compress/decompress batch size for IAA's batch compression
+ * and batch decompression functionality.
+ */
+#define IAA_CRYPTO_MAX_BATCH_SIZE 8U
+
+/*
+ * Used to create per-CPU structure comprising of IAA_CRYPTO_MAX_BATCH_SIZE
+ * reqs for batch [de]compressions.
+ */
+struct iaa_batch_ctx {
+ struct iaa_req **reqs;
+ struct mutex mutex;
+};
+
/* Representation of IAA workqueue */
struct iaa_wq {
struct list_head list;
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index 7395822430b1..0a620f2dc58e 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -55,6 +55,9 @@ static struct wq_table_entry **pkg_global_comp_wqs;
/* For software deflate fallback compress/decompress. */
static struct crypto_acomp *deflate_crypto_acomp;
+/* Per-cpu iaa_reqs for batching. */
+static struct iaa_batch_ctx __percpu *iaa_batch_ctx;
+
LIST_HEAD(iaa_devices);
DEFINE_MUTEX(iaa_devices_lock);
@@ -2190,7 +2193,7 @@ static int iaa_comp_adecompress(struct iaa_compression_ctx *ctx, struct iaa_req
return ret;
}
-static int __maybe_unused iaa_comp_poll(struct iaa_compression_ctx *ctx, struct iaa_req *req)
+static int iaa_comp_poll(struct iaa_compression_ctx *ctx, struct iaa_req *req)
{
struct idxd_desc *idxd_desc;
struct idxd_device *idxd;
@@ -2255,6 +2258,234 @@ static int __maybe_unused iaa_comp_poll(struct iaa_compression_ctx *ctx, struct
return ret;
}
+static __always_inline void iaa_set_req_poll(
+ struct iaa_req *reqs[],
+ int nr_reqs,
+ bool set_flag)
+{
+ int i;
+
+ for (i = 0; i < nr_reqs; ++i) {
+ set_flag ? (reqs[i]->flags |= IAA_REQ_POLL_FLAG) :
+ (reqs[i]->flags &= ~IAA_REQ_POLL_FLAG);
+ }
+}
+
+/**
+ * This API provides IAA compress batching functionality for use by swap
+ * modules.
+ *
+ * @ctx: compression ctx for the requested IAA mode (fixed/dynamic).
+ * @parent_req: The "parent" iaa_req that contains SG lists for the batch's
+ * inputs and outputs.
+ * @unit_size: The unit size to apply to @parent_req->slen to get the number of
+ * scatterlists it contains.
+ *
+ * The caller should check the individual sg->lengths in the @parent_req for
+ * errors, including incompressible page errors.
+ *
+ * Returns 0 if all compress requests in the batch complete successfully,
+ * -EINVAL otherwise.
+ */
+static int iaa_comp_acompress_batch(
+ struct iaa_compression_ctx *ctx,
+ struct iaa_req *parent_req,
+ unsigned int unit_size)
+{
+ struct iaa_batch_ctx *cpu_ctx = raw_cpu_ptr(iaa_batch_ctx);
+ int nr_reqs = parent_req->slen / unit_size;
+ int errors[IAA_CRYPTO_MAX_BATCH_SIZE];
+ bool compressions_done = false;
+ struct sg_page_iter sgiter;
+ struct scatterlist *sg;
+ struct iaa_req **reqs;
+ int i, err = 0;
+
+ mutex_lock(&cpu_ctx->mutex);
+
+ reqs = cpu_ctx->reqs;
+
+ __sg_page_iter_start(&sgiter, parent_req->src, nr_reqs,
+ parent_req->src->offset/unit_size);
+
+ for (i = 0; i < nr_reqs; ++i, ++sgiter.sg_pgoffset) {
+ sg_set_page(reqs[i]->src, sg_page_iter_page(&sgiter), PAGE_SIZE, 0);
+ reqs[i]->slen = PAGE_SIZE;
+ }
+
+ for_each_sg(parent_req->dst, sg, nr_reqs, i) {
+ sg->length = PAGE_SIZE;
+ parent_req->dlens[i] = &sg->length;
+ reqs[i]->dst = sg;
+ reqs[i]->dlen = PAGE_SIZE;
+ }
+
+ iaa_set_req_poll(reqs, nr_reqs, true);
+
+ /*
+ * Prepare and submit the batch of iaa_reqs to IAA. IAA will process
+ * these compress jobs in parallel.
+ */
+ for (i = 0; i < nr_reqs; ++i) {
+ errors[i] = iaa_comp_acompress(ctx, reqs[i]);
+
+ if (likely(errors[i] == -EINPROGRESS)) {
+ errors[i] = -EAGAIN;
+ } else if (unlikely(errors[i])) {
+ *parent_req->dlens[i] = errors[i];
+ err = -EINVAL;
+ } else {
+ *parent_req->dlens[i] = reqs[i]->dlen;
+ }
+ }
+
+ /*
+ * Asynchronously poll for and process IAA compress job completions.
+ */
+ while (!compressions_done) {
+ compressions_done = true;
+
+ for (i = 0; i < nr_reqs; ++i) {
+ /*
+ * Skip, if the compression has already completed
+ * successfully or with an error.
+ */
+ if (errors[i] != -EAGAIN)
+ continue;
+
+ errors[i] = iaa_comp_poll(ctx, reqs[i]);
+
+ if (errors[i]) {
+ if (likely(errors[i] == -EAGAIN)) {
+ compressions_done = false;
+ } else {
+ *parent_req->dlens[i] = errors[i];
+ err = -EINVAL;
+ }
+ } else {
+ *parent_req->dlens[i] = reqs[i]->dlen;
+ }
+ }
+ }
+
+ /*
+ * For the same 'reqs[]' to be usable by
+ * iaa_comp_acompress()/iaa_comp_adecompress(),
+ * clear the IAA_REQ_POLL_FLAG bit on all iaa_reqs.
+ */
+ iaa_set_req_poll(reqs, nr_reqs, false);
+
+ mutex_unlock(&cpu_ctx->mutex);
+ return err;
+}
+
+/**
+ * This API provides IAA decompress batching functionality for use by swap
+ * modules.
+ *
+ * @ctx: compression ctx for the requested IAA mode (fixed/dynamic).
+ * @parent_req: The "parent" iaa_req that contains SG lists for the batch's
+ * inputs and outputs.
+ * @unit_size: The unit size to apply to @parent_req->dlen to get the number of
+ * scatterlists it contains.
+ *
+ * The caller should check @parent_req->dst scatterlist's component SG lists'
+ * @length for errors and handle @length != PAGE_SIZE.
+ *
+ * Returns 0 if all decompress requests complete successfully,
+ * -EINVAL otherwise.
+ */
+static int iaa_comp_adecompress_batch(
+ struct iaa_compression_ctx *ctx,
+ struct iaa_req *parent_req,
+ unsigned int unit_size)
+{
+ struct iaa_batch_ctx *cpu_ctx = raw_cpu_ptr(iaa_batch_ctx);
+ int nr_reqs = parent_req->dlen / unit_size;
+ int errors[IAA_CRYPTO_MAX_BATCH_SIZE];
+ bool decompressions_done = false;
+ struct scatterlist *sg;
+ struct iaa_req **reqs;
+ int i, err = 0;
+
+ mutex_lock(&cpu_ctx->mutex);
+
+ reqs = cpu_ctx->reqs;
+
+ for_each_sg(parent_req->src, sg, nr_reqs, i) {
+ reqs[i]->src = sg;
+ reqs[i]->slen = sg->length;
+ }
+
+ for_each_sg(parent_req->dst, sg, nr_reqs, i) {
+ parent_req->dlens[i] = &sg->length;
+ reqs[i]->dst = sg;
+ reqs[i]->dlen = PAGE_SIZE;
+ }
+
+ iaa_set_req_poll(reqs, nr_reqs, true);
+
+ /*
+ * Prepare and submit the batch of iaa_reqs to IAA. IAA will process
+ * these decompress jobs in parallel.
+ */
+ for (i = 0; i < nr_reqs; ++i) {
+ errors[i] = iaa_comp_adecompress(ctx, reqs[i]);
+
+ /*
+ * If it failed desc allocation/submission, errors[i] can
+ * be 0 or error value from software decompress.
+ */
+ if (likely(errors[i] == -EINPROGRESS)) {
+ errors[i] = -EAGAIN;
+ } else if (unlikely(errors[i])) {
+ *parent_req->dlens[i] = errors[i];
+ err = -EINVAL;
+ } else {
+ *parent_req->dlens[i] = reqs[i]->dlen;
+ }
+ }
+
+ /*
+ * Asynchronously poll for and process IAA decompress job completions.
+ */
+ while (!decompressions_done) {
+ decompressions_done = true;
+
+ for (i = 0; i < nr_reqs; ++i) {
+ /*
+ * Skip, if the decompression has already completed
+ * successfully or with an error.
+ */
+ if (errors[i] != -EAGAIN)
+ continue;
+
+ errors[i] = iaa_comp_poll(ctx, reqs[i]);
+
+ if (errors[i]) {
+ if (likely(errors[i] == -EAGAIN)) {
+ decompressions_done = false;
+ } else {
+ *parent_req->dlens[i] = errors[i];
+ err = -EINVAL;
+ }
+ } else {
+ *parent_req->dlens[i] = reqs[i]->dlen;
+ }
+ }
+ }
+
+ /*
+ * For the same 'reqs[]' to be usable by
+ * iaa_comp_acompress()/iaa_comp_adecompress(),
+ * clear the IAA_REQ_POLL_FLAG bit on all iaa_reqs.
+ */
+ iaa_set_req_poll(reqs, nr_reqs, false);
+
+ mutex_unlock(&cpu_ctx->mutex);
+ return err;
+}
+
static void compression_ctx_init(struct iaa_compression_ctx *ctx, enum iaa_mode mode)
{
ctx->mode = mode;
@@ -2357,6 +2588,12 @@ u8 iaa_comp_get_modes(char **iaa_mode_names, enum iaa_mode *iaa_modes)
}
EXPORT_SYMBOL_GPL(iaa_comp_get_modes);
+__always_inline unsigned int iaa_comp_get_max_batch_size(void)
+{
+ return IAA_CRYPTO_MAX_BATCH_SIZE;
+}
+EXPORT_SYMBOL_GPL(iaa_comp_get_max_batch_size);
+
__always_inline int iaa_comp_compress(enum iaa_mode mode, struct iaa_req *req)
{
return iaa_comp_acompress(iaa_ctx[mode], req);
@@ -2369,6 +2606,24 @@ __always_inline int iaa_comp_decompress(enum iaa_mode mode, struct iaa_req *req)
}
EXPORT_SYMBOL_GPL(iaa_comp_decompress);
+__always_inline int iaa_comp_compress_batch(
+ enum iaa_mode mode,
+ struct iaa_req *parent_req,
+ unsigned int unit_size)
+{
+ return iaa_comp_acompress_batch(iaa_ctx[mode], parent_req, unit_size);
+}
+EXPORT_SYMBOL_GPL(iaa_comp_compress_batch);
+
+__always_inline int iaa_comp_decompress_batch(
+ enum iaa_mode mode,
+ struct iaa_req *parent_req,
+ unsigned int unit_size)
+{
+ return iaa_comp_adecompress_batch(iaa_ctx[mode], parent_req, unit_size);
+}
+EXPORT_SYMBOL_GPL(iaa_comp_decompress_batch);
+
/*********************************************
* Interfaces to crypto_alg and crypto_acomp.
*********************************************/
@@ -2383,9 +2638,16 @@ static int iaa_comp_acompress_main(struct acomp_req *areq)
if (iaa_alg_is_registered(crypto_tfm_alg_driver_name(tfm), &idx)) {
ctx = iaa_ctx[idx];
- acomp_to_iaa(areq, &req, ctx);
- ret = iaa_comp_acompress(ctx, &req);
- iaa_to_acomp(unlikely(ret) ? ret : req.dlen, areq);
+ if (likely(areq->slen == areq->unit_size)) {
+ acomp_to_iaa(areq, &req, ctx);
+ ret = iaa_comp_acompress(ctx, &req);
+ iaa_to_acomp(unlikely(ret) ? ret : req.dlen, areq);
+ return ret;
+ } else {
+ acomp_to_iaa(areq, &req, ctx);
+ ret = iaa_comp_acompress_batch(ctx, &req, areq->unit_size);
+ return ret;
+ }
}
return ret;
@@ -2401,9 +2663,16 @@ static int iaa_comp_adecompress_main(struct acomp_req *areq)
if (iaa_alg_is_registered(crypto_tfm_alg_driver_name(tfm), &idx)) {
ctx = iaa_ctx[idx];
- acomp_to_iaa(areq, &req, ctx);
- ret = iaa_comp_adecompress(ctx, &req);
- iaa_to_acomp(unlikely(ret) ? ret : req.dlen, areq);
+ if (likely(areq->dlen == areq->unit_size)) {
+ acomp_to_iaa(areq, &req, ctx);
+ ret = iaa_comp_adecompress(ctx, &req);
+ iaa_to_acomp(unlikely(ret) ? ret : req.dlen, areq);
+ return ret;
+ } else {
+ acomp_to_iaa(areq, &req, ctx);
+ ret = iaa_comp_adecompress_batch(ctx, &req, areq->unit_size);
+ return ret;
+ }
}
return ret;
@@ -2699,9 +2968,31 @@ static struct idxd_device_driver iaa_crypto_driver = {
* Module init/exit.
********************/
+static void iaa_batch_ctx_dealloc(void)
+{
+ int cpu;
+ u8 i;
+
+ if (!iaa_batch_ctx)
+ return;
+
+ for (cpu = 0; cpu < nr_cpus; cpu++) {
+ struct iaa_batch_ctx *cpu_ctx = per_cpu_ptr(iaa_batch_ctx, cpu);
+
+ if (cpu_ctx && cpu_ctx->reqs) {
+ for (i = 0; i < IAA_CRYPTO_MAX_BATCH_SIZE; ++i)
+ kfree(cpu_ctx->reqs[i]);
+ kfree(cpu_ctx->reqs);
+ }
+ }
+
+ free_percpu(iaa_batch_ctx);
+}
+
static int __init iaa_crypto_init_module(void)
{
- int ret = 0;
+ int cpu, ret = 0;
+ u8 i;
INIT_LIST_HEAD(&iaa_devices);
@@ -2756,6 +3047,39 @@ static int __init iaa_crypto_init_module(void)
goto err_sync_attr_create;
}
+ /* Allocate batching resources for iaa_crypto. */
+ iaa_batch_ctx = alloc_percpu_gfp(struct iaa_batch_ctx, GFP_KERNEL | __GFP_ZERO);
+ if (!iaa_batch_ctx) {
+ pr_debug("Failed to allocate per-cpu iaa_batch_ctx\n");
+ goto batch_ctx_fail;
+ }
+
+ for (cpu = 0; cpu < nr_cpus; cpu++) {
+ struct iaa_batch_ctx *cpu_ctx = per_cpu_ptr(iaa_batch_ctx, cpu);
+ int cpu_node = cpu_to_node(cpu);
+
+ cpu_ctx->reqs = kcalloc_node(IAA_CRYPTO_MAX_BATCH_SIZE,
+ sizeof(struct iaa_req *),
+ GFP_KERNEL, cpu_node);
+
+ if (!cpu_ctx->reqs)
+ goto reqs_fail;
+
+ for (i = 0; i < IAA_CRYPTO_MAX_BATCH_SIZE; ++i) {
+ cpu_ctx->reqs[i] = kzalloc_node(sizeof(struct iaa_req),
+ GFP_KERNEL, cpu_node);
+ if (!cpu_ctx->reqs[i]) {
+ pr_debug("Could not alloc iaa_req reqs[%d]\n", i);
+ goto reqs_fail;
+ }
+
+ sg_init_table(&cpu_ctx->reqs[i]->sg_src, 1);
+ cpu_ctx->reqs[i]->src = &cpu_ctx->reqs[i]->sg_src;
+ }
+
+ mutex_init(&cpu_ctx->mutex);
+ }
+
if (iaa_crypto_debugfs_init())
pr_warn("debugfs init failed, stats not available\n");
@@ -2763,6 +3087,11 @@ static int __init iaa_crypto_init_module(void)
out:
return ret;
+reqs_fail:
+ iaa_batch_ctx_dealloc();
+batch_ctx_fail:
+ driver_remove_file(&iaa_crypto_driver.drv,
+ &driver_attr_sync_mode);
err_sync_attr_create:
driver_remove_file(&iaa_crypto_driver.drv,
&driver_attr_verify_compress);
@@ -2789,6 +3118,7 @@ static void __exit iaa_crypto_cleanup_module(void)
iaa_unregister_acomp_compression_device();
iaa_unregister_compression_device();
+ iaa_batch_ctx_dealloc();
iaa_crypto_debugfs_cleanup();
driver_remove_file(&iaa_crypto_driver.drv,
&driver_attr_sync_mode);
diff --git a/include/linux/iaa_comp.h b/include/linux/iaa_comp.h
index ec061315f477..7b765760485c 100644
--- a/include/linux/iaa_comp.h
+++ b/include/linux/iaa_comp.h
@@ -18,11 +18,13 @@ enum iaa_mode {
struct iaa_req {
struct scatterlist *src;
struct scatterlist *dst;
+ struct scatterlist sg_src;
unsigned int slen;
unsigned int dlen;
u32 flags;
u32 compression_crc;
void *drv_data; /* for driver internal use */
+ int **dlens;
};
extern bool iaa_comp_enabled(void);
@@ -35,10 +37,22 @@ extern u8 iaa_comp_get_modes(char **iaa_mode_names, enum iaa_mode *iaa_modes);
extern void iaa_comp_put_modes(char **iaa_mode_names, enum iaa_mode *iaa_modes, u8 nr_modes);
+extern unsigned int iaa_comp_get_max_batch_size(void);
+
extern int iaa_comp_compress(enum iaa_mode mode, struct iaa_req *req);
extern int iaa_comp_decompress(enum iaa_mode mode, struct iaa_req *req);
+extern int iaa_comp_compress_batch(
+ enum iaa_mode mode,
+ struct iaa_req *parent_req,
+ unsigned int unit_size);
+
+extern int iaa_comp_decompress_batch(
+ enum iaa_mode mode,
+ struct iaa_req *parent_req,
+ unsigned int unit_size);
+
#else /* CONFIG_CRYPTO_DEV_IAA_CRYPTO */
enum iaa_mode {
@@ -71,6 +85,11 @@ static inline void iaa_comp_put_modes(char **iaa_mode_names, enum iaa_mode *iaa_
{
}
+static inline unsigned int iaa_comp_get_max_batch_size(void)
+{
+ return 0;
+}
+
static inline int iaa_comp_compress(enum iaa_mode mode, struct iaa_req *req)
{
return -EINVAL;
@@ -81,6 +100,22 @@ static inline int iaa_comp_decompress(enum iaa_mode mode, struct iaa_req *req)
return -EINVAL;
}
+static inline int iaa_comp_compress_batch(
+ enum iaa_mode mode,
+ struct iaa_req *parent_req,
+ unsigned int unit_size)
+{
+ return -EINVAL;
+}
+
+static inline int iaa_comp_decompress_batch(
+ enum iaa_mode mode,
+ struct iaa_req *parent_req,
+ unsigned int unit_size)
+{
+ return -EINVAL;
+}
+
#endif /* CONFIG_CRYPTO_DEV_IAA_CRYPTO */
#endif
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 15/23] crypto: iaa - Enable async mode and make it the default.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (13 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 14/23] crypto: iaa - IAA Batching for parallel compressions/decompressions Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 16/23] crypto: iaa - Disable iaa_verify_compress by default Kanchana P Sridhar
` (8 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch enables the 'async' sync_mode in the driver. Further, it sets
the default sync_mode to 'async', which makes it easier for IAA hardware
acceleration in the iaa_crypto driver to be loaded by default in the most
efficient/recommended 'async' mode for parallel
compressions/decompressions, namely, asynchronous submission of
descriptors, followed by polling for job completions. Earlier, the
"sync" mode used to be the default.
The iaa_crypto driver documentation has been updated with these
changes.
This way, anyone who wants to use IAA for zswap/zram can do so after
building the kernel, and without having to go through these steps to use
async mode:
1) disable all the IAA device/wq bindings that happen at boot time
2) rmmod iaa_crypto
3) modprobe iaa_crypto
4) echo async > /sys/bus/dsa/drivers/crypto/sync_mode
5) re-run initialization of the IAA devices and wqs
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
Documentation/driver-api/crypto/iaa/iaa-crypto.rst | 11 ++---------
drivers/crypto/intel/iaa/iaa_crypto_main.c | 4 ++--
2 files changed, 4 insertions(+), 11 deletions(-)
diff --git a/Documentation/driver-api/crypto/iaa/iaa-crypto.rst b/Documentation/driver-api/crypto/iaa/iaa-crypto.rst
index 0ff4ec603b43..d5e610ef4612 100644
--- a/Documentation/driver-api/crypto/iaa/iaa-crypto.rst
+++ b/Documentation/driver-api/crypto/iaa/iaa-crypto.rst
@@ -272,7 +272,7 @@ The available attributes are:
echo async_irq > /sys/bus/dsa/drivers/crypto/sync_mode
Async mode without interrupts (caller must poll) can be enabled by
- writing 'async' to it (please see Caveat)::
+ writing 'async' to it::
echo async > /sys/bus/dsa/drivers/crypto/sync_mode
@@ -281,14 +281,7 @@ The available attributes are:
echo sync > /sys/bus/dsa/drivers/crypto/sync_mode
- The default mode is 'sync'.
-
- Caveat: since the only mechanism that iaa_crypto currently implements
- for async polling without interrupts is via the 'sync' mode as
- described earlier, writing 'async' to
- '/sys/bus/dsa/drivers/crypto/sync_mode' will internally enable the
- 'sync' mode. This is to ensure correct iaa_crypto behavior until true
- async polling without interrupts is enabled in iaa_crypto.
+ The default mode is 'async'.
- g_comp_wqs_per_iaa
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index 0a620f2dc58e..c4f40984e9bf 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -152,7 +152,7 @@ static bool iaa_verify_compress = true;
*/
/* Use async mode */
-static bool async_mode;
+static bool async_mode = true;
/* Use interrupts */
static bool use_irq;
@@ -206,7 +206,7 @@ static int set_iaa_sync_mode(const char *name)
async_mode = false;
use_irq = false;
} else if (sysfs_streq(name, "async")) {
- async_mode = false;
+ async_mode = true;
use_irq = false;
} else if (sysfs_streq(name, "async_irq")) {
async_mode = true;
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 16/23] crypto: iaa - Disable iaa_verify_compress by default.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (14 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 15/23] crypto: iaa - Enable async mode and make it the default Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 17/23] crypto: iaa - Submit the two largest source buffers first in decompress batching Kanchana P Sridhar
` (7 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch makes it easier for IAA hardware acceleration in the iaa_crypto
driver to be loaded by default with "iaa_verify_compress" disabled, to
facilitate performance comparisons with software compressors (which also
do not run compress verification by default). Earlier, iaa_crypto compress
verification used to be enabled by default.
The iaa_crypto driver documentation has been updated with this change.
With this patch, if users want to enable compress verification, they can do
so with these steps:
1) disable all the IAA device/wq bindings that happen at boot time
2) rmmod iaa_crypto
3) modprobe iaa_crypto
4) echo 1 > /sys/bus/dsa/drivers/crypto/verify_compress
5) re-run initialization of the IAA devices and wqs
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
Documentation/driver-api/crypto/iaa/iaa-crypto.rst | 2 +-
drivers/crypto/intel/iaa/iaa_crypto_main.c | 2 +-
2 files changed, 2 insertions(+), 2 deletions(-)
diff --git a/Documentation/driver-api/crypto/iaa/iaa-crypto.rst b/Documentation/driver-api/crypto/iaa/iaa-crypto.rst
index d5e610ef4612..81a7dbd15f8b 100644
--- a/Documentation/driver-api/crypto/iaa/iaa-crypto.rst
+++ b/Documentation/driver-api/crypto/iaa/iaa-crypto.rst
@@ -239,7 +239,7 @@ The available attributes are:
echo 0 > /sys/bus/dsa/drivers/crypto/verify_compress
- The default setting is '1' - verify all compresses.
+ The default setting is '0' - to not verify compresses.
- sync_mode
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index c4f40984e9bf..5b933c138e50 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -120,7 +120,7 @@ static bool iaa_distribute_decomps;
static bool iaa_distribute_comps = true;
/* Verify results of IAA compress or not */
-static bool iaa_verify_compress = true;
+static bool iaa_verify_compress;
/*
* The iaa crypto driver supports three 'sync' methods determining how
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 17/23] crypto: iaa - Submit the two largest source buffers first in decompress batching.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (15 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 16/23] crypto: iaa - Disable iaa_verify_compress by default Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 18/23] crypto: iaa - Add deflate-iaa-dynamic compression mode Kanchana P Sridhar
` (6 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch finds the two largest source buffers in a given decompression
batch, and submits them first to the IAA decompress engines.
This improves decompress batching latency because the hardware has a
head start on decompressing the highest latency source buffers in the
batch. Workload performance is also significantly improved as a result
of this optimization.
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
drivers/crypto/intel/iaa/iaa_crypto_main.c | 61 +++++++++++++++++++++-
1 file changed, 59 insertions(+), 2 deletions(-)
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index 5b933c138e50..0669ae155e90 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -2379,6 +2379,36 @@ static int iaa_comp_acompress_batch(
return err;
}
+/*
+ * Find the two largest source buffers in @slens for a decompress batch,
+ * and pass their indices back in @idx_max and @idx_next_max.
+ *
+ * Returns true if there is no second largest source buffer, only a max buffer.
+ */
+static bool decomp_batch_get_max_slens_idx(
+ struct iaa_req *reqs[],
+ int nr_pages,
+ int *idx_max,
+ int *idx_next_max)
+{
+ int i, max_i = 0, next_max_i = 0;
+
+ for (i = 0; i < nr_pages; ++i) {
+ if (reqs[i]->slen >= reqs[max_i]->slen) {
+ next_max_i = max_i;
+ max_i = i;
+ } else if ((next_max_i == max_i) ||
+ (reqs[i]->slen > reqs[next_max_i]->slen)) {
+ next_max_i = i;
+ }
+ }
+
+ *idx_max = max_i;
+ *idx_next_max = next_max_i;
+
+ return (next_max_i == max_i);
+}
+
/**
* This API provides IAA decompress batching functionality for use by swap
* modules.
@@ -2401,12 +2431,13 @@ static int iaa_comp_adecompress_batch(
unsigned int unit_size)
{
struct iaa_batch_ctx *cpu_ctx = raw_cpu_ptr(iaa_batch_ctx);
+ bool max_processed = false, next_max_processed = false;
int nr_reqs = parent_req->dlen / unit_size;
int errors[IAA_CRYPTO_MAX_BATCH_SIZE];
+ int i = 0, max_i, next_max_i, err = 0;
bool decompressions_done = false;
struct scatterlist *sg;
struct iaa_req **reqs;
- int i, err = 0;
mutex_lock(&cpu_ctx->mutex);
@@ -2425,11 +2456,28 @@ static int iaa_comp_adecompress_batch(
iaa_set_req_poll(reqs, nr_reqs, true);
+ /*
+ * Get the indices of the two largest decomp buffers in the batch.
+ * Submit them first. This improves latency of the batch.
+ */
+ next_max_processed = decomp_batch_get_max_slens_idx(reqs, nr_reqs,
+ &max_i, &next_max_i);
+
+ i = max_i;
+
/*
* Prepare and submit the batch of iaa_reqs to IAA. IAA will process
* these decompress jobs in parallel.
*/
- for (i = 0; i < nr_reqs; ++i) {
+ for (; i < nr_reqs; ++i) {
+ if ((i == max_i) && max_processed)
+ continue;
+ if ((i == next_max_i) && max_processed && next_max_processed)
+ continue;
+
+ if (max_processed && !next_max_processed)
+ i = next_max_i;
+
errors[i] = iaa_comp_adecompress(ctx, reqs[i]);
/*
@@ -2444,6 +2492,15 @@ static int iaa_comp_adecompress_batch(
} else {
*parent_req->dlens[i] = reqs[i]->dlen;
}
+
+ if (i == max_i) {
+ max_processed = true;
+ i = -1;
+ }
+ if (i == next_max_i) {
+ next_max_processed = true;
+ i = -1;
+ }
}
/*
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 18/23] crypto: iaa - Add deflate-iaa-dynamic compression mode.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (16 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 17/23] crypto: iaa - Submit the two largest source buffers first in decompress batching Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 19/23] crypto: acomp - Add crypto_acomp_batch_size() to get an algorithm's batch-size Kanchana P Sridhar
` (5 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
Some versions of Intel IAA support dynamic compression where the hardware
dynamically computes the Huffman tables and generates a Deflate header
if the input size is no larger than 4KB. This patch will use IAA for
dynamic compression if an appropriate IAA is present and the input size is
not too big. If an IAA is not present, the algorithm will not
be available. Otherwise, if the size of the input is greater than
PAGE_SIZE, zlib is used to do the compression. If the algorithm is
selected, IAA will be used for decompression. If the compressed stream
contains a reference whose distance is greater than 4KB, hardware
decompression will fail, and the decompression will be done with zlib.
Intel IAA dynamic compression results in a compression ratio that is
better than or equal to the currently supported "fixed" compression mode
on the same data set. Compressing a data set of 4300 4KB pages sampled
from SPEC CPU17 workloads produces a compression ratio of 3.14 for IAA
dynamic compression and 2.69 for IAA fixed compression.
If an appropriate IAA exists, dynamic mode can be chosen as the IAA
compression mode by selecting the corresponding algorithm.
For example, to use IAA dynamic mode in zswap:
echo deflate-iaa-dynamic > /sys/module/zswap/parameters/compressor
This patch also adds a deflate_generic_compress() fallback when dynamic
mode is selected and the input size is over 4KB; along with stats
support that will count these software fallback calls as
"total_sw_comp_calls" in the driver's global_stats.
Furthermore, we define IAA_DYN_ALLOC_DESC_COMP_TIMEOUT as 2000 for
dynamic mode compression on Granite Rapids.
Signed-off-by: Andre Glover <andre.glover@linux.intel.com>
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
.../driver-api/crypto/iaa/iaa-crypto.rst | 21 ++++
crypto/testmgr.c | 10 ++
crypto/testmgr.h | 74 ++++++++++++++
drivers/crypto/intel/iaa/Makefile | 2 +-
drivers/crypto/intel/iaa/iaa_crypto.h | 5 +
.../intel/iaa/iaa_crypto_comp_dynamic.c | 22 +++++
drivers/crypto/intel/iaa/iaa_crypto_main.c | 98 +++++++++++++++++--
drivers/crypto/intel/iaa/iaa_crypto_stats.c | 8 ++
drivers/crypto/intel/iaa/iaa_crypto_stats.h | 2 +
include/linux/iaa_comp.h | 5 +-
10 files changed, 236 insertions(+), 11 deletions(-)
create mode 100644 drivers/crypto/intel/iaa/iaa_crypto_comp_dynamic.c
diff --git a/Documentation/driver-api/crypto/iaa/iaa-crypto.rst b/Documentation/driver-api/crypto/iaa/iaa-crypto.rst
index 81a7dbd15f8b..e841a33564db 100644
--- a/Documentation/driver-api/crypto/iaa/iaa-crypto.rst
+++ b/Documentation/driver-api/crypto/iaa/iaa-crypto.rst
@@ -33,6 +33,8 @@ compresses and decompresses.
Currently, there is only one compression modes available, 'fixed'
mode.
+'dynamic' mode is available on certain generations of IAA hardware.
+
The 'fixed' compression mode implements the compression scheme
specified by RFC 1951 and is given the crypto algorithm name
'deflate-iaa'. (Because the IAA hardware has a 4k history-window
@@ -43,6 +45,25 @@ the IAA fixed mode deflate algorithm is given its own algorithm name
rather than simply 'deflate').
+The 'dynamic' compression mode implements a compression scheme where
+the IAA hardware will internally do one pass through the data, compute the
+Huffman tables and generate a Deflate header, then automatically do a
+second pass through the data, generating the final compressed output. IAA
+dynamic compression can be used if an appropriate IAA is present and the
+input size is not too big. If an appropriate IAA is not present, the
+algorithm will not be available. Otherwise, if the size of the input is too
+big, zlib is used to do the compression. If the algorithm is selected,
+IAA will be used for decompression. If the compressed stream contains a
+reference whose distance is greater than 4KB, hardware decompression will
+fail, and the decompression will be done with zlib. If an appropriate IAA
+exists, 'dynamic' compression, it is implemented by the
+'deflate-iaa-dynamic' crypto algorithm.
+
+A zswap device can select the IAA 'dynamic' mode represented by
+selecting the 'deflate-iaa-dynamic' crypto compression algorithm::
+
+ # echo deflate-iaa-dynamic> /sys/module/zswap/parameters/compressor
+
Config options and other setup
==============================
diff --git a/crypto/testmgr.c b/crypto/testmgr.c
index ee33ba21ae2b..d0f271ea1201 100644
--- a/crypto/testmgr.c
+++ b/crypto/testmgr.c
@@ -4665,6 +4665,16 @@ static const struct alg_test_desc alg_test_descs[] = {
.decomp = __VECS(deflate_decomp_tv_template)
}
}
+ }, {
+ .alg = "deflate-iaa-dynamic",
+ .test = alg_test_comp,
+ .fips_allowed = 1,
+ .suite = {
+ .comp = {
+ .comp = __VECS(deflate_iaa_dynamic_comp_tv_template),
+ .decomp = __VECS(deflate_iaa_dynamic_decomp_tv_template)
+ }
+ }
}, {
.alg = "dh",
.test = alg_test_kpp,
diff --git a/crypto/testmgr.h b/crypto/testmgr.h
index 32d099ac9e73..42db2399013e 100644
--- a/crypto/testmgr.h
+++ b/crypto/testmgr.h
@@ -34575,6 +34575,80 @@ static const struct comp_testvec deflate_decomp_tv_template[] = {
},
};
+static const struct comp_testvec deflate_iaa_dynamic_comp_tv_template[] = {
+ {
+ .inlen = 70,
+ .outlen = 46,
+ .input = "Join us now and share the software "
+ "Join us now and share the software ",
+ .output = "\x85\xca\xc1\x09\x00\x20\x08\x05"
+ "\xd0\x55\xfe\x3c\x6e\x21\x64\xd8"
+ "\x45\x21\x0d\xd7\xb7\x26\xe8\xf8"
+ "\xe0\x91\x2f\xc3\x09\x98\x17\xd8"
+ "\x06\x42\x79\x0b\x52\x05\xe1\x33"
+ "\xeb\x81\x3e\xe5\xa2\x01",
+ }, {
+ .inlen = 191,
+ .outlen = 121,
+ .input = "This document describes a compression method based on the DEFLATE"
+ "compression algorithm. This document defines the application of "
+ "the DEFLATE algorithm to the IP Payload Compression Protocol.",
+ .output = "\x5d\x8d\xc1\x0d\xc2\x30\x10\x04"
+ "\x5b\xd9\x0a\xd2\x03\x82\x20\x21"
+ "\xf1\xf0\x23\x0d\x5c\xec\x0b\xb6"
+ "\x64\xfb\x2c\xdf\xf1\xa0\x7b\x12"
+ "\x3e\x58\x79\xae\x76\x67\x76\x89"
+ "\x49\x11\xc4\xbf\x0b\x57\x43\x60"
+ "\xf5\x3d\xad\xac\x20\x78\x29\xad"
+ "\xb3\x6a\x92\x8a\xc2\x16\x25\x60"
+ "\x25\xe5\x80\x3d\x5b\x64\xdc\xe6"
+ "\xfb\xf3\xb2\xcc\xe3\x8c\xf2\x4b"
+ "\x7a\xb2\x58\x26\xe0\x2c\xde\x52"
+ "\xdd\xb5\x07\x48\xad\xe5\xe4\xc9"
+ "\x0e\x42\xb6\xd1\xf5\x17\xc0\xe4"
+ "\x57\x3c\x1c\x1c\x7d\xb2\x50\xc0"
+ "\x75\x38\x72\x5d\x4c\xbc\xe4\xe9"
+ "\x0b",
+ },
+};
+
+static const struct comp_testvec deflate_iaa_dynamic_decomp_tv_template[] = {
+ {
+ .inlen = 121,
+ .outlen = 191,
+ .input = "\x5d\x8d\xc1\x0d\xc2\x30\x10\x04"
+ "\x5b\xd9\x0a\xd2\x03\x82\x20\x21"
+ "\xf1\xf0\x23\x0d\x5c\xec\x0b\xb6"
+ "\x64\xfb\x2c\xdf\xf1\xa0\x7b\x12"
+ "\x3e\x58\x79\xae\x76\x67\x76\x89"
+ "\x49\x11\xc4\xbf\x0b\x57\x43\x60"
+ "\xf5\x3d\xad\xac\x20\x78\x29\xad"
+ "\xb3\x6a\x92\x8a\xc2\x16\x25\x60"
+ "\x25\xe5\x80\x3d\x5b\x64\xdc\xe6"
+ "\xfb\xf3\xb2\xcc\xe3\x8c\xf2\x4b"
+ "\x7a\xb2\x58\x26\xe0\x2c\xde\x52"
+ "\xdd\xb5\x07\x48\xad\xe5\xe4\xc9"
+ "\x0e\x42\xb6\xd1\xf5\x17\xc0\xe4"
+ "\x57\x3c\x1c\x1c\x7d\xb2\x50\xc0"
+ "\x75\x38\x72\x5d\x4c\xbc\xe4\xe9"
+ "\x0b",
+ .output = "This document describes a compression method based on the DEFLATE"
+ "compression algorithm. This document defines the application of "
+ "the DEFLATE algorithm to the IP Payload Compression Protocol.",
+ }, {
+ .inlen = 46,
+ .outlen = 70,
+ .input = "\x85\xca\xc1\x09\x00\x20\x08\x05"
+ "\xd0\x55\xfe\x3c\x6e\x21\x64\xd8"
+ "\x45\x21\x0d\xd7\xb7\x26\xe8\xf8"
+ "\xe0\x91\x2f\xc3\x09\x98\x17\xd8"
+ "\x06\x42\x79\x0b\x52\x05\xe1\x33"
+ "\xeb\x81\x3e\xe5\xa2\x01",
+ .output = "Join us now and share the software "
+ "Join us now and share the software ",
+ },
+};
+
/*
* LZO test vectors (null-terminated strings).
*/
diff --git a/drivers/crypto/intel/iaa/Makefile b/drivers/crypto/intel/iaa/Makefile
index ebfa1a425f80..96f22cd39924 100644
--- a/drivers/crypto/intel/iaa/Makefile
+++ b/drivers/crypto/intel/iaa/Makefile
@@ -7,6 +7,6 @@ ccflags-y += -I $(srctree)/drivers/dma/idxd -DDEFAULT_SYMBOL_NAMESPACE='"CRYPTO_
obj-$(CONFIG_CRYPTO_DEV_IAA_CRYPTO) := iaa_crypto.o
-iaa_crypto-y := iaa_crypto_main.o iaa_crypto_comp_fixed.o
+iaa_crypto-y := iaa_crypto_main.o iaa_crypto_comp_fixed.o iaa_crypto_comp_dynamic.o
iaa_crypto-$(CONFIG_CRYPTO_DEV_IAA_CRYPTO_STATS) += iaa_crypto_stats.o
diff --git a/drivers/crypto/intel/iaa/iaa_crypto.h b/drivers/crypto/intel/iaa/iaa_crypto.h
index db1e50574662..5a8dec3be9fa 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto.h
+++ b/drivers/crypto/intel/iaa/iaa_crypto.h
@@ -19,12 +19,15 @@
#define IAA_COMP_FLUSH_OUTPUT BIT(1)
#define IAA_COMP_APPEND_EOB BIT(2)
+#define IAA_COMP_GEN_HDR_1_PASS (BIT(12) | BIT(13))
#define IAA_COMPLETION_TIMEOUT 1000000
#define IAA_ALLOC_DESC_COMP_TIMEOUT 1000
#define IAA_ALLOC_DESC_DECOMP_TIMEOUT 500
+#define IAA_DYN_ALLOC_DESC_COMP_TIMEOUT 2000
+
#define IAA_ANALYTICS_ERROR 0x0a
#define IAA_ERROR_DECOMP_BUF_OVERFLOW 0x0b
#define IAA_ERROR_COMP_BUF_OVERFLOW 0x19
@@ -134,6 +137,8 @@ struct aecs_comp_table_record {
int iaa_aecs_init_fixed(void);
void iaa_aecs_cleanup_fixed(void);
+int iaa_aecs_init_dynamic(void);
+void iaa_aecs_cleanup_dynamic(void);
typedef int (*iaa_dev_comp_init_fn_t) (struct iaa_device_compression_mode *mode);
typedef int (*iaa_dev_comp_free_fn_t) (struct iaa_device_compression_mode *mode);
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_comp_dynamic.c b/drivers/crypto/intel/iaa/iaa_crypto_comp_dynamic.c
new file mode 100644
index 000000000000..3a93d7913443
--- /dev/null
+++ b/drivers/crypto/intel/iaa/iaa_crypto_comp_dynamic.c
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright(c) 2023 Intel Corporation. All rights rsvd. */
+
+#include "idxd.h"
+#include "iaa_crypto.h"
+
+int iaa_aecs_init_dynamic(void)
+{
+ int ret;
+
+ ret = add_iaa_compression_mode("dynamic", NULL, 0, NULL, 0, NULL, NULL);
+
+ if (!ret)
+ pr_debug("IAA dynamic compression mode initialized\n");
+
+ return ret;
+}
+
+void iaa_aecs_cleanup_dynamic(void)
+{
+ remove_iaa_compression_mode("dynamic");
+}
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index 0669ae155e90..cbe3a2457253 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -102,10 +102,12 @@ DEFINE_MUTEX(first_wq_found_lock);
const char *iaa_compression_mode_names[IAA_COMP_MODES_MAX] = {
"fixed",
+ "dynamic",
};
const char *iaa_compression_alg_names[IAA_COMP_MODES_MAX] = {
"deflate-iaa",
+ "deflate-iaa-dynamic",
};
static struct iaa_compression_mode *iaa_compression_modes[IAA_COMP_MODES_MAX];
@@ -1482,6 +1484,23 @@ static int deflate_generic_decompress(struct iaa_req *req)
return ret;
}
+static int deflate_generic_compress(struct iaa_req *req)
+{
+ ACOMP_REQUEST_ON_STACK(fbreq, deflate_crypto_acomp);
+ int ret;
+
+ acomp_request_set_callback(fbreq, 0, NULL, NULL);
+ acomp_request_set_params(fbreq, req->src, req->dst, req->slen,
+ PAGE_SIZE);
+
+ ret = crypto_acomp_compress(fbreq);
+ req->dlen = fbreq->dlen;
+
+ update_total_sw_comp_calls();
+
+ return ret;
+}
+
static __always_inline void acomp_to_iaa(struct acomp_req *areq,
struct iaa_req *req,
struct iaa_compression_ctx *ctx)
@@ -1808,9 +1827,13 @@ iaa_setup_compress_hw_desc(struct idxd_desc *idxd_desc,
desc->src1_size = slen;
desc->dst_addr = (u64)dst_addr;
desc->max_dst_size = dlen;
- desc->flags |= IDXD_OP_FLAG_RD_SRC2_AECS;
- desc->src2_addr = active_compression_mode->aecs_comp_table_dma_addr;
- desc->src2_size = sizeof(struct aecs_comp_table_record);
+ if (mode == IAA_MODE_DYNAMIC) {
+ desc->compr_flags |= IAA_COMP_GEN_HDR_1_PASS;
+ } else {
+ desc->flags |= IDXD_OP_FLAG_RD_SRC2_AECS;
+ desc->src2_addr = active_compression_mode->aecs_comp_table_dma_addr;
+ desc->src2_size = sizeof(struct aecs_comp_table_record);
+ }
desc->completion_addr = idxd_desc->compl_dma;
return desc;
@@ -2064,6 +2087,9 @@ static int iaa_comp_acompress(struct iaa_compression_ctx *ctx, struct iaa_req *r
return -EINVAL;
}
+ if (ctx->mode == IAA_MODE_DYNAMIC && req->slen > PAGE_SIZE)
+ return deflate_generic_compress(req);
+
cpu = get_cpu();
wq = comp_wq_table_next_wq(cpu);
put_cpu();
@@ -2546,7 +2572,9 @@ static int iaa_comp_adecompress_batch(
static void compression_ctx_init(struct iaa_compression_ctx *ctx, enum iaa_mode mode)
{
ctx->mode = mode;
- ctx->alloc_comp_desc_timeout = IAA_ALLOC_DESC_COMP_TIMEOUT;
+ ctx->alloc_comp_desc_timeout = (mode == IAA_MODE_DYNAMIC ?
+ IAA_DYN_ALLOC_DESC_COMP_TIMEOUT :
+ IAA_ALLOC_DESC_COMP_TIMEOUT);
ctx->alloc_decomp_desc_timeout = IAA_ALLOC_DESC_DECOMP_TIMEOUT;
ctx->verify_compress = iaa_verify_compress;
ctx->async_mode = async_mode;
@@ -2760,6 +2788,30 @@ static struct acomp_alg iaa_acomp_fixed_deflate = {
}
};
+static int iaa_comp_init_dynamic(struct crypto_acomp *acomp_tfm)
+{
+ struct crypto_tfm *tfm = crypto_acomp_tfm(acomp_tfm);
+ struct iaa_compression_ctx *ctx = crypto_tfm_ctx(tfm);
+
+ ctx = iaa_ctx[IAA_MODE_DYNAMIC];
+
+ return 0;
+}
+
+static struct acomp_alg iaa_acomp_dynamic_deflate = {
+ .init = iaa_comp_init_dynamic,
+ .compress = iaa_comp_acompress_main,
+ .decompress = iaa_comp_adecompress_main,
+ .base = {
+ .cra_name = "deflate",
+ .cra_driver_name = "deflate-iaa-dynamic",
+ .cra_flags = CRYPTO_ALG_ASYNC,
+ .cra_ctxsize = sizeof(struct iaa_compression_ctx),
+ .cra_module = THIS_MODULE,
+ .cra_priority = IAA_ALG_PRIORITY + 1,
+ }
+};
+
/*******************************************
* Implement idxd_device_driver interfaces.
*******************************************/
@@ -2779,7 +2831,7 @@ static void iaa_unregister_compression_device(void)
num_iaa_modes_registered = 0;
}
-static int iaa_register_compression_device(void)
+static int iaa_register_compression_device(struct idxd_device *idxd)
{
struct iaa_compression_mode *mode;
int i, idx;
@@ -2788,6 +2840,13 @@ static int iaa_register_compression_device(void)
iaa_mode_registered[i] = false;
mode = find_iaa_compression_mode(iaa_compression_mode_names[i], &idx);
if (mode) {
+ /* Header Generation Capability is required for the dynamic algorithm. */
+ if ((!strcmp(mode->name, "dynamic")) && !idxd->hw.iaa_cap.header_gen) {
+ if (num_iaa_modes_registered > 0)
+ --num_iaa_modes_registered;
+ continue;
+ }
+
iaa_ctx[i] = kmalloc(sizeof(struct iaa_compression_ctx), GFP_KERNEL);
if (!iaa_ctx[i])
goto err;
@@ -2805,7 +2864,7 @@ static int iaa_register_compression_device(void)
return -ENODEV;
}
-static int iaa_register_acomp_compression_device(void)
+static int iaa_register_acomp_compression_device(struct idxd_device *idxd)
{
int ret = -ENOMEM;
@@ -2819,8 +2878,19 @@ static int iaa_register_acomp_compression_device(void)
goto err_fixed;
}
+ if (iaa_mode_registered[IAA_MODE_DYNAMIC]) {
+ ret = crypto_register_acomp(&iaa_acomp_dynamic_deflate);
+ if (ret) {
+ pr_err("deflate algorithm acomp dynamic registration failed (%d)\n", ret);
+ goto err_dynamic;
+ }
+ }
+
return 0;
+err_dynamic:
+ crypto_unregister_acomp(&iaa_acomp_fixed_deflate);
+
err_fixed:
if (!IS_ERR_OR_NULL(deflate_crypto_acomp)) {
crypto_free_acomp(deflate_crypto_acomp);
@@ -2839,6 +2909,9 @@ static void iaa_unregister_acomp_compression_device(void)
if (iaa_mode_registered[IAA_MODE_FIXED])
crypto_unregister_acomp(&iaa_acomp_fixed_deflate);
+ if (iaa_mode_registered[IAA_MODE_DYNAMIC])
+ crypto_unregister_acomp(&iaa_acomp_dynamic_deflate);
+
if (!IS_ERR_OR_NULL(deflate_crypto_acomp)) {
crypto_free_acomp(deflate_crypto_acomp);
deflate_crypto_acomp = NULL;
@@ -2906,13 +2979,13 @@ static int iaa_crypto_probe(struct idxd_dev *idxd_dev)
atomic_set(&iaa_crypto_enabled, 1);
if (first_wq) {
- ret = iaa_register_compression_device();
+ ret = iaa_register_compression_device(idxd);
if (ret != 0) {
dev_dbg(dev, "IAA compression device registration failed\n");
goto err_register;
}
- ret = iaa_register_acomp_compression_device();
+ ret = iaa_register_acomp_compression_device(idxd);
if (ret != 0) {
dev_dbg(dev, "IAA compression device acomp registration failed\n");
goto err_register;
@@ -3063,6 +3136,12 @@ static int __init iaa_crypto_init_module(void)
goto err_aecs_init;
}
+ ret = iaa_aecs_init_dynamic();
+ if (ret < 0) {
+ pr_debug("IAA dynamic compression mode init failed\n");
+ goto err_dynamic;
+ }
+
ret = idxd_driver_register(&iaa_crypto_driver);
if (ret) {
pr_debug("IAA wq sub-driver registration failed\n");
@@ -3164,6 +3243,8 @@ static int __init iaa_crypto_init_module(void)
err_g_comp_wqs_per_iaa_attr_create:
idxd_driver_unregister(&iaa_crypto_driver);
err_driver_reg:
+ iaa_aecs_cleanup_dynamic();
+err_dynamic:
iaa_aecs_cleanup_fixed();
err_aecs_init:
@@ -3188,6 +3269,7 @@ static void __exit iaa_crypto_cleanup_module(void)
driver_remove_file(&iaa_crypto_driver.drv,
&driver_attr_g_comp_wqs_per_iaa);
idxd_driver_unregister(&iaa_crypto_driver);
+ iaa_aecs_cleanup_dynamic();
iaa_aecs_cleanup_fixed();
pr_debug("cleaned up\n");
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_stats.c b/drivers/crypto/intel/iaa/iaa_crypto_stats.c
index f5cc3d29ca19..42aae8a738ac 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_stats.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_stats.c
@@ -19,6 +19,7 @@
static atomic64_t total_comp_calls;
static atomic64_t total_decomp_calls;
+static atomic64_t total_sw_comp_calls;
static atomic64_t total_sw_decomp_calls;
static atomic64_t total_comp_bytes_out;
static atomic64_t total_decomp_bytes_in;
@@ -43,6 +44,11 @@ void update_total_decomp_calls(void)
atomic64_inc(&total_decomp_calls);
}
+void update_total_sw_comp_calls(void)
+{
+ atomic64_inc(&total_sw_comp_calls);
+}
+
void update_total_sw_decomp_calls(void)
{
atomic64_inc(&total_sw_decomp_calls);
@@ -174,6 +180,8 @@ static int global_stats_show(struct seq_file *m, void *v)
atomic64_read(&total_comp_calls));
seq_printf(m, " total_decomp_calls: %llu\n",
atomic64_read(&total_decomp_calls));
+ seq_printf(m, " total_sw_comp_calls: %llu\n",
+ atomic64_read(&total_sw_comp_calls));
seq_printf(m, " total_sw_decomp_calls: %llu\n",
atomic64_read(&total_sw_decomp_calls));
seq_printf(m, " total_comp_bytes_out: %llu\n",
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_stats.h b/drivers/crypto/intel/iaa/iaa_crypto_stats.h
index 3787a5f507eb..6e0c6f9939bf 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_stats.h
+++ b/drivers/crypto/intel/iaa/iaa_crypto_stats.h
@@ -11,6 +11,7 @@ void iaa_crypto_debugfs_cleanup(void);
void update_total_comp_calls(void);
void update_total_comp_bytes_out(int n);
void update_total_decomp_calls(void);
+void update_total_sw_comp_calls(void);
void update_total_sw_decomp_calls(void);
void update_total_decomp_bytes_in(int n);
void update_completion_einval_errs(void);
@@ -29,6 +30,7 @@ static inline void iaa_crypto_debugfs_cleanup(void) {}
static inline void update_total_comp_calls(void) {}
static inline void update_total_comp_bytes_out(int n) {}
static inline void update_total_decomp_calls(void) {}
+static inline void update_total_sw_comp_calls(void) {}
static inline void update_total_sw_decomp_calls(void) {}
static inline void update_total_decomp_bytes_in(int n) {}
static inline void update_completion_einval_errs(void) {}
diff --git a/include/linux/iaa_comp.h b/include/linux/iaa_comp.h
index 7b765760485c..ec28a9d9f2b4 100644
--- a/include/linux/iaa_comp.h
+++ b/include/linux/iaa_comp.h
@@ -12,7 +12,8 @@
enum iaa_mode {
IAA_MODE_FIXED = 0,
- IAA_MODE_NONE = 1,
+ IAA_MODE_DYNAMIC = 1,
+ IAA_MODE_NONE = 2,
};
struct iaa_req {
@@ -56,7 +57,7 @@ extern int iaa_comp_decompress_batch(
#else /* CONFIG_CRYPTO_DEV_IAA_CRYPTO */
enum iaa_mode {
- IAA_MODE_NONE = 1,
+ IAA_MODE_NONE = 2,
};
struct iaa_req {};
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 19/23] crypto: acomp - Add crypto_acomp_batch_size() to get an algorithm's batch-size.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (17 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 18/23] crypto: iaa - Add deflate-iaa-dynamic compression mode Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-10-17 1:04 ` Herbert Xu
2025-09-26 3:34 ` [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources exist from pool creation to deletion Kanchana P Sridhar
` (4 subsequent siblings)
23 siblings, 1 reply; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This commit adds a @batch_size data member to:
struct acomp_alg
struct crypto_acomp
A crypto_acomp compression algorithm that supports batching of
compressions and decompressions must provide a @batch_size greater than
one, representing the maximum batch-size that the compressor supports,
so that kernel users of crypto_acomp, such as zswap, can allocate
resources for submitting multiple compress/decompress jobs that can be
batched, and invoke batching of [de]compressions.
A new helper function acomp_has_async_batching() can be invoked to query
if a crypto_acomp has defined a @batch_size.
The new crypto_acomp_batch_size() API uses this helper function to return
the batch-size for compressors that have registered a @batch_size. If the
algorithm does not define a @batch_size, a default of "1" is returned.
zswap can invoke crypto_acomp_batch_size() to query the maximum number
of requests that can be batch [de]compressed. Based on this, zswap
can use the minimum of any zswap-specific upper limits for batch-size
and the compressor's max @batch_size, to allocate batching resources.
The IAA acomp_algs Fixed ("deflate-iaa") and Dynamic
("deflate-iaa-dynamic") register @batch_size as
IAA_CRYPTO_MAX_BATCH_SIZE.
This enables zswap to compress/decompress pages in parallel in the IAA
hardware accelerator to improve swapout/swapin performance and memory
savings.
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
crypto/acompress.c | 1 +
drivers/crypto/intel/iaa/iaa_crypto_main.c | 2 ++
include/crypto/acompress.h | 27 ++++++++++++++++++++++
include/crypto/internal/acompress.h | 3 +++
4 files changed, 33 insertions(+)
diff --git a/crypto/acompress.c b/crypto/acompress.c
index be28cbfd22e3..a1bdfa21e688 100644
--- a/crypto/acompress.c
+++ b/crypto/acompress.c
@@ -105,6 +105,7 @@ static int crypto_acomp_init_tfm(struct crypto_tfm *tfm)
acomp->compress = alg->compress;
acomp->decompress = alg->decompress;
+ acomp->batch_size = alg->batch_size;
acomp->reqsize = alg->base.cra_reqsize;
acomp->base.exit = crypto_acomp_exit_tfm;
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index cbe3a2457253..2fa38176034d 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -2777,6 +2777,7 @@ static struct acomp_alg iaa_acomp_fixed_deflate = {
.init = iaa_comp_init_fixed,
.compress = iaa_comp_acompress_main,
.decompress = iaa_comp_adecompress_main,
+ .batch_size = IAA_CRYPTO_MAX_BATCH_SIZE,
.base = {
.cra_name = "deflate",
.cra_driver_name = "deflate-iaa",
@@ -2802,6 +2803,7 @@ static struct acomp_alg iaa_acomp_dynamic_deflate = {
.init = iaa_comp_init_dynamic,
.compress = iaa_comp_acompress_main,
.decompress = iaa_comp_adecompress_main,
+ .batch_size = IAA_CRYPTO_MAX_BATCH_SIZE,
.base = {
.cra_name = "deflate",
.cra_driver_name = "deflate-iaa-dynamic",
diff --git a/include/crypto/acompress.h b/include/crypto/acompress.h
index 0f1334168f1b..e94046529e46 100644
--- a/include/crypto/acompress.h
+++ b/include/crypto/acompress.h
@@ -108,6 +108,8 @@ struct acomp_req {
*
* @compress: Function performs a compress operation
* @decompress: Function performs a de-compress operation
+ * @batch_size: Maximum batch-size for batching compress/decompress
+ * operations.
* @reqsize: Context size for (de)compression requests
* @fb: Synchronous fallback tfm
* @base: Common crypto API algorithm data structure
@@ -115,6 +117,7 @@ struct acomp_req {
struct crypto_acomp {
int (*compress)(struct acomp_req *req);
int (*decompress)(struct acomp_req *req);
+ unsigned int batch_size;
unsigned int reqsize;
struct crypto_tfm base;
};
@@ -205,6 +208,13 @@ static inline bool acomp_is_async(struct crypto_acomp *tfm)
CRYPTO_ALG_ASYNC;
}
+static inline bool acomp_has_async_batching(struct crypto_acomp *tfm)
+{
+ return (acomp_is_async(tfm) &&
+ (crypto_comp_alg_common(tfm)->base.cra_flags & CRYPTO_ALG_TYPE_ACOMPRESS) &&
+ (tfm->batch_size > 1));
+}
+
static inline struct crypto_acomp *crypto_acomp_reqtfm(struct acomp_req *req)
{
return __crypto_acomp_tfm(req->base.tfm);
@@ -578,6 +588,23 @@ int crypto_acomp_compress(struct acomp_req *req);
*/
int crypto_acomp_decompress(struct acomp_req *req);
+/**
+ * crypto_acomp_batch_size() -- Get the algorithm's batch size
+ *
+ * Function returns the algorithm's batch size for batching operations
+ *
+ * @tfm: ACOMPRESS tfm handle allocated with crypto_alloc_acomp()
+ *
+ * Return: crypto_acomp's batch size.
+ */
+static inline unsigned int crypto_acomp_batch_size(struct crypto_acomp *tfm)
+{
+ if (acomp_has_async_batching(tfm))
+ return tfm->batch_size;
+
+ return 1;
+}
+
static inline struct acomp_req *acomp_request_on_stack_init(
char *buf, struct crypto_acomp *tfm)
{
diff --git a/include/crypto/internal/acompress.h b/include/crypto/internal/acompress.h
index 2d97440028ff..e451e0ae3b9b 100644
--- a/include/crypto/internal/acompress.h
+++ b/include/crypto/internal/acompress.h
@@ -28,6 +28,8 @@
*
* @compress: Function performs a compress operation
* @decompress: Function performs a de-compress operation
+ * @batch_size: Maximum batch-size for batching compress/decompress
+ * operations.
* @init: Initialize the cryptographic transformation object.
* This function is used to initialize the cryptographic
* transformation object. This function is called only once at
@@ -46,6 +48,7 @@
struct acomp_alg {
int (*compress)(struct acomp_req *req);
int (*decompress)(struct acomp_req *req);
+ unsigned int batch_size;
int (*init)(struct crypto_acomp *tfm);
void (*exit)(struct crypto_acomp *tfm);
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources exist from pool creation to deletion.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (18 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 19/23] crypto: acomp - Add crypto_acomp_batch_size() to get an algorithm's batch-size Kanchana P Sridhar
@ 2025-09-26 3:34 ` Kanchana P Sridhar
2025-09-30 15:49 ` Yosry Ahmed
2025-09-26 3:35 ` [PATCH v12 21/23] mm: zswap: Consistently use IS_ERR_OR_NULL() to check acomp_ctx resources Kanchana P Sridhar
` (3 subsequent siblings)
23 siblings, 1 reply; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:34 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch simplifies the zswap_pool's per-CPU acomp_ctx resource
management. Similar to the per-CPU acomp_ctx itself, the per-CPU
acomp_ctx's resources' (acomp, req, buffer) lifetime will also be from
pool creation to pool deletion. These resources will persist through CPU
hotplug operations. The zswap_cpu_comp_dead() teardown callback has been
deleted from the call to
cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE). As a result, CPU
offline hotplug operations will be no-ops as far as the acomp_ctx
resources are concerned.
This commit refactors the code from zswap_cpu_comp_dead() into a
new function acomp_ctx_dealloc() that preserves the IS_ERR_OR_NULL()
checks on acomp_ctx, req and acomp from the existing mainline
implementation of zswap_cpu_comp_dead(). acomp_ctx_dealloc() is called
to clean up acomp_ctx resources from all these procedures:
1) zswap_cpu_comp_prepare() when an error is encountered,
2) zswap_pool_create() when an error is encountered, and
3) from zswap_pool_destroy().
The main benefit of using the CPU hotplug multi state instance startup
callback to allocate the acomp_ctx resources is that it prevents the
cores from being offlined until the multi state instance addition call
returns.
From Documentation/core-api/cpu_hotplug.rst:
"The node list add/remove operations and the callback invocations are
serialized against CPU hotplug operations."
Furthermore, zswap_[de]compress() cannot contend with
zswap_cpu_comp_prepare() because:
- During pool creation/deletion, the pool is not in the zswap_pools
list.
- During CPU hot[un]plug, the CPU is not yet online, as Yosry pointed
out. zswap_cpu_comp_prepare() will be executed on a control CPU,
since CPUHP_MM_ZSWP_POOL_PREPARE is in the PREPARE section of "enum
cpuhp_state". Thanks Yosry for sharing this observation!
In both these cases, any recursions into zswap reclaim from
zswap_cpu_comp_prepare() will be handled by the old pool.
The above two observations enable the following simplifications:
1) zswap_cpu_comp_prepare(): CPU cannot be offlined. Reclaim cannot use
the pool. Considerations for mutex init/locking and handling
subsequent CPU hotplug online-offlines:
Should we lock the mutex of current CPU's acomp_ctx from start to
end? It doesn't seem like this is required. The CPU hotplug
operations acquire a "cpuhp_state_mutex" before proceeding, hence
they are serialized against CPU hotplug operations.
If the process gets migrated while zswap_cpu_comp_prepare() is
running, it will complete on the new CPU. In case of failures, we
pass the acomp_ctx pointer obtained at the start of
zswap_cpu_comp_prepare() to acomp_ctx_dealloc(), which again, can
only undergo migration. There appear to be no contention scenarios
that might cause inconsistent values of acomp_ctx's members. Hence,
it seems there is no need for mutex_lock(&acomp_ctx->mutex) in
zswap_cpu_comp_prepare().
Since the pool is not yet on zswap_pools list, we don't need to
initialize the per-CPU acomp_ctx mutex in zswap_pool_create(). This
has been restored to occur in zswap_cpu_comp_prepare().
zswap_cpu_comp_prepare() checks upfront if acomp_ctx->acomp is
valid. If so, it returns success. This should handle any CPU
hotplug online-offline transitions after pool creation is done.
2) CPU offline vis-a-vis zswap ops: Let's suppose the process is
migrated to another CPU before the current CPU is dysfunctional. If
zswap_[de]compress() holds the acomp_ctx->mutex lock of the offlined
CPU, that mutex will be released once it completes on the new
CPU. Since there is no teardown callback, there is no possibility of
UAF.
3) Pool creation/deletion and process migration to another CPU:
- During pool creation/deletion, the pool is not in the zswap_pools
list. Hence it cannot contend with zswap ops on that CPU. However,
the process can get migrated.
Pool creation --> zswap_cpu_comp_prepare()
--> process migrated:
* CPU offline: no-op.
* zswap_cpu_comp_prepare() continues
to run on the new CPU to finish
allocating acomp_ctx resources for
the offlined CPU.
Pool deletion --> acomp_ctx_dealloc()
--> process migrated:
* CPU offline: no-op.
* acomp_ctx_dealloc() continues
to run on the new CPU to finish
de-allocating acomp_ctx resources
for the offlined CPU.
4) Pool deletion vis-a-vis CPU onlining:
To prevent possibility of race conditions between
acomp_ctx_dealloc() freeing the acomp_ctx resources and the initial
check for a valid acomp_ctx->acomp in zswap_cpu_comp_prepare(), we
need to delete the multi state instance right after it is added, in
zswap_pool_create().
Summary of changes based on the above:
--------------------------------------
1) Zero-initialization of pool->acomp_ctx in zswap_pool_create() to
simplify and share common code for different error handling/cleanup
related to the acomp_ctx.
2) Remove the node list instance right after node list add function
call in zswap_pool_create(). This prevents race conditions between
CPU onlining after initial pool creation, and acomp_ctx_dealloc()
freeing the acomp_ctx resources.
3) zswap_pool_destroy() will call acomp_ctx_dealloc() to de-allocate
the per-CPU acomp_ctx resources.
4) Changes to zswap_cpu_comp_prepare():
a) Check if acomp_ctx->acomp is valid at the beginning and return,
because the acomp_ctx is already initialized.
b) Move the mutex_init to happen in this procedure, before it
returns.
c) All error conditions handled by calling acomp_ctx_dealloc().
5) New procedure acomp_ctx_dealloc() for common error/cleanup code.
6) No more multi state instance teardown callback. CPU offlining is a
no-op as far as acomp_ctx resources are concerned.
7) Delete acomp_ctx_get_cpu_lock()/acomp_ctx_put_unlock(). Directly
call mutex_lock(&acomp_ctx->mutex)/mutex_unlock(&acomp_ctx->mutex)
in zswap_[de]compress().
The per-CPU memory cost of not deleting the acomp_ctx resources upon CPU
offlining, and only deleting them when the pool is destroyed, is as
follows, on x86_64:
IAA with 8 dst buffers for batching: 64.34 KB
Software compressors with 1 dst buffer: 8.28 KB
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
mm/zswap.c | 194 +++++++++++++++++++++++++----------------------------
1 file changed, 93 insertions(+), 101 deletions(-)
diff --git a/mm/zswap.c b/mm/zswap.c
index c1af782e54ec..27665eaa3f89 100644
--- a/mm/zswap.c
+++ b/mm/zswap.c
@@ -242,6 +242,30 @@ static inline struct xarray *swap_zswap_tree(swp_entry_t swp)
**********************************/
static void __zswap_pool_empty(struct percpu_ref *ref);
+/*
+ * The per-cpu pool->acomp_ctx is zero-initialized on allocation. This makes
+ * it easy for different error conditions/cleanup related to the acomp_ctx
+ * to be handled by acomp_ctx_dealloc():
+ * - Errors during zswap_cpu_comp_prepare().
+ * - Partial success/error of cpuhp_state_add_instance() call in
+ * zswap_pool_create(). Only some cores could have executed
+ * zswap_cpu_comp_prepare(), not others.
+ * - Cleanup acomp_ctx resources on all cores in zswap_pool_destroy().
+ */
+static void acomp_ctx_dealloc(struct crypto_acomp_ctx *acomp_ctx)
+{
+ if (IS_ERR_OR_NULL(acomp_ctx))
+ return;
+
+ if (!IS_ERR_OR_NULL(acomp_ctx->req))
+ acomp_request_free(acomp_ctx->req);
+
+ if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
+ crypto_free_acomp(acomp_ctx->acomp);
+
+ kfree(acomp_ctx->buffer);
+}
+
static struct zswap_pool *zswap_pool_create(char *compressor)
{
struct zswap_pool *pool;
@@ -263,19 +287,43 @@ static struct zswap_pool *zswap_pool_create(char *compressor)
strscpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
- pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
+ /* Many things rely on the zero-initialization. */
+ pool->acomp_ctx = alloc_percpu_gfp(*pool->acomp_ctx,
+ GFP_KERNEL | __GFP_ZERO);
if (!pool->acomp_ctx) {
pr_err("percpu alloc failed\n");
goto error;
}
- for_each_possible_cpu(cpu)
- mutex_init(&per_cpu_ptr(pool->acomp_ctx, cpu)->mutex);
-
+ /*
+ * This is serialized against CPU hotplug operations. Hence, cores
+ * cannot be offlined until this finishes.
+ * In case of errors, we need to goto "ref_fail" instead of "error"
+ * because there is no teardown callback registered anymore, for
+ * cpuhp_state_add_instance() to de-allocate resources as it rolls back
+ * state on cores before the CPU on which error was encountered.
+ */
ret = cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
&pool->node);
+
+ /*
+ * We only needed the multi state instance add operation to invoke the
+ * startup callback for all cores without cores getting offlined. Since
+ * the acomp_ctx resources will now only be de-allocated when the pool
+ * is destroyed, we can safely remove the multi state instance. This
+ * minimizes (but does not eliminate) the possibility of
+ * zswap_cpu_comp_prepare() being invoked again due to a CPU
+ * offline-online transition. Removing the instance also prevents race
+ * conditions between CPU onlining after initial pool creation, and
+ * acomp_ctx_dealloc() freeing the acomp_ctx resources.
+ * Note that we delete the instance before checking the error status of
+ * the node list add operation because we want the instance removal even
+ * in case of errors in the former.
+ */
+ cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
+
if (ret)
- goto error;
+ goto ref_fail;
/* being the current pool takes 1 ref; this func expects the
* caller to always add the new pool as the current pool
@@ -291,7 +339,8 @@ static struct zswap_pool *zswap_pool_create(char *compressor)
return pool;
ref_fail:
- cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
+ for_each_possible_cpu(cpu)
+ acomp_ctx_dealloc(per_cpu_ptr(pool->acomp_ctx, cpu));
error:
if (pool->acomp_ctx)
free_percpu(pool->acomp_ctx);
@@ -322,9 +371,13 @@ static struct zswap_pool *__zswap_pool_create_fallback(void)
static void zswap_pool_destroy(struct zswap_pool *pool)
{
+ int cpu;
+
zswap_pool_debug("destroying", pool);
- cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
+ for_each_possible_cpu(cpu)
+ acomp_ctx_dealloc(per_cpu_ptr(pool->acomp_ctx, cpu));
+
free_percpu(pool->acomp_ctx);
zs_destroy_pool(pool->zs_pool);
@@ -736,39 +789,39 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
{
struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
- struct crypto_acomp *acomp = NULL;
- struct acomp_req *req = NULL;
- u8 *buffer = NULL;
- int ret;
+ int ret = -ENOMEM;
- buffer = kmalloc_node(PAGE_SIZE, GFP_KERNEL, cpu_to_node(cpu));
- if (!buffer) {
- ret = -ENOMEM;
- goto fail;
- }
+ /*
+ * The per-CPU pool->acomp_ctx is zero-initialized on allocation.
+ * Even though we delete the multi state instance right after successful
+ * addition of the instance in zswap_pool_create(), we cannot eliminate
+ * the possibility of the CPU going through offline-online transitions.
+ * If this does happen, we check if the acomp_ctx has already been
+ * initialized, and return.
+ */
+ if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
+ return 0;
- acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_to_node(cpu));
- if (IS_ERR(acomp)) {
+ acomp_ctx->buffer = kmalloc_node(PAGE_SIZE, GFP_KERNEL, cpu_to_node(cpu));
+ if (!acomp_ctx->buffer)
+ return ret;
+
+ acomp_ctx->acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_to_node(cpu));
+ if (IS_ERR(acomp_ctx->acomp)) {
pr_err("could not alloc crypto acomp %s : %ld\n",
- pool->tfm_name, PTR_ERR(acomp));
- ret = PTR_ERR(acomp);
+ pool->tfm_name, PTR_ERR(acomp_ctx->acomp));
+ ret = PTR_ERR(acomp_ctx->acomp);
goto fail;
}
+ acomp_ctx->is_sleepable = acomp_is_async(acomp_ctx->acomp);
- req = acomp_request_alloc(acomp);
- if (!req) {
+ acomp_ctx->req = acomp_request_alloc(acomp_ctx->acomp);
+ if (!acomp_ctx->req) {
pr_err("could not alloc crypto acomp_request %s\n",
pool->tfm_name);
- ret = -ENOMEM;
goto fail;
}
- /*
- * Only hold the mutex after completing allocations, otherwise we may
- * recurse into zswap through reclaim and attempt to hold the mutex
- * again resulting in a deadlock.
- */
- mutex_lock(&acomp_ctx->mutex);
crypto_init_wait(&acomp_ctx->wait);
/*
@@ -776,81 +829,17 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
* crypto_wait_req(); if the backend of acomp is scomp, the callback
* won't be called, crypto_wait_req() will return without blocking.
*/
- acomp_request_set_callback(req, CRYPTO_TFM_REQ_MAY_BACKLOG,
+ acomp_request_set_callback(acomp_ctx->req, CRYPTO_TFM_REQ_MAY_BACKLOG,
crypto_req_done, &acomp_ctx->wait);
- acomp_ctx->buffer = buffer;
- acomp_ctx->acomp = acomp;
- acomp_ctx->is_sleepable = acomp_is_async(acomp);
- acomp_ctx->req = req;
- mutex_unlock(&acomp_ctx->mutex);
+ mutex_init(&acomp_ctx->mutex);
return 0;
fail:
- if (acomp)
- crypto_free_acomp(acomp);
- kfree(buffer);
+ acomp_ctx_dealloc(acomp_ctx);
return ret;
}
-static int zswap_cpu_comp_dead(unsigned int cpu, struct hlist_node *node)
-{
- struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
- struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
- struct acomp_req *req;
- struct crypto_acomp *acomp;
- u8 *buffer;
-
- if (IS_ERR_OR_NULL(acomp_ctx))
- return 0;
-
- mutex_lock(&acomp_ctx->mutex);
- req = acomp_ctx->req;
- acomp = acomp_ctx->acomp;
- buffer = acomp_ctx->buffer;
- acomp_ctx->req = NULL;
- acomp_ctx->acomp = NULL;
- acomp_ctx->buffer = NULL;
- mutex_unlock(&acomp_ctx->mutex);
-
- /*
- * Do the actual freeing after releasing the mutex to avoid subtle
- * locking dependencies causing deadlocks.
- */
- if (!IS_ERR_OR_NULL(req))
- acomp_request_free(req);
- if (!IS_ERR_OR_NULL(acomp))
- crypto_free_acomp(acomp);
- kfree(buffer);
-
- return 0;
-}
-
-static struct crypto_acomp_ctx *acomp_ctx_get_cpu_lock(struct zswap_pool *pool)
-{
- struct crypto_acomp_ctx *acomp_ctx;
-
- for (;;) {
- acomp_ctx = raw_cpu_ptr(pool->acomp_ctx);
- mutex_lock(&acomp_ctx->mutex);
- if (likely(acomp_ctx->req))
- return acomp_ctx;
- /*
- * It is possible that we were migrated to a different CPU after
- * getting the per-CPU ctx but before the mutex was acquired. If
- * the old CPU got offlined, zswap_cpu_comp_dead() could have
- * already freed ctx->req (among other things) and set it to
- * NULL. Just try again on the new CPU that we ended up on.
- */
- mutex_unlock(&acomp_ctx->mutex);
- }
-}
-
-static void acomp_ctx_put_unlock(struct crypto_acomp_ctx *acomp_ctx)
-{
- mutex_unlock(&acomp_ctx->mutex);
-}
-
static bool zswap_compress(struct page *page, struct zswap_entry *entry,
struct zswap_pool *pool)
{
@@ -863,7 +852,9 @@ static bool zswap_compress(struct page *page, struct zswap_entry *entry,
u8 *dst;
bool mapped = false;
- acomp_ctx = acomp_ctx_get_cpu_lock(pool);
+ acomp_ctx = raw_cpu_ptr(pool->acomp_ctx);
+ mutex_lock(&acomp_ctx->mutex);
+
dst = acomp_ctx->buffer;
sg_init_table(&input, 1);
sg_set_page(&input, page, PAGE_SIZE, 0);
@@ -927,7 +918,7 @@ static bool zswap_compress(struct page *page, struct zswap_entry *entry,
else if (alloc_ret)
zswap_reject_alloc_fail++;
- acomp_ctx_put_unlock(acomp_ctx);
+ mutex_unlock(&acomp_ctx->mutex);
return comp_ret == 0 && alloc_ret == 0;
}
@@ -939,7 +930,8 @@ static bool zswap_decompress(struct zswap_entry *entry, struct folio *folio)
int decomp_ret = 0, dlen = PAGE_SIZE;
u8 *src, *obj;
- acomp_ctx = acomp_ctx_get_cpu_lock(pool);
+ acomp_ctx = raw_cpu_ptr(pool->acomp_ctx);
+ mutex_lock(&acomp_ctx->mutex);
obj = zs_obj_read_begin(pool->zs_pool, entry->handle, acomp_ctx->buffer);
/* zswap entries of length PAGE_SIZE are not compressed. */
@@ -970,7 +962,7 @@ static bool zswap_decompress(struct zswap_entry *entry, struct folio *folio)
read_done:
zs_obj_read_end(pool->zs_pool, entry->handle, obj);
- acomp_ctx_put_unlock(acomp_ctx);
+ mutex_unlock(&acomp_ctx->mutex);
if (!decomp_ret && dlen == PAGE_SIZE)
return true;
@@ -1796,7 +1788,7 @@ static int zswap_setup(void)
ret = cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE,
"mm/zswap_pool:prepare",
zswap_cpu_comp_prepare,
- zswap_cpu_comp_dead);
+ NULL);
if (ret)
goto hp_fail;
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 21/23] mm: zswap: Consistently use IS_ERR_OR_NULL() to check acomp_ctx resources.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (19 preceding siblings ...)
2025-09-26 3:34 ` [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources exist from pool creation to deletion Kanchana P Sridhar
@ 2025-09-26 3:35 ` Kanchana P Sridhar
2025-09-26 3:35 ` [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches Kanchana P Sridhar
` (2 subsequent siblings)
23 siblings, 0 replies; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:35 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch uses IS_ERR_OR_NULL() in zswap_cpu_comp_prepare() to check
for valid acomp/req, thereby making it consistent with
acomp_ctx_dealloc().
This is based on this earlier comment [1] from Yosry, when reviewing v8.
[1] https://patchwork.kernel.org/comment/26282128/
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
mm/zswap.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/mm/zswap.c b/mm/zswap.c
index 27665eaa3f89..3b3716808d7d 100644
--- a/mm/zswap.c
+++ b/mm/zswap.c
@@ -807,7 +807,7 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
return ret;
acomp_ctx->acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_to_node(cpu));
- if (IS_ERR(acomp_ctx->acomp)) {
+ if (IS_ERR_OR_NULL(acomp_ctx->acomp)) {
pr_err("could not alloc crypto acomp %s : %ld\n",
pool->tfm_name, PTR_ERR(acomp_ctx->acomp));
ret = PTR_ERR(acomp_ctx->acomp);
@@ -816,7 +816,7 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
acomp_ctx->is_sleepable = acomp_is_async(acomp_ctx->acomp);
acomp_ctx->req = acomp_request_alloc(acomp_ctx->acomp);
- if (!acomp_ctx->req) {
+ if (IS_ERR_OR_NULL(acomp_ctx->req)) {
pr_err("could not alloc crypto acomp_request %s\n",
pool->tfm_name);
goto fail;
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (20 preceding siblings ...)
2025-09-26 3:35 ` [PATCH v12 21/23] mm: zswap: Consistently use IS_ERR_OR_NULL() to check acomp_ctx resources Kanchana P Sridhar
@ 2025-09-26 3:35 ` Kanchana P Sridhar
2025-10-01 16:19 ` Yosry Ahmed
2025-09-26 3:35 ` [PATCH v12 23/23] mm: zswap: Batched zswap_compress() with compress batching of large folios Kanchana P Sridhar
2025-10-13 18:03 ` [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Sridhar, Kanchana P
23 siblings, 1 reply; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:35 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
This patch makes two major changes:
First, we allocate pool batching resources if the compressor supports
batching:
This patch sets up zswap for allocating per-CPU resources optimally
for non-batching and batching compressors.
A new ZSWAP_MAX_BATCH_SIZE constant is defined as 8U, to set an upper
limit on the number of pages in large folios that will be batch
compressed.
It is up to the compressor to manage multiple requests, as needed, to
accomplish batch parallelism. zswap only needs to allocate the per-CPU
dst buffers according to the batch size supported by the compressor.
A "u8 compr_batch_size" member is added to "struct zswap_pool", as per
Yosry's suggestion. pool->compr_batch_size is set as the minimum of
the compressor's max batch-size and ZSWAP_MAX_BATCH_SIZE. Accordingly,
it proceeds to allocate the necessary compression dst buffers in the
per-CPU acomp_ctx.
Another "u8 store_batch_size" member is added to "struct zswap_pool"
to store the unit for batching large folio stores: for batching
compressors, this is the pool->compr_batch_size. For non-batching
compressors, this is ZSWAP_MAX_BATCH_SIZE.
zswap does not use more than one dst buffer yet. Follow-up patches
will actually utilize the multiple acomp_ctx buffers for batch
compression/decompression of multiple pages.
Thus, ZSWAP_MAX_BATCH_SIZE limits the amount of extra memory used for
batching. There is a small extra memory overhead of allocating
the acomp_ctx->buffers array for compressors that do not support
batching: On x86_64, the overhead is 1 pointer per-CPU (i.e. 8 bytes).
Next, we store the folio in batches:
This patch modifies zswap_store() to store a batch of pages in large
folios at a time, instead of storing one page at a time. It does this by
calling a new procedure zswap_store_pages() with a range of
"pool->store_batch_size" indices in the folio.
zswap_store_pages() implements all the computes done earlier in
zswap_store_page() for a single-page, for multiple pages in a folio,
namely the "batch":
1) It starts by allocating all zswap entries required to store the
batch. New procedures, zswap_entries_cache_alloc_batch() and
zswap_entries_cache_free_batch() call kmem_cache_[free]alloc_bulk()
to optimize the performance of this step.
2) Next, the entries fields are written, computes that need to be happen
anyway, without modifying the zswap xarray/LRU publishing order. This
improves latency by avoiding having to bring the entries into the
cache for writing in different code blocks within this procedure.
3) Next, it calls zswap_compress() to sequentially compress each page in
the batch.
4) Finally, it adds the batch's zswap entries to the xarray and LRU,
charges zswap memory and increments zswap stats.
5) The error handling and cleanup required for all failure scenarios
that can occur while storing a batch in zswap are consolidated to a
single "store_pages_failed" label in zswap_store_pages(). Here again,
we optimize performance by calling kmem_cache_free_bulk().
This commit also makes a minor optimization in zswap_compress(), for the
info on whether or not the page's folio has memcg writeback enabled to
be passed in via a "bool folio_wb" flag from zswap_store(). The intent
is to not re-compute this for every page in a folio. Since
zswap_compress() is a static function, I figured this should be safe.
A repetition of "dlen = PAGE_SIZE" is deleted.
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
mm/zswap.c | 319 +++++++++++++++++++++++++++++++++++++----------------
1 file changed, 224 insertions(+), 95 deletions(-)
diff --git a/mm/zswap.c b/mm/zswap.c
index 3b3716808d7d..9e0e7887de33 100644
--- a/mm/zswap.c
+++ b/mm/zswap.c
@@ -82,6 +82,9 @@ static bool zswap_pool_reached_full;
#define ZSWAP_PARAM_UNSET ""
+/* Limit the batch size to limit per-CPU memory usage for dst buffers. */
+#define ZSWAP_MAX_BATCH_SIZE 8U
+
static int zswap_setup(void);
/* Enable/disable zswap */
@@ -139,7 +142,7 @@ struct crypto_acomp_ctx {
struct crypto_acomp *acomp;
struct acomp_req *req;
struct crypto_wait wait;
- u8 *buffer;
+ u8 **buffers;
struct mutex mutex;
bool is_sleepable;
};
@@ -158,6 +161,8 @@ struct zswap_pool {
struct work_struct release_work;
struct hlist_node node;
char tfm_name[CRYPTO_MAX_ALG_NAME];
+ u8 compr_batch_size;
+ u8 store_batch_size;
};
/* Global LRU lists shared by all zswap pools. */
@@ -252,8 +257,10 @@ static void __zswap_pool_empty(struct percpu_ref *ref);
* zswap_cpu_comp_prepare(), not others.
* - Cleanup acomp_ctx resources on all cores in zswap_pool_destroy().
*/
-static void acomp_ctx_dealloc(struct crypto_acomp_ctx *acomp_ctx)
+static void acomp_ctx_dealloc(struct crypto_acomp_ctx *acomp_ctx, u8 nr_buffers)
{
+ u8 i;
+
if (IS_ERR_OR_NULL(acomp_ctx))
return;
@@ -263,7 +270,11 @@ static void acomp_ctx_dealloc(struct crypto_acomp_ctx *acomp_ctx)
if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
crypto_free_acomp(acomp_ctx->acomp);
- kfree(acomp_ctx->buffer);
+ if (acomp_ctx->buffers) {
+ for (i = 0; i < nr_buffers; ++i)
+ kfree(acomp_ctx->buffers[i]);
+ kfree(acomp_ctx->buffers);
+ }
}
static struct zswap_pool *zswap_pool_create(char *compressor)
@@ -275,6 +286,7 @@ static struct zswap_pool *zswap_pool_create(char *compressor)
if (!zswap_has_pool && !strcmp(compressor, ZSWAP_PARAM_UNSET))
return NULL;
+ /* Many things rely on the zero-initialization. */
pool = kzalloc(sizeof(*pool), GFP_KERNEL);
if (!pool)
return NULL;
@@ -334,13 +346,28 @@ static struct zswap_pool *zswap_pool_create(char *compressor)
goto ref_fail;
INIT_LIST_HEAD(&pool->list);
+ /*
+ * Set the unit of compress batching for large folios, for quick
+ * retrieval in the zswap_compress() fast path:
+ * If the compressor is sequential (@pool->compr_batch_size is 1),
+ * large folios will be compressed in batches of ZSWAP_MAX_BATCH_SIZE
+ * pages, where each page in the batch is compressed sequentially.
+ * We see better performance by processing the folio in batches of
+ * ZSWAP_MAX_BATCH_SIZE, due to cache locality of working set
+ * structures.
+ */
+ pool->store_batch_size = (pool->compr_batch_size > 1) ?
+ pool->compr_batch_size : ZSWAP_MAX_BATCH_SIZE;
+
zswap_pool_debug("created", pool);
return pool;
ref_fail:
for_each_possible_cpu(cpu)
- acomp_ctx_dealloc(per_cpu_ptr(pool->acomp_ctx, cpu));
+ acomp_ctx_dealloc(per_cpu_ptr(pool->acomp_ctx, cpu),
+ pool->compr_batch_size);
+
error:
if (pool->acomp_ctx)
free_percpu(pool->acomp_ctx);
@@ -376,7 +403,8 @@ static void zswap_pool_destroy(struct zswap_pool *pool)
zswap_pool_debug("destroying", pool);
for_each_possible_cpu(cpu)
- acomp_ctx_dealloc(per_cpu_ptr(pool->acomp_ctx, cpu));
+ acomp_ctx_dealloc(per_cpu_ptr(pool->acomp_ctx, cpu),
+ pool->compr_batch_size);
free_percpu(pool->acomp_ctx);
@@ -763,6 +791,24 @@ static void zswap_entry_cache_free(struct zswap_entry *entry)
kmem_cache_free(zswap_entry_cache, entry);
}
+/*
+ * Returns 0 if kmem_cache_alloc_bulk() failed and a positive number otherwise.
+ * The code for __kmem_cache_alloc_bulk() indicates that this positive number
+ * will be the @size requested, i.e., @nr_entries.
+ */
+static __always_inline int zswap_entries_cache_alloc_batch(void **entries,
+ unsigned int nr_entries,
+ gfp_t gfp)
+{
+ return kmem_cache_alloc_bulk(zswap_entry_cache, gfp, nr_entries, entries);
+}
+
+static __always_inline void zswap_entries_cache_free_batch(void **entries,
+ unsigned int nr_entries)
+{
+ kmem_cache_free_bulk(zswap_entry_cache, nr_entries, entries);
+}
+
/*
* Carries out the common pattern of freeing an entry's zsmalloc allocation,
* freeing the entry itself, and decrementing the number of stored pages.
@@ -789,7 +835,9 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
{
struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
+ int cpu_node = cpu_to_node(cpu);
int ret = -ENOMEM;
+ u8 i;
/*
* The per-CPU pool->acomp_ctx is zero-initialized on allocation.
@@ -802,11 +850,7 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
return 0;
- acomp_ctx->buffer = kmalloc_node(PAGE_SIZE, GFP_KERNEL, cpu_to_node(cpu));
- if (!acomp_ctx->buffer)
- return ret;
-
- acomp_ctx->acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_to_node(cpu));
+ acomp_ctx->acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_node);
if (IS_ERR_OR_NULL(acomp_ctx->acomp)) {
pr_err("could not alloc crypto acomp %s : %ld\n",
pool->tfm_name, PTR_ERR(acomp_ctx->acomp));
@@ -815,20 +859,40 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
}
acomp_ctx->is_sleepable = acomp_is_async(acomp_ctx->acomp);
+ /*
+ * Allocate up to ZSWAP_MAX_BATCH_SIZE dst buffers if the
+ * compressor supports batching.
+ */
+ pool->compr_batch_size = min(ZSWAP_MAX_BATCH_SIZE,
+ crypto_acomp_batch_size(acomp_ctx->acomp));
+
acomp_ctx->req = acomp_request_alloc(acomp_ctx->acomp);
+
if (IS_ERR_OR_NULL(acomp_ctx->req)) {
pr_err("could not alloc crypto acomp_request %s\n",
- pool->tfm_name);
+ pool->tfm_name);
goto fail;
}
- crypto_init_wait(&acomp_ctx->wait);
+ acomp_ctx->buffers = kcalloc_node(pool->compr_batch_size, sizeof(u8 *),
+ GFP_KERNEL, cpu_node);
+ if (!acomp_ctx->buffers)
+ goto fail;
+
+ for (i = 0; i < pool->compr_batch_size; ++i) {
+ acomp_ctx->buffers[i] = kmalloc_node(PAGE_SIZE, GFP_KERNEL,
+ cpu_node);
+ if (!acomp_ctx->buffers[i])
+ goto fail;
+ }
/*
* if the backend of acomp is async zip, crypto_req_done() will wakeup
* crypto_wait_req(); if the backend of acomp is scomp, the callback
* won't be called, crypto_wait_req() will return without blocking.
*/
+ crypto_init_wait(&acomp_ctx->wait);
+
acomp_request_set_callback(acomp_ctx->req, CRYPTO_TFM_REQ_MAY_BACKLOG,
crypto_req_done, &acomp_ctx->wait);
@@ -836,12 +900,12 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
return 0;
fail:
- acomp_ctx_dealloc(acomp_ctx);
+ acomp_ctx_dealloc(acomp_ctx, pool->compr_batch_size);
return ret;
}
static bool zswap_compress(struct page *page, struct zswap_entry *entry,
- struct zswap_pool *pool)
+ struct zswap_pool *pool, bool folio_wb)
{
struct crypto_acomp_ctx *acomp_ctx;
struct scatterlist input, output;
@@ -855,7 +919,7 @@ static bool zswap_compress(struct page *page, struct zswap_entry *entry,
acomp_ctx = raw_cpu_ptr(pool->acomp_ctx);
mutex_lock(&acomp_ctx->mutex);
- dst = acomp_ctx->buffer;
+ dst = acomp_ctx->buffers[0];
sg_init_table(&input, 1);
sg_set_page(&input, page, PAGE_SIZE, 0);
@@ -886,13 +950,11 @@ static bool zswap_compress(struct page *page, struct zswap_entry *entry,
*/
if (comp_ret || !dlen || dlen >= PAGE_SIZE) {
dlen = PAGE_SIZE;
- if (!mem_cgroup_zswap_writeback_enabled(
- folio_memcg(page_folio(page)))) {
+ if (!folio_wb) {
comp_ret = comp_ret ? comp_ret : -EINVAL;
goto unlock;
}
comp_ret = 0;
- dlen = PAGE_SIZE;
dst = kmap_local_page(page);
mapped = true;
}
@@ -932,7 +994,7 @@ static bool zswap_decompress(struct zswap_entry *entry, struct folio *folio)
acomp_ctx = raw_cpu_ptr(pool->acomp_ctx);
mutex_lock(&acomp_ctx->mutex);
- obj = zs_obj_read_begin(pool->zs_pool, entry->handle, acomp_ctx->buffer);
+ obj = zs_obj_read_begin(pool->zs_pool, entry->handle, acomp_ctx->buffers[0]);
/* zswap entries of length PAGE_SIZE are not compressed. */
if (entry->length == PAGE_SIZE) {
@@ -942,15 +1004,15 @@ static bool zswap_decompress(struct zswap_entry *entry, struct folio *folio)
/*
* zs_obj_read_begin() might return a kmap address of highmem when
- * acomp_ctx->buffer is not used. However, sg_init_one() does not
- * handle highmem addresses, so copy the object to acomp_ctx->buffer.
+ * acomp_ctx->buffers[0] is not used. However, sg_init_one() does not
+ * handle highmem addresses, so copy the object to acomp_ctx->buffers[0].
*/
if (virt_addr_valid(obj)) {
src = obj;
} else {
- WARN_ON_ONCE(obj == acomp_ctx->buffer);
- memcpy(acomp_ctx->buffer, obj, entry->length);
- src = acomp_ctx->buffer;
+ WARN_ON_ONCE(obj == acomp_ctx->buffers[0]);
+ memcpy(acomp_ctx->buffers[0], obj, entry->length);
+ src = acomp_ctx->buffers[0];
}
sg_init_one(&input, src, entry->length);
@@ -1404,95 +1466,160 @@ static void shrink_worker(struct work_struct *w)
* main API
**********************************/
-static bool zswap_store_page(struct page *page,
- struct obj_cgroup *objcg,
- struct zswap_pool *pool)
+/*
+ * Store multiple pages in @folio, starting from the page at index @start up to
+ * the page at index @end-1.
+ */
+static bool zswap_store_pages(struct folio *folio,
+ long start,
+ long end,
+ struct obj_cgroup *objcg,
+ struct zswap_pool *pool,
+ int node_id,
+ bool folio_wb)
{
- swp_entry_t page_swpentry = page_swap_entry(page);
- struct zswap_entry *entry, *old;
-
- /* allocate entry */
- entry = zswap_entry_cache_alloc(GFP_KERNEL, page_to_nid(page));
- if (!entry) {
- zswap_reject_kmemcache_fail++;
- return false;
+ struct zswap_entry *entries[ZSWAP_MAX_BATCH_SIZE];
+ u8 i, store_fail_idx = 0, nr_pages = end - start;
+
+ VM_WARN_ON_ONCE(nr_pages > ZSWAP_MAX_BATCH_SIZE);
+
+ if (unlikely(!zswap_entries_cache_alloc_batch((void **)&entries[0],
+ nr_pages, GFP_KERNEL))) {
+ for (i = 0; i < nr_pages; ++i) {
+ entries[i] = zswap_entry_cache_alloc(GFP_KERNEL, node_id);
+
+ if (unlikely(!entries[i])) {
+ zswap_reject_kmemcache_fail++;
+ /*
+ * While handling this error, we only need to
+ * call zswap_entries_cache_free_batch() for
+ * entries[0 .. i-1].
+ */
+ nr_pages = i;
+ goto store_pages_failed;
+ }
+ }
}
- if (!zswap_compress(page, entry, pool))
- goto compress_failed;
+ /*
+ * Three sets of initializations are done to minimize bringing
+ * @entries into the cache for writing at different parts of this
+ * procedure, since doing so regresses performance:
+ *
+ * 1) Do all the writes to each entry in one code block. These
+ * writes need to be done anyway upon success which is more likely
+ * than not.
+ *
+ * 2) Initialize the handle to an error value. This facilitates
+ * having a consolidated failure handling
+ * 'goto store_pages_failed' that can inspect the value of the
+ * handle to determine whether zsmalloc memory needs to be
+ * de-allocated.
+ *
+ * 3) The page_swap_entry() is obtained once and stored in the entry.
+ * Subsequent store in xarray gets the entry->swpentry instead of
+ * calling page_swap_entry(), minimizing computes.
+ */
+ for (i = 0; i < nr_pages; ++i) {
+ entries[i]->handle = (unsigned long)ERR_PTR(-EINVAL);
+ entries[i]->pool = pool;
+ entries[i]->swpentry = page_swap_entry(folio_page(folio, start + i));
+ entries[i]->objcg = objcg;
+ entries[i]->referenced = true;
+ INIT_LIST_HEAD(&entries[i]->lru);
+ }
- old = xa_store(swap_zswap_tree(page_swpentry),
- swp_offset(page_swpentry),
- entry, GFP_KERNEL);
- if (xa_is_err(old)) {
- int err = xa_err(old);
+ for (i = 0; i < nr_pages; ++i) {
+ struct page *page = folio_page(folio, start + i);
- WARN_ONCE(err != -ENOMEM, "unexpected xarray error: %d\n", err);
- zswap_reject_alloc_fail++;
- goto store_failed;
+ if (!zswap_compress(page, entries[i], pool, folio_wb))
+ goto store_pages_failed;
}
- /*
- * We may have had an existing entry that became stale when
- * the folio was redirtied and now the new version is being
- * swapped out. Get rid of the old.
- */
- if (old)
- zswap_entry_free(old);
+ for (i = 0; i < nr_pages; ++i) {
+ struct zswap_entry *old, *entry = entries[i];
- /*
- * The entry is successfully compressed and stored in the tree, there is
- * no further possibility of failure. Grab refs to the pool and objcg,
- * charge zswap memory, and increment zswap_stored_pages.
- * The opposite actions will be performed by zswap_entry_free()
- * when the entry is removed from the tree.
- */
- zswap_pool_get(pool);
- if (objcg) {
- obj_cgroup_get(objcg);
- obj_cgroup_charge_zswap(objcg, entry->length);
- }
- atomic_long_inc(&zswap_stored_pages);
- if (entry->length == PAGE_SIZE)
- atomic_long_inc(&zswap_stored_incompressible_pages);
+ old = xa_store(swap_zswap_tree(entry->swpentry),
+ swp_offset(entry->swpentry),
+ entry, GFP_KERNEL);
+ if (unlikely(xa_is_err(old))) {
+ int err = xa_err(old);
- /*
- * We finish initializing the entry while it's already in xarray.
- * This is safe because:
- *
- * 1. Concurrent stores and invalidations are excluded by folio lock.
- *
- * 2. Writeback is excluded by the entry not being on the LRU yet.
- * The publishing order matters to prevent writeback from seeing
- * an incoherent entry.
- */
- entry->pool = pool;
- entry->swpentry = page_swpentry;
- entry->objcg = objcg;
- entry->referenced = true;
- if (entry->length) {
- INIT_LIST_HEAD(&entry->lru);
- zswap_lru_add(&zswap_list_lru, entry);
+ WARN_ONCE(err != -ENOMEM, "unexpected xarray error: %d\n", err);
+ zswap_reject_alloc_fail++;
+ /*
+ * Entries up to this point have been stored in the
+ * xarray. zswap_store() will erase them from the xarray
+ * and call zswap_entry_free(). Local cleanup in
+ * 'store_pages_failed' only needs to happen for
+ * entries from [@i to @nr_pages).
+ */
+ store_fail_idx = i;
+ goto store_pages_failed;
+ }
+
+ /*
+ * We may have had an existing entry that became stale when
+ * the folio was redirtied and now the new version is being
+ * swapped out. Get rid of the old.
+ */
+ if (unlikely(old))
+ zswap_entry_free(old);
+
+ /*
+ * The entry is successfully compressed and stored in the tree, there is
+ * no further possibility of failure. Grab refs to the pool and objcg,
+ * charge zswap memory, and increment zswap_stored_pages.
+ * The opposite actions will be performed by zswap_entry_free()
+ * when the entry is removed from the tree.
+ */
+ zswap_pool_get(pool);
+ if (objcg) {
+ obj_cgroup_get(objcg);
+ obj_cgroup_charge_zswap(objcg, entry->length);
+ }
+ atomic_long_inc(&zswap_stored_pages);
+ if (entry->length == PAGE_SIZE)
+ atomic_long_inc(&zswap_stored_incompressible_pages);
+
+ /*
+ * We finish by adding the entry to the LRU while it's already
+ * in xarray. This is safe because:
+ *
+ * 1. Concurrent stores and invalidations are excluded by folio lock.
+ *
+ * 2. Writeback is excluded by the entry not being on the LRU yet.
+ * The publishing order matters to prevent writeback from seeing
+ * an incoherent entry.
+ */
+ if (likely(entry->length))
+ zswap_lru_add(&zswap_list_lru, entry);
}
return true;
-store_failed:
- zs_free(pool->zs_pool, entry->handle);
-compress_failed:
- zswap_entry_cache_free(entry);
+store_pages_failed:
+ for (i = store_fail_idx; i < nr_pages; ++i) {
+ if (!IS_ERR_VALUE(entries[i]->handle))
+ zs_free(pool->zs_pool, entries[i]->handle);
+ }
+ zswap_entries_cache_free_batch((void **)&entries[store_fail_idx],
+ nr_pages - store_fail_idx);
+
return false;
}
bool zswap_store(struct folio *folio)
{
+ bool folio_wb = mem_cgroup_zswap_writeback_enabled(folio_memcg(folio));
long nr_pages = folio_nr_pages(folio);
+ int node_id = folio_nid(folio);
swp_entry_t swp = folio->swap;
struct obj_cgroup *objcg = NULL;
struct mem_cgroup *memcg = NULL;
struct zswap_pool *pool;
bool ret = false;
- long index;
+ long start, end;
VM_WARN_ON_ONCE(!folio_test_locked(folio));
VM_WARN_ON_ONCE(!folio_test_swapcache(folio));
@@ -1526,10 +1653,12 @@ bool zswap_store(struct folio *folio)
mem_cgroup_put(memcg);
}
- for (index = 0; index < nr_pages; ++index) {
- struct page *page = folio_page(folio, index);
+ /* Store the folio in batches of @pool->store_batch_size pages. */
+ for (start = 0; start < nr_pages; start += pool->store_batch_size) {
+ end = min(start + pool->store_batch_size, nr_pages);
- if (!zswap_store_page(page, objcg, pool))
+ if (!zswap_store_pages(folio, start, end, objcg, pool,
+ node_id, folio_wb))
goto put_pool;
}
@@ -1559,9 +1688,9 @@ bool zswap_store(struct folio *folio)
struct zswap_entry *entry;
struct xarray *tree;
- for (index = 0; index < nr_pages; ++index) {
- tree = swap_zswap_tree(swp_entry(type, offset + index));
- entry = xa_erase(tree, offset + index);
+ for (start = 0; start < nr_pages; ++start) {
+ tree = swap_zswap_tree(swp_entry(type, offset + start));
+ entry = xa_erase(tree, offset + start);
if (entry)
zswap_entry_free(entry);
}
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* [PATCH v12 23/23] mm: zswap: Batched zswap_compress() with compress batching of large folios.
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (21 preceding siblings ...)
2025-09-26 3:35 ` [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches Kanchana P Sridhar
@ 2025-09-26 3:35 ` Kanchana P Sridhar
2025-10-22 5:18 ` Herbert Xu
2025-10-13 18:03 ` [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Sridhar, Kanchana P
23 siblings, 1 reply; 56+ messages in thread
From: Kanchana P Sridhar @ 2025-09-26 3:35 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, kristen.c.accardi,
vinicius.gomes
Cc: wajdi.k.feghali, vinodh.gopal, kanchana.p.sridhar
crypto: iaa - Use the memory allocated in acomp_req->__ctx[] for batching.
This patch introduces a new unified implementation of zswap_compress()
for compressors that do and do not support batching. This eliminates
code duplication and facilitates maintainability of the code with the
introduction of compress batching.
The vectorized implementation of calling the earlier zswap_compress()
sequentially, one page at a time in zswap_store_pages(), is replaced
with this new version of zswap_compress() that accepts multiple pages to
compress as a batch.
If the compressor does not support batching, each page in the batch is
compressed and stored sequentially. If the compressor supports batching,
for e.g., 'deflate-iaa', the Intel IAA hardware accelerator, the batch
is compressed in parallel in hardware. If the batch is compressed
without errors, the compressed buffers are then stored in zsmalloc. In
case of compression errors, the current behavior is preserved for the
batching zswap_compress(): if the folio's memcg is writeback enabled,
pages with compression errors are store uncompressed in zsmalloc; if
not, we return an error for the folio in zswap_store().
As per Herbert's suggestion in [1] for batching to be based on SG lists
to interface with the crypto API, an "struct sg_table *sg_outputs" is
added to the per-CPU acomp_ctx. In zswap_cpu_comp_prepare(), memory is
allocated for @pool->compr_batch_size scatterlists in
@acomp_ctx->sg_outputs. The per-CPU @acomp_ctx->buffers' addresses are
mapped to the respective SG in @acomp_ctx->sg_outputs. This is done once
and does not need to be repeated in zswap_compress(). The existing
non-NUMA sg_alloc_table() was found to give better performance than a
NUMA-aware allocation function, hence is used in this patch.
All that zswap_compress() needs to do for non-batching software
compressors is to set the singular output SG's length to PAGE_SIZE.
Batching compressors should initialize the output SG lengths to
PAGE_SIZE as part of the internal compress batching setup, to avoid
having to do multiple traversals over the @acomp_ctx->sg_outputs->sgl.
This is exactly how batching is implemented in the iaa_crypto driver's
compress batching procedure, iaa_comp_acompress_batch().
Another initialization level optimization that happens in
zswap_cpu_comp_prepare() is, only for batching compressors, we allocate
extra memory for "batch-size" int pointers in the
@acomp_ctx->req->__ctx[] that are statically set to track the output SG
lists' lengths. This optimization was necessary to avoid the latency
cost of multiple @acomp_ctx->sg_outputs->sgl traversals in
zswap_compress() and in iaa_comp_acompress_batch() and to recover
batching performance with the SG lists based architecture. This
optimization's per-CPU memory cost for a batching compressor with
batch-size of 8 is 64 bytes. There is no memory cost for software
compressors.
Consequently, batching compressors can use the memory allocated in
@acomp_ctx->req->__ctx[] to internally manage updates to the output
@sg->lengths for the batch. zswap_compress() does not need to
traverse @pool->compr_batch_size number of output SG list elements to
get the compressed output length/error for each page in the batch.
This is readily available in the per-CPU acomp_ctx->req->__ctx[].
On a related note, the code in zswap_compress() is generalized as much
as possible for software compressors, by introducing a local @dstlen int
pointer to track @acomp_ctx->req->dlen, and assigning it the @err return
value from crypto_acomp_compress(), so that the subsequent
incompressible page handling, zs_pool writes, and error handling code is
seamless for software and batching compressors, i.e., avoids
conditionals to switch to specialized code for either.
Finally, again as per Herbert's suggestion, we call:
acomp_request_set_unit_size(acomp_ctx->req, PAGE_SIZE);
to set the unit size for zswap to PAGE_SIZE, after the call to
acomp_request_set_callback() in zswap_cpu_comp_prepare().
The new batching implementation of zswap_compress() is called with a
batch of @nr_pages sent to zswap_store_pages(). It sets up the
acomp_ctx->req's src/dst SG lists to contain the folio pages and
@nr_comps output buffers, where @nr_comps is min(nr_pages,
pool->compr_batch_size); before calling crypto_acomp_compress().
An upfront mapping of @dlens to the @acomp_ctx->req->__ctx for batching
compressors, and to @acomp_ctx->req->dlen for software compressors,
simplifies the zs_pool writes and error handling after compression.
Some important requirements of this batching architecture for batching
compressors:
1) The output SG lengths for each sg in the acomp_req->dst should be
intialized to PAGE_SIZE as part of other batch setup in the batch
compression function. zswap will not take care of this in the
interest of avoiding repetitive traversals of the
@acomp_ctx->sg_outputs->sgl so as to not lose the benefits of
batching.
2) In case of a compression error for any page in the batch, the
batching compressor should set the corresponding @sg->length to a
negative error number, as suggested by Herbert. Otherwise, the
@sg->length will contain the compressed output length.
Another important change this patch makes is with the acomp_ctx mutex
locking in zswap_compress(). Earlier, the mutex was held per page's
compression. With the new code, [un]locking the mutex per page caused
regressions for software compressors when testing with usemem
(30 processes) and also kernel compilation with 'allmod' config. The
regressions were more eggregious when PMD folios were stored. The
implementation in this commit locks/unlocks the mutex once per batch,
that resolves the regression.
The use of likely()/unlikely() annotations prevent regressions with
software compressors like zstd, and generally improve non-batching
compressors' performance with the batching code by ~3%.
Architectural considerations for the zswap batching framework:
==============================================================
We have designed the zswap batching framework to be
hardware-agnostic. It has no dependencies on Intel-specific features and
can be leveraged by any hardware accelerator or software-based
compressor. In other words, the framework is open and inclusive by
design.
Other ongoing work that can use batching:
=========================================
This patch-series demonstrates the performance benefits of compress
batching when used in zswap_store() of large folios. shrink_folio_list()
"reclaim batching" of any-order folios is the major next work that uses
the zswap compress batching framework: our testing of kernel_compilation
with writeback and the zswap shrinker indicates 10X fewer pages get
written back when we reclaim 32 folios as a batch, as compared to one
folio at a time: this is with deflate-iaa and with zstd. We expect to
submit a patch-series with this data and the resulting performance
improvements shortly. Reclaim batching relieves memory pressure faster
than reclaiming one folio at a time, hence alleviates the need to scan
slab memory for writeback.
Nhat has given ideas on using batching with the ongoing kcompressd work,
as well as beneficially using decompression batching & block IO batching
to improve zswap writeback efficiency.
Experiments that combine zswap compress batching, reclaim batching,
swapin_readahead() decompression batching of prefetched pages, and
writeback batching show that 0 pages are written back with deflate-iaa
and zstd. For comparison, the baselines for these compressors see
200K-800K pages written to disk (kernel compilation 'allmod' config).
To summarize, these are future clients of the batching framework:
- shrink_folio_list() reclaim batching of multiple folios:
Implemented, will submit patch-series.
- zswap writeback with decompress batching:
Implemented, will submit patch-series.
- zram:
Implemented, will submit patch-series.
- kcompressd:
Not yet implemented.
- file systems:
Not yet implemented.
- swapin_readahead() decompression batching of prefetched pages:
Implemented, will submit patch-series.
Additionally, any place we have folios that need to be compressed, can
potentially be parallelized.
Performance data:
=================
As suggested by Barry, this is the performance data gathered on Intel
Sapphire Rapids with usemem 30 processes running at 50% memory pressure
and kernel_compilation/allmod config run with 2G limit using 32
threads. To keep comparisons simple, all testing was done without the
zswap shrinker.
usemem30, 64K folios:
=====================
------------------------------------------------------------------
deflate-iaa
mm-unstable-9-18-2025 v12 Change
------------------------------------------------------------------
Total throughput (KB/s) 7,191,949 10,702,115 49%
Average throughput (KB/s) 239,731 356,737 49%
elapsed time (sec) 93.21 69.98 -25%
sys time (sec) 2,190.52 1,651.51 -25%
------------------------------------------------------------------
------------------------------------------------------------------
zstd
mm-unstable-9-18-2025 v12 Change
------------------------------------------------------------------
Total throughput (KB/s) 6,258,312 6,269,511 0.2%
Average throughput (KB/s) 208,610 208,983 0.2%
elapsed time (sec) 100.01 100.50 0.5%
sys time (sec) 2,505.14 2,490.00 -1%
------------------------------------------------------------------
usemem30, PMD folios:
=====================
------------------------------------------------------------------
deflate-iaa
mm-unstable-9-18-2025 v12 Change
------------------------------------------------------------------
Total throughput (KB/s) 7,237,676 11,228,928 55%
Average throughput (KB/s) 241,255 374,297 55%
elapsed time (sec) 82.26 69.30 -16%
sys time (sec) 1,901.90 1,634.78 -14%
------------------------------------------------------------------
------------------------------------------------------------------
zstd
mm-unstable-9-18-2025 v12 Change
------------------------------------------------------------------
Total throughput (KB/s) 6,796,376 6,865,781 1.0%
Average throughput (KB/s) 226,545 228,859 1.0%
elapsed time (sec) 94.07 88.80 -6%
sys time (sec) 2,261.67 2,082.91 -8%
------------------------------------------------------------------
The main takeaway from usemem, a workload that is mostly compression
dominated (very few swapins) is that the higher the number of batches,
such as, with larger folios, the more the benefit of batching cost
amortization, as shown by the PMD usemem data. This aligns well with the
future direction noted earlier.
kernel_compilation/allmodconfig, 64K folios:
============================================
-------------------------------------------------------
deflate-iaa
mm-unstable-9-18-2025 v12 Change
-------------------------------------------------------
real_sec 874.74 821.59 -6.1%
sys_sec 3,834.35 3,791.12 -1%
-------------------------------------------------------
-------------------------------------------------------
zstd
mm-unstable-9-18-2025 v12 Change
-------------------------------------------------------
real_sec 925.08 853.14 -7.8%
sys_sec 5,318.65 5,172.23 -2.8%
-------------------------------------------------------
kernel_compilation/allmodconfig, PMD folios:
============================================
-------------------------------------------------------
deflate-iaa
mm-unstable-9-18-2025 v12 Change
-------------------------------------------------------
real_sec 808.10 794.85 -1.6%
sys_sec 4,351.01 4,266.95 -2%
-------------------------------------------------------
-------------------------------------------------------
zstd
mm-unstable-9-18-2025 v12 Change
-------------------------------------------------------
real_sec 848.06 845.42 -0.3%
sys_sec 5,898.58 5,741.31 -2.7%
-------------------------------------------------------
[1]: https://lore.kernel.org/all/aJ7Fk6RpNc815Ivd@gondor.apana.org.au/T/#m99aea2ce3d284e6c5a3253061d97b08c4752a798
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
---
drivers/crypto/intel/iaa/iaa_crypto_main.c | 4 +-
mm/zswap.c | 311 ++++++++++++++++-----
2 files changed, 247 insertions(+), 68 deletions(-)
diff --git a/drivers/crypto/intel/iaa/iaa_crypto_main.c b/drivers/crypto/intel/iaa/iaa_crypto_main.c
index 2fa38176034d..019f80f87993 100644
--- a/drivers/crypto/intel/iaa/iaa_crypto_main.c
+++ b/drivers/crypto/intel/iaa/iaa_crypto_main.c
@@ -1510,7 +1510,9 @@ static __always_inline void acomp_to_iaa(struct acomp_req *areq,
req->slen = areq->slen;
req->dlen = areq->dlen;
req->flags = areq->base.flags;
- if (unlikely(ctx->use_irq))
+ if (likely(!ctx->use_irq))
+ req->dlens = (int **)&areq->__ctx[0];
+ else
req->drv_data = areq;
}
diff --git a/mm/zswap.c b/mm/zswap.c
index 9e0e7887de33..8104ecb30a1a 100644
--- a/mm/zswap.c
+++ b/mm/zswap.c
@@ -143,6 +143,7 @@ struct crypto_acomp_ctx {
struct acomp_req *req;
struct crypto_wait wait;
u8 **buffers;
+ struct sg_table *sg_outputs;
struct mutex mutex;
bool is_sleepable;
};
@@ -275,6 +276,11 @@ static void acomp_ctx_dealloc(struct crypto_acomp_ctx *acomp_ctx, u8 nr_buffers)
kfree(acomp_ctx->buffers[i]);
kfree(acomp_ctx->buffers);
}
+
+ if (acomp_ctx->sg_outputs) {
+ sg_free_table(acomp_ctx->sg_outputs);
+ kfree(acomp_ctx->sg_outputs);
+ }
}
static struct zswap_pool *zswap_pool_create(char *compressor)
@@ -836,6 +842,7 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
int cpu_node = cpu_to_node(cpu);
+ struct scatterlist *sg;
int ret = -ENOMEM;
u8 i;
@@ -866,7 +873,28 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
pool->compr_batch_size = min(ZSWAP_MAX_BATCH_SIZE,
crypto_acomp_batch_size(acomp_ctx->acomp));
- acomp_ctx->req = acomp_request_alloc(acomp_ctx->acomp);
+ /*
+ * For batching compressors, we allocate extra memory for "batch-size"
+ * int pointers that will be statically set to track the output
+ * SG lists' lengths later in this procedure. This optimization was
+ * required to avoid the latency cost of SG lists traversal in
+ * zswap_compress().
+ *
+ * Consequently, batching compressors can use the memory allocated in
+ * acomp_ctx->req->__ctx[] to internally manage updates to the output
+ * @sg->lengths for the batch. zswap_compress() does not need to
+ * traverse ZSWAP_MAX_BATCH_SIZE number of output SG list elements to
+ * get the compressed output length/error for each page in the batch.
+ * This is readily available in the per-CPU acomp_ctx->req->__ctx[].
+ * This optimization's per-CPU memory cost for a batching compressor
+ * with batch-size of 8 is 64 bytes.
+ */
+ if (pool->compr_batch_size > 1)
+ acomp_ctx->req = acomp_request_alloc_extra(acomp_ctx->acomp,
+ pool->compr_batch_size * sizeof(int *),
+ GFP_KERNEL | __GFP_ZERO);
+ else
+ acomp_ctx->req = acomp_request_alloc(acomp_ctx->acomp);
if (IS_ERR_OR_NULL(acomp_ctx->req)) {
pr_err("could not alloc crypto acomp_request %s\n",
@@ -886,6 +914,37 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
goto fail;
}
+ acomp_ctx->sg_outputs = kmalloc(sizeof(*acomp_ctx->sg_outputs),
+ GFP_KERNEL);
+ if (!acomp_ctx->sg_outputs)
+ goto fail;
+
+ if (sg_alloc_table(acomp_ctx->sg_outputs, pool->compr_batch_size,
+ GFP_KERNEL))
+ goto fail;
+
+ /*
+ * Map the per-CPU destination buffers to the per-CPU SG list.
+ * This only needs to be done once.
+ */
+ for_each_sg(acomp_ctx->sg_outputs->sgl, sg, pool->compr_batch_size, i)
+ sg_set_buf(sg, acomp_ctx->buffers[i], PAGE_SIZE);
+
+ /*
+ * Use the @pool->compr_batch_size number of int pointers for
+ * which we allocated extra memory in the @acomp_ctx->req above, to
+ * track the addresses of the @sg->length members of the individual
+ * SG lists in @acomp_ctx->sg_outputs->sgl. This is a static mapping
+ * that needs to be done only once, and saves latency by avoiding
+ * traversing the SG lists in zswap_compress().
+ */
+ if (pool->compr_batch_size > 1) {
+ for_each_sg(acomp_ctx->sg_outputs->sgl, sg, pool->compr_batch_size, i)
+ acomp_ctx->req->__ctx[i] = &sg->length;
+ } else {
+ acomp_ctx->req->dlen = PAGE_SIZE;
+ }
+
/*
* if the backend of acomp is async zip, crypto_req_done() will wakeup
* crypto_wait_req(); if the backend of acomp is scomp, the callback
@@ -896,6 +955,8 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
acomp_request_set_callback(acomp_ctx->req, CRYPTO_TFM_REQ_MAY_BACKLOG,
crypto_req_done, &acomp_ctx->wait);
+ acomp_request_set_unit_size(acomp_ctx->req, PAGE_SIZE);
+
mutex_init(&acomp_ctx->mutex);
return 0;
@@ -904,84 +965,203 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
return ret;
}
-static bool zswap_compress(struct page *page, struct zswap_entry *entry,
- struct zswap_pool *pool, bool folio_wb)
+/*
+ * Unified code path for compressors that do and do not support batching. This
+ * procedure will compress multiple @nr_pages in @folio starting from the
+ * @start index.
+ *
+ * It is assumed that @nr_pages <= ZSWAP_MAX_BATCH_SIZE. zswap_store() makes
+ * sure of this by design and zswap_store_pages() warns once if this is not
+ * true.
+ *
+ * @nr_pages can be in (1, ZSWAP_MAX_BATCH_SIZE] even if the compressor does not
+ * support batching.
+ *
+ * If @pool->compr_batch_size is 1, each page is processed sequentially.
+ *
+ * If @pool->compr_batch_size is > 1, compression batching is invoked within
+ * the algorithm's driver, except if @nr_pages is 1: if so, the driver can
+ * choose to call the sequential/non-batching compress API.
+ *
+ * In both cases, if all compressions are successful, the compressed buffers
+ * are stored in zsmalloc.
+ *
+ * Traversing multiple SG lists when @nr_comps is > 1 is expensive, and impacts
+ * batching performance if we were to repeat this operation multiple times,
+ * such as:
+ * - to map destination buffers to each SG list in the @acomp_ctx->sg_outputs
+ * sg_table.
+ * - to initialize each output SG list @sg->length to PAGE_SIZE.
+ * - to get the compressed output length in each @sg->length.
+ *
+ * These are some design choices made to optimize batching with SG lists:
+ *
+ * 1) The source folio pages in the batch are directly submitted to
+ * crypto_acomp via acomp_request_set_src_folio().
+ *
+ * 2) The per-CPU @acomp_ctx->sg_outputs scatterlists are used to set up
+ * destination buffers for interfacing with crypto_acomp.
+ *
+ * 3) To optimize performance, we map the per-CPU @acomp_ctx->buffers to the
+ * @acomp_ctx->sg_outputs->sgl SG lists at pool creation time. The only task
+ * remaining to be done for the output SG lists in zswap_compress() is to
+ * set the @sg->length to PAGE_SIZE. This is done in zswap_compress()
+ * for non-batching compressors. This needs to be done within the compress
+ * batching driver procedure as part of iterating through the SG lists for
+ * batch setup, so as to minimize traversals through the SG lists.
+ *
+ * 4) For batching compressors, we allocate extra memory in the
+ * @acomp_ctx->req->__ctx[] to store @pool->compr_batch_size number of
+ * int pointers (the @dlens). These pointers are initialized to the
+ * individual @sg->lengths' addresses in @acomp_ctx->sg_outputs->sgl at pool
+ * creation time. zswap_compress() has this readily available without having
+ * to re-traverse @acomp_ctx->sg_outputs->sgl to get the compressed output
+ * lengths after batch compression.
+ *
+ * 5) An important requirement for compression errors and batching compressors:
+ * the individual @sg->lengths in @acomp_ctx->req->sg_outputs->sgl should be
+ * set to the error value for the respective batch page.
+ *
+ * A few important changes made to not regress and in fact improve
+ * compression performance with non-batching software compressors, using this
+ * new batching code:
+ *
+ * 1) acomp_ctx mutex locking:
+ * Earlier, the mutex was held per page compression. With the new code,
+ * [un]locking the mutex per page caused regressions for software
+ * compressors. We now lock the mutex once per batch, which resolves the
+ * regression.
+ *
+ * 2) The likely()/unlikely() annotations prevent regressions with software
+ * compressors like zstd, and generally improve non-batching compressors'
+ * performance with the batching code by ~3%.
+ */
+static bool zswap_compress(struct folio *folio, long start, unsigned int nr_pages,
+ struct zswap_entry *entries[], struct zswap_pool *pool,
+ int node_id, bool folio_wb)
{
+ unsigned int nr_comps = min(nr_pages, pool->compr_batch_size);
struct crypto_acomp_ctx *acomp_ctx;
- struct scatterlist input, output;
- int comp_ret = 0, alloc_ret = 0;
- unsigned int dlen = PAGE_SIZE;
+ int *dstlen[1], **dlens;
+ struct scatterlist *sg;
unsigned long handle;
+ unsigned int i, j, k;
+ void *dst;
gfp_t gfp;
- u8 *dst;
- bool mapped = false;
+ int err;
+
+ gfp = GFP_NOWAIT | __GFP_NORETRY | __GFP_HIGHMEM | __GFP_MOVABLE;
acomp_ctx = raw_cpu_ptr(pool->acomp_ctx);
mutex_lock(&acomp_ctx->mutex);
- dst = acomp_ctx->buffers[0];
- sg_init_table(&input, 1);
- sg_set_page(&input, page, PAGE_SIZE, 0);
-
- sg_init_one(&output, dst, PAGE_SIZE);
- acomp_request_set_params(acomp_ctx->req, &input, &output, PAGE_SIZE, dlen);
+ dstlen[0] = &acomp_ctx->req->dlen;
+ if (pool->compr_batch_size == 1)
+ dlens = &dstlen[0];
+ else
+ dlens = (int **)&acomp_ctx->req->__ctx[0];
/*
- * it maybe looks a little bit silly that we send an asynchronous request,
- * then wait for its completion synchronously. This makes the process look
- * synchronous in fact.
- * Theoretically, acomp supports users send multiple acomp requests in one
- * acomp instance, then get those requests done simultaneously. but in this
- * case, zswap actually does store and load page by page, there is no
- * existing method to send the second page before the first page is done
- * in one thread doing zwap.
- * but in different threads running on different cpu, we have different
- * acomp instance, so multiple threads can do (de)compression in parallel.
+ * [i] refers to the incoming batch space and is used to
+ * index into the folio pages.
+ *
+ * [j] refers to the incoming batch space and is used to
+ * index into the @entries for the folio's pages in this
+ * batch, per compress call while iterating over the output SG
+ * lists. Also used to index into the folio's pages from @start, in
+ * case of compress errors.
+ *
+ * [k] refers to the @acomp_ctx space, as determined by
+ * @pool->compr_batch_size, and is used to index into
+ * @acomp_ctx->sg_outputs->sgl, @acomp_ctx->buffers and @dlens.
*/
- comp_ret = crypto_wait_req(crypto_acomp_compress(acomp_ctx->req), &acomp_ctx->wait);
- dlen = acomp_ctx->req->dlen;
+ for (i = 0; i < nr_pages; i += nr_comps) {
+ acomp_request_set_src_folio(acomp_ctx->req, folio,
+ (start + i) * PAGE_SIZE,
+ nr_comps * PAGE_SIZE);
- /*
- * If a page cannot be compressed into a size smaller than PAGE_SIZE,
- * save the content as is without a compression, to keep the LRU order
- * of writebacks. If writeback is disabled, reject the page since it
- * only adds metadata overhead. swap_writeout() will put the page back
- * to the active LRU list in the case.
- */
- if (comp_ret || !dlen || dlen >= PAGE_SIZE) {
- dlen = PAGE_SIZE;
- if (!folio_wb) {
- comp_ret = comp_ret ? comp_ret : -EINVAL;
- goto unlock;
+ acomp_ctx->sg_outputs->sgl->length = nr_comps * PAGE_SIZE;
+
+ acomp_request_set_dst_sg(acomp_ctx->req,
+ acomp_ctx->sg_outputs->sgl,
+ nr_comps * PAGE_SIZE);
+
+ err = crypto_wait_req(crypto_acomp_compress(acomp_ctx->req),
+ &acomp_ctx->wait);
+
+ /*
+ * If a page cannot be compressed into a size smaller than
+ * PAGE_SIZE, save the content as is without a compression, to
+ * keep the LRU order of writebacks. If writeback is disabled,
+ * reject the page since it only adds metadata overhead.
+ * swap_writeout() will put the page back to the active LRU list
+ * in the case.
+ *
+ * It is assumed that any compressor that sets the output length
+ * to 0 or a value >= PAGE_SIZE will also return a negative
+ * error status in @err; i.e, will not return a successful
+ * compression status in @err in this case.
+ */
+ if (unlikely(err)) {
+ *dstlen[0] = err;
+ if (!folio_wb)
+ goto compress_error;
}
- comp_ret = 0;
- dst = kmap_local_page(page);
- mapped = true;
- }
- gfp = GFP_NOWAIT | __GFP_NORETRY | __GFP_HIGHMEM | __GFP_MOVABLE;
- handle = zs_malloc(pool->zs_pool, dlen, gfp, page_to_nid(page));
- if (IS_ERR_VALUE(handle)) {
- alloc_ret = PTR_ERR((void *)handle);
- goto unlock;
- }
+ /*
+ * All @nr_comps pages were either successfully compressed, or,
+ * writeback is enabled for the folio's memcg and there were
+ * compression errors. In either case, store the pages in
+ * zsmalloc:
+ *
+ * - For successful compressions, store the compressed outputs.
+ * - For errors, store the page uncompressed.
+ */
+ for_each_sg(acomp_ctx->sg_outputs->sgl, sg, nr_comps, k) {
+ j = k + i;
+
+ dst = acomp_ctx->buffers[k];
+
+ if (unlikely(*dlens[k] < 0)) {
+ *dlens[k] = PAGE_SIZE;
+ dst = kmap_local_page(folio_page(folio, start + j));
+ }
+
+ handle = zs_malloc(pool->zs_pool, *dlens[k], gfp, node_id);
- zs_obj_write(pool->zs_pool, handle, dst, dlen);
- entry->handle = handle;
- entry->length = dlen;
+ if (unlikely(IS_ERR_VALUE(handle))) {
+ if (PTR_ERR((void *)handle) == -ENOSPC)
+ zswap_reject_compress_poor++;
+ else
+ zswap_reject_alloc_fail++;
-unlock:
- if (mapped)
- kunmap_local(dst);
- if (comp_ret == -ENOSPC || alloc_ret == -ENOSPC)
- zswap_reject_compress_poor++;
- else if (comp_ret)
- zswap_reject_compress_fail++;
- else if (alloc_ret)
- zswap_reject_alloc_fail++;
+ goto err_unlock;
+ }
+
+ zs_obj_write(pool->zs_pool, handle, dst, *dlens[k]);
+ entries[j]->handle = handle;
+ entries[j]->length = *dlens[k];
+ if (unlikely(dst != acomp_ctx->buffers[k]))
+ kunmap_local(dst);
+ }
+ } /* finished compress and store nr_pages. */
mutex_unlock(&acomp_ctx->mutex);
- return comp_ret == 0 && alloc_ret == 0;
+ return true;
+
+compress_error:
+ for (k = 0; k < nr_comps; ++k) {
+ if (*dlens[k] < 0) {
+ if (*dlens[k] == -ENOSPC)
+ zswap_reject_compress_poor++;
+ else
+ zswap_reject_compress_fail++;
+ }
+ }
+
+err_unlock:
+ mutex_unlock(&acomp_ctx->mutex);
+ return false;
}
static bool zswap_decompress(struct zswap_entry *entry, struct folio *folio)
@@ -1529,12 +1709,9 @@ static bool zswap_store_pages(struct folio *folio,
INIT_LIST_HEAD(&entries[i]->lru);
}
- for (i = 0; i < nr_pages; ++i) {
- struct page *page = folio_page(folio, start + i);
-
- if (!zswap_compress(page, entries[i], pool, folio_wb))
- goto store_pages_failed;
- }
+ if (unlikely(!zswap_compress(folio, start, nr_pages, entries, pool,
+ node_id, folio_wb)))
+ goto store_pages_failed;
for (i = 0; i < nr_pages; ++i) {
struct zswap_entry *old, *entry = entries[i];
--
2.27.0
^ permalink raw reply [flat|nested] 56+ messages in thread
* Re: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources exist from pool creation to deletion.
2025-09-26 3:34 ` [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources exist from pool creation to deletion Kanchana P Sridhar
@ 2025-09-30 15:49 ` Yosry Ahmed
2025-09-30 18:20 ` Sridhar, Kanchana P
0 siblings, 1 reply; 56+ messages in thread
From: Yosry Ahmed @ 2025-09-30 15:49 UTC (permalink / raw)
To: Kanchana P Sridhar
Cc: linux-kernel, linux-mm, hannes, nphamcs, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, kristen.c.accardi, vinicius.gomes,
wajdi.k.feghali, vinodh.gopal
On Thu, Sep 25, 2025 at 08:34:59PM -0700, Kanchana P Sridhar wrote:
> This patch simplifies the zswap_pool's per-CPU acomp_ctx resource
> management. Similar to the per-CPU acomp_ctx itself, the per-CPU
> acomp_ctx's resources' (acomp, req, buffer) lifetime will also be from
> pool creation to pool deletion. These resources will persist through CPU
> hotplug operations. The zswap_cpu_comp_dead() teardown callback has been
> deleted from the call to
> cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE). As a result, CPU
> offline hotplug operations will be no-ops as far as the acomp_ctx
> resources are concerned.
>
> This commit refactors the code from zswap_cpu_comp_dead() into a
> new function acomp_ctx_dealloc() that preserves the IS_ERR_OR_NULL()
> checks on acomp_ctx, req and acomp from the existing mainline
> implementation of zswap_cpu_comp_dead(). acomp_ctx_dealloc() is called
> to clean up acomp_ctx resources from all these procedures:
>
> 1) zswap_cpu_comp_prepare() when an error is encountered,
> 2) zswap_pool_create() when an error is encountered, and
> 3) from zswap_pool_destroy().
>
> The main benefit of using the CPU hotplug multi state instance startup
> callback to allocate the acomp_ctx resources is that it prevents the
> cores from being offlined until the multi state instance addition call
> returns.
>
> From Documentation/core-api/cpu_hotplug.rst:
>
> "The node list add/remove operations and the callback invocations are
> serialized against CPU hotplug operations."
>
> Furthermore, zswap_[de]compress() cannot contend with
> zswap_cpu_comp_prepare() because:
>
> - During pool creation/deletion, the pool is not in the zswap_pools
> list.
>
> - During CPU hot[un]plug, the CPU is not yet online, as Yosry pointed
> out. zswap_cpu_comp_prepare() will be executed on a control CPU,
> since CPUHP_MM_ZSWP_POOL_PREPARE is in the PREPARE section of "enum
> cpuhp_state". Thanks Yosry for sharing this observation!
>
> In both these cases, any recursions into zswap reclaim from
> zswap_cpu_comp_prepare() will be handled by the old pool.
>
> The above two observations enable the following simplifications:
>
> 1) zswap_cpu_comp_prepare(): CPU cannot be offlined. Reclaim cannot use
> the pool. Considerations for mutex init/locking and handling
> subsequent CPU hotplug online-offlines:
>
> Should we lock the mutex of current CPU's acomp_ctx from start to
> end? It doesn't seem like this is required. The CPU hotplug
> operations acquire a "cpuhp_state_mutex" before proceeding, hence
> they are serialized against CPU hotplug operations.
>
> If the process gets migrated while zswap_cpu_comp_prepare() is
> running, it will complete on the new CPU. In case of failures, we
> pass the acomp_ctx pointer obtained at the start of
> zswap_cpu_comp_prepare() to acomp_ctx_dealloc(), which again, can
> only undergo migration. There appear to be no contention scenarios
> that might cause inconsistent values of acomp_ctx's members. Hence,
> it seems there is no need for mutex_lock(&acomp_ctx->mutex) in
> zswap_cpu_comp_prepare().
>
> Since the pool is not yet on zswap_pools list, we don't need to
> initialize the per-CPU acomp_ctx mutex in zswap_pool_create(). This
> has been restored to occur in zswap_cpu_comp_prepare().
>
> zswap_cpu_comp_prepare() checks upfront if acomp_ctx->acomp is
> valid. If so, it returns success. This should handle any CPU
> hotplug online-offline transitions after pool creation is done.
>
> 2) CPU offline vis-a-vis zswap ops: Let's suppose the process is
> migrated to another CPU before the current CPU is dysfunctional. If
> zswap_[de]compress() holds the acomp_ctx->mutex lock of the offlined
> CPU, that mutex will be released once it completes on the new
> CPU. Since there is no teardown callback, there is no possibility of
> UAF.
>
> 3) Pool creation/deletion and process migration to another CPU:
>
> - During pool creation/deletion, the pool is not in the zswap_pools
> list. Hence it cannot contend with zswap ops on that CPU. However,
> the process can get migrated.
>
> Pool creation --> zswap_cpu_comp_prepare()
> --> process migrated:
> * CPU offline: no-op.
> * zswap_cpu_comp_prepare() continues
> to run on the new CPU to finish
> allocating acomp_ctx resources for
> the offlined CPU.
>
> Pool deletion --> acomp_ctx_dealloc()
> --> process migrated:
> * CPU offline: no-op.
> * acomp_ctx_dealloc() continues
> to run on the new CPU to finish
> de-allocating acomp_ctx resources
> for the offlined CPU.
>
> 4) Pool deletion vis-a-vis CPU onlining:
> To prevent possibility of race conditions between
> acomp_ctx_dealloc() freeing the acomp_ctx resources and the initial
> check for a valid acomp_ctx->acomp in zswap_cpu_comp_prepare(), we
> need to delete the multi state instance right after it is added, in
> zswap_pool_create().
>
> Summary of changes based on the above:
> --------------------------------------
> 1) Zero-initialization of pool->acomp_ctx in zswap_pool_create() to
> simplify and share common code for different error handling/cleanup
> related to the acomp_ctx.
>
> 2) Remove the node list instance right after node list add function
> call in zswap_pool_create(). This prevents race conditions between
> CPU onlining after initial pool creation, and acomp_ctx_dealloc()
> freeing the acomp_ctx resources.
>
> 3) zswap_pool_destroy() will call acomp_ctx_dealloc() to de-allocate
> the per-CPU acomp_ctx resources.
>
> 4) Changes to zswap_cpu_comp_prepare():
>
> a) Check if acomp_ctx->acomp is valid at the beginning and return,
> because the acomp_ctx is already initialized.
> b) Move the mutex_init to happen in this procedure, before it
> returns.
> c) All error conditions handled by calling acomp_ctx_dealloc().
>
> 5) New procedure acomp_ctx_dealloc() for common error/cleanup code.
>
> 6) No more multi state instance teardown callback. CPU offlining is a
> no-op as far as acomp_ctx resources are concerned.
>
> 7) Delete acomp_ctx_get_cpu_lock()/acomp_ctx_put_unlock(). Directly
> call mutex_lock(&acomp_ctx->mutex)/mutex_unlock(&acomp_ctx->mutex)
> in zswap_[de]compress().
>
> The per-CPU memory cost of not deleting the acomp_ctx resources upon CPU
> offlining, and only deleting them when the pool is destroyed, is as
> follows, on x86_64:
>
> IAA with 8 dst buffers for batching: 64.34 KB
> Software compressors with 1 dst buffer: 8.28 KB
>
> Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
Please try to make the commit logs a bit more summarized. Details are
helpful, but it's easy to lose track of things when it gets too long.
> ---
> mm/zswap.c | 194 +++++++++++++++++++++++++----------------------------
> 1 file changed, 93 insertions(+), 101 deletions(-)
>
> diff --git a/mm/zswap.c b/mm/zswap.c
> index c1af782e54ec..27665eaa3f89 100644
> --- a/mm/zswap.c
> +++ b/mm/zswap.c
> @@ -242,6 +242,30 @@ static inline struct xarray *swap_zswap_tree(swp_entry_t swp)
> **********************************/
> static void __zswap_pool_empty(struct percpu_ref *ref);
>
> +/*
> + * The per-cpu pool->acomp_ctx is zero-initialized on allocation. This makes
> + * it easy for different error conditions/cleanup related to the acomp_ctx
> + * to be handled by acomp_ctx_dealloc():
> + * - Errors during zswap_cpu_comp_prepare().
> + * - Partial success/error of cpuhp_state_add_instance() call in
> + * zswap_pool_create(). Only some cores could have executed
> + * zswap_cpu_comp_prepare(), not others.
> + * - Cleanup acomp_ctx resources on all cores in zswap_pool_destroy().
> + */
Comments describing specific code paths go out of date really fast. The
comment is probably unnecessary, it's easy to check the allocation path
to figure out that these are zero-initialized.
Also in general, please keep the comments as summarized as possible, and
only when the logic is not clear from the code.
> +static void acomp_ctx_dealloc(struct crypto_acomp_ctx *acomp_ctx)
> +{
> + if (IS_ERR_OR_NULL(acomp_ctx))
> + return;
> +
> + if (!IS_ERR_OR_NULL(acomp_ctx->req))
> + acomp_request_free(acomp_ctx->req);
> +
> + if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
> + crypto_free_acomp(acomp_ctx->acomp);
> +
> + kfree(acomp_ctx->buffer);
> +}
> +
> static struct zswap_pool *zswap_pool_create(char *compressor)
> {
> struct zswap_pool *pool;
> @@ -263,19 +287,43 @@ static struct zswap_pool *zswap_pool_create(char *compressor)
>
> strscpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
>
> - pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
> + /* Many things rely on the zero-initialization. */
> + pool->acomp_ctx = alloc_percpu_gfp(*pool->acomp_ctx,
> + GFP_KERNEL | __GFP_ZERO);
> if (!pool->acomp_ctx) {
> pr_err("percpu alloc failed\n");
> goto error;
> }
>
> - for_each_possible_cpu(cpu)
> - mutex_init(&per_cpu_ptr(pool->acomp_ctx, cpu)->mutex);
> -
> + /*
> + * This is serialized against CPU hotplug operations. Hence, cores
> + * cannot be offlined until this finishes.
> + * In case of errors, we need to goto "ref_fail" instead of "error"
> + * because there is no teardown callback registered anymore, for
> + * cpuhp_state_add_instance() to de-allocate resources as it rolls back
> + * state on cores before the CPU on which error was encountered.
> + */
> ret = cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
> &pool->node);
> +
> + /*
> + * We only needed the multi state instance add operation to invoke the
> + * startup callback for all cores without cores getting offlined. Since
> + * the acomp_ctx resources will now only be de-allocated when the pool
> + * is destroyed, we can safely remove the multi state instance. This
> + * minimizes (but does not eliminate) the possibility of
> + * zswap_cpu_comp_prepare() being invoked again due to a CPU
> + * offline-online transition. Removing the instance also prevents race
> + * conditions between CPU onlining after initial pool creation, and
> + * acomp_ctx_dealloc() freeing the acomp_ctx resources.
> + * Note that we delete the instance before checking the error status of
> + * the node list add operation because we want the instance removal even
> + * in case of errors in the former.
> + */
> + cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
> +
I don't understand what's wrong with the current flow? We call
cpuhp_state_remove_instance() in pool deletion before freeing up the
per-CPU resources. Why is this not enough?
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources exist from pool creation to deletion.
2025-09-30 15:49 ` Yosry Ahmed
@ 2025-09-30 18:20 ` Sridhar, Kanchana P
2025-09-30 18:29 ` Yosry Ahmed
0 siblings, 1 reply; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-09-30 18:20 UTC (permalink / raw)
To: Yosry Ahmed
Cc: linux-kernel, linux-mm, hannes, nphamcs, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius,
Feghali, Wajdi K, Gopal, Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Yosry Ahmed <yosry.ahmed@linux.dev>
> Sent: Tuesday, September 30, 2025 8:49 AM
> To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> hannes@cmpxchg.org; nphamcs@gmail.com; chengming.zhou@linux.dev;
> usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>;
> Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> <vinodh.gopal@intel.com>
> Subject: Re: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources
> exist from pool creation to deletion.
>
> On Thu, Sep 25, 2025 at 08:34:59PM -0700, Kanchana P Sridhar wrote:
> > This patch simplifies the zswap_pool's per-CPU acomp_ctx resource
> > management. Similar to the per-CPU acomp_ctx itself, the per-CPU
> > acomp_ctx's resources' (acomp, req, buffer) lifetime will also be from
> > pool creation to pool deletion. These resources will persist through CPU
> > hotplug operations. The zswap_cpu_comp_dead() teardown callback has
> been
> > deleted from the call to
> > cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE). As a result,
> CPU
> > offline hotplug operations will be no-ops as far as the acomp_ctx
> > resources are concerned.
> >
> > This commit refactors the code from zswap_cpu_comp_dead() into a
> > new function acomp_ctx_dealloc() that preserves the IS_ERR_OR_NULL()
> > checks on acomp_ctx, req and acomp from the existing mainline
> > implementation of zswap_cpu_comp_dead(). acomp_ctx_dealloc() is called
> > to clean up acomp_ctx resources from all these procedures:
> >
> > 1) zswap_cpu_comp_prepare() when an error is encountered,
> > 2) zswap_pool_create() when an error is encountered, and
> > 3) from zswap_pool_destroy().
> >
> > The main benefit of using the CPU hotplug multi state instance startup
> > callback to allocate the acomp_ctx resources is that it prevents the
> > cores from being offlined until the multi state instance addition call
> > returns.
> >
> > From Documentation/core-api/cpu_hotplug.rst:
> >
> > "The node list add/remove operations and the callback invocations are
> > serialized against CPU hotplug operations."
> >
> > Furthermore, zswap_[de]compress() cannot contend with
> > zswap_cpu_comp_prepare() because:
> >
> > - During pool creation/deletion, the pool is not in the zswap_pools
> > list.
> >
> > - During CPU hot[un]plug, the CPU is not yet online, as Yosry pointed
> > out. zswap_cpu_comp_prepare() will be executed on a control CPU,
> > since CPUHP_MM_ZSWP_POOL_PREPARE is in the PREPARE section of
> "enum
> > cpuhp_state". Thanks Yosry for sharing this observation!
> >
> > In both these cases, any recursions into zswap reclaim from
> > zswap_cpu_comp_prepare() will be handled by the old pool.
> >
> > The above two observations enable the following simplifications:
> >
> > 1) zswap_cpu_comp_prepare(): CPU cannot be offlined. Reclaim cannot
> use
> > the pool. Considerations for mutex init/locking and handling
> > subsequent CPU hotplug online-offlines:
> >
> > Should we lock the mutex of current CPU's acomp_ctx from start to
> > end? It doesn't seem like this is required. The CPU hotplug
> > operations acquire a "cpuhp_state_mutex" before proceeding, hence
> > they are serialized against CPU hotplug operations.
> >
> > If the process gets migrated while zswap_cpu_comp_prepare() is
> > running, it will complete on the new CPU. In case of failures, we
> > pass the acomp_ctx pointer obtained at the start of
> > zswap_cpu_comp_prepare() to acomp_ctx_dealloc(), which again, can
> > only undergo migration. There appear to be no contention scenarios
> > that might cause inconsistent values of acomp_ctx's members. Hence,
> > it seems there is no need for mutex_lock(&acomp_ctx->mutex) in
> > zswap_cpu_comp_prepare().
> >
> > Since the pool is not yet on zswap_pools list, we don't need to
> > initialize the per-CPU acomp_ctx mutex in zswap_pool_create(). This
> > has been restored to occur in zswap_cpu_comp_prepare().
> >
> > zswap_cpu_comp_prepare() checks upfront if acomp_ctx->acomp is
> > valid. If so, it returns success. This should handle any CPU
> > hotplug online-offline transitions after pool creation is done.
> >
> > 2) CPU offline vis-a-vis zswap ops: Let's suppose the process is
> > migrated to another CPU before the current CPU is dysfunctional. If
> > zswap_[de]compress() holds the acomp_ctx->mutex lock of the offlined
> > CPU, that mutex will be released once it completes on the new
> > CPU. Since there is no teardown callback, there is no possibility of
> > UAF.
> >
> > 3) Pool creation/deletion and process migration to another CPU:
> >
> > - During pool creation/deletion, the pool is not in the zswap_pools
> > list. Hence it cannot contend with zswap ops on that CPU. However,
> > the process can get migrated.
> >
> > Pool creation --> zswap_cpu_comp_prepare()
> > --> process migrated:
> > * CPU offline: no-op.
> > * zswap_cpu_comp_prepare() continues
> > to run on the new CPU to finish
> > allocating acomp_ctx resources for
> > the offlined CPU.
> >
> > Pool deletion --> acomp_ctx_dealloc()
> > --> process migrated:
> > * CPU offline: no-op.
> > * acomp_ctx_dealloc() continues
> > to run on the new CPU to finish
> > de-allocating acomp_ctx resources
> > for the offlined CPU.
> >
> > 4) Pool deletion vis-a-vis CPU onlining:
> > To prevent possibility of race conditions between
> > acomp_ctx_dealloc() freeing the acomp_ctx resources and the initial
> > check for a valid acomp_ctx->acomp in zswap_cpu_comp_prepare(), we
> > need to delete the multi state instance right after it is added, in
> > zswap_pool_create().
> >
> > Summary of changes based on the above:
> > --------------------------------------
> > 1) Zero-initialization of pool->acomp_ctx in zswap_pool_create() to
> > simplify and share common code for different error handling/cleanup
> > related to the acomp_ctx.
> >
> > 2) Remove the node list instance right after node list add function
> > call in zswap_pool_create(). This prevents race conditions between
> > CPU onlining after initial pool creation, and acomp_ctx_dealloc()
> > freeing the acomp_ctx resources.
> >
> > 3) zswap_pool_destroy() will call acomp_ctx_dealloc() to de-allocate
> > the per-CPU acomp_ctx resources.
> >
> > 4) Changes to zswap_cpu_comp_prepare():
> >
> > a) Check if acomp_ctx->acomp is valid at the beginning and return,
> > because the acomp_ctx is already initialized.
> > b) Move the mutex_init to happen in this procedure, before it
> > returns.
> > c) All error conditions handled by calling acomp_ctx_dealloc().
> >
> > 5) New procedure acomp_ctx_dealloc() for common error/cleanup code.
> >
> > 6) No more multi state instance teardown callback. CPU offlining is a
> > no-op as far as acomp_ctx resources are concerned.
> >
> > 7) Delete acomp_ctx_get_cpu_lock()/acomp_ctx_put_unlock(). Directly
> > call mutex_lock(&acomp_ctx->mutex)/mutex_unlock(&acomp_ctx-
> >mutex)
> > in zswap_[de]compress().
> >
> > The per-CPU memory cost of not deleting the acomp_ctx resources upon
> CPU
> > offlining, and only deleting them when the pool is destroyed, is as
> > follows, on x86_64:
> >
> > IAA with 8 dst buffers for batching: 64.34 KB
> > Software compressors with 1 dst buffer: 8.28 KB
> >
> > Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
>
> Please try to make the commit logs a bit more summarized. Details are
> helpful, but it's easy to lose track of things when it gets too long.
Thanks Yosry, for the feedback.
>
> > ---
> > mm/zswap.c | 194 +++++++++++++++++++++++++----------------------------
> > 1 file changed, 93 insertions(+), 101 deletions(-)
> >
> > diff --git a/mm/zswap.c b/mm/zswap.c
> > index c1af782e54ec..27665eaa3f89 100644
> > --- a/mm/zswap.c
> > +++ b/mm/zswap.c
> > @@ -242,6 +242,30 @@ static inline struct xarray
> *swap_zswap_tree(swp_entry_t swp)
> > **********************************/
> > static void __zswap_pool_empty(struct percpu_ref *ref);
> >
> > +/*
> > + * The per-cpu pool->acomp_ctx is zero-initialized on allocation. This
> makes
> > + * it easy for different error conditions/cleanup related to the acomp_ctx
> > + * to be handled by acomp_ctx_dealloc():
> > + * - Errors during zswap_cpu_comp_prepare().
> > + * - Partial success/error of cpuhp_state_add_instance() call in
> > + * zswap_pool_create(). Only some cores could have executed
> > + * zswap_cpu_comp_prepare(), not others.
> > + * - Cleanup acomp_ctx resources on all cores in zswap_pool_destroy().
> > + */
>
> Comments describing specific code paths go out of date really fast. The
> comment is probably unnecessary, it's easy to check the allocation path
> to figure out that these are zero-initialized.
>
> Also in general, please keep the comments as summarized as possible, and
> only when the logic is not clear from the code.
Sure. I have tried to explain the rationale for significant changes, but I can
look for opportunities to summarize. I was sort of hoping that v12 would
be it, but I can work on the comments being concise if this is crucial.
>
> > +static void acomp_ctx_dealloc(struct crypto_acomp_ctx *acomp_ctx)
> > +{
> > + if (IS_ERR_OR_NULL(acomp_ctx))
> > + return;
> > +
> > + if (!IS_ERR_OR_NULL(acomp_ctx->req))
> > + acomp_request_free(acomp_ctx->req);
> > +
> > + if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
> > + crypto_free_acomp(acomp_ctx->acomp);
> > +
> > + kfree(acomp_ctx->buffer);
> > +}
> > +
> > static struct zswap_pool *zswap_pool_create(char *compressor)
> > {
> > struct zswap_pool *pool;
> > @@ -263,19 +287,43 @@ static struct zswap_pool
> *zswap_pool_create(char *compressor)
> >
> > strscpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
> >
> > - pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
> > + /* Many things rely on the zero-initialization. */
> > + pool->acomp_ctx = alloc_percpu_gfp(*pool->acomp_ctx,
> > + GFP_KERNEL | __GFP_ZERO);
> > if (!pool->acomp_ctx) {
> > pr_err("percpu alloc failed\n");
> > goto error;
> > }
> >
> > - for_each_possible_cpu(cpu)
> > - mutex_init(&per_cpu_ptr(pool->acomp_ctx, cpu)->mutex);
> > -
> > + /*
> > + * This is serialized against CPU hotplug operations. Hence, cores
> > + * cannot be offlined until this finishes.
> > + * In case of errors, we need to goto "ref_fail" instead of "error"
> > + * because there is no teardown callback registered anymore, for
> > + * cpuhp_state_add_instance() to de-allocate resources as it rolls
> back
> > + * state on cores before the CPU on which error was encountered.
> > + */
> > ret =
> cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
> > &pool->node);
> > +
> > + /*
> > + * We only needed the multi state instance add operation to invoke
> the
> > + * startup callback for all cores without cores getting offlined. Since
> > + * the acomp_ctx resources will now only be de-allocated when the
> pool
> > + * is destroyed, we can safely remove the multi state instance. This
> > + * minimizes (but does not eliminate) the possibility of
> > + * zswap_cpu_comp_prepare() being invoked again due to a CPU
> > + * offline-online transition. Removing the instance also prevents race
> > + * conditions between CPU onlining after initial pool creation, and
> > + * acomp_ctx_dealloc() freeing the acomp_ctx resources.
> > + * Note that we delete the instance before checking the error status
> of
> > + * the node list add operation because we want the instance removal
> even
> > + * in case of errors in the former.
> > + */
> > + cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
> &pool->node);
> > +
>
> I don't understand what's wrong with the current flow? We call
> cpuhp_state_remove_instance() in pool deletion before freeing up the
> per-CPU resources. Why is this not enough?
This is because with the changes proposed in this commit, the multi state
add instance is used during pool creation as a way to create acomp_ctx
resources correctly with just the offline/online state transitions guaranteed
by CPU hotplug, without needing additional mutex locking as in the mainline.
In other words, the consistency wrt safely creating/deleting acomp_ctx
resources with the changes being proposed is accomplished by the hotplug
state transitions guarantee. Stated differently, the hotplug framework
helps enforce the new design during pool creation without relying on the
mutex and subsequent simplifications during zswap_[de]compress()
proposed in this commit.
Once this is done, deleting the CPU hotplug state seems cleaner, and reflects
the change in policy of the resources' lifetime. It also prevents race conditions
between zswap_cpu_comp_prepare() and acomp_ctx_dealloc() called from
zswap_pool_destroy().
The only cleaner design I can think of is to not use CPU hotplug callbacks
at all, instead use a for_each_possible_cpu() to allocate acomp_ctx
resources. The one benefit of the current design is that it saves memory
if a considerable number of CPUs are offlined to begin with, for some
reason.
Thanks,
Kanchana
^ permalink raw reply [flat|nested] 56+ messages in thread
* Re: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources exist from pool creation to deletion.
2025-09-30 18:20 ` Sridhar, Kanchana P
@ 2025-09-30 18:29 ` Yosry Ahmed
2025-09-30 21:00 ` Sridhar, Kanchana P
0 siblings, 1 reply; 56+ messages in thread
From: Yosry Ahmed @ 2025-09-30 18:29 UTC (permalink / raw)
To: Sridhar, Kanchana P
Cc: linux-kernel, linux-mm, hannes, nphamcs, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius,
Feghali, Wajdi K, Gopal, Vinodh
On Tue, Sep 30, 2025 at 06:20:13PM +0000, Sridhar, Kanchana P wrote:
>
> > -----Original Message-----
> > From: Yosry Ahmed <yosry.ahmed@linux.dev>
> > Sent: Tuesday, September 30, 2025 8:49 AM
> > To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> > Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> > hannes@cmpxchg.org; nphamcs@gmail.com; chengming.zhou@linux.dev;
> > usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> > ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> > senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> > crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> > davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> > ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> > <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>;
> > Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> > <vinodh.gopal@intel.com>
> > Subject: Re: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources
> > exist from pool creation to deletion.
> >
> > On Thu, Sep 25, 2025 at 08:34:59PM -0700, Kanchana P Sridhar wrote:
> > > This patch simplifies the zswap_pool's per-CPU acomp_ctx resource
> > > management. Similar to the per-CPU acomp_ctx itself, the per-CPU
> > > acomp_ctx's resources' (acomp, req, buffer) lifetime will also be from
> > > pool creation to pool deletion. These resources will persist through CPU
> > > hotplug operations. The zswap_cpu_comp_dead() teardown callback has
> > been
> > > deleted from the call to
> > > cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE). As a result,
> > CPU
> > > offline hotplug operations will be no-ops as far as the acomp_ctx
> > > resources are concerned.
> > >
> > > This commit refactors the code from zswap_cpu_comp_dead() into a
> > > new function acomp_ctx_dealloc() that preserves the IS_ERR_OR_NULL()
> > > checks on acomp_ctx, req and acomp from the existing mainline
> > > implementation of zswap_cpu_comp_dead(). acomp_ctx_dealloc() is called
> > > to clean up acomp_ctx resources from all these procedures:
> > >
> > > 1) zswap_cpu_comp_prepare() when an error is encountered,
> > > 2) zswap_pool_create() when an error is encountered, and
> > > 3) from zswap_pool_destroy().
> > >
> > > The main benefit of using the CPU hotplug multi state instance startup
> > > callback to allocate the acomp_ctx resources is that it prevents the
> > > cores from being offlined until the multi state instance addition call
> > > returns.
> > >
> > > From Documentation/core-api/cpu_hotplug.rst:
> > >
> > > "The node list add/remove operations and the callback invocations are
> > > serialized against CPU hotplug operations."
> > >
> > > Furthermore, zswap_[de]compress() cannot contend with
> > > zswap_cpu_comp_prepare() because:
> > >
> > > - During pool creation/deletion, the pool is not in the zswap_pools
> > > list.
> > >
> > > - During CPU hot[un]plug, the CPU is not yet online, as Yosry pointed
> > > out. zswap_cpu_comp_prepare() will be executed on a control CPU,
> > > since CPUHP_MM_ZSWP_POOL_PREPARE is in the PREPARE section of
> > "enum
> > > cpuhp_state". Thanks Yosry for sharing this observation!
> > >
> > > In both these cases, any recursions into zswap reclaim from
> > > zswap_cpu_comp_prepare() will be handled by the old pool.
> > >
> > > The above two observations enable the following simplifications:
> > >
> > > 1) zswap_cpu_comp_prepare(): CPU cannot be offlined. Reclaim cannot
> > use
> > > the pool. Considerations for mutex init/locking and handling
> > > subsequent CPU hotplug online-offlines:
> > >
> > > Should we lock the mutex of current CPU's acomp_ctx from start to
> > > end? It doesn't seem like this is required. The CPU hotplug
> > > operations acquire a "cpuhp_state_mutex" before proceeding, hence
> > > they are serialized against CPU hotplug operations.
> > >
> > > If the process gets migrated while zswap_cpu_comp_prepare() is
> > > running, it will complete on the new CPU. In case of failures, we
> > > pass the acomp_ctx pointer obtained at the start of
> > > zswap_cpu_comp_prepare() to acomp_ctx_dealloc(), which again, can
> > > only undergo migration. There appear to be no contention scenarios
> > > that might cause inconsistent values of acomp_ctx's members. Hence,
> > > it seems there is no need for mutex_lock(&acomp_ctx->mutex) in
> > > zswap_cpu_comp_prepare().
> > >
> > > Since the pool is not yet on zswap_pools list, we don't need to
> > > initialize the per-CPU acomp_ctx mutex in zswap_pool_create(). This
> > > has been restored to occur in zswap_cpu_comp_prepare().
> > >
> > > zswap_cpu_comp_prepare() checks upfront if acomp_ctx->acomp is
> > > valid. If so, it returns success. This should handle any CPU
> > > hotplug online-offline transitions after pool creation is done.
> > >
> > > 2) CPU offline vis-a-vis zswap ops: Let's suppose the process is
> > > migrated to another CPU before the current CPU is dysfunctional. If
> > > zswap_[de]compress() holds the acomp_ctx->mutex lock of the offlined
> > > CPU, that mutex will be released once it completes on the new
> > > CPU. Since there is no teardown callback, there is no possibility of
> > > UAF.
> > >
> > > 3) Pool creation/deletion and process migration to another CPU:
> > >
> > > - During pool creation/deletion, the pool is not in the zswap_pools
> > > list. Hence it cannot contend with zswap ops on that CPU. However,
> > > the process can get migrated.
> > >
> > > Pool creation --> zswap_cpu_comp_prepare()
> > > --> process migrated:
> > > * CPU offline: no-op.
> > > * zswap_cpu_comp_prepare() continues
> > > to run on the new CPU to finish
> > > allocating acomp_ctx resources for
> > > the offlined CPU.
> > >
> > > Pool deletion --> acomp_ctx_dealloc()
> > > --> process migrated:
> > > * CPU offline: no-op.
> > > * acomp_ctx_dealloc() continues
> > > to run on the new CPU to finish
> > > de-allocating acomp_ctx resources
> > > for the offlined CPU.
> > >
> > > 4) Pool deletion vis-a-vis CPU onlining:
> > > To prevent possibility of race conditions between
> > > acomp_ctx_dealloc() freeing the acomp_ctx resources and the initial
> > > check for a valid acomp_ctx->acomp in zswap_cpu_comp_prepare(), we
> > > need to delete the multi state instance right after it is added, in
> > > zswap_pool_create().
> > >
> > > Summary of changes based on the above:
> > > --------------------------------------
> > > 1) Zero-initialization of pool->acomp_ctx in zswap_pool_create() to
> > > simplify and share common code for different error handling/cleanup
> > > related to the acomp_ctx.
> > >
> > > 2) Remove the node list instance right after node list add function
> > > call in zswap_pool_create(). This prevents race conditions between
> > > CPU onlining after initial pool creation, and acomp_ctx_dealloc()
> > > freeing the acomp_ctx resources.
> > >
> > > 3) zswap_pool_destroy() will call acomp_ctx_dealloc() to de-allocate
> > > the per-CPU acomp_ctx resources.
> > >
> > > 4) Changes to zswap_cpu_comp_prepare():
> > >
> > > a) Check if acomp_ctx->acomp is valid at the beginning and return,
> > > because the acomp_ctx is already initialized.
> > > b) Move the mutex_init to happen in this procedure, before it
> > > returns.
> > > c) All error conditions handled by calling acomp_ctx_dealloc().
> > >
> > > 5) New procedure acomp_ctx_dealloc() for common error/cleanup code.
> > >
> > > 6) No more multi state instance teardown callback. CPU offlining is a
> > > no-op as far as acomp_ctx resources are concerned.
> > >
> > > 7) Delete acomp_ctx_get_cpu_lock()/acomp_ctx_put_unlock(). Directly
> > > call mutex_lock(&acomp_ctx->mutex)/mutex_unlock(&acomp_ctx-
> > >mutex)
> > > in zswap_[de]compress().
> > >
> > > The per-CPU memory cost of not deleting the acomp_ctx resources upon
> > CPU
> > > offlining, and only deleting them when the pool is destroyed, is as
> > > follows, on x86_64:
> > >
> > > IAA with 8 dst buffers for batching: 64.34 KB
> > > Software compressors with 1 dst buffer: 8.28 KB
> > >
> > > Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
> >
> > Please try to make the commit logs a bit more summarized. Details are
> > helpful, but it's easy to lose track of things when it gets too long.
>
> Thanks Yosry, for the feedback.
>
> >
> > > ---
> > > mm/zswap.c | 194 +++++++++++++++++++++++++----------------------------
> > > 1 file changed, 93 insertions(+), 101 deletions(-)
> > >
> > > diff --git a/mm/zswap.c b/mm/zswap.c
> > > index c1af782e54ec..27665eaa3f89 100644
> > > --- a/mm/zswap.c
> > > +++ b/mm/zswap.c
> > > @@ -242,6 +242,30 @@ static inline struct xarray
> > *swap_zswap_tree(swp_entry_t swp)
> > > **********************************/
> > > static void __zswap_pool_empty(struct percpu_ref *ref);
> > >
> > > +/*
> > > + * The per-cpu pool->acomp_ctx is zero-initialized on allocation. This
> > makes
> > > + * it easy for different error conditions/cleanup related to the acomp_ctx
> > > + * to be handled by acomp_ctx_dealloc():
> > > + * - Errors during zswap_cpu_comp_prepare().
> > > + * - Partial success/error of cpuhp_state_add_instance() call in
> > > + * zswap_pool_create(). Only some cores could have executed
> > > + * zswap_cpu_comp_prepare(), not others.
> > > + * - Cleanup acomp_ctx resources on all cores in zswap_pool_destroy().
> > > + */
> >
> > Comments describing specific code paths go out of date really fast. The
> > comment is probably unnecessary, it's easy to check the allocation path
> > to figure out that these are zero-initialized.
> >
> > Also in general, please keep the comments as summarized as possible, and
> > only when the logic is not clear from the code.
>
> Sure. I have tried to explain the rationale for significant changes, but I can
> look for opportunities to summarize. I was sort of hoping that v12 would
> be it, but I can work on the comments being concise if this is crucial.
>
> >
> > > +static void acomp_ctx_dealloc(struct crypto_acomp_ctx *acomp_ctx)
> > > +{
> > > + if (IS_ERR_OR_NULL(acomp_ctx))
> > > + return;
> > > +
> > > + if (!IS_ERR_OR_NULL(acomp_ctx->req))
> > > + acomp_request_free(acomp_ctx->req);
> > > +
> > > + if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
> > > + crypto_free_acomp(acomp_ctx->acomp);
> > > +
> > > + kfree(acomp_ctx->buffer);
> > > +}
> > > +
> > > static struct zswap_pool *zswap_pool_create(char *compressor)
> > > {
> > > struct zswap_pool *pool;
> > > @@ -263,19 +287,43 @@ static struct zswap_pool
> > *zswap_pool_create(char *compressor)
> > >
> > > strscpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
> > >
> > > - pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
> > > + /* Many things rely on the zero-initialization. */
> > > + pool->acomp_ctx = alloc_percpu_gfp(*pool->acomp_ctx,
> > > + GFP_KERNEL | __GFP_ZERO);
> > > if (!pool->acomp_ctx) {
> > > pr_err("percpu alloc failed\n");
> > > goto error;
> > > }
> > >
> > > - for_each_possible_cpu(cpu)
> > > - mutex_init(&per_cpu_ptr(pool->acomp_ctx, cpu)->mutex);
> > > -
> > > + /*
> > > + * This is serialized against CPU hotplug operations. Hence, cores
> > > + * cannot be offlined until this finishes.
> > > + * In case of errors, we need to goto "ref_fail" instead of "error"
> > > + * because there is no teardown callback registered anymore, for
> > > + * cpuhp_state_add_instance() to de-allocate resources as it rolls
> > back
> > > + * state on cores before the CPU on which error was encountered.
> > > + */
> > > ret =
> > cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
> > > &pool->node);
> > > +
> > > + /*
> > > + * We only needed the multi state instance add operation to invoke
> > the
> > > + * startup callback for all cores without cores getting offlined. Since
> > > + * the acomp_ctx resources will now only be de-allocated when the
> > pool
> > > + * is destroyed, we can safely remove the multi state instance. This
> > > + * minimizes (but does not eliminate) the possibility of
> > > + * zswap_cpu_comp_prepare() being invoked again due to a CPU
> > > + * offline-online transition. Removing the instance also prevents race
> > > + * conditions between CPU onlining after initial pool creation, and
> > > + * acomp_ctx_dealloc() freeing the acomp_ctx resources.
> > > + * Note that we delete the instance before checking the error status
> > of
> > > + * the node list add operation because we want the instance removal
> > even
> > > + * in case of errors in the former.
> > > + */
> > > + cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
> > &pool->node);
> > > +
> >
> > I don't understand what's wrong with the current flow? We call
> > cpuhp_state_remove_instance() in pool deletion before freeing up the
> > per-CPU resources. Why is this not enough?
>
> This is because with the changes proposed in this commit, the multi state
> add instance is used during pool creation as a way to create acomp_ctx
> resources correctly with just the offline/online state transitions guaranteed
> by CPU hotplug, without needing additional mutex locking as in the mainline.
> In other words, the consistency wrt safely creating/deleting acomp_ctx
> resources with the changes being proposed is accomplished by the hotplug
> state transitions guarantee. Stated differently, the hotplug framework
> helps enforce the new design during pool creation without relying on the
> mutex and subsequent simplifications during zswap_[de]compress()
> proposed in this commit.
>
> Once this is done, deleting the CPU hotplug state seems cleaner, and reflects
> the change in policy of the resources' lifetime. It also prevents race conditions
> between zswap_cpu_comp_prepare() and acomp_ctx_dealloc() called from
> zswap_pool_destroy().
How is a race with zswap_cpu_comp_prepare() possible if we call
cpuhp_state_remove_instance() before acomp_ctx_dealloc() in the pool
deletion path?
>
> The only cleaner design I can think of is to not use CPU hotplug callbacks
> at all, instead use a for_each_possible_cpu() to allocate acomp_ctx
> resources. The one benefit of the current design is that it saves memory
> if a considerable number of CPUs are offlined to begin with, for some
> reason.
>
> Thanks,
> Kanchana
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources exist from pool creation to deletion.
2025-09-30 18:29 ` Yosry Ahmed
@ 2025-09-30 21:00 ` Sridhar, Kanchana P
2025-09-30 21:20 ` Yosry Ahmed
0 siblings, 1 reply; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-09-30 21:00 UTC (permalink / raw)
To: Yosry Ahmed
Cc: linux-kernel, linux-mm, hannes, nphamcs, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius,
Feghali, Wajdi K, Gopal, Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Yosry Ahmed <yosry.ahmed@linux.dev>
> Sent: Tuesday, September 30, 2025 11:30 AM
> To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> hannes@cmpxchg.org; nphamcs@gmail.com; chengming.zhou@linux.dev;
> usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>;
> Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> <vinodh.gopal@intel.com>
> Subject: Re: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources
> exist from pool creation to deletion.
>
> On Tue, Sep 30, 2025 at 06:20:13PM +0000, Sridhar, Kanchana P wrote:
> >
> > > -----Original Message-----
> > > From: Yosry Ahmed <yosry.ahmed@linux.dev>
> > > Sent: Tuesday, September 30, 2025 8:49 AM
> > > To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> > > Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> > > hannes@cmpxchg.org; nphamcs@gmail.com;
> chengming.zhou@linux.dev;
> > > usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> > > ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> > > senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> > > crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> > > davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> > > ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> > > <kristen.c.accardi@intel.com>; Gomes, Vinicius
> <vinicius.gomes@intel.com>;
> > > Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> > > <vinodh.gopal@intel.com>
> > > Subject: Re: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx
> resources
> > > exist from pool creation to deletion.
> > >
> > > On Thu, Sep 25, 2025 at 08:34:59PM -0700, Kanchana P Sridhar wrote:
> > > > This patch simplifies the zswap_pool's per-CPU acomp_ctx resource
> > > > management. Similar to the per-CPU acomp_ctx itself, the per-CPU
> > > > acomp_ctx's resources' (acomp, req, buffer) lifetime will also be from
> > > > pool creation to pool deletion. These resources will persist through CPU
> > > > hotplug operations. The zswap_cpu_comp_dead() teardown callback
> has
> > > been
> > > > deleted from the call to
> > > > cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE). As a
> result,
> > > CPU
> > > > offline hotplug operations will be no-ops as far as the acomp_ctx
> > > > resources are concerned.
> > > >
> > > > This commit refactors the code from zswap_cpu_comp_dead() into a
> > > > new function acomp_ctx_dealloc() that preserves the
> IS_ERR_OR_NULL()
> > > > checks on acomp_ctx, req and acomp from the existing mainline
> > > > implementation of zswap_cpu_comp_dead(). acomp_ctx_dealloc() is
> called
> > > > to clean up acomp_ctx resources from all these procedures:
> > > >
> > > > 1) zswap_cpu_comp_prepare() when an error is encountered,
> > > > 2) zswap_pool_create() when an error is encountered, and
> > > > 3) from zswap_pool_destroy().
> > > >
> > > > The main benefit of using the CPU hotplug multi state instance startup
> > > > callback to allocate the acomp_ctx resources is that it prevents the
> > > > cores from being offlined until the multi state instance addition call
> > > > returns.
> > > >
> > > > From Documentation/core-api/cpu_hotplug.rst:
> > > >
> > > > "The node list add/remove operations and the callback invocations are
> > > > serialized against CPU hotplug operations."
> > > >
> > > > Furthermore, zswap_[de]compress() cannot contend with
> > > > zswap_cpu_comp_prepare() because:
> > > >
> > > > - During pool creation/deletion, the pool is not in the zswap_pools
> > > > list.
> > > >
> > > > - During CPU hot[un]plug, the CPU is not yet online, as Yosry pointed
> > > > out. zswap_cpu_comp_prepare() will be executed on a control CPU,
> > > > since CPUHP_MM_ZSWP_POOL_PREPARE is in the PREPARE section
> of
> > > "enum
> > > > cpuhp_state". Thanks Yosry for sharing this observation!
> > > >
> > > > In both these cases, any recursions into zswap reclaim from
> > > > zswap_cpu_comp_prepare() will be handled by the old pool.
> > > >
> > > > The above two observations enable the following simplifications:
> > > >
> > > > 1) zswap_cpu_comp_prepare(): CPU cannot be offlined. Reclaim cannot
> > > use
> > > > the pool. Considerations for mutex init/locking and handling
> > > > subsequent CPU hotplug online-offlines:
> > > >
> > > > Should we lock the mutex of current CPU's acomp_ctx from start to
> > > > end? It doesn't seem like this is required. The CPU hotplug
> > > > operations acquire a "cpuhp_state_mutex" before proceeding, hence
> > > > they are serialized against CPU hotplug operations.
> > > >
> > > > If the process gets migrated while zswap_cpu_comp_prepare() is
> > > > running, it will complete on the new CPU. In case of failures, we
> > > > pass the acomp_ctx pointer obtained at the start of
> > > > zswap_cpu_comp_prepare() to acomp_ctx_dealloc(), which again, can
> > > > only undergo migration. There appear to be no contention scenarios
> > > > that might cause inconsistent values of acomp_ctx's members. Hence,
> > > > it seems there is no need for mutex_lock(&acomp_ctx->mutex) in
> > > > zswap_cpu_comp_prepare().
> > > >
> > > > Since the pool is not yet on zswap_pools list, we don't need to
> > > > initialize the per-CPU acomp_ctx mutex in zswap_pool_create(). This
> > > > has been restored to occur in zswap_cpu_comp_prepare().
> > > >
> > > > zswap_cpu_comp_prepare() checks upfront if acomp_ctx->acomp is
> > > > valid. If so, it returns success. This should handle any CPU
> > > > hotplug online-offline transitions after pool creation is done.
> > > >
> > > > 2) CPU offline vis-a-vis zswap ops: Let's suppose the process is
> > > > migrated to another CPU before the current CPU is dysfunctional. If
> > > > zswap_[de]compress() holds the acomp_ctx->mutex lock of the
> offlined
> > > > CPU, that mutex will be released once it completes on the new
> > > > CPU. Since there is no teardown callback, there is no possibility of
> > > > UAF.
> > > >
> > > > 3) Pool creation/deletion and process migration to another CPU:
> > > >
> > > > - During pool creation/deletion, the pool is not in the zswap_pools
> > > > list. Hence it cannot contend with zswap ops on that CPU. However,
> > > > the process can get migrated.
> > > >
> > > > Pool creation --> zswap_cpu_comp_prepare()
> > > > --> process migrated:
> > > > * CPU offline: no-op.
> > > > * zswap_cpu_comp_prepare() continues
> > > > to run on the new CPU to finish
> > > > allocating acomp_ctx resources for
> > > > the offlined CPU.
> > > >
> > > > Pool deletion --> acomp_ctx_dealloc()
> > > > --> process migrated:
> > > > * CPU offline: no-op.
> > > > * acomp_ctx_dealloc() continues
> > > > to run on the new CPU to finish
> > > > de-allocating acomp_ctx resources
> > > > for the offlined CPU.
> > > >
> > > > 4) Pool deletion vis-a-vis CPU onlining:
> > > > To prevent possibility of race conditions between
> > > > acomp_ctx_dealloc() freeing the acomp_ctx resources and the initial
> > > > check for a valid acomp_ctx->acomp in zswap_cpu_comp_prepare(),
> we
> > > > need to delete the multi state instance right after it is added, in
> > > > zswap_pool_create().
> > > >
> > > > Summary of changes based on the above:
> > > > --------------------------------------
> > > > 1) Zero-initialization of pool->acomp_ctx in zswap_pool_create() to
> > > > simplify and share common code for different error handling/cleanup
> > > > related to the acomp_ctx.
> > > >
> > > > 2) Remove the node list instance right after node list add function
> > > > call in zswap_pool_create(). This prevents race conditions between
> > > > CPU onlining after initial pool creation, and acomp_ctx_dealloc()
> > > > freeing the acomp_ctx resources.
> > > >
> > > > 3) zswap_pool_destroy() will call acomp_ctx_dealloc() to de-allocate
> > > > the per-CPU acomp_ctx resources.
> > > >
> > > > 4) Changes to zswap_cpu_comp_prepare():
> > > >
> > > > a) Check if acomp_ctx->acomp is valid at the beginning and return,
> > > > because the acomp_ctx is already initialized.
> > > > b) Move the mutex_init to happen in this procedure, before it
> > > > returns.
> > > > c) All error conditions handled by calling acomp_ctx_dealloc().
> > > >
> > > > 5) New procedure acomp_ctx_dealloc() for common error/cleanup
> code.
> > > >
> > > > 6) No more multi state instance teardown callback. CPU offlining is a
> > > > no-op as far as acomp_ctx resources are concerned.
> > > >
> > > > 7) Delete acomp_ctx_get_cpu_lock()/acomp_ctx_put_unlock(). Directly
> > > > call mutex_lock(&acomp_ctx->mutex)/mutex_unlock(&acomp_ctx-
> > > >mutex)
> > > > in zswap_[de]compress().
> > > >
> > > > The per-CPU memory cost of not deleting the acomp_ctx resources
> upon
> > > CPU
> > > > offlining, and only deleting them when the pool is destroyed, is as
> > > > follows, on x86_64:
> > > >
> > > > IAA with 8 dst buffers for batching: 64.34 KB
> > > > Software compressors with 1 dst buffer: 8.28 KB
> > > >
> > > > Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
> > >
> > > Please try to make the commit logs a bit more summarized. Details are
> > > helpful, but it's easy to lose track of things when it gets too long.
> >
> > Thanks Yosry, for the feedback.
> >
> > >
> > > > ---
> > > > mm/zswap.c | 194 +++++++++++++++++++++++++-------------------------
> ---
> > > > 1 file changed, 93 insertions(+), 101 deletions(-)
> > > >
> > > > diff --git a/mm/zswap.c b/mm/zswap.c
> > > > index c1af782e54ec..27665eaa3f89 100644
> > > > --- a/mm/zswap.c
> > > > +++ b/mm/zswap.c
> > > > @@ -242,6 +242,30 @@ static inline struct xarray
> > > *swap_zswap_tree(swp_entry_t swp)
> > > > **********************************/
> > > > static void __zswap_pool_empty(struct percpu_ref *ref);
> > > >
> > > > +/*
> > > > + * The per-cpu pool->acomp_ctx is zero-initialized on allocation. This
> > > makes
> > > > + * it easy for different error conditions/cleanup related to the
> acomp_ctx
> > > > + * to be handled by acomp_ctx_dealloc():
> > > > + * - Errors during zswap_cpu_comp_prepare().
> > > > + * - Partial success/error of cpuhp_state_add_instance() call in
> > > > + * zswap_pool_create(). Only some cores could have executed
> > > > + * zswap_cpu_comp_prepare(), not others.
> > > > + * - Cleanup acomp_ctx resources on all cores in zswap_pool_destroy().
> > > > + */
> > >
> > > Comments describing specific code paths go out of date really fast. The
> > > comment is probably unnecessary, it's easy to check the allocation path
> > > to figure out that these are zero-initialized.
> > >
> > > Also in general, please keep the comments as summarized as possible,
> and
> > > only when the logic is not clear from the code.
> >
> > Sure. I have tried to explain the rationale for significant changes, but I can
> > look for opportunities to summarize. I was sort of hoping that v12 would
> > be it, but I can work on the comments being concise if this is crucial.
> >
> > >
> > > > +static void acomp_ctx_dealloc(struct crypto_acomp_ctx *acomp_ctx)
> > > > +{
> > > > + if (IS_ERR_OR_NULL(acomp_ctx))
> > > > + return;
> > > > +
> > > > + if (!IS_ERR_OR_NULL(acomp_ctx->req))
> > > > + acomp_request_free(acomp_ctx->req);
> > > > +
> > > > + if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
> > > > + crypto_free_acomp(acomp_ctx->acomp);
> > > > +
> > > > + kfree(acomp_ctx->buffer);
> > > > +}
> > > > +
> > > > static struct zswap_pool *zswap_pool_create(char *compressor)
> > > > {
> > > > struct zswap_pool *pool;
> > > > @@ -263,19 +287,43 @@ static struct zswap_pool
> > > *zswap_pool_create(char *compressor)
> > > >
> > > > strscpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
> > > >
> > > > - pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
> > > > + /* Many things rely on the zero-initialization. */
> > > > + pool->acomp_ctx = alloc_percpu_gfp(*pool->acomp_ctx,
> > > > + GFP_KERNEL | __GFP_ZERO);
> > > > if (!pool->acomp_ctx) {
> > > > pr_err("percpu alloc failed\n");
> > > > goto error;
> > > > }
> > > >
> > > > - for_each_possible_cpu(cpu)
> > > > - mutex_init(&per_cpu_ptr(pool->acomp_ctx, cpu)->mutex);
> > > > -
> > > > + /*
> > > > + * This is serialized against CPU hotplug operations. Hence, cores
> > > > + * cannot be offlined until this finishes.
> > > > + * In case of errors, we need to goto "ref_fail" instead of "error"
> > > > + * because there is no teardown callback registered anymore, for
> > > > + * cpuhp_state_add_instance() to de-allocate resources as it rolls
> > > back
> > > > + * state on cores before the CPU on which error was encountered.
> > > > + */
> > > > ret =
> > > cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
> > > > &pool->node);
> > > > +
> > > > + /*
> > > > + * We only needed the multi state instance add operation to invoke
> > > the
> > > > + * startup callback for all cores without cores getting offlined. Since
> > > > + * the acomp_ctx resources will now only be de-allocated when the
> > > pool
> > > > + * is destroyed, we can safely remove the multi state instance. This
> > > > + * minimizes (but does not eliminate) the possibility of
> > > > + * zswap_cpu_comp_prepare() being invoked again due to a CPU
> > > > + * offline-online transition. Removing the instance also prevents race
> > > > + * conditions between CPU onlining after initial pool creation, and
> > > > + * acomp_ctx_dealloc() freeing the acomp_ctx resources.
> > > > + * Note that we delete the instance before checking the error status
> > > of
> > > > + * the node list add operation because we want the instance removal
> > > even
> > > > + * in case of errors in the former.
> > > > + */
> > > > + cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
> > > &pool->node);
> > > > +
> > >
> > > I don't understand what's wrong with the current flow? We call
> > > cpuhp_state_remove_instance() in pool deletion before freeing up the
> > > per-CPU resources. Why is this not enough?
> >
> > This is because with the changes proposed in this commit, the multi state
> > add instance is used during pool creation as a way to create acomp_ctx
> > resources correctly with just the offline/online state transitions guaranteed
> > by CPU hotplug, without needing additional mutex locking as in the
> mainline.
> > In other words, the consistency wrt safely creating/deleting acomp_ctx
> > resources with the changes being proposed is accomplished by the hotplug
> > state transitions guarantee. Stated differently, the hotplug framework
> > helps enforce the new design during pool creation without relying on the
> > mutex and subsequent simplifications during zswap_[de]compress()
> > proposed in this commit.
> >
> > Once this is done, deleting the CPU hotplug state seems cleaner, and
> reflects
> > the change in policy of the resources' lifetime. It also prevents race
> conditions
> > between zswap_cpu_comp_prepare() and acomp_ctx_dealloc() called from
> > zswap_pool_destroy().
>
> How is a race with zswap_cpu_comp_prepare() possible if we call
> cpuhp_state_remove_instance() before acomp_ctx_dealloc() in the pool
> deletion path?
Good point. I agree, calling cpuhp_state_remove_instance() before
acomp_ctx_dealloc() will not cause a race. However, if we consider the
time from pool creation to deletion: if there is an online-offline-online
transition, can zswap_cpu_comp_prepare() race with the call to
cpuhp_state_remove_instance()? If so, wouldn't this cause unpredictable
behavior?
I agree, this can occur even with the code in this commit, but there is
less risk of things going wrong because we remove the CPU hotplug
instance before the pool is added to zswap_pools.
Further, removing the CPU hotplug instance directly codifies the
intent of this commit, i.e., to use this as a facilitator and manage memory
allotted to acomp_ctx, but not to manage those resources' lifetime
thereafter.
Do you see any advantage of keeping the call to cpuhp_state_remove_instance()
occur before acomp_ctx_dealloc() in zswap_pool_destroy()? Please let me know
if I am missing something.
Thanks,
Kanchana
>
> >
> > The only cleaner design I can think of is to not use CPU hotplug callbacks
> > at all, instead use a for_each_possible_cpu() to allocate acomp_ctx
> > resources. The one benefit of the current design is that it saves memory
> > if a considerable number of CPUs are offlined to begin with, for some
> > reason.
> >
> > Thanks,
> > Kanchana
^ permalink raw reply [flat|nested] 56+ messages in thread
* Re: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources exist from pool creation to deletion.
2025-09-30 21:00 ` Sridhar, Kanchana P
@ 2025-09-30 21:20 ` Yosry Ahmed
2025-09-30 21:56 ` Sridhar, Kanchana P
0 siblings, 1 reply; 56+ messages in thread
From: Yosry Ahmed @ 2025-09-30 21:20 UTC (permalink / raw)
To: Sridhar, Kanchana P
Cc: linux-kernel, linux-mm, hannes, nphamcs, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius,
Feghali, Wajdi K, Gopal, Vinodh
> > > > > static struct zswap_pool *zswap_pool_create(char *compressor)
> > > > > {
> > > > > struct zswap_pool *pool;
> > > > > @@ -263,19 +287,43 @@ static struct zswap_pool
> > > > *zswap_pool_create(char *compressor)
> > > > >
> > > > > strscpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
> > > > >
> > > > > - pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
> > > > > + /* Many things rely on the zero-initialization. */
> > > > > + pool->acomp_ctx = alloc_percpu_gfp(*pool->acomp_ctx,
> > > > > + GFP_KERNEL | __GFP_ZERO);
> > > > > if (!pool->acomp_ctx) {
> > > > > pr_err("percpu alloc failed\n");
> > > > > goto error;
> > > > > }
> > > > >
> > > > > - for_each_possible_cpu(cpu)
> > > > > - mutex_init(&per_cpu_ptr(pool->acomp_ctx, cpu)->mutex);
> > > > > -
> > > > > + /*
> > > > > + * This is serialized against CPU hotplug operations. Hence, cores
> > > > > + * cannot be offlined until this finishes.
> > > > > + * In case of errors, we need to goto "ref_fail" instead of "error"
> > > > > + * because there is no teardown callback registered anymore, for
> > > > > + * cpuhp_state_add_instance() to de-allocate resources as it rolls
> > > > back
> > > > > + * state on cores before the CPU on which error was encountered.
> > > > > + */
> > > > > ret =
> > > > cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
> > > > > &pool->node);
> > > > > +
> > > > > + /*
> > > > > + * We only needed the multi state instance add operation to invoke
> > > > the
> > > > > + * startup callback for all cores without cores getting offlined. Since
> > > > > + * the acomp_ctx resources will now only be de-allocated when the
> > > > pool
> > > > > + * is destroyed, we can safely remove the multi state instance. This
> > > > > + * minimizes (but does not eliminate) the possibility of
> > > > > + * zswap_cpu_comp_prepare() being invoked again due to a CPU
> > > > > + * offline-online transition. Removing the instance also prevents race
> > > > > + * conditions between CPU onlining after initial pool creation, and
> > > > > + * acomp_ctx_dealloc() freeing the acomp_ctx resources.
> > > > > + * Note that we delete the instance before checking the error status
> > > > of
> > > > > + * the node list add operation because we want the instance removal
> > > > even
> > > > > + * in case of errors in the former.
> > > > > + */
> > > > > + cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
> > > > &pool->node);
> > > > > +
> > > >
> > > > I don't understand what's wrong with the current flow? We call
> > > > cpuhp_state_remove_instance() in pool deletion before freeing up the
> > > > per-CPU resources. Why is this not enough?
> > >
> > > This is because with the changes proposed in this commit, the multi state
> > > add instance is used during pool creation as a way to create acomp_ctx
> > > resources correctly with just the offline/online state transitions guaranteed
> > > by CPU hotplug, without needing additional mutex locking as in the
> > mainline.
> > > In other words, the consistency wrt safely creating/deleting acomp_ctx
> > > resources with the changes being proposed is accomplished by the hotplug
> > > state transitions guarantee. Stated differently, the hotplug framework
> > > helps enforce the new design during pool creation without relying on the
> > > mutex and subsequent simplifications during zswap_[de]compress()
> > > proposed in this commit.
> > >
> > > Once this is done, deleting the CPU hotplug state seems cleaner, and
> > reflects
> > > the change in policy of the resources' lifetime. It also prevents race
> > conditions
> > > between zswap_cpu_comp_prepare() and acomp_ctx_dealloc() called from
> > > zswap_pool_destroy().
> >
> > How is a race with zswap_cpu_comp_prepare() possible if we call
> > cpuhp_state_remove_instance() before acomp_ctx_dealloc() in the pool
> > deletion path?
>
> Good point. I agree, calling cpuhp_state_remove_instance() before
> acomp_ctx_dealloc() will not cause a race. However, if we consider the
> time from pool creation to deletion: if there is an online-offline-online
> transition, can zswap_cpu_comp_prepare() race with the call to
> cpuhp_state_remove_instance()? If so, wouldn't this cause unpredictable
> behavior?
How will this race happen?
cpuhp_state_remove_instance() is called while a pool is being destroyed,
while zswap_cpu_comp_prepare() while the pool is being created or during
CPU onlining.
The former cannot race, and the latter should be synchronized by hotplug
code.
>
> I agree, this can occur even with the code in this commit, but there is
> less risk of things going wrong because we remove the CPU hotplug
> instance before the pool is added to zswap_pools.
>
> Further, removing the CPU hotplug instance directly codifies the
> intent of this commit, i.e., to use this as a facilitator and manage memory
> allotted to acomp_ctx, but not to manage those resources' lifetime
> thereafter.
>
> Do you see any advantage of keeping the call to cpuhp_state_remove_instance()
> occur before acomp_ctx_dealloc() in zswap_pool_destroy()? Please let me know
> if I am missing something.
What about more CPUs going online? Without the hotplug instance we don't
get per-CPU resources for those. We are not using the hotplug mechanism
just to facilitate per-CPU resource allocation, we use it to
automatically allocate resources for newly onlined CPUs without having
to preallocate for all possible CPUs.
Also, this makes the code more difficult to reason about, and is an
unncessary change from the current behavior.
The only change needed is to drop the teardown callback and do the
freeing in the pool destruction path instead.
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources exist from pool creation to deletion.
2025-09-30 21:20 ` Yosry Ahmed
@ 2025-09-30 21:56 ` Sridhar, Kanchana P
2025-10-01 15:33 ` Yosry Ahmed
0 siblings, 1 reply; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-09-30 21:56 UTC (permalink / raw)
To: Yosry Ahmed
Cc: linux-kernel, linux-mm, hannes, nphamcs, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius,
Feghali, Wajdi K, Gopal, Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Yosry Ahmed <yosry.ahmed@linux.dev>
> Sent: Tuesday, September 30, 2025 2:20 PM
> To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> hannes@cmpxchg.org; nphamcs@gmail.com; chengming.zhou@linux.dev;
> usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>;
> Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> <vinodh.gopal@intel.com>
> Subject: Re: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources
> exist from pool creation to deletion.
>
> > > > > > static struct zswap_pool *zswap_pool_create(char *compressor)
> > > > > > {
> > > > > > struct zswap_pool *pool;
> > > > > > @@ -263,19 +287,43 @@ static struct zswap_pool
> > > > > *zswap_pool_create(char *compressor)
> > > > > >
> > > > > > strscpy(pool->tfm_name, compressor, sizeof(pool-
> >tfm_name));
> > > > > >
> > > > > > - pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
> > > > > > + /* Many things rely on the zero-initialization. */
> > > > > > + pool->acomp_ctx = alloc_percpu_gfp(*pool->acomp_ctx,
> > > > > > + GFP_KERNEL |
> __GFP_ZERO);
> > > > > > if (!pool->acomp_ctx) {
> > > > > > pr_err("percpu alloc failed\n");
> > > > > > goto error;
> > > > > > }
> > > > > >
> > > > > > - for_each_possible_cpu(cpu)
> > > > > > - mutex_init(&per_cpu_ptr(pool->acomp_ctx, cpu)-
> >mutex);
> > > > > > -
> > > > > > + /*
> > > > > > + * This is serialized against CPU hotplug operations. Hence,
> cores
> > > > > > + * cannot be offlined until this finishes.
> > > > > > + * In case of errors, we need to goto "ref_fail" instead of
> "error"
> > > > > > + * because there is no teardown callback registered anymore,
> for
> > > > > > + * cpuhp_state_add_instance() to de-allocate resources as it
> rolls
> > > > > back
> > > > > > + * state on cores before the CPU on which error was
> encountered.
> > > > > > + */
> > > > > > ret =
> > > > > cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
> > > > > > &pool->node);
> > > > > > +
> > > > > > + /*
> > > > > > + * We only needed the multi state instance add operation to
> invoke
> > > > > the
> > > > > > + * startup callback for all cores without cores getting offlined.
> Since
> > > > > > + * the acomp_ctx resources will now only be de-allocated
> when the
> > > > > pool
> > > > > > + * is destroyed, we can safely remove the multi state
> instance. This
> > > > > > + * minimizes (but does not eliminate) the possibility of
> > > > > > + * zswap_cpu_comp_prepare() being invoked again due to a
> CPU
> > > > > > + * offline-online transition. Removing the instance also
> prevents race
> > > > > > + * conditions between CPU onlining after initial pool creation,
> and
> > > > > > + * acomp_ctx_dealloc() freeing the acomp_ctx resources.
> > > > > > + * Note that we delete the instance before checking the error
> status
> > > > > of
> > > > > > + * the node list add operation because we want the instance
> removal
> > > > > even
> > > > > > + * in case of errors in the former.
> > > > > > + */
> > > > > > +
> cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
> > > > > &pool->node);
> > > > > > +
> > > > >
> > > > > I don't understand what's wrong with the current flow? We call
> > > > > cpuhp_state_remove_instance() in pool deletion before freeing up the
> > > > > per-CPU resources. Why is this not enough?
> > > >
> > > > This is because with the changes proposed in this commit, the multi
> state
> > > > add instance is used during pool creation as a way to create acomp_ctx
> > > > resources correctly with just the offline/online state transitions
> guaranteed
> > > > by CPU hotplug, without needing additional mutex locking as in the
> > > mainline.
> > > > In other words, the consistency wrt safely creating/deleting acomp_ctx
> > > > resources with the changes being proposed is accomplished by the
> hotplug
> > > > state transitions guarantee. Stated differently, the hotplug framework
> > > > helps enforce the new design during pool creation without relying on the
> > > > mutex and subsequent simplifications during zswap_[de]compress()
> > > > proposed in this commit.
> > > >
> > > > Once this is done, deleting the CPU hotplug state seems cleaner, and
> > > reflects
> > > > the change in policy of the resources' lifetime. It also prevents race
> > > conditions
> > > > between zswap_cpu_comp_prepare() and acomp_ctx_dealloc() called
> from
> > > > zswap_pool_destroy().
> > >
> > > How is a race with zswap_cpu_comp_prepare() possible if we call
> > > cpuhp_state_remove_instance() before acomp_ctx_dealloc() in the pool
> > > deletion path?
> >
> > Good point. I agree, calling cpuhp_state_remove_instance() before
> > acomp_ctx_dealloc() will not cause a race. However, if we consider the
> > time from pool creation to deletion: if there is an online-offline-online
> > transition, can zswap_cpu_comp_prepare() race with the call to
> > cpuhp_state_remove_instance()? If so, wouldn't this cause unpredictable
> > behavior?
>
> How will this race happen?
>
> cpuhp_state_remove_instance() is called while a pool is being destroyed,
> while zswap_cpu_comp_prepare() while the pool is being created or during
> CPU onlining.
>
> The former cannot race, and the latter should be synchronized by hotplug
> code.
>
> >
> > I agree, this can occur even with the code in this commit, but there is
> > less risk of things going wrong because we remove the CPU hotplug
> > instance before the pool is added to zswap_pools.
> >
> > Further, removing the CPU hotplug instance directly codifies the
> > intent of this commit, i.e., to use this as a facilitator and manage memory
> > allotted to acomp_ctx, but not to manage those resources' lifetime
> > thereafter.
> >
> > Do you see any advantage of keeping the call to
> cpuhp_state_remove_instance()
> > occur before acomp_ctx_dealloc() in zswap_pool_destroy()? Please let me
> know
> > if I am missing something.
>
> What about more CPUs going online? Without the hotplug instance we don't
> get per-CPU resources for those. We are not using the hotplug mechanism
> just to facilitate per-CPU resource allocation, we use it to
> automatically allocate resources for newly onlined CPUs without having
> to preallocate for all possible CPUs.
This is an excellent point! It makes sense, I will move the call to
cpuhp_state_remove_instance() to be before the call to
acomp_ctx_dealloc() in zswap_pool_destroy(). Thanks for catching this.
>
> Also, this makes the code more difficult to reason about, and is an
> unncessary change from the current behavior.
Ok.
>
> The only change needed is to drop the teardown callback and do the
> freeing in the pool destruction path instead.
Just to summarize: besides moving the call to cpuhp_state_remove_instance()
to zswap_pool_destroy() and more concise comments/commit logs, are there
other changes to be made in patch 20?
Thanks,
Kanchana
^ permalink raw reply [flat|nested] 56+ messages in thread
* Re: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources exist from pool creation to deletion.
2025-09-30 21:56 ` Sridhar, Kanchana P
@ 2025-10-01 15:33 ` Yosry Ahmed
2025-10-01 17:37 ` Sridhar, Kanchana P
0 siblings, 1 reply; 56+ messages in thread
From: Yosry Ahmed @ 2025-10-01 15:33 UTC (permalink / raw)
To: Sridhar, Kanchana P
Cc: linux-kernel, linux-mm, hannes, nphamcs, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius,
Feghali, Wajdi K, Gopal, Vinodh
On Tue, Sep 30, 2025 at 09:56:33PM +0000, Sridhar, Kanchana P wrote:
>
> > -----Original Message-----
> > From: Yosry Ahmed <yosry.ahmed@linux.dev>
> > Sent: Tuesday, September 30, 2025 2:20 PM
> > To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> > Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> > hannes@cmpxchg.org; nphamcs@gmail.com; chengming.zhou@linux.dev;
> > usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> > ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> > senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> > crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> > davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> > ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> > <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>;
> > Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> > <vinodh.gopal@intel.com>
> > Subject: Re: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources
> > exist from pool creation to deletion.
> >
> > > > > > > static struct zswap_pool *zswap_pool_create(char *compressor)
> > > > > > > {
> > > > > > > struct zswap_pool *pool;
> > > > > > > @@ -263,19 +287,43 @@ static struct zswap_pool
> > > > > > *zswap_pool_create(char *compressor)
> > > > > > >
> > > > > > > strscpy(pool->tfm_name, compressor, sizeof(pool-
> > >tfm_name));
> > > > > > >
> > > > > > > - pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
> > > > > > > + /* Many things rely on the zero-initialization. */
> > > > > > > + pool->acomp_ctx = alloc_percpu_gfp(*pool->acomp_ctx,
> > > > > > > + GFP_KERNEL |
> > __GFP_ZERO);
> > > > > > > if (!pool->acomp_ctx) {
> > > > > > > pr_err("percpu alloc failed\n");
> > > > > > > goto error;
> > > > > > > }
> > > > > > >
> > > > > > > - for_each_possible_cpu(cpu)
> > > > > > > - mutex_init(&per_cpu_ptr(pool->acomp_ctx, cpu)-
> > >mutex);
> > > > > > > -
> > > > > > > + /*
> > > > > > > + * This is serialized against CPU hotplug operations. Hence,
> > cores
> > > > > > > + * cannot be offlined until this finishes.
> > > > > > > + * In case of errors, we need to goto "ref_fail" instead of
> > "error"
> > > > > > > + * because there is no teardown callback registered anymore,
> > for
> > > > > > > + * cpuhp_state_add_instance() to de-allocate resources as it
> > rolls
> > > > > > back
> > > > > > > + * state on cores before the CPU on which error was
> > encountered.
> > > > > > > + */
> > > > > > > ret =
> > > > > > cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
> > > > > > > &pool->node);
> > > > > > > +
> > > > > > > + /*
> > > > > > > + * We only needed the multi state instance add operation to
> > invoke
> > > > > > the
> > > > > > > + * startup callback for all cores without cores getting offlined.
> > Since
> > > > > > > + * the acomp_ctx resources will now only be de-allocated
> > when the
> > > > > > pool
> > > > > > > + * is destroyed, we can safely remove the multi state
> > instance. This
> > > > > > > + * minimizes (but does not eliminate) the possibility of
> > > > > > > + * zswap_cpu_comp_prepare() being invoked again due to a
> > CPU
> > > > > > > + * offline-online transition. Removing the instance also
> > prevents race
> > > > > > > + * conditions between CPU onlining after initial pool creation,
> > and
> > > > > > > + * acomp_ctx_dealloc() freeing the acomp_ctx resources.
> > > > > > > + * Note that we delete the instance before checking the error
> > status
> > > > > > of
> > > > > > > + * the node list add operation because we want the instance
> > removal
> > > > > > even
> > > > > > > + * in case of errors in the former.
> > > > > > > + */
> > > > > > > +
> > cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
> > > > > > &pool->node);
> > > > > > > +
> > > > > >
> > > > > > I don't understand what's wrong with the current flow? We call
> > > > > > cpuhp_state_remove_instance() in pool deletion before freeing up the
> > > > > > per-CPU resources. Why is this not enough?
> > > > >
> > > > > This is because with the changes proposed in this commit, the multi
> > state
> > > > > add instance is used during pool creation as a way to create acomp_ctx
> > > > > resources correctly with just the offline/online state transitions
> > guaranteed
> > > > > by CPU hotplug, without needing additional mutex locking as in the
> > > > mainline.
> > > > > In other words, the consistency wrt safely creating/deleting acomp_ctx
> > > > > resources with the changes being proposed is accomplished by the
> > hotplug
> > > > > state transitions guarantee. Stated differently, the hotplug framework
> > > > > helps enforce the new design during pool creation without relying on the
> > > > > mutex and subsequent simplifications during zswap_[de]compress()
> > > > > proposed in this commit.
> > > > >
> > > > > Once this is done, deleting the CPU hotplug state seems cleaner, and
> > > > reflects
> > > > > the change in policy of the resources' lifetime. It also prevents race
> > > > conditions
> > > > > between zswap_cpu_comp_prepare() and acomp_ctx_dealloc() called
> > from
> > > > > zswap_pool_destroy().
> > > >
> > > > How is a race with zswap_cpu_comp_prepare() possible if we call
> > > > cpuhp_state_remove_instance() before acomp_ctx_dealloc() in the pool
> > > > deletion path?
> > >
> > > Good point. I agree, calling cpuhp_state_remove_instance() before
> > > acomp_ctx_dealloc() will not cause a race. However, if we consider the
> > > time from pool creation to deletion: if there is an online-offline-online
> > > transition, can zswap_cpu_comp_prepare() race with the call to
> > > cpuhp_state_remove_instance()? If so, wouldn't this cause unpredictable
> > > behavior?
> >
> > How will this race happen?
> >
> > cpuhp_state_remove_instance() is called while a pool is being destroyed,
> > while zswap_cpu_comp_prepare() while the pool is being created or during
> > CPU onlining.
> >
> > The former cannot race, and the latter should be synchronized by hotplug
> > code.
> >
> > >
> > > I agree, this can occur even with the code in this commit, but there is
> > > less risk of things going wrong because we remove the CPU hotplug
> > > instance before the pool is added to zswap_pools.
> > >
> > > Further, removing the CPU hotplug instance directly codifies the
> > > intent of this commit, i.e., to use this as a facilitator and manage memory
> > > allotted to acomp_ctx, but not to manage those resources' lifetime
> > > thereafter.
> > >
> > > Do you see any advantage of keeping the call to
> > cpuhp_state_remove_instance()
> > > occur before acomp_ctx_dealloc() in zswap_pool_destroy()? Please let me
> > know
> > > if I am missing something.
> >
> > What about more CPUs going online? Without the hotplug instance we don't
> > get per-CPU resources for those. We are not using the hotplug mechanism
> > just to facilitate per-CPU resource allocation, we use it to
> > automatically allocate resources for newly onlined CPUs without having
> > to preallocate for all possible CPUs.
>
> This is an excellent point! It makes sense, I will move the call to
> cpuhp_state_remove_instance() to be before the call to
> acomp_ctx_dealloc() in zswap_pool_destroy(). Thanks for catching this.
>
> >
> > Also, this makes the code more difficult to reason about, and is an
> > unncessary change from the current behavior.
>
> Ok.
>
> >
> > The only change needed is to drop the teardown callback and do the
> > freeing in the pool destruction path instead.
>
> Just to summarize: besides moving the call to cpuhp_state_remove_instance()
> to zswap_pool_destroy() and more concise comments/commit logs, are there
> other changes to be made in patch 20?
I don't believe so. Thanks!
^ permalink raw reply [flat|nested] 56+ messages in thread
* Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches.
2025-09-26 3:35 ` [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches Kanchana P Sridhar
@ 2025-10-01 16:19 ` Yosry Ahmed
2025-10-01 21:20 ` Sridhar, Kanchana P
0 siblings, 1 reply; 56+ messages in thread
From: Yosry Ahmed @ 2025-10-01 16:19 UTC (permalink / raw)
To: Kanchana P Sridhar
Cc: linux-kernel, linux-mm, hannes, nphamcs, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, kristen.c.accardi, vinicius.gomes,
wajdi.k.feghali, vinodh.gopal
On Thu, Sep 25, 2025 at 08:35:01PM -0700, Kanchana P Sridhar wrote:
> This patch makes two major changes:
>
> First, we allocate pool batching resources if the compressor supports
> batching:
>
> This patch sets up zswap for allocating per-CPU resources optimally
> for non-batching and batching compressors.
>
> A new ZSWAP_MAX_BATCH_SIZE constant is defined as 8U, to set an upper
> limit on the number of pages in large folios that will be batch
> compressed.
>
> It is up to the compressor to manage multiple requests, as needed, to
> accomplish batch parallelism. zswap only needs to allocate the per-CPU
> dst buffers according to the batch size supported by the compressor.
>
> A "u8 compr_batch_size" member is added to "struct zswap_pool", as per
> Yosry's suggestion. pool->compr_batch_size is set as the minimum of
> the compressor's max batch-size and ZSWAP_MAX_BATCH_SIZE. Accordingly,
> it proceeds to allocate the necessary compression dst buffers in the
> per-CPU acomp_ctx.
>
> Another "u8 store_batch_size" member is added to "struct zswap_pool"
> to store the unit for batching large folio stores: for batching
> compressors, this is the pool->compr_batch_size. For non-batching
> compressors, this is ZSWAP_MAX_BATCH_SIZE.
>
> zswap does not use more than one dst buffer yet. Follow-up patches
> will actually utilize the multiple acomp_ctx buffers for batch
> compression/decompression of multiple pages.
>
> Thus, ZSWAP_MAX_BATCH_SIZE limits the amount of extra memory used for
> batching. There is a small extra memory overhead of allocating
> the acomp_ctx->buffers array for compressors that do not support
> batching: On x86_64, the overhead is 1 pointer per-CPU (i.e. 8 bytes).
>
> Next, we store the folio in batches:
>
> This patch modifies zswap_store() to store a batch of pages in large
> folios at a time, instead of storing one page at a time. It does this by
> calling a new procedure zswap_store_pages() with a range of
> "pool->store_batch_size" indices in the folio.
>
> zswap_store_pages() implements all the computes done earlier in
> zswap_store_page() for a single-page, for multiple pages in a folio,
> namely the "batch":
>
> 1) It starts by allocating all zswap entries required to store the
> batch. New procedures, zswap_entries_cache_alloc_batch() and
> zswap_entries_cache_free_batch() call kmem_cache_[free]alloc_bulk()
> to optimize the performance of this step.
>
> 2) Next, the entries fields are written, computes that need to be happen
> anyway, without modifying the zswap xarray/LRU publishing order. This
> improves latency by avoiding having to bring the entries into the
> cache for writing in different code blocks within this procedure.
>
> 3) Next, it calls zswap_compress() to sequentially compress each page in
> the batch.
>
> 4) Finally, it adds the batch's zswap entries to the xarray and LRU,
> charges zswap memory and increments zswap stats.
>
> 5) The error handling and cleanup required for all failure scenarios
> that can occur while storing a batch in zswap are consolidated to a
> single "store_pages_failed" label in zswap_store_pages(). Here again,
> we optimize performance by calling kmem_cache_free_bulk().
>
> This commit also makes a minor optimization in zswap_compress(), for the
> info on whether or not the page's folio has memcg writeback enabled to
> be passed in via a "bool folio_wb" flag from zswap_store(). The intent
> is to not re-compute this for every page in a folio. Since
> zswap_compress() is a static function, I figured this should be safe.
> A repetition of "dlen = PAGE_SIZE" is deleted.
>
> Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
> ---
> mm/zswap.c | 319 +++++++++++++++++++++++++++++++++++++----------------
> 1 file changed, 224 insertions(+), 95 deletions(-)
>
> diff --git a/mm/zswap.c b/mm/zswap.c
> index 3b3716808d7d..9e0e7887de33 100644
> --- a/mm/zswap.c
> +++ b/mm/zswap.c
> @@ -82,6 +82,9 @@ static bool zswap_pool_reached_full;
>
> #define ZSWAP_PARAM_UNSET ""
>
> +/* Limit the batch size to limit per-CPU memory usage for dst buffers. */
> +#define ZSWAP_MAX_BATCH_SIZE 8U
> +
> static int zswap_setup(void);
>
> /* Enable/disable zswap */
> @@ -139,7 +142,7 @@ struct crypto_acomp_ctx {
> struct crypto_acomp *acomp;
> struct acomp_req *req;
> struct crypto_wait wait;
> - u8 *buffer;
> + u8 **buffers;
> struct mutex mutex;
> bool is_sleepable;
> };
> @@ -158,6 +161,8 @@ struct zswap_pool {
> struct work_struct release_work;
> struct hlist_node node;
> char tfm_name[CRYPTO_MAX_ALG_NAME];
> + u8 compr_batch_size;
> + u8 store_batch_size;
I don't think we need to store store_batch_size, seems trivial to
calculate at store time (perhaps in a helper).
Taking a step back, is there any benefit to limiting store_batch_size to
compr_batch_size? Is there a disadvantage to using ZSWAP_MAX_BATCH_SIZE
even if it's higher than the HW compression batch size?
> };
>
> /* Global LRU lists shared by all zswap pools. */
> @@ -252,8 +257,10 @@ static void __zswap_pool_empty(struct percpu_ref *ref);
> * zswap_cpu_comp_prepare(), not others.
> * - Cleanup acomp_ctx resources on all cores in zswap_pool_destroy().
> */
> -static void acomp_ctx_dealloc(struct crypto_acomp_ctx *acomp_ctx)
> +static void acomp_ctx_dealloc(struct crypto_acomp_ctx *acomp_ctx, u8 nr_buffers)
> {
> + u8 i;
> +
> if (IS_ERR_OR_NULL(acomp_ctx))
> return;
>
> @@ -263,7 +270,11 @@ static void acomp_ctx_dealloc(struct crypto_acomp_ctx *acomp_ctx)
> if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
> crypto_free_acomp(acomp_ctx->acomp);
>
> - kfree(acomp_ctx->buffer);
> + if (acomp_ctx->buffers) {
> + for (i = 0; i < nr_buffers; ++i)
> + kfree(acomp_ctx->buffers[i]);
> + kfree(acomp_ctx->buffers);
> + }
> }
>
> static struct zswap_pool *zswap_pool_create(char *compressor)
> @@ -275,6 +286,7 @@ static struct zswap_pool *zswap_pool_create(char *compressor)
> if (!zswap_has_pool && !strcmp(compressor, ZSWAP_PARAM_UNSET))
> return NULL;
>
> + /* Many things rely on the zero-initialization. */
> pool = kzalloc(sizeof(*pool), GFP_KERNEL);
> if (!pool)
> return NULL;
> @@ -334,13 +346,28 @@ static struct zswap_pool *zswap_pool_create(char *compressor)
> goto ref_fail;
> INIT_LIST_HEAD(&pool->list);
>
> + /*
> + * Set the unit of compress batching for large folios, for quick
> + * retrieval in the zswap_compress() fast path:
> + * If the compressor is sequential (@pool->compr_batch_size is 1),
> + * large folios will be compressed in batches of ZSWAP_MAX_BATCH_SIZE
> + * pages, where each page in the batch is compressed sequentially.
> + * We see better performance by processing the folio in batches of
> + * ZSWAP_MAX_BATCH_SIZE, due to cache locality of working set
> + * structures.
> + */
> + pool->store_batch_size = (pool->compr_batch_size > 1) ?
> + pool->compr_batch_size : ZSWAP_MAX_BATCH_SIZE;
> +
> zswap_pool_debug("created", pool);
>
> return pool;
>
> ref_fail:
> for_each_possible_cpu(cpu)
> - acomp_ctx_dealloc(per_cpu_ptr(pool->acomp_ctx, cpu));
> + acomp_ctx_dealloc(per_cpu_ptr(pool->acomp_ctx, cpu),
> + pool->compr_batch_size);
> +
> error:
> if (pool->acomp_ctx)
> free_percpu(pool->acomp_ctx);
> @@ -376,7 +403,8 @@ static void zswap_pool_destroy(struct zswap_pool *pool)
> zswap_pool_debug("destroying", pool);
>
> for_each_possible_cpu(cpu)
> - acomp_ctx_dealloc(per_cpu_ptr(pool->acomp_ctx, cpu));
> + acomp_ctx_dealloc(per_cpu_ptr(pool->acomp_ctx, cpu),
> + pool->compr_batch_size);
>
> free_percpu(pool->acomp_ctx);
>
> @@ -763,6 +791,24 @@ static void zswap_entry_cache_free(struct zswap_entry *entry)
> kmem_cache_free(zswap_entry_cache, entry);
> }
>
> +/*
> + * Returns 0 if kmem_cache_alloc_bulk() failed and a positive number otherwise.
> + * The code for __kmem_cache_alloc_bulk() indicates that this positive number
> + * will be the @size requested, i.e., @nr_entries.
The behavior is not documented tho, and other users seem like they don't
all assume the return has to be 0 or nr_entries. Maybe we should add a
WARN if the returned size is not nr_entries or 0?
> + */
> +static __always_inline int zswap_entries_cache_alloc_batch(void **entries,
> + unsigned int nr_entries,
> + gfp_t gfp)
> +{
> + return kmem_cache_alloc_bulk(zswap_entry_cache, gfp, nr_entries, entries);
We currently use kmem_cache_alloc_node() in zswap_entry_cache_alloc() to
allocate the entry on the same node as the compressed page. We use
entry_to_nid() to get the node for LRU operations.
This breaks that assumption.
> +}
> +
> +static __always_inline void zswap_entries_cache_free_batch(void **entries,
> + unsigned int nr_entries)
> +{
> + kmem_cache_free_bulk(zswap_entry_cache, nr_entries, entries);
> +}
> +
> /*
> * Carries out the common pattern of freeing an entry's zsmalloc allocation,
> * freeing the entry itself, and decrementing the number of stored pages.
> @@ -789,7 +835,9 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
> {
> struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
> struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
> + int cpu_node = cpu_to_node(cpu);
"nid" is a more common name.
> int ret = -ENOMEM;
> + u8 i;
>
> /*
> * The per-CPU pool->acomp_ctx is zero-initialized on allocation.
> @@ -802,11 +850,7 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
> if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
> return 0;
>
> - acomp_ctx->buffer = kmalloc_node(PAGE_SIZE, GFP_KERNEL, cpu_to_node(cpu));
> - if (!acomp_ctx->buffer)
> - return ret;
> -
> - acomp_ctx->acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_to_node(cpu));
> + acomp_ctx->acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_node);
> if (IS_ERR_OR_NULL(acomp_ctx->acomp)) {
> pr_err("could not alloc crypto acomp %s : %ld\n",
> pool->tfm_name, PTR_ERR(acomp_ctx->acomp));
> @@ -815,20 +859,40 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
> }
> acomp_ctx->is_sleepable = acomp_is_async(acomp_ctx->acomp);
>
> + /*
> + * Allocate up to ZSWAP_MAX_BATCH_SIZE dst buffers if the
> + * compressor supports batching.
> + */
> + pool->compr_batch_size = min(ZSWAP_MAX_BATCH_SIZE,
> + crypto_acomp_batch_size(acomp_ctx->acomp));
> +
> acomp_ctx->req = acomp_request_alloc(acomp_ctx->acomp);
> +
> if (IS_ERR_OR_NULL(acomp_ctx->req)) {
> pr_err("could not alloc crypto acomp_request %s\n",
> - pool->tfm_name);
> + pool->tfm_name);
Unrelated change.
> goto fail;
> }
>
> - crypto_init_wait(&acomp_ctx->wait);
> + acomp_ctx->buffers = kcalloc_node(pool->compr_batch_size, sizeof(u8 *),
> + GFP_KERNEL, cpu_node);
> + if (!acomp_ctx->buffers)
> + goto fail;
> +
> + for (i = 0; i < pool->compr_batch_size; ++i) {
> + acomp_ctx->buffers[i] = kmalloc_node(PAGE_SIZE, GFP_KERNEL,
> + cpu_node);
> + if (!acomp_ctx->buffers[i])
> + goto fail;
> + }
>
> /*
> * if the backend of acomp is async zip, crypto_req_done() will wakeup
> * crypto_wait_req(); if the backend of acomp is scomp, the callback
> * won't be called, crypto_wait_req() will return without blocking.
> */
> + crypto_init_wait(&acomp_ctx->wait);
> +
> acomp_request_set_callback(acomp_ctx->req, CRYPTO_TFM_REQ_MAY_BACKLOG,
> crypto_req_done, &acomp_ctx->wait);
>
> @@ -836,12 +900,12 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
> return 0;
>
> fail:
> - acomp_ctx_dealloc(acomp_ctx);
> + acomp_ctx_dealloc(acomp_ctx, pool->compr_batch_size);
> return ret;
> }
>
> static bool zswap_compress(struct page *page, struct zswap_entry *entry,
> - struct zswap_pool *pool)
> + struct zswap_pool *pool, bool folio_wb)
Maybe "wb_enabled" instead of folio_wb?
> {
> struct crypto_acomp_ctx *acomp_ctx;
> struct scatterlist input, output;
> @@ -855,7 +919,7 @@ static bool zswap_compress(struct page *page, struct zswap_entry *entry,
> acomp_ctx = raw_cpu_ptr(pool->acomp_ctx);
> mutex_lock(&acomp_ctx->mutex);
>
> - dst = acomp_ctx->buffer;
> + dst = acomp_ctx->buffers[0];
> sg_init_table(&input, 1);
> sg_set_page(&input, page, PAGE_SIZE, 0);
>
> @@ -886,13 +950,11 @@ static bool zswap_compress(struct page *page, struct zswap_entry *entry,
> */
> if (comp_ret || !dlen || dlen >= PAGE_SIZE) {
> dlen = PAGE_SIZE;
> - if (!mem_cgroup_zswap_writeback_enabled(
> - folio_memcg(page_folio(page)))) {
> + if (!folio_wb) {
> comp_ret = comp_ret ? comp_ret : -EINVAL;
> goto unlock;
> }
> comp_ret = 0;
> - dlen = PAGE_SIZE;
Unrelated change.
> dst = kmap_local_page(page);
> mapped = true;
> }
> @@ -932,7 +994,7 @@ static bool zswap_decompress(struct zswap_entry *entry, struct folio *folio)
>
> acomp_ctx = raw_cpu_ptr(pool->acomp_ctx);
> mutex_lock(&acomp_ctx->mutex);
> - obj = zs_obj_read_begin(pool->zs_pool, entry->handle, acomp_ctx->buffer);
> + obj = zs_obj_read_begin(pool->zs_pool, entry->handle, acomp_ctx->buffers[0]);
>
> /* zswap entries of length PAGE_SIZE are not compressed. */
> if (entry->length == PAGE_SIZE) {
> @@ -942,15 +1004,15 @@ static bool zswap_decompress(struct zswap_entry *entry, struct folio *folio)
>
> /*
> * zs_obj_read_begin() might return a kmap address of highmem when
> - * acomp_ctx->buffer is not used. However, sg_init_one() does not
> - * handle highmem addresses, so copy the object to acomp_ctx->buffer.
> + * acomp_ctx->buffers[0] is not used. However, sg_init_one() does not
> + * handle highmem addresses, so copy the object to acomp_ctx->buffers[0].
> */
> if (virt_addr_valid(obj)) {
> src = obj;
> } else {
> - WARN_ON_ONCE(obj == acomp_ctx->buffer);
> - memcpy(acomp_ctx->buffer, obj, entry->length);
> - src = acomp_ctx->buffer;
> + WARN_ON_ONCE(obj == acomp_ctx->buffers[0]);
> + memcpy(acomp_ctx->buffers[0], obj, entry->length);
> + src = acomp_ctx->buffers[0];
> }
>
> sg_init_one(&input, src, entry->length);
> @@ -1404,95 +1466,160 @@ static void shrink_worker(struct work_struct *w)
> * main API
> **********************************/
>
> -static bool zswap_store_page(struct page *page,
> - struct obj_cgroup *objcg,
> - struct zswap_pool *pool)
> +/*
> + * Store multiple pages in @folio, starting from the page at index @start up to
> + * the page at index @end-1.
> + */
> +static bool zswap_store_pages(struct folio *folio,
> + long start,
> + long end,
> + struct obj_cgroup *objcg,
> + struct zswap_pool *pool,
> + int node_id,
> + bool folio_wb)
> {
> - swp_entry_t page_swpentry = page_swap_entry(page);
> - struct zswap_entry *entry, *old;
> -
> - /* allocate entry */
> - entry = zswap_entry_cache_alloc(GFP_KERNEL, page_to_nid(page));
> - if (!entry) {
> - zswap_reject_kmemcache_fail++;
> - return false;
> + struct zswap_entry *entries[ZSWAP_MAX_BATCH_SIZE];
> + u8 i, store_fail_idx = 0, nr_pages = end - start;
> +
> + VM_WARN_ON_ONCE(nr_pages > ZSWAP_MAX_BATCH_SIZE);
> +
> + if (unlikely(!zswap_entries_cache_alloc_batch((void **)&entries[0],
> + nr_pages, GFP_KERNEL))) {
> + for (i = 0; i < nr_pages; ++i) {
> + entries[i] = zswap_entry_cache_alloc(GFP_KERNEL, node_id);
> +
> + if (unlikely(!entries[i])) {
> + zswap_reject_kmemcache_fail++;
> + /*
> + * While handling this error, we only need to
> + * call zswap_entries_cache_free_batch() for
> + * entries[0 .. i-1].
> + */
> + nr_pages = i;
> + goto store_pages_failed;
> + }
Is it okay to use kmem_cache_free_bulk() to free slab objects that were
not allocated with kmem_cache_alloc_bulk()?
> + }
> }
>
> - if (!zswap_compress(page, entry, pool))
> - goto compress_failed;
> + /*
> + * Three sets of initializations are done to minimize bringing
> + * @entries into the cache for writing at different parts of this
> + * procedure, since doing so regresses performance:
> + *
> + * 1) Do all the writes to each entry in one code block. These
> + * writes need to be done anyway upon success which is more likely
> + * than not.
> + *
> + * 2) Initialize the handle to an error value. This facilitates
> + * having a consolidated failure handling
> + * 'goto store_pages_failed' that can inspect the value of the
> + * handle to determine whether zsmalloc memory needs to be
> + * de-allocated.
> + *
> + * 3) The page_swap_entry() is obtained once and stored in the entry.
> + * Subsequent store in xarray gets the entry->swpentry instead of
> + * calling page_swap_entry(), minimizing computes.
> + */
Very long comment, and I am not sure what it is trying to say. We don't
need to describe what the code is doing like that.
The only thing that may be worth pointing out is that we are colocating
initialization as much as possible here to minimize potential cache
misses.
Does it actually matter if we do the initializations here vs. right
before inserting to the LRU (current behavior)?
> + for (i = 0; i < nr_pages; ++i) {
> + entries[i]->handle = (unsigned long)ERR_PTR(-EINVAL);
> + entries[i]->pool = pool;
> + entries[i]->swpentry = page_swap_entry(folio_page(folio, start + i));
> + entries[i]->objcg = objcg;
> + entries[i]->referenced = true;
> + INIT_LIST_HEAD(&entries[i]->lru);
> + }
>
> - old = xa_store(swap_zswap_tree(page_swpentry),
> - swp_offset(page_swpentry),
> - entry, GFP_KERNEL);
> - if (xa_is_err(old)) {
> - int err = xa_err(old);
> + for (i = 0; i < nr_pages; ++i) {
> + struct page *page = folio_page(folio, start + i);
>
> - WARN_ONCE(err != -ENOMEM, "unexpected xarray error: %d\n", err);
> - zswap_reject_alloc_fail++;
> - goto store_failed;
> + if (!zswap_compress(page, entries[i], pool, folio_wb))
> + goto store_pages_failed;
> }
>
> - /*
> - * We may have had an existing entry that became stale when
> - * the folio was redirtied and now the new version is being
> - * swapped out. Get rid of the old.
> - */
> - if (old)
> - zswap_entry_free(old);
> + for (i = 0; i < nr_pages; ++i) {
> + struct zswap_entry *old, *entry = entries[i];
>
> - /*
> - * The entry is successfully compressed and stored in the tree, there is
> - * no further possibility of failure. Grab refs to the pool and objcg,
> - * charge zswap memory, and increment zswap_stored_pages.
> - * The opposite actions will be performed by zswap_entry_free()
> - * when the entry is removed from the tree.
> - */
> - zswap_pool_get(pool);
> - if (objcg) {
> - obj_cgroup_get(objcg);
> - obj_cgroup_charge_zswap(objcg, entry->length);
> - }
> - atomic_long_inc(&zswap_stored_pages);
> - if (entry->length == PAGE_SIZE)
> - atomic_long_inc(&zswap_stored_incompressible_pages);
> + old = xa_store(swap_zswap_tree(entry->swpentry),
> + swp_offset(entry->swpentry),
> + entry, GFP_KERNEL);
> + if (unlikely(xa_is_err(old))) {
> + int err = xa_err(old);
>
> - /*
> - * We finish initializing the entry while it's already in xarray.
> - * This is safe because:
> - *
> - * 1. Concurrent stores and invalidations are excluded by folio lock.
> - *
> - * 2. Writeback is excluded by the entry not being on the LRU yet.
> - * The publishing order matters to prevent writeback from seeing
> - * an incoherent entry.
> - */
> - entry->pool = pool;
> - entry->swpentry = page_swpentry;
> - entry->objcg = objcg;
> - entry->referenced = true;
> - if (entry->length) {
> - INIT_LIST_HEAD(&entry->lru);
> - zswap_lru_add(&zswap_list_lru, entry);
> + WARN_ONCE(err != -ENOMEM, "unexpected xarray error: %d\n", err);
> + zswap_reject_alloc_fail++;
> + /*
> + * Entries up to this point have been stored in the
> + * xarray. zswap_store() will erase them from the xarray
> + * and call zswap_entry_free(). Local cleanup in
> + * 'store_pages_failed' only needs to happen for
> + * entries from [@i to @nr_pages).
> + */
> + store_fail_idx = i;
> + goto store_pages_failed;
> + }
> +
> + /*
> + * We may have had an existing entry that became stale when
> + * the folio was redirtied and now the new version is being
> + * swapped out. Get rid of the old.
> + */
> + if (unlikely(old))
> + zswap_entry_free(old);
> +
> + /*
> + * The entry is successfully compressed and stored in the tree, there is
> + * no further possibility of failure. Grab refs to the pool and objcg,
> + * charge zswap memory, and increment zswap_stored_pages.
> + * The opposite actions will be performed by zswap_entry_free()
> + * when the entry is removed from the tree.
> + */
But there *is* further possibility of failure if a subsequent entry
xa_store() fails, right?
Seems like if xa_store() fails we do not cleanup previously charged
objects, pool references, zswap_stored_pages, etc. Instead of rolling
all this back on failure, can we do all the xarray stores first and only
do the rest when we're at a point where no failure can happen? Would
that cause a performance regression?
> + zswap_pool_get(pool);
> + if (objcg) {
> + obj_cgroup_get(objcg);
> + obj_cgroup_charge_zswap(objcg, entry->length);
> + }
> + atomic_long_inc(&zswap_stored_pages);
> + if (entry->length == PAGE_SIZE)
> + atomic_long_inc(&zswap_stored_incompressible_pages);
> +
> + /*
> + * We finish by adding the entry to the LRU while it's already
> + * in xarray. This is safe because:
> + *
> + * 1. Concurrent stores and invalidations are excluded by folio lock.
> + *
> + * 2. Writeback is excluded by the entry not being on the LRU yet.
> + * The publishing order matters to prevent writeback from seeing
> + * an incoherent entry.
> + */
> + if (likely(entry->length))
> + zswap_lru_add(&zswap_list_lru, entry);
> }
>
> return true;
>
> -store_failed:
> - zs_free(pool->zs_pool, entry->handle);
> -compress_failed:
> - zswap_entry_cache_free(entry);
> +store_pages_failed:
> + for (i = store_fail_idx; i < nr_pages; ++i) {
> + if (!IS_ERR_VALUE(entries[i]->handle))
> + zs_free(pool->zs_pool, entries[i]->handle);
> + }
> + zswap_entries_cache_free_batch((void **)&entries[store_fail_idx],
> + nr_pages - store_fail_idx);
> +
> return false;
> }
>
> bool zswap_store(struct folio *folio)
> {
> + bool folio_wb = mem_cgroup_zswap_writeback_enabled(folio_memcg(folio));
Ditto renaming folio_wb.
> long nr_pages = folio_nr_pages(folio);
> + int node_id = folio_nid(folio);
Ditto nid.
> swp_entry_t swp = folio->swap;
> struct obj_cgroup *objcg = NULL;
> struct mem_cgroup *memcg = NULL;
> struct zswap_pool *pool;
> bool ret = false;
> - long index;
> + long start, end;
>
> VM_WARN_ON_ONCE(!folio_test_locked(folio));
> VM_WARN_ON_ONCE(!folio_test_swapcache(folio));
> @@ -1526,10 +1653,12 @@ bool zswap_store(struct folio *folio)
> mem_cgroup_put(memcg);
> }
>
> - for (index = 0; index < nr_pages; ++index) {
> - struct page *page = folio_page(folio, index);
> + /* Store the folio in batches of @pool->store_batch_size pages. */
> + for (start = 0; start < nr_pages; start += pool->store_batch_size) {
> + end = min(start + pool->store_batch_size, nr_pages);
>
> - if (!zswap_store_page(page, objcg, pool))
> + if (!zswap_store_pages(folio, start, end, objcg, pool,
> + node_id, folio_wb))
> goto put_pool;
> }
>
> @@ -1559,9 +1688,9 @@ bool zswap_store(struct folio *folio)
> struct zswap_entry *entry;
> struct xarray *tree;
>
> - for (index = 0; index < nr_pages; ++index) {
> - tree = swap_zswap_tree(swp_entry(type, offset + index));
> - entry = xa_erase(tree, offset + index);
> + for (start = 0; start < nr_pages; ++start) {
> + tree = swap_zswap_tree(swp_entry(type, offset + start));
> + entry = xa_erase(tree, offset + start);
> if (entry)
> zswap_entry_free(entry);
> }
> --
> 2.27.0
>
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources exist from pool creation to deletion.
2025-10-01 15:33 ` Yosry Ahmed
@ 2025-10-01 17:37 ` Sridhar, Kanchana P
0 siblings, 0 replies; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-10-01 17:37 UTC (permalink / raw)
To: Yosry Ahmed
Cc: linux-kernel, linux-mm, hannes, nphamcs, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius,
Feghali, Wajdi K, Gopal, Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Yosry Ahmed <yosry.ahmed@linux.dev>
> Sent: Wednesday, October 1, 2025 8:33 AM
> To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> hannes@cmpxchg.org; nphamcs@gmail.com; chengming.zhou@linux.dev;
> usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>;
> Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> <vinodh.gopal@intel.com>
> Subject: Re: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources
> exist from pool creation to deletion.
>
> On Tue, Sep 30, 2025 at 09:56:33PM +0000, Sridhar, Kanchana P wrote:
> >
> > > -----Original Message-----
> > > From: Yosry Ahmed <yosry.ahmed@linux.dev>
> > > Sent: Tuesday, September 30, 2025 2:20 PM
> > > To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> > > Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> > > hannes@cmpxchg.org; nphamcs@gmail.com;
> chengming.zhou@linux.dev;
> > > usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> > > ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> > > senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> > > crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> > > davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> > > ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> > > <kristen.c.accardi@intel.com>; Gomes, Vinicius
> <vinicius.gomes@intel.com>;
> > > Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> > > <vinodh.gopal@intel.com>
> > > Subject: Re: [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx
> resources
> > > exist from pool creation to deletion.
> > >
> > > > > > > > static struct zswap_pool *zswap_pool_create(char *compressor)
> > > > > > > > {
> > > > > > > > struct zswap_pool *pool;
> > > > > > > > @@ -263,19 +287,43 @@ static struct zswap_pool
> > > > > > > *zswap_pool_create(char *compressor)
> > > > > > > >
> > > > > > > > strscpy(pool->tfm_name, compressor, sizeof(pool-
> > > >tfm_name));
> > > > > > > >
> > > > > > > > - pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
> > > > > > > > + /* Many things rely on the zero-initialization. */
> > > > > > > > + pool->acomp_ctx = alloc_percpu_gfp(*pool->acomp_ctx,
> > > > > > > > + GFP_KERNEL |
> > > __GFP_ZERO);
> > > > > > > > if (!pool->acomp_ctx) {
> > > > > > > > pr_err("percpu alloc failed\n");
> > > > > > > > goto error;
> > > > > > > > }
> > > > > > > >
> > > > > > > > - for_each_possible_cpu(cpu)
> > > > > > > > - mutex_init(&per_cpu_ptr(pool->acomp_ctx, cpu)-
> > > >mutex);
> > > > > > > > -
> > > > > > > > + /*
> > > > > > > > + * This is serialized against CPU hotplug operations. Hence,
> > > cores
> > > > > > > > + * cannot be offlined until this finishes.
> > > > > > > > + * In case of errors, we need to goto "ref_fail" instead of
> > > "error"
> > > > > > > > + * because there is no teardown callback registered anymore,
> > > for
> > > > > > > > + * cpuhp_state_add_instance() to de-allocate resources as it
> > > rolls
> > > > > > > back
> > > > > > > > + * state on cores before the CPU on which error was
> > > encountered.
> > > > > > > > + */
> > > > > > > > ret =
> > > > > > > cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
> > > > > > > > &pool->node);
> > > > > > > > +
> > > > > > > > + /*
> > > > > > > > + * We only needed the multi state instance add operation to
> > > invoke
> > > > > > > the
> > > > > > > > + * startup callback for all cores without cores getting offlined.
> > > Since
> > > > > > > > + * the acomp_ctx resources will now only be de-allocated
> > > when the
> > > > > > > pool
> > > > > > > > + * is destroyed, we can safely remove the multi state
> > > instance. This
> > > > > > > > + * minimizes (but does not eliminate) the possibility of
> > > > > > > > + * zswap_cpu_comp_prepare() being invoked again due to a
> > > CPU
> > > > > > > > + * offline-online transition. Removing the instance also
> > > prevents race
> > > > > > > > + * conditions between CPU onlining after initial pool creation,
> > > and
> > > > > > > > + * acomp_ctx_dealloc() freeing the acomp_ctx resources.
> > > > > > > > + * Note that we delete the instance before checking the error
> > > status
> > > > > > > of
> > > > > > > > + * the node list add operation because we want the instance
> > > removal
> > > > > > > even
> > > > > > > > + * in case of errors in the former.
> > > > > > > > + */
> > > > > > > > +
> > > cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
> > > > > > > &pool->node);
> > > > > > > > +
> > > > > > >
> > > > > > > I don't understand what's wrong with the current flow? We call
> > > > > > > cpuhp_state_remove_instance() in pool deletion before freeing up
> the
> > > > > > > per-CPU resources. Why is this not enough?
> > > > > >
> > > > > > This is because with the changes proposed in this commit, the multi
> > > state
> > > > > > add instance is used during pool creation as a way to create
> acomp_ctx
> > > > > > resources correctly with just the offline/online state transitions
> > > guaranteed
> > > > > > by CPU hotplug, without needing additional mutex locking as in the
> > > > > mainline.
> > > > > > In other words, the consistency wrt safely creating/deleting
> acomp_ctx
> > > > > > resources with the changes being proposed is accomplished by the
> > > hotplug
> > > > > > state transitions guarantee. Stated differently, the hotplug
> framework
> > > > > > helps enforce the new design during pool creation without relying on
> the
> > > > > > mutex and subsequent simplifications during zswap_[de]compress()
> > > > > > proposed in this commit.
> > > > > >
> > > > > > Once this is done, deleting the CPU hotplug state seems cleaner, and
> > > > > reflects
> > > > > > the change in policy of the resources' lifetime. It also prevents race
> > > > > conditions
> > > > > > between zswap_cpu_comp_prepare() and acomp_ctx_dealloc()
> called
> > > from
> > > > > > zswap_pool_destroy().
> > > > >
> > > > > How is a race with zswap_cpu_comp_prepare() possible if we call
> > > > > cpuhp_state_remove_instance() before acomp_ctx_dealloc() in the
> pool
> > > > > deletion path?
> > > >
> > > > Good point. I agree, calling cpuhp_state_remove_instance() before
> > > > acomp_ctx_dealloc() will not cause a race. However, if we consider the
> > > > time from pool creation to deletion: if there is an online-offline-online
> > > > transition, can zswap_cpu_comp_prepare() race with the call to
> > > > cpuhp_state_remove_instance()? If so, wouldn't this cause
> unpredictable
> > > > behavior?
> > >
> > > How will this race happen?
> > >
> > > cpuhp_state_remove_instance() is called while a pool is being destroyed,
> > > while zswap_cpu_comp_prepare() while the pool is being created or
> during
> > > CPU onlining.
> > >
> > > The former cannot race, and the latter should be synchronized by hotplug
> > > code.
> > >
> > > >
> > > > I agree, this can occur even with the code in this commit, but there is
> > > > less risk of things going wrong because we remove the CPU hotplug
> > > > instance before the pool is added to zswap_pools.
> > > >
> > > > Further, removing the CPU hotplug instance directly codifies the
> > > > intent of this commit, i.e., to use this as a facilitator and manage
> memory
> > > > allotted to acomp_ctx, but not to manage those resources' lifetime
> > > > thereafter.
> > > >
> > > > Do you see any advantage of keeping the call to
> > > cpuhp_state_remove_instance()
> > > > occur before acomp_ctx_dealloc() in zswap_pool_destroy()? Please let
> me
> > > know
> > > > if I am missing something.
> > >
> > > What about more CPUs going online? Without the hotplug instance we
> don't
> > > get per-CPU resources for those. We are not using the hotplug mechanism
> > > just to facilitate per-CPU resource allocation, we use it to
> > > automatically allocate resources for newly onlined CPUs without having
> > > to preallocate for all possible CPUs.
> >
> > This is an excellent point! It makes sense, I will move the call to
> > cpuhp_state_remove_instance() to be before the call to
> > acomp_ctx_dealloc() in zswap_pool_destroy(). Thanks for catching this.
> >
> > >
> > > Also, this makes the code more difficult to reason about, and is an
> > > unncessary change from the current behavior.
> >
> > Ok.
> >
> > >
> > > The only change needed is to drop the teardown callback and do the
> > > freeing in the pool destruction path instead.
> >
> > Just to summarize: besides moving the call to
> cpuhp_state_remove_instance()
> > to zswap_pool_destroy() and more concise comments/commit logs, are
> there
> > other changes to be made in patch 20?
>
> I don't believe so. Thanks!
Sounds good, thanks!
Thanks,
Kanchana
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches.
2025-10-01 16:19 ` Yosry Ahmed
@ 2025-10-01 21:20 ` Sridhar, Kanchana P
2025-10-03 19:10 ` Sridhar, Kanchana P
` (2 more replies)
0 siblings, 3 replies; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-10-01 21:20 UTC (permalink / raw)
To: Yosry Ahmed
Cc: linux-kernel, linux-mm, hannes, nphamcs, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius,
Feghali, Wajdi K, Gopal, Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Yosry Ahmed <yosry.ahmed@linux.dev>
> Sent: Wednesday, October 1, 2025 9:19 AM
> To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> hannes@cmpxchg.org; nphamcs@gmail.com; chengming.zhou@linux.dev;
> usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>;
> Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> <vinodh.gopal@intel.com>
> Subject: Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a
> large folio in batches.
>
> On Thu, Sep 25, 2025 at 08:35:01PM -0700, Kanchana P Sridhar wrote:
> > This patch makes two major changes:
> >
> > First, we allocate pool batching resources if the compressor supports
> > batching:
> >
> > This patch sets up zswap for allocating per-CPU resources optimally
> > for non-batching and batching compressors.
> >
> > A new ZSWAP_MAX_BATCH_SIZE constant is defined as 8U, to set an upper
> > limit on the number of pages in large folios that will be batch
> > compressed.
> >
> > It is up to the compressor to manage multiple requests, as needed, to
> > accomplish batch parallelism. zswap only needs to allocate the per-CPU
> > dst buffers according to the batch size supported by the compressor.
> >
> > A "u8 compr_batch_size" member is added to "struct zswap_pool", as per
> > Yosry's suggestion. pool->compr_batch_size is set as the minimum of
> > the compressor's max batch-size and ZSWAP_MAX_BATCH_SIZE.
> Accordingly,
> > it proceeds to allocate the necessary compression dst buffers in the
> > per-CPU acomp_ctx.
> >
> > Another "u8 store_batch_size" member is added to "struct zswap_pool"
> > to store the unit for batching large folio stores: for batching
> > compressors, this is the pool->compr_batch_size. For non-batching
> > compressors, this is ZSWAP_MAX_BATCH_SIZE.
> >
> > zswap does not use more than one dst buffer yet. Follow-up patches
> > will actually utilize the multiple acomp_ctx buffers for batch
> > compression/decompression of multiple pages.
> >
> > Thus, ZSWAP_MAX_BATCH_SIZE limits the amount of extra memory used
> for
> > batching. There is a small extra memory overhead of allocating
> > the acomp_ctx->buffers array for compressors that do not support
> > batching: On x86_64, the overhead is 1 pointer per-CPU (i.e. 8 bytes).
> >
> > Next, we store the folio in batches:
> >
> > This patch modifies zswap_store() to store a batch of pages in large
> > folios at a time, instead of storing one page at a time. It does this by
> > calling a new procedure zswap_store_pages() with a range of
> > "pool->store_batch_size" indices in the folio.
> >
> > zswap_store_pages() implements all the computes done earlier in
> > zswap_store_page() for a single-page, for multiple pages in a folio,
> > namely the "batch":
> >
> > 1) It starts by allocating all zswap entries required to store the
> > batch. New procedures, zswap_entries_cache_alloc_batch() and
> > zswap_entries_cache_free_batch() call kmem_cache_[free]alloc_bulk()
> > to optimize the performance of this step.
> >
> > 2) Next, the entries fields are written, computes that need to be happen
> > anyway, without modifying the zswap xarray/LRU publishing order. This
> > improves latency by avoiding having to bring the entries into the
> > cache for writing in different code blocks within this procedure.
> >
> > 3) Next, it calls zswap_compress() to sequentially compress each page in
> > the batch.
> >
> > 4) Finally, it adds the batch's zswap entries to the xarray and LRU,
> > charges zswap memory and increments zswap stats.
> >
> > 5) The error handling and cleanup required for all failure scenarios
> > that can occur while storing a batch in zswap are consolidated to a
> > single "store_pages_failed" label in zswap_store_pages(). Here again,
> > we optimize performance by calling kmem_cache_free_bulk().
> >
> > This commit also makes a minor optimization in zswap_compress(), for the
> > info on whether or not the page's folio has memcg writeback enabled to
> > be passed in via a "bool folio_wb" flag from zswap_store(). The intent
> > is to not re-compute this for every page in a folio. Since
> > zswap_compress() is a static function, I figured this should be safe.
> > A repetition of "dlen = PAGE_SIZE" is deleted.
> >
> > Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@intel.com>
> > ---
> > mm/zswap.c | 319 +++++++++++++++++++++++++++++++++++++-----------
> -----
> > 1 file changed, 224 insertions(+), 95 deletions(-)
> >
> > diff --git a/mm/zswap.c b/mm/zswap.c
> > index 3b3716808d7d..9e0e7887de33 100644
> > --- a/mm/zswap.c
> > +++ b/mm/zswap.c
> > @@ -82,6 +82,9 @@ static bool zswap_pool_reached_full;
> >
> > #define ZSWAP_PARAM_UNSET ""
> >
> > +/* Limit the batch size to limit per-CPU memory usage for dst buffers. */
> > +#define ZSWAP_MAX_BATCH_SIZE 8U
> > +
> > static int zswap_setup(void);
> >
> > /* Enable/disable zswap */
> > @@ -139,7 +142,7 @@ struct crypto_acomp_ctx {
> > struct crypto_acomp *acomp;
> > struct acomp_req *req;
> > struct crypto_wait wait;
> > - u8 *buffer;
> > + u8 **buffers;
> > struct mutex mutex;
> > bool is_sleepable;
> > };
> > @@ -158,6 +161,8 @@ struct zswap_pool {
> > struct work_struct release_work;
> > struct hlist_node node;
> > char tfm_name[CRYPTO_MAX_ALG_NAME];
> > + u8 compr_batch_size;
> > + u8 store_batch_size;
>
> I don't think we need to store store_batch_size, seems trivial to
> calculate at store time (perhaps in a helper).
>
> Taking a step back, is there any benefit to limiting store_batch_size to
> compr_batch_size? Is there a disadvantage to using
> ZSWAP_MAX_BATCH_SIZE
> even if it's higher than the HW compression batch size?
Thanks Yosry, for the code review comments. I had a discussion with
Barry earlier on these very same topics as follow up to his review comments
for v11, starting with [1]. Can you please go through the rationale for
these design choices, and let me know if you have any questions:
[1]: https://patchwork.kernel.org/comment/26530319/
>
> > };
> >
> > /* Global LRU lists shared by all zswap pools. */
> > @@ -252,8 +257,10 @@ static void __zswap_pool_empty(struct
> percpu_ref *ref);
> > * zswap_cpu_comp_prepare(), not others.
> > * - Cleanup acomp_ctx resources on all cores in zswap_pool_destroy().
> > */
> > -static void acomp_ctx_dealloc(struct crypto_acomp_ctx *acomp_ctx)
> > +static void acomp_ctx_dealloc(struct crypto_acomp_ctx *acomp_ctx, u8
> nr_buffers)
> > {
> > + u8 i;
> > +
> > if (IS_ERR_OR_NULL(acomp_ctx))
> > return;
> >
> > @@ -263,7 +270,11 @@ static void acomp_ctx_dealloc(struct
> crypto_acomp_ctx *acomp_ctx)
> > if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
> > crypto_free_acomp(acomp_ctx->acomp);
> >
> > - kfree(acomp_ctx->buffer);
> > + if (acomp_ctx->buffers) {
> > + for (i = 0; i < nr_buffers; ++i)
> > + kfree(acomp_ctx->buffers[i]);
> > + kfree(acomp_ctx->buffers);
> > + }
> > }
> >
> > static struct zswap_pool *zswap_pool_create(char *compressor)
> > @@ -275,6 +286,7 @@ static struct zswap_pool *zswap_pool_create(char
> *compressor)
> > if (!zswap_has_pool && !strcmp(compressor,
> ZSWAP_PARAM_UNSET))
> > return NULL;
> >
> > + /* Many things rely on the zero-initialization. */
> > pool = kzalloc(sizeof(*pool), GFP_KERNEL);
> > if (!pool)
> > return NULL;
> > @@ -334,13 +346,28 @@ static struct zswap_pool
> *zswap_pool_create(char *compressor)
> > goto ref_fail;
> > INIT_LIST_HEAD(&pool->list);
> >
> > + /*
> > + * Set the unit of compress batching for large folios, for quick
> > + * retrieval in the zswap_compress() fast path:
> > + * If the compressor is sequential (@pool->compr_batch_size is 1),
> > + * large folios will be compressed in batches of
> ZSWAP_MAX_BATCH_SIZE
> > + * pages, where each page in the batch is compressed sequentially.
> > + * We see better performance by processing the folio in batches of
> > + * ZSWAP_MAX_BATCH_SIZE, due to cache locality of working set
> > + * structures.
> > + */
> > + pool->store_batch_size = (pool->compr_batch_size > 1) ?
> > + pool->compr_batch_size :
> ZSWAP_MAX_BATCH_SIZE;
> > +
> > zswap_pool_debug("created", pool);
> >
> > return pool;
> >
> > ref_fail:
> > for_each_possible_cpu(cpu)
> > - acomp_ctx_dealloc(per_cpu_ptr(pool->acomp_ctx, cpu));
> > + acomp_ctx_dealloc(per_cpu_ptr(pool->acomp_ctx, cpu),
> > + pool->compr_batch_size);
> > +
> > error:
> > if (pool->acomp_ctx)
> > free_percpu(pool->acomp_ctx);
> > @@ -376,7 +403,8 @@ static void zswap_pool_destroy(struct zswap_pool
> *pool)
> > zswap_pool_debug("destroying", pool);
> >
> > for_each_possible_cpu(cpu)
> > - acomp_ctx_dealloc(per_cpu_ptr(pool->acomp_ctx, cpu));
> > + acomp_ctx_dealloc(per_cpu_ptr(pool->acomp_ctx, cpu),
> > + pool->compr_batch_size);
> >
> > free_percpu(pool->acomp_ctx);
> >
> > @@ -763,6 +791,24 @@ static void zswap_entry_cache_free(struct
> zswap_entry *entry)
> > kmem_cache_free(zswap_entry_cache, entry);
> > }
> >
> > +/*
> > + * Returns 0 if kmem_cache_alloc_bulk() failed and a positive number
> otherwise.
> > + * The code for __kmem_cache_alloc_bulk() indicates that this positive
> number
> > + * will be the @size requested, i.e., @nr_entries.
>
> The behavior is not documented tho, and other users seem like they don't
> all assume the return has to be 0 or nr_entries. Maybe we should add a
> WARN if the returned size is not nr_entries or 0?
Sure, I will do so.
>
> > + */
> > +static __always_inline int zswap_entries_cache_alloc_batch(void
> **entries,
> > + unsigned int
> nr_entries,
> > + gfp_t gfp)
> > +{
> > + return kmem_cache_alloc_bulk(zswap_entry_cache, gfp, nr_entries,
> entries);
>
> We currently use kmem_cache_alloc_node() in zswap_entry_cache_alloc() to
> allocate the entry on the same node as the compressed page. We use
> entry_to_nid() to get the node for LRU operations.
>
> This breaks that assumption.
You bring up a good point. I was looking at the code in slub.c and my
understanding thus far is that both, bulk allocations and kmem_cache_alloc_node()
allocations are made from a per-CPU "cpu_slab" that is allocated by SLUB.
IIUC, the concern you are raising is in the mainline, the entry is allocated on
the same node as the compressed page, and gets added to the LRU list of that
node. IOW, the node to which the compressed page belongs is the one to whose
LRU the entry will be added.
With this patch, with kmem_cache_alloc_bulk(), the entry will be created on
the per-CPU slab of the CPU on which zswap_store() is called and will be
added to the LRU of that per-CPU slab's NUMA node. Hence, the end result
could potentially be that the zswap_entry for a page could potentially be
on a different NUMA node/memcg than the page's NUMA node.
This is my thinking as to how this will impact the zswap shrinker:
1) memcg shrinker: if the memcg the entry ends up in is on the zswap_list_lru,
the entry will be written back.
2) Global shrinker: will cycle through all memcg's that have pages in the
zswap_list_lru, and the entry will be written back.
Based on this, it is not clear to me if there is a problem, and would like to
request you, Nhat and others to provide insights as well.
Interestingly, most of the code in slub.c has unlikely(!node_match(slab, node)).
Does this imply some higher level mm slab allocation requirements?
I am Ok with just calling zswap_entry_cache_alloc() for "nr_pages" if we
think this would be more correct.
>
> > +}
> > +
> > +static __always_inline void zswap_entries_cache_free_batch(void
> **entries,
> > + unsigned int
> nr_entries)
> > +{
> > + kmem_cache_free_bulk(zswap_entry_cache, nr_entries, entries);
> > +}
> > +
> > /*
> > * Carries out the common pattern of freeing an entry's zsmalloc allocation,
> > * freeing the entry itself, and decrementing the number of stored pages.
> > @@ -789,7 +835,9 @@ static int zswap_cpu_comp_prepare(unsigned int
> cpu, struct hlist_node *node)
> > {
> > struct zswap_pool *pool = hlist_entry(node, struct zswap_pool,
> node);
> > struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool-
> >acomp_ctx, cpu);
> > + int cpu_node = cpu_to_node(cpu);
>
> "nid" is a more common name.
Sure, I can change this.
>
> > int ret = -ENOMEM;
> > + u8 i;
> >
> > /*
> > * The per-CPU pool->acomp_ctx is zero-initialized on allocation.
> > @@ -802,11 +850,7 @@ static int zswap_cpu_comp_prepare(unsigned int
> cpu, struct hlist_node *node)
> > if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
> > return 0;
> >
> > - acomp_ctx->buffer = kmalloc_node(PAGE_SIZE, GFP_KERNEL,
> cpu_to_node(cpu));
> > - if (!acomp_ctx->buffer)
> > - return ret;
> > -
> > - acomp_ctx->acomp = crypto_alloc_acomp_node(pool->tfm_name, 0,
> 0, cpu_to_node(cpu));
> > + acomp_ctx->acomp = crypto_alloc_acomp_node(pool->tfm_name, 0,
> 0, cpu_node);
> > if (IS_ERR_OR_NULL(acomp_ctx->acomp)) {
> > pr_err("could not alloc crypto acomp %s : %ld\n",
> > pool->tfm_name, PTR_ERR(acomp_ctx-
> >acomp));
> > @@ -815,20 +859,40 @@ static int zswap_cpu_comp_prepare(unsigned int
> cpu, struct hlist_node *node)
> > }
> > acomp_ctx->is_sleepable = acomp_is_async(acomp_ctx->acomp);
> >
> > + /*
> > + * Allocate up to ZSWAP_MAX_BATCH_SIZE dst buffers if the
> > + * compressor supports batching.
> > + */
> > + pool->compr_batch_size = min(ZSWAP_MAX_BATCH_SIZE,
> > + crypto_acomp_batch_size(acomp_ctx-
> >acomp));
> > +
> > acomp_ctx->req = acomp_request_alloc(acomp_ctx->acomp);
> > +
> > if (IS_ERR_OR_NULL(acomp_ctx->req)) {
> > pr_err("could not alloc crypto acomp_request %s\n",
> > - pool->tfm_name);
> > + pool->tfm_name);
>
> Unrelated change.
Ok, will submit this separately.
>
> > goto fail;
> > }
> >
> > - crypto_init_wait(&acomp_ctx->wait);
> > + acomp_ctx->buffers = kcalloc_node(pool->compr_batch_size,
> sizeof(u8 *),
> > + GFP_KERNEL, cpu_node);
> > + if (!acomp_ctx->buffers)
> > + goto fail;
> > +
> > + for (i = 0; i < pool->compr_batch_size; ++i) {
> > + acomp_ctx->buffers[i] = kmalloc_node(PAGE_SIZE,
> GFP_KERNEL,
> > + cpu_node);
> > + if (!acomp_ctx->buffers[i])
> > + goto fail;
> > + }
> >
> > /*
> > * if the backend of acomp is async zip, crypto_req_done() will
> wakeup
> > * crypto_wait_req(); if the backend of acomp is scomp, the callback
> > * won't be called, crypto_wait_req() will return without blocking.
> > */
> > + crypto_init_wait(&acomp_ctx->wait);
> > +
> > acomp_request_set_callback(acomp_ctx->req,
> CRYPTO_TFM_REQ_MAY_BACKLOG,
> > crypto_req_done, &acomp_ctx->wait);
> >
> > @@ -836,12 +900,12 @@ static int zswap_cpu_comp_prepare(unsigned int
> cpu, struct hlist_node *node)
> > return 0;
> >
> > fail:
> > - acomp_ctx_dealloc(acomp_ctx);
> > + acomp_ctx_dealloc(acomp_ctx, pool->compr_batch_size);
> > return ret;
> > }
> >
> > static bool zswap_compress(struct page *page, struct zswap_entry *entry,
> > - struct zswap_pool *pool)
> > + struct zswap_pool *pool, bool folio_wb)
>
> Maybe "wb_enabled" instead of folio_wb?
Ok.
>
> > {
> > struct crypto_acomp_ctx *acomp_ctx;
> > struct scatterlist input, output;
> > @@ -855,7 +919,7 @@ static bool zswap_compress(struct page *page,
> struct zswap_entry *entry,
> > acomp_ctx = raw_cpu_ptr(pool->acomp_ctx);
> > mutex_lock(&acomp_ctx->mutex);
> >
> > - dst = acomp_ctx->buffer;
> > + dst = acomp_ctx->buffers[0];
> > sg_init_table(&input, 1);
> > sg_set_page(&input, page, PAGE_SIZE, 0);
> >
> > @@ -886,13 +950,11 @@ static bool zswap_compress(struct page *page,
> struct zswap_entry *entry,
> > */
> > if (comp_ret || !dlen || dlen >= PAGE_SIZE) {
> > dlen = PAGE_SIZE;
> > - if (!mem_cgroup_zswap_writeback_enabled(
> > - folio_memcg(page_folio(page)))) {
> > + if (!folio_wb) {
> > comp_ret = comp_ret ? comp_ret : -EINVAL;
> > goto unlock;
> > }
> > comp_ret = 0;
> > - dlen = PAGE_SIZE;
>
> Unrelated change.
Will submit a separate patch.
>
> > dst = kmap_local_page(page);
> > mapped = true;
> > }
> > @@ -932,7 +994,7 @@ static bool zswap_decompress(struct zswap_entry
> *entry, struct folio *folio)
> >
> > acomp_ctx = raw_cpu_ptr(pool->acomp_ctx);
> > mutex_lock(&acomp_ctx->mutex);
> > - obj = zs_obj_read_begin(pool->zs_pool, entry->handle, acomp_ctx-
> >buffer);
> > + obj = zs_obj_read_begin(pool->zs_pool, entry->handle, acomp_ctx-
> >buffers[0]);
> >
> > /* zswap entries of length PAGE_SIZE are not compressed. */
> > if (entry->length == PAGE_SIZE) {
> > @@ -942,15 +1004,15 @@ static bool zswap_decompress(struct
> zswap_entry *entry, struct folio *folio)
> >
> > /*
> > * zs_obj_read_begin() might return a kmap address of highmem
> when
> > - * acomp_ctx->buffer is not used. However, sg_init_one() does not
> > - * handle highmem addresses, so copy the object to acomp_ctx-
> >buffer.
> > + * acomp_ctx->buffers[0] is not used. However, sg_init_one() does
> not
> > + * handle highmem addresses, so copy the object to acomp_ctx-
> >buffers[0].
> > */
> > if (virt_addr_valid(obj)) {
> > src = obj;
> > } else {
> > - WARN_ON_ONCE(obj == acomp_ctx->buffer);
> > - memcpy(acomp_ctx->buffer, obj, entry->length);
> > - src = acomp_ctx->buffer;
> > + WARN_ON_ONCE(obj == acomp_ctx->buffers[0]);
> > + memcpy(acomp_ctx->buffers[0], obj, entry->length);
> > + src = acomp_ctx->buffers[0];
> > }
> >
> > sg_init_one(&input, src, entry->length);
> > @@ -1404,95 +1466,160 @@ static void shrink_worker(struct work_struct
> *w)
> > * main API
> > **********************************/
> >
> > -static bool zswap_store_page(struct page *page,
> > - struct obj_cgroup *objcg,
> > - struct zswap_pool *pool)
> > +/*
> > + * Store multiple pages in @folio, starting from the page at index @start up
> to
> > + * the page at index @end-1.
> > + */
> > +static bool zswap_store_pages(struct folio *folio,
> > + long start,
> > + long end,
> > + struct obj_cgroup *objcg,
> > + struct zswap_pool *pool,
> > + int node_id,
> > + bool folio_wb)
> > {
> > - swp_entry_t page_swpentry = page_swap_entry(page);
> > - struct zswap_entry *entry, *old;
> > -
> > - /* allocate entry */
> > - entry = zswap_entry_cache_alloc(GFP_KERNEL, page_to_nid(page));
> > - if (!entry) {
> > - zswap_reject_kmemcache_fail++;
> > - return false;
> > + struct zswap_entry *entries[ZSWAP_MAX_BATCH_SIZE];
> > + u8 i, store_fail_idx = 0, nr_pages = end - start;
> > +
> > + VM_WARN_ON_ONCE(nr_pages > ZSWAP_MAX_BATCH_SIZE);
> > +
> > + if (unlikely(!zswap_entries_cache_alloc_batch((void **)&entries[0],
> > + nr_pages, GFP_KERNEL)))
> {
> > + for (i = 0; i < nr_pages; ++i) {
> > + entries[i] = zswap_entry_cache_alloc(GFP_KERNEL,
> node_id);
> > +
> > + if (unlikely(!entries[i])) {
> > + zswap_reject_kmemcache_fail++;
> > + /*
> > + * While handling this error, we only need to
> > + * call zswap_entries_cache_free_batch() for
> > + * entries[0 .. i-1].
> > + */
> > + nr_pages = i;
> > + goto store_pages_failed;
> > + }
>
> Is it okay to use kmem_cache_free_bulk() to free slab objects that were
> not allocated with kmem_cache_alloc_bulk()?
I recall verifying that it should be Ok, but can check again.
>
> > + }
> > }
> >
> > - if (!zswap_compress(page, entry, pool))
> > - goto compress_failed;
> > + /*
> > + * Three sets of initializations are done to minimize bringing
> > + * @entries into the cache for writing at different parts of this
> > + * procedure, since doing so regresses performance:
> > + *
> > + * 1) Do all the writes to each entry in one code block. These
> > + * writes need to be done anyway upon success which is more likely
> > + * than not.
> > + *
> > + * 2) Initialize the handle to an error value. This facilitates
> > + * having a consolidated failure handling
> > + * 'goto store_pages_failed' that can inspect the value of the
> > + * handle to determine whether zsmalloc memory needs to be
> > + * de-allocated.
> > + *
> > + * 3) The page_swap_entry() is obtained once and stored in the entry.
> > + * Subsequent store in xarray gets the entry->swpentry instead of
> > + * calling page_swap_entry(), minimizing computes.
> > + */
>
> Very long comment, and I am not sure what it is trying to say. We don't
> need to describe what the code is doing like that.
>
> The only thing that may be worth pointing out is that we are colocating
> initialization as much as possible here to minimize potential cache
> misses.
Sounds good.
>
> Does it actually matter if we do the initializations here vs. right
> before inserting to the LRU (current behavior)?
Yes, it impacts batching performance with software quite a bit.
>
> > + for (i = 0; i < nr_pages; ++i) {
> > + entries[i]->handle = (unsigned long)ERR_PTR(-EINVAL);
> > + entries[i]->pool = pool;
> > + entries[i]->swpentry = page_swap_entry(folio_page(folio,
> start + i));
> > + entries[i]->objcg = objcg;
> > + entries[i]->referenced = true;
> > + INIT_LIST_HEAD(&entries[i]->lru);
> > + }
> >
> > - old = xa_store(swap_zswap_tree(page_swpentry),
> > - swp_offset(page_swpentry),
> > - entry, GFP_KERNEL);
> > - if (xa_is_err(old)) {
> > - int err = xa_err(old);
> > + for (i = 0; i < nr_pages; ++i) {
> > + struct page *page = folio_page(folio, start + i);
> >
> > - WARN_ONCE(err != -ENOMEM, "unexpected xarray error:
> %d\n", err);
> > - zswap_reject_alloc_fail++;
> > - goto store_failed;
> > + if (!zswap_compress(page, entries[i], pool, folio_wb))
> > + goto store_pages_failed;
> > }
> >
> > - /*
> > - * We may have had an existing entry that became stale when
> > - * the folio was redirtied and now the new version is being
> > - * swapped out. Get rid of the old.
> > - */
> > - if (old)
> > - zswap_entry_free(old);
> > + for (i = 0; i < nr_pages; ++i) {
> > + struct zswap_entry *old, *entry = entries[i];
> >
> > - /*
> > - * The entry is successfully compressed and stored in the tree, there is
> > - * no further possibility of failure. Grab refs to the pool and objcg,
> > - * charge zswap memory, and increment zswap_stored_pages.
> > - * The opposite actions will be performed by zswap_entry_free()
> > - * when the entry is removed from the tree.
> > - */
> > - zswap_pool_get(pool);
> > - if (objcg) {
> > - obj_cgroup_get(objcg);
> > - obj_cgroup_charge_zswap(objcg, entry->length);
> > - }
> > - atomic_long_inc(&zswap_stored_pages);
> > - if (entry->length == PAGE_SIZE)
> > - atomic_long_inc(&zswap_stored_incompressible_pages);
> > + old = xa_store(swap_zswap_tree(entry->swpentry),
> > + swp_offset(entry->swpentry),
> > + entry, GFP_KERNEL);
> > + if (unlikely(xa_is_err(old))) {
> > + int err = xa_err(old);
> >
> > - /*
> > - * We finish initializing the entry while it's already in xarray.
> > - * This is safe because:
> > - *
> > - * 1. Concurrent stores and invalidations are excluded by folio lock.
> > - *
> > - * 2. Writeback is excluded by the entry not being on the LRU yet.
> > - * The publishing order matters to prevent writeback from seeing
> > - * an incoherent entry.
> > - */
> > - entry->pool = pool;
> > - entry->swpentry = page_swpentry;
> > - entry->objcg = objcg;
> > - entry->referenced = true;
> > - if (entry->length) {
> > - INIT_LIST_HEAD(&entry->lru);
> > - zswap_lru_add(&zswap_list_lru, entry);
> > + WARN_ONCE(err != -ENOMEM, "unexpected xarray
> error: %d\n", err);
> > + zswap_reject_alloc_fail++;
> > + /*
> > + * Entries up to this point have been stored in the
> > + * xarray. zswap_store() will erase them from the
> xarray
> > + * and call zswap_entry_free(). Local cleanup in
> > + * 'store_pages_failed' only needs to happen for
> > + * entries from [@i to @nr_pages).
> > + */
> > + store_fail_idx = i;
> > + goto store_pages_failed;
> > + }
> > +
> > + /*
> > + * We may have had an existing entry that became stale when
> > + * the folio was redirtied and now the new version is being
> > + * swapped out. Get rid of the old.
> > + */
> > + if (unlikely(old))
> > + zswap_entry_free(old);
> > +
> > + /*
> > + * The entry is successfully compressed and stored in the tree,
> there is
> > + * no further possibility of failure. Grab refs to the pool and
> objcg,
> > + * charge zswap memory, and increment
> zswap_stored_pages.
> > + * The opposite actions will be performed by
> zswap_entry_free()
> > + * when the entry is removed from the tree.
> > + */
>
> But there *is* further possibility of failure if a subsequent entry
> xa_store() fails, right?
This comment is referring to this specific entry.
>
> Seems like if xa_store() fails we do not cleanup previously charged
> objects, pool references, zswap_stored_pages, etc. Instead of rolling
> all this back on failure, can we do all the xarray stores first and only
> do the rest when we're at a point where no failure can happen? Would
> that cause a performance regression?
It would make the code more complicated and thus cause a performance
regression. I have tried to balance code simplicity (which impacts performance)
with correctness here. The "store_failed_idx" ensures that all partial work done
in zswap_store_pages() for this batch, is cleaned up.
The rest is handled in zswap_store() when it sees an error returned by
zswap_store_pages().
>
> > + zswap_pool_get(pool);
> > + if (objcg) {
> > + obj_cgroup_get(objcg);
> > + obj_cgroup_charge_zswap(objcg, entry->length);
> > + }
> > + atomic_long_inc(&zswap_stored_pages);
> > + if (entry->length == PAGE_SIZE)
> > +
> atomic_long_inc(&zswap_stored_incompressible_pages);
> > +
> > + /*
> > + * We finish by adding the entry to the LRU while it's already
> > + * in xarray. This is safe because:
> > + *
> > + * 1. Concurrent stores and invalidations are excluded by folio
> lock.
> > + *
> > + * 2. Writeback is excluded by the entry not being on the LRU
> yet.
> > + * The publishing order matters to prevent writeback from
> seeing
> > + * an incoherent entry.
> > + */
> > + if (likely(entry->length))
> > + zswap_lru_add(&zswap_list_lru, entry);
> > }
> >
> > return true;
> >
> > -store_failed:
> > - zs_free(pool->zs_pool, entry->handle);
> > -compress_failed:
> > - zswap_entry_cache_free(entry);
> > +store_pages_failed:
> > + for (i = store_fail_idx; i < nr_pages; ++i) {
> > + if (!IS_ERR_VALUE(entries[i]->handle))
> > + zs_free(pool->zs_pool, entries[i]->handle);
> > + }
> > + zswap_entries_cache_free_batch((void **)&entries[store_fail_idx],
> > + nr_pages - store_fail_idx);
> > +
> > return false;
> > }
> >
> > bool zswap_store(struct folio *folio)
> > {
> > + bool folio_wb =
> mem_cgroup_zswap_writeback_enabled(folio_memcg(folio));
>
> Ditto renaming folio_wb.
Yes.
>
> > long nr_pages = folio_nr_pages(folio);
> > + int node_id = folio_nid(folio);
>
> Ditto nid.
Ok.
>
> > swp_entry_t swp = folio->swap;
> > struct obj_cgroup *objcg = NULL;
> > struct mem_cgroup *memcg = NULL;
> > struct zswap_pool *pool;
> > bool ret = false;
> > - long index;
> > + long start, end;
> >
> > VM_WARN_ON_ONCE(!folio_test_locked(folio));
> > VM_WARN_ON_ONCE(!folio_test_swapcache(folio));
> > @@ -1526,10 +1653,12 @@ bool zswap_store(struct folio *folio)
> > mem_cgroup_put(memcg);
> > }
> >
> > - for (index = 0; index < nr_pages; ++index) {
> > - struct page *page = folio_page(folio, index);
> > + /* Store the folio in batches of @pool->store_batch_size pages. */
> > + for (start = 0; start < nr_pages; start += pool->store_batch_size) {
> > + end = min(start + pool->store_batch_size, nr_pages);
> >
> > - if (!zswap_store_page(page, objcg, pool))
> > + if (!zswap_store_pages(folio, start, end, objcg, pool,
> > + node_id, folio_wb))
> > goto put_pool;
> > }
> >
> > @@ -1559,9 +1688,9 @@ bool zswap_store(struct folio *folio)
> > struct zswap_entry *entry;
> > struct xarray *tree;
> >
> > - for (index = 0; index < nr_pages; ++index) {
> > - tree = swap_zswap_tree(swp_entry(type, offset +
> index));
> > - entry = xa_erase(tree, offset + index);
> > + for (start = 0; start < nr_pages; ++start) {
> > + tree = swap_zswap_tree(swp_entry(type, offset +
> start));
> > + entry = xa_erase(tree, offset + start);
> > if (entry)
> > zswap_entry_free(entry);
> > }
> > --
> > 2.27.0
> >
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches.
2025-10-01 21:20 ` Sridhar, Kanchana P
@ 2025-10-03 19:10 ` Sridhar, Kanchana P
2025-10-13 17:47 ` Sridhar, Kanchana P
2025-10-13 17:58 ` Sridhar, Kanchana P
2025-10-14 15:29 ` Yosry Ahmed
2 siblings, 1 reply; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-10-03 19:10 UTC (permalink / raw)
To: Yosry Ahmed
Cc: linux-kernel, linux-mm, hannes, nphamcs, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius,
Feghali, Wajdi K, Gopal, Vinodh, Sridhar, Kanchana P
> > -----Original Message-----
> > From: Yosry Ahmed <yosry.ahmed@linux.dev>
> > Sent: Wednesday, October 1, 2025 9:19 AM
> > To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> > Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> > hannes@cmpxchg.org; nphamcs@gmail.com; chengming.zhou@linux.dev;
> > usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> > ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> > senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> > crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> > davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> > ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> > <kristen.c.accardi@intel.com>; Gomes, Vinicius
> <vinicius.gomes@intel.com>;
> > Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> > <vinodh.gopal@intel.com>
> > Subject: Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a
> > large folio in batches.
[...]
> >
> > > + */
> > > +static __always_inline int zswap_entries_cache_alloc_batch(void
> > **entries,
> > > + unsigned int
> > nr_entries,
> > > + gfp_t gfp)
> > > +{
> > > + return kmem_cache_alloc_bulk(zswap_entry_cache, gfp, nr_entries,
> > entries);
> >
> > We currently use kmem_cache_alloc_node() in zswap_entry_cache_alloc()
> to
> > allocate the entry on the same node as the compressed page. We use
> > entry_to_nid() to get the node for LRU operations.
> >
> > This breaks that assumption.
>
> You bring up a good point. I was looking at the code in slub.c and my
> understanding thus far is that both, bulk allocations and
> kmem_cache_alloc_node()
> allocations are made from a per-CPU "cpu_slab" that is allocated by SLUB.
>
> IIUC, the concern you are raising is in the mainline, the entry is allocated on
> the same node as the compressed page, and gets added to the LRU list of that
> node. IOW, the node to which the compressed page belongs is the one to
> whose
> LRU the entry will be added.
>
> With this patch, with kmem_cache_alloc_bulk(), the entry will be created on
> the per-CPU slab of the CPU on which zswap_store() is called and will be
> added to the LRU of that per-CPU slab's NUMA node. Hence, the end result
> could potentially be that the zswap_entry for a page could potentially be
> on a different NUMA node/memcg than the page's NUMA node.
>
> This is my thinking as to how this will impact the zswap shrinker:
>
> 1) memcg shrinker: if the memcg the entry ends up in is on the zswap_list_lru,
> the entry will be written back.
> 2) Global shrinker: will cycle through all memcg's that have pages in the
> zswap_list_lru, and the entry will be written back.
>
> Based on this, it is not clear to me if there is a problem, and would like to
> request you, Nhat and others to provide insights as well.
>
> Interestingly, most of the code in slub.c has unlikely(!node_match(slab,
> node)).
> Does this imply some higher level mm slab allocation requirements?
>
> I am Ok with just calling zswap_entry_cache_alloc() for "nr_pages" if we
> think this would be more correct.
I wanted to share some updates and summarize my understanding from
looking some more at slub.c. The "zswap_entry_cache" kmem_cache
has slab memory created for each node. The main change with v12 is:
before:
The page being compressed and its zswap_entry are in the same memcg,
thus, implicitly on the same slab node.
v12:
The page being compressed and its zswap_entry could be in different
memcgs if the process that owns the page gets migrated to a different
node.
Impact to zswap LRU list vis-à-vis the memcg shrinker:
before:
The original NUMA node memcg [hierarchy] would need to face memory
pressure for the page's zswap_entry to be written back. Hence it is possible
that the node on which the process is currently running may not see the
benefit of memory being freed up.
v12:
The memcg whose node on which the process was running when the page
was compressed would have to face memory pressure for the page's zswap_entry
to be written back. This node will see the benefit of memory being freed up due
to writeback. If the process has migrated to a different node than the one
on which it was running when the page was compressed, we have the same issue
as "before".
Is my understanding correct? Please let me know if I am missing something.
The bulk allocation does improve batching performance, especially with 64K folios:
kernel_compile, 64K folios, deflate-iaa:
========================================
mm-unstable-9-18-2025 v12 without bulk alloc/free
------------------------------------------------------------------------------
real_sec 874.74 821.59 870.82
sys_sec 3,834.35 3,791.12 3,794.06
------------------------------------------------------------------------------
kernel_compile, 64K folios, zstd:
=================================
mm-unstable-9-18-2025 v12 without bulk alloc/free
------------------------------------------------------------------------------
real_sec 925.08 853.14 899.40
sys_sec 5,318.65 5,172.23 5,415.20
------------------------------------------------------------------------------
kernel_compile, PMD folios, deflate-iaa:
========================================
mm-unstable-9-18-2025 v12 without bulk alloc/free
------------------------------------------------------------------------------
real_sec 808.10 794.85 809.33
sys_sec 4,351.01 4,266.95 4,169.07
------------------------------------------------------------------------------
kernel_compile, PMD folios, zstd:
=================================
mm-unstable-9-18-2025 v12 without bulk alloc/free
------------------------------------------------------------------------------
real_sec 848.06 845.42 836.06
sys_sec 5,898.58 5,741.31 5,795.75
------------------------------------------------------------------------------
Based on this, and if my understanding of the v12 change's impact on the
zswap shrinker is correct, I am not sure if the change in the assumption
is necessarily a concern.
>
> >
> > > +}
> > > +
> > > +static __always_inline void zswap_entries_cache_free_batch(void
> > **entries,
> > > + unsigned int
> > nr_entries)
> > > +{
> > > + kmem_cache_free_bulk(zswap_entry_cache, nr_entries, entries);
> > > +}
[...]
> > > +static bool zswap_store_pages(struct folio *folio,
> > > + long start,
> > > + long end,
> > > + struct obj_cgroup *objcg,
> > > + struct zswap_pool *pool,
> > > + int node_id,
> > > + bool folio_wb)
> > > {
> > > - swp_entry_t page_swpentry = page_swap_entry(page);
> > > - struct zswap_entry *entry, *old;
> > > -
> > > - /* allocate entry */
> > > - entry = zswap_entry_cache_alloc(GFP_KERNEL, page_to_nid(page));
> > > - if (!entry) {
> > > - zswap_reject_kmemcache_fail++;
> > > - return false;
> > > + struct zswap_entry *entries[ZSWAP_MAX_BATCH_SIZE];
> > > + u8 i, store_fail_idx = 0, nr_pages = end - start;
> > > +
> > > + VM_WARN_ON_ONCE(nr_pages > ZSWAP_MAX_BATCH_SIZE);
> > > +
> > > + if (unlikely(!zswap_entries_cache_alloc_batch((void **)&entries[0],
> > > + nr_pages, GFP_KERNEL)))
> > {
> > > + for (i = 0; i < nr_pages; ++i) {
> > > + entries[i] = zswap_entry_cache_alloc(GFP_KERNEL,
> > node_id);
> > > +
> > > + if (unlikely(!entries[i])) {
> > > + zswap_reject_kmemcache_fail++;
> > > + /*
> > > + * While handling this error, we only need to
> > > + * call zswap_entries_cache_free_batch() for
> > > + * entries[0 .. i-1].
> > > + */
> > > + nr_pages = i;
> > > + goto store_pages_failed;
> > > + }
> >
> > Is it okay to use kmem_cache_free_bulk() to free slab objects that were
> > not allocated with kmem_cache_alloc_bulk()?
>
> I recall verifying that it should be Ok, but can check again.
>
I verified the code again, and yes, it is Ok for slab objects allocated by
either kmem_cache_alloc_bulk() or kmem_cache_alloc_node() to be
freed by calling kmem_cache_free_bulk(). kmem_cache_free_bulk()
opportunistically looks to create freelists from the list of slab objects
to then "bulk transfer" them to the slab's freelists, with optimizations
in synchronization as compared to calling kmem_cache_free().
I verified this with usemem30 with the following code change in
zswap_store_pages(), and saw no issues with deflate-iaa/zstd:
for each of nr_pages: kmem_cache_alloc_node()
kmem_cache_free_bulk()
kmem_cache_alloc_bulk()
kmem_cache_free_bulk()
kmem_cache_alloc _bulk()
[proceed]
Yosry, please let me know how I should proceed and if there are
other experiments that I can run to verify this change to bulk alloc/free
to address any concerns.
Thanks,
Kanchana
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches.
2025-10-03 19:10 ` Sridhar, Kanchana P
@ 2025-10-13 17:47 ` Sridhar, Kanchana P
0 siblings, 0 replies; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-10-13 17:47 UTC (permalink / raw)
To: Yosry Ahmed
Cc: linux-kernel, linux-mm, hannes, nphamcs, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius,
Feghali, Wajdi K, Gopal, Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> Sent: Friday, October 3, 2025 12:11 PM
> To: Yosry Ahmed <yosry.ahmed@linux.dev>
> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> hannes@cmpxchg.org; nphamcs@gmail.com; chengming.zhou@linux.dev;
> usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>;
> Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> <vinodh.gopal@intel.com>; Sridhar, Kanchana P
> <kanchana.p.sridhar@intel.com>
> Subject: RE: [PATCH v12 22/23] mm: zswap: zswap_store() will process a
> large folio in batches.
>
>
> > > -----Original Message-----
> > > From: Yosry Ahmed <yosry.ahmed@linux.dev>
> > > Sent: Wednesday, October 1, 2025 9:19 AM
> > > To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> > > Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> > > hannes@cmpxchg.org; nphamcs@gmail.com;
> chengming.zhou@linux.dev;
> > > usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> > > ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> > > senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> > > crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> > > davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> > > ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> > > <kristen.c.accardi@intel.com>; Gomes, Vinicius
> > <vinicius.gomes@intel.com>;
> > > Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> > > <vinodh.gopal@intel.com>
> > > Subject: Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a
> > > large folio in batches.
> [...]
> > >
> > > > + */
> > > > +static __always_inline int zswap_entries_cache_alloc_batch(void
> > > **entries,
> > > > + unsigned int
> > > nr_entries,
> > > > + gfp_t gfp)
> > > > +{
> > > > + return kmem_cache_alloc_bulk(zswap_entry_cache, gfp, nr_entries,
> > > entries);
> > >
> > > We currently use kmem_cache_alloc_node() in
> zswap_entry_cache_alloc()
> > to
> > > allocate the entry on the same node as the compressed page. We use
> > > entry_to_nid() to get the node for LRU operations.
> > >
> > > This breaks that assumption.
> >
> > You bring up a good point. I was looking at the code in slub.c and my
> > understanding thus far is that both, bulk allocations and
> > kmem_cache_alloc_node()
> > allocations are made from a per-CPU "cpu_slab" that is allocated by SLUB.
> >
> > IIUC, the concern you are raising is in the mainline, the entry is allocated on
> > the same node as the compressed page, and gets added to the LRU list of
> that
> > node. IOW, the node to which the compressed page belongs is the one to
> > whose
> > LRU the entry will be added.
> >
> > With this patch, with kmem_cache_alloc_bulk(), the entry will be created on
> > the per-CPU slab of the CPU on which zswap_store() is called and will be
> > added to the LRU of that per-CPU slab's NUMA node. Hence, the end result
> > could potentially be that the zswap_entry for a page could potentially be
> > on a different NUMA node/memcg than the page's NUMA node.
> >
> > This is my thinking as to how this will impact the zswap shrinker:
> >
> > 1) memcg shrinker: if the memcg the entry ends up in is on the
> zswap_list_lru,
> > the entry will be written back.
> > 2) Global shrinker: will cycle through all memcg's that have pages in the
> > zswap_list_lru, and the entry will be written back.
> >
> > Based on this, it is not clear to me if there is a problem, and would like to
> > request you, Nhat and others to provide insights as well.
> >
> > Interestingly, most of the code in slub.c has unlikely(!node_match(slab,
> > node)).
> > Does this imply some higher level mm slab allocation requirements?
> >
> > I am Ok with just calling zswap_entry_cache_alloc() for "nr_pages" if we
> > think this would be more correct.
>
> I wanted to share some updates and summarize my understanding from
> looking some more at slub.c. The "zswap_entry_cache" kmem_cache
> has slab memory created for each node. The main change with v12 is:
>
> before:
> The page being compressed and its zswap_entry are in the same memcg,
> thus, implicitly on the same slab node.
>
> v12:
> The page being compressed and its zswap_entry could be in different
> memcgs if the process that owns the page gets migrated to a different
> node.
>
> Impact to zswap LRU list vis-à-vis the memcg shrinker:
>
> before:
> The original NUMA node memcg [hierarchy] would need to face memory
> pressure for the page's zswap_entry to be written back. Hence it is possible
> that the node on which the process is currently running may not see the
> benefit of memory being freed up.
>
> v12:
> The memcg whose node on which the process was running when the page
> was compressed would have to face memory pressure for the page's
> zswap_entry
> to be written back. This node will see the benefit of memory being freed up
> due
> to writeback. If the process has migrated to a different node than the one
> on which it was running when the page was compressed, we have the same
> issue
> as "before".
>
> Is my understanding correct? Please let me know if I am missing something.
>
> The bulk allocation does improve batching performance, especially with 64K
> folios:
>
> kernel_compile, 64K folios, deflate-iaa:
> ========================================
> mm-unstable-9-18-2025 v12 without bulk alloc/free
> ------------------------------------------------------------------------------
> real_sec 874.74 821.59 870.82
> sys_sec 3,834.35 3,791.12 3,794.06
> ------------------------------------------------------------------------------
>
> kernel_compile, 64K folios, zstd:
> =================================
> mm-unstable-9-18-2025 v12 without bulk alloc/free
> ------------------------------------------------------------------------------
> real_sec 925.08 853.14 899.40
> sys_sec 5,318.65 5,172.23 5,415.20
> ------------------------------------------------------------------------------
>
> kernel_compile, PMD folios, deflate-iaa:
> ========================================
> mm-unstable-9-18-2025 v12 without bulk alloc/free
> ------------------------------------------------------------------------------
> real_sec 808.10 794.85 809.33
> sys_sec 4,351.01 4,266.95 4,169.07
> ------------------------------------------------------------------------------
>
> kernel_compile, PMD folios, zstd:
> =================================
> mm-unstable-9-18-2025 v12 without bulk alloc/free
> ------------------------------------------------------------------------------
> real_sec 848.06 845.42 836.06
> sys_sec 5,898.58 5,741.31 5,795.75
> ------------------------------------------------------------------------------
>
> Based on this, and if my understanding of the v12 change's impact on the
> zswap shrinker is correct, I am not sure if the change in the assumption
> is necessarily a concern.
>
> >
> > >
> > > > +}
> > > > +
> > > > +static __always_inline void zswap_entries_cache_free_batch(void
> > > **entries,
> > > > + unsigned int
> > > nr_entries)
> > > > +{
> > > > + kmem_cache_free_bulk(zswap_entry_cache, nr_entries, entries);
> > > > +}
> [...]
> > > > +static bool zswap_store_pages(struct folio *folio,
> > > > + long start,
> > > > + long end,
> > > > + struct obj_cgroup *objcg,
> > > > + struct zswap_pool *pool,
> > > > + int node_id,
> > > > + bool folio_wb)
> > > > {
> > > > - swp_entry_t page_swpentry = page_swap_entry(page);
> > > > - struct zswap_entry *entry, *old;
> > > > -
> > > > - /* allocate entry */
> > > > - entry = zswap_entry_cache_alloc(GFP_KERNEL, page_to_nid(page));
> > > > - if (!entry) {
> > > > - zswap_reject_kmemcache_fail++;
> > > > - return false;
> > > > + struct zswap_entry *entries[ZSWAP_MAX_BATCH_SIZE];
> > > > + u8 i, store_fail_idx = 0, nr_pages = end - start;
> > > > +
> > > > + VM_WARN_ON_ONCE(nr_pages > ZSWAP_MAX_BATCH_SIZE);
> > > > +
> > > > + if (unlikely(!zswap_entries_cache_alloc_batch((void **)&entries[0],
> > > > + nr_pages, GFP_KERNEL)))
> > > {
> > > > + for (i = 0; i < nr_pages; ++i) {
> > > > + entries[i] = zswap_entry_cache_alloc(GFP_KERNEL,
> > > node_id);
> > > > +
> > > > + if (unlikely(!entries[i])) {
> > > > + zswap_reject_kmemcache_fail++;
> > > > + /*
> > > > + * While handling this error, we only need to
> > > > + * call zswap_entries_cache_free_batch() for
> > > > + * entries[0 .. i-1].
> > > > + */
> > > > + nr_pages = i;
> > > > + goto store_pages_failed;
> > > > + }
> > >
> > > Is it okay to use kmem_cache_free_bulk() to free slab objects that were
> > > not allocated with kmem_cache_alloc_bulk()?
> >
> > I recall verifying that it should be Ok, but can check again.
> >
>
> I verified the code again, and yes, it is Ok for slab objects allocated by
> either kmem_cache_alloc_bulk() or kmem_cache_alloc_node() to be
> freed by calling kmem_cache_free_bulk(). kmem_cache_free_bulk()
> opportunistically looks to create freelists from the list of slab objects
> to then "bulk transfer" them to the slab's freelists, with optimizations
> in synchronization as compared to calling kmem_cache_free().
>
> I verified this with usemem30 with the following code change in
> zswap_store_pages(), and saw no issues with deflate-iaa/zstd:
>
> for each of nr_pages: kmem_cache_alloc_node()
>
> kmem_cache_free_bulk()
>
> kmem_cache_alloc_bulk()
>
> kmem_cache_free_bulk()
>
> kmem_cache_alloc _bulk()
>
> [proceed]
>
> Yosry, please let me know how I should proceed and if there are
> other experiments that I can run to verify this change to bulk alloc/free
> to address any concerns.
Hi Yosry,
Just wanted to check in to get your suggestions on next steps as I start
working on v13.
Thanks for taking the time to provide code review comments!
Best regards,
Kanchana
>
> Thanks,
> Kanchana
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches.
2025-10-01 21:20 ` Sridhar, Kanchana P
2025-10-03 19:10 ` Sridhar, Kanchana P
@ 2025-10-13 17:58 ` Sridhar, Kanchana P
2025-10-14 15:29 ` Yosry Ahmed
2 siblings, 0 replies; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-10-13 17:58 UTC (permalink / raw)
To: Yosry Ahmed
Cc: linux-kernel, linux-mm, hannes, nphamcs, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius,
Feghali, Wajdi K, Gopal, Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> Sent: Wednesday, October 1, 2025 2:21 PM
> To: Yosry Ahmed <yosry.ahmed@linux.dev>
> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> hannes@cmpxchg.org; nphamcs@gmail.com; chengming.zhou@linux.dev;
> usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>;
> Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> <vinodh.gopal@intel.com>; Sridhar, Kanchana P
> <kanchana.p.sridhar@intel.com>
> Subject: RE: [PATCH v12 22/23] mm: zswap: zswap_store() will process a
> large folio in batches.
>
>
> > -----Original Message-----
> > From: Yosry Ahmed <yosry.ahmed@linux.dev>
> > Sent: Wednesday, October 1, 2025 9:19 AM
> > To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> > Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> > hannes@cmpxchg.org; nphamcs@gmail.com; chengming.zhou@linux.dev;
> > usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> > ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> > senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> > crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> > davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> > ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> > <kristen.c.accardi@intel.com>; Gomes, Vinicius
> <vinicius.gomes@intel.com>;
> > Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> > <vinodh.gopal@intel.com>
> > Subject: Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a
> > large folio in batches.
> >
> > On Thu, Sep 25, 2025 at 08:35:01PM -0700, Kanchana P Sridhar wrote:
[...]
> > > @@ -158,6 +161,8 @@ struct zswap_pool {
> > > struct work_struct release_work;
> > > struct hlist_node node;
> > > char tfm_name[CRYPTO_MAX_ALG_NAME];
> > > + u8 compr_batch_size;
> > > + u8 store_batch_size;
> >
> > I don't think we need to store store_batch_size, seems trivial to
> > calculate at store time (perhaps in a helper).
> >
> > Taking a step back, is there any benefit to limiting store_batch_size to
> > compr_batch_size? Is there a disadvantage to using
> > ZSWAP_MAX_BATCH_SIZE
> > even if it's higher than the HW compression batch size?
>
> Thanks Yosry, for the code review comments. I had a discussion with
> Barry earlier on these very same topics as follow up to his review comments
> for v11, starting with [1]. Can you please go through the rationale for
> these design choices, and let me know if you have any questions:
>
> [1]: https://patchwork.kernel.org/comment/26530319/
>
[...]
> >
> > Does it actually matter if we do the initializations here vs. right
> > before inserting to the LRU (current behavior)?
>
> Yes, it impacts batching performance with software quite a bit.
[...]
> > Seems like if xa_store() fails we do not cleanup previously charged
> > objects, pool references, zswap_stored_pages, etc. Instead of rolling
> > all this back on failure, can we do all the xarray stores first and only
> > do the rest when we're at a point where no failure can happen? Would
> > that cause a performance regression?
>
> It would make the code more complicated and thus cause a performance
> regression. I have tried to balance code simplicity (which impacts
> performance)
> with correctness here. The "store_failed_idx" ensures that all partial work
> done
> in zswap_store_pages() for this batch, is cleaned up.
>
> The rest is handled in zswap_store() when it sees an error returned by
> zswap_store_pages().
>
Hi Yosry,
I was wondering if my explanations to the above comments answer
your questions? Please let me know.
The bulk entries alloc/free comments' follow-up is covered in the other
email.
Thanks,
Kanchana
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
` (22 preceding siblings ...)
2025-09-26 3:35 ` [PATCH v12 23/23] mm: zswap: Batched zswap_compress() with compress batching of large folios Kanchana P Sridhar
@ 2025-10-13 18:03 ` Sridhar, Kanchana P
23 siblings, 0 replies; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-10-13 18:03 UTC (permalink / raw)
To: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, Accardi, Kristen C, Gomes,
Vinicius
Cc: Feghali, Wajdi K, Gopal, Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> Sent: Thursday, September 25, 2025 8:35 PM
> To: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> hannes@cmpxchg.org; yosry.ahmed@linux.dev; nphamcs@gmail.com;
> chengming.zhou@linux.dev; usamaarif642@gmail.com;
> ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>
> Cc: Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> <vinodh.gopal@intel.com>; Sridhar, Kanchana P
> <kanchana.p.sridhar@intel.com>
> Subject: [PATCH v12 00/23] zswap compression batching with optimized
> iaa_crypto driver
>
>
> v12: zswap compression batching with optimized iaa_crypto driver
>
> Based on Herbert's suggestions, the most significant change in v12 is, the
> batching interfaces from zswap to crypto, from crypto to iaa_crypto, and
> the batching implementation within iaa_crypto now use the folio directly as
> the
> source (sg_page_iter for retrieving pages) and destination SG lists. A unit_size
> has been added to struct acomp_req, with kernel users such as zswap using
> the
> new acomp_request_set_unit_size() API to set the unit size to use while
> breaking
> down the request's src/dst scatterlists. zswap sets the unit-size to PAGE_SIZE.
>
Hi Herbert,
Just wanted to follow up to see if you have any code review comments on
the v12 changes made to the batching interfaces as per your suggestions?
Please let me know, I would appreciate it.
Thanks,
Kanchana
>
> Changes since v11:
> ==================
> 1) Rebased to mm-unstable as of 9-18-2025, commit 1f98191f08b4.
> 2) Incorporated Herbert's suggestions on submitting the folio as the source
> and
> SG lists for the destination to create the compress batching interface from
> zswap to crypto.
> 3) As per Herbert's suggestion, added a new unit_size member to struct
> acomp_req, along with a acomp_request_set_unit_size() API for kernel
> modules
> to set the unit size to use while breaking down the request's src/dst
> scatterlists.
> 4) Implemented iaa_crypto batching using the new SG lists based architecture
> and
> crypto interfaces.
> 5) To make the SG lists based approach functional and performant for IAA, I
> have
> changed all the calls to dma_map_sg() to use nents of 1. This should not be
> a
> concern, since it eliminates redundant computes to scan an SG list with only
> one scatterlist for existing kernel users, i.e. zswap with the
> zswap_compress() modifications in this series. This will continue to hold
> true with the zram IAA batching support I am developing. There are no
> kernel
> use cases for the iaa_crypto driver that will break this assumption.
> 6) Addressed Herbert's comment about batch_size being a statically defined
> data
> member in struct acomp_alg and struct crypto_acomp.
> 7) Addressed Nhat's comment about VM_WARN_ON_ONCE(nr_pages >
> ZSWAP_MAX_BATCH_SIZE) in zswap_store_pages().
> 8) Nhat's comment about deleting struct swap_batch_decomp_data is
> automatically
> addressed by the SG lists based rewrite of the crypto batching interface.
> 9) Addressed Barry's comment about renaming pool->batch_size to
> pool->store_batch_size.
> 10) Incorporated Barry's suggestion to merge patches that introduce data
> members
> to structures and/or API and their usage.
> 11) Added performance data to patch 0023's commit log, as suggested by
> Barry.
>
^ permalink raw reply [flat|nested] 56+ messages in thread
* Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches.
2025-10-01 21:20 ` Sridhar, Kanchana P
2025-10-03 19:10 ` Sridhar, Kanchana P
2025-10-13 17:58 ` Sridhar, Kanchana P
@ 2025-10-14 15:29 ` Yosry Ahmed
2025-10-14 16:35 ` Nhat Pham
2 siblings, 1 reply; 56+ messages in thread
From: Yosry Ahmed @ 2025-10-14 15:29 UTC (permalink / raw)
To: Sridhar, Kanchana P
Cc: linux-kernel, linux-mm, hannes, nphamcs, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius,
Feghali, Wajdi K, Gopal, Vinodh
[..]
> > > @@ -158,6 +161,8 @@ struct zswap_pool {
> > > struct work_struct release_work;
> > > struct hlist_node node;
> > > char tfm_name[CRYPTO_MAX_ALG_NAME];
> > > + u8 compr_batch_size;
> > > + u8 store_batch_size;
> >
> > I don't think we need to store store_batch_size, seems trivial to
> > calculate at store time (perhaps in a helper).
> >
> > Taking a step back, is there any benefit to limiting store_batch_size to
> > compr_batch_size? Is there a disadvantage to using
> > ZSWAP_MAX_BATCH_SIZE
> > even if it's higher than the HW compression batch size?
>
> Thanks Yosry, for the code review comments. I had a discussion with
> Barry earlier on these very same topics as follow up to his review comments
> for v11, starting with [1]. Can you please go through the rationale for
> these design choices, and let me know if you have any questions:
>
> [1]: https://patchwork.kernel.org/comment/26530319/
I am surprised that calculating the value in zswap_store() causes a
regression, but I am fine with keeping the precalculation in this case.
I think there's a bigger problem here tho, more below.
> > > + */
> > > +static __always_inline int zswap_entries_cache_alloc_batch(void
> > **entries,
> > > + unsigned int
> > nr_entries,
> > > + gfp_t gfp)
> > > +{
> > > + return kmem_cache_alloc_bulk(zswap_entry_cache, gfp, nr_entries,
> > entries);
> >
> > We currently use kmem_cache_alloc_node() in zswap_entry_cache_alloc() to
> > allocate the entry on the same node as the compressed page. We use
> > entry_to_nid() to get the node for LRU operations.
> >
> > This breaks that assumption.
>
> You bring up a good point. I was looking at the code in slub.c and my
> understanding thus far is that both, bulk allocations and kmem_cache_alloc_node()
> allocations are made from a per-CPU "cpu_slab" that is allocated by SLUB.
>
> IIUC, the concern you are raising is in the mainline, the entry is allocated on
> the same node as the compressed page, and gets added to the LRU list of that
> node. IOW, the node to which the compressed page belongs is the one to whose
> LRU the entry will be added.
>
> With this patch, with kmem_cache_alloc_bulk(), the entry will be created on
> the per-CPU slab of the CPU on which zswap_store() is called and will be
> added to the LRU of that per-CPU slab's NUMA node. Hence, the end result
> could potentially be that the zswap_entry for a page could potentially be
> on a different NUMA node/memcg than the page's NUMA node.
>
> This is my thinking as to how this will impact the zswap shrinker:
>
> 1) memcg shrinker: if the memcg the entry ends up in is on the zswap_list_lru,
> the entry will be written back.
> 2) Global shrinker: will cycle through all memcg's that have pages in the
> zswap_list_lru, and the entry will be written back.
>
> Based on this, it is not clear to me if there is a problem, and would like to
> request you, Nhat and others to provide insights as well.
>
> Interestingly, most of the code in slub.c has unlikely(!node_match(slab, node)).
> Does this imply some higher level mm slab allocation requirements?
>
> I am Ok with just calling zswap_entry_cache_alloc() for "nr_pages" if we
> think this would be more correct.
I saw your other response as well, but I think one thing is not clear
here. The zswap entry will get written back "eventually", sure, but
that's not the problem.
If the zswap entry is on the wrong node lru, two things happen:
(a) When the "right" node is under memory pressure, we cannot free this
entry by writing it back since it's not available in the lru.
(b) When the "wrong" node is under memory pressure, it will potentially
writeback entries from other nodes AND report them as being freed
from this node.
Both (a) and (b) cause less effective reclaim from the zswap shrinker.
Additionally (b) causes the shrinker to report the wrong amount of freed
memory from the node. While this may not be significant today, it's very
possible that more heuristics start relying on this number in the
future.
I don't believe we should put zswap entries on the wrong LRU, but I will
defer to Nhat for the final verdict if he has a different opinion.
> >
> > Does it actually matter if we do the initializations here vs. right
> > before inserting to the LRU (current behavior)?
>
> Yes, it impacts batching performance with software quite a bit.
Taking a step back, based on discussions in this thread and others, it
seems like the performance of zswap_store() is very sensitive to minor
changes like this one, calculating the store size, etc.
I don't recall this being the case before this patch series. It seems
like an artifact of batching affecting performance negatively and
a collection of micro-optimizations and "correct" placement of code/data
to fix it. This seems to be very fragile.
It is very possible that an incoming change will move the
initializations or make other changes to the code, that will seemingly
cause regressions when they really shouldn't.
Additionally, what may seem optimal on the CPU you are testing on maybe
be suboptimal for other CPUs. I think the big problem here is not where
to initialize the entry or whether to precompute the store batch size,
it's why the code has become extremely sensitive to these changes when
it shouldn't be. zswap_store() is not a fast path by any means, and
people will not expect such changes to cause meaningful regressions.
>
> >
> > > + for (i = 0; i < nr_pages; ++i) {
> > > + entries[i]->handle = (unsigned long)ERR_PTR(-EINVAL);
> > > + entries[i]->pool = pool;
> > > + entries[i]->swpentry = page_swap_entry(folio_page(folio,
> > start + i));
> > > + entries[i]->objcg = objcg;
> > > + entries[i]->referenced = true;
> > > + INIT_LIST_HEAD(&entries[i]->lru);
> > > + }
> > >
> > > - old = xa_store(swap_zswap_tree(page_swpentry),
> > > - swp_offset(page_swpentry),
> > > - entry, GFP_KERNEL);
> > > - if (xa_is_err(old)) {
> > > - int err = xa_err(old);
> > > + for (i = 0; i < nr_pages; ++i) {
> > > + struct page *page = folio_page(folio, start + i);
> > >
> > > - WARN_ONCE(err != -ENOMEM, "unexpected xarray error:
> > %d\n", err);
> > > - zswap_reject_alloc_fail++;
> > > - goto store_failed;
> > > + if (!zswap_compress(page, entries[i], pool, folio_wb))
> > > + goto store_pages_failed;
> > > }
> > >
> > > - /*
> > > - * We may have had an existing entry that became stale when
> > > - * the folio was redirtied and now the new version is being
> > > - * swapped out. Get rid of the old.
> > > - */
> > > - if (old)
> > > - zswap_entry_free(old);
> > > + for (i = 0; i < nr_pages; ++i) {
> > > + struct zswap_entry *old, *entry = entries[i];
> > >
> > > - /*
> > > - * The entry is successfully compressed and stored in the tree, there is
> > > - * no further possibility of failure. Grab refs to the pool and objcg,
> > > - * charge zswap memory, and increment zswap_stored_pages.
> > > - * The opposite actions will be performed by zswap_entry_free()
> > > - * when the entry is removed from the tree.
> > > - */
> > > - zswap_pool_get(pool);
> > > - if (objcg) {
> > > - obj_cgroup_get(objcg);
> > > - obj_cgroup_charge_zswap(objcg, entry->length);
> > > - }
> > > - atomic_long_inc(&zswap_stored_pages);
> > > - if (entry->length == PAGE_SIZE)
> > > - atomic_long_inc(&zswap_stored_incompressible_pages);
> > > + old = xa_store(swap_zswap_tree(entry->swpentry),
> > > + swp_offset(entry->swpentry),
> > > + entry, GFP_KERNEL);
> > > + if (unlikely(xa_is_err(old))) {
> > > + int err = xa_err(old);
> > >
> > > - /*
> > > - * We finish initializing the entry while it's already in xarray.
> > > - * This is safe because:
> > > - *
> > > - * 1. Concurrent stores and invalidations are excluded by folio lock.
> > > - *
> > > - * 2. Writeback is excluded by the entry not being on the LRU yet.
> > > - * The publishing order matters to prevent writeback from seeing
> > > - * an incoherent entry.
> > > - */
> > > - entry->pool = pool;
> > > - entry->swpentry = page_swpentry;
> > > - entry->objcg = objcg;
> > > - entry->referenced = true;
> > > - if (entry->length) {
> > > - INIT_LIST_HEAD(&entry->lru);
> > > - zswap_lru_add(&zswap_list_lru, entry);
> > > + WARN_ONCE(err != -ENOMEM, "unexpected xarray
> > error: %d\n", err);
> > > + zswap_reject_alloc_fail++;
> > > + /*
> > > + * Entries up to this point have been stored in the
> > > + * xarray. zswap_store() will erase them from the
> > xarray
> > > + * and call zswap_entry_free(). Local cleanup in
> > > + * 'store_pages_failed' only needs to happen for
> > > + * entries from [@i to @nr_pages).
> > > + */
> > > + store_fail_idx = i;
> > > + goto store_pages_failed;
> > > + }
> > > +
> > > + /*
> > > + * We may have had an existing entry that became stale when
> > > + * the folio was redirtied and now the new version is being
> > > + * swapped out. Get rid of the old.
> > > + */
> > > + if (unlikely(old))
> > > + zswap_entry_free(old);
> > > +
> > > + /*
> > > + * The entry is successfully compressed and stored in the tree,
> > there is
> > > + * no further possibility of failure. Grab refs to the pool and
> > objcg,
> > > + * charge zswap memory, and increment
> > zswap_stored_pages.
> > > + * The opposite actions will be performed by
> > zswap_entry_free()
> > > + * when the entry is removed from the tree.
> > > + */
> >
> > But there *is* further possibility of failure if a subsequent entry
> > xa_store() fails, right?
>
> This comment is referring to this specific entry.
I think it's currently misleading in its current form. Perhaps:
The entry is successfully compressed and stored in the tree, and further
failures will be cleaned up in zswap_store().
>
> >
> > Seems like if xa_store() fails we do not cleanup previously charged
> > objects, pool references, zswap_stored_pages, etc. Instead of rolling
> > all this back on failure, can we do all the xarray stores first and only
> > do the rest when we're at a point where no failure can happen? Would
> > that cause a performance regression?
>
> It would make the code more complicated and thus cause a performance
> regression. I have tried to balance code simplicity (which impacts performance)
> with correctness here. The "store_failed_idx" ensures that all partial work done
> in zswap_store_pages() for this batch, is cleaned up.
>
> The rest is handled in zswap_store() when it sees an error returned by
> zswap_store_pages().
Right, I missed the part about zswap_store() handling this, the comment
above confused me.
^ permalink raw reply [flat|nested] 56+ messages in thread
* Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches.
2025-10-14 15:29 ` Yosry Ahmed
@ 2025-10-14 16:35 ` Nhat Pham
2025-10-15 3:42 ` Sridhar, Kanchana P
0 siblings, 1 reply; 56+ messages in thread
From: Nhat Pham @ 2025-10-14 16:35 UTC (permalink / raw)
To: Yosry Ahmed
Cc: Sridhar, Kanchana P, linux-kernel, linux-mm, hannes,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, Accardi, Kristen C, Gomes,
Vinicius, Feghali, Wajdi K, Gopal, Vinodh
On Tue, Oct 14, 2025 at 8:29 AM Yosry Ahmed <yosry.ahmed@linux.dev> wrote:
>
> [..]
> > > > @@ -158,6 +161,8 @@ struct zswap_pool {
> > > > struct work_struct release_work;
> > > > struct hlist_node node;
> > > > char tfm_name[CRYPTO_MAX_ALG_NAME];
> > > > + u8 compr_batch_size;
> > > > + u8 store_batch_size;
> > >
> > > I don't think we need to store store_batch_size, seems trivial to
> > > calculate at store time (perhaps in a helper).
> > >
> > > Taking a step back, is there any benefit to limiting store_batch_size to
> > > compr_batch_size? Is there a disadvantage to using
> > > ZSWAP_MAX_BATCH_SIZE
> > > even if it's higher than the HW compression batch size?
> >
> > Thanks Yosry, for the code review comments. I had a discussion with
> > Barry earlier on these very same topics as follow up to his review comments
> > for v11, starting with [1]. Can you please go through the rationale for
> > these design choices, and let me know if you have any questions:
> >
> > [1]: https://patchwork.kernel.org/comment/26530319/
>
> I am surprised that calculating the value in zswap_store() causes a
> regression, but I am fine with keeping the precalculation in this case.
>
> I think there's a bigger problem here tho, more below.
>
> > > > + */
> > > > +static __always_inline int zswap_entries_cache_alloc_batch(void
> > > **entries,
> > > > + unsigned int
> > > nr_entries,
> > > > + gfp_t gfp)
> > > > +{
> > > > + return kmem_cache_alloc_bulk(zswap_entry_cache, gfp, nr_entries,
> > > entries);
> > >
> > > We currently use kmem_cache_alloc_node() in zswap_entry_cache_alloc() to
> > > allocate the entry on the same node as the compressed page. We use
> > > entry_to_nid() to get the node for LRU operations.
> > >
> > > This breaks that assumption.
> >
> > You bring up a good point. I was looking at the code in slub.c and my
> > understanding thus far is that both, bulk allocations and kmem_cache_alloc_node()
> > allocations are made from a per-CPU "cpu_slab" that is allocated by SLUB.
> >
> > IIUC, the concern you are raising is in the mainline, the entry is allocated on
> > the same node as the compressed page, and gets added to the LRU list of that
> > node. IOW, the node to which the compressed page belongs is the one to whose
> > LRU the entry will be added.
> >
> > With this patch, with kmem_cache_alloc_bulk(), the entry will be created on
> > the per-CPU slab of the CPU on which zswap_store() is called and will be
> > added to the LRU of that per-CPU slab's NUMA node. Hence, the end result
> > could potentially be that the zswap_entry for a page could potentially be
> > on a different NUMA node/memcg than the page's NUMA node.
I think only the NUMA node is the problem, not the memcg.
> >
> > This is my thinking as to how this will impact the zswap shrinker:
> >
> > 1) memcg shrinker: if the memcg the entry ends up in is on the zswap_list_lru,
> > the entry will be written back.
> > 2) Global shrinker: will cycle through all memcg's that have pages in the
> > zswap_list_lru, and the entry will be written back.
> >
> > Based on this, it is not clear to me if there is a problem, and would like to
> > request you, Nhat and others to provide insights as well.
> >
> > Interestingly, most of the code in slub.c has unlikely(!node_match(slab, node)).
> > Does this imply some higher level mm slab allocation requirements?
> >
> > I am Ok with just calling zswap_entry_cache_alloc() for "nr_pages" if we
> > think this would be more correct.
>
> I saw your other response as well, but I think one thing is not clear
> here. The zswap entry will get written back "eventually", sure, but
> that's not the problem.
>
> If the zswap entry is on the wrong node lru, two things happen:
> (a) When the "right" node is under memory pressure, we cannot free this
> entry by writing it back since it's not available in the lru.
> (b) When the "wrong" node is under memory pressure, it will potentially
> writeback entries from other nodes AND report them as being freed
> from this node.
>
> Both (a) and (b) cause less effective reclaim from the zswap shrinker.
> Additionally (b) causes the shrinker to report the wrong amount of freed
> memory from the node. While this may not be significant today, it's very
> possible that more heuristics start relying on this number in the
> future.
>
> I don't believe we should put zswap entries on the wrong LRU, but I will
> defer to Nhat for the final verdict if he has a different opinion.
Oh shoot. Yeah I missed that part.
In the past, we sort of did not care - zswap was very poorly designed
for NUMA architecture in general, and most of our test setups have
been single-node, so these kinds of discrepancies did not show up in
performance numbers.
But we are getting more multi-node systems:
1. Bigger hosts (memory-wise) tend to also have more than one nodes.
It scales better that way (especially because a lot of structures and
locks protecting them are node-partitioned).
2. We have also seen different memory media that are often expressed
to the kernel as nodes: CXL, GPU memory, etc.
This will necessitate tightening memory placement. We recently had to
fix one such issue:
https://github.com/torvalds/linux/commit/56e5a103a721d0ef139bba7ff3d3ada6c8217d5b
So I'm a bit nervous about this change, which will make us use the wrong LRU...
Some work around:
1. Can we squeeze an extra int field anywhere in struct zswap_entry?
2. Can we pump nid all the way to zswap_lru_add()?
This is still not 100% ideal - the metadata (struct zswap_entry) will
still be allocated on the wrong node. But at least the data are
properly managed, i.e on the right LRU.
>
> > >
> > > Does it actually matter if we do the initializations here vs. right
> > > before inserting to the LRU (current behavior)?
> >
> > Yes, it impacts batching performance with software quite a bit.
>
> Taking a step back, based on discussions in this thread and others, it
> seems like the performance of zswap_store() is very sensitive to minor
> changes like this one, calculating the store size, etc.
>
> I don't recall this being the case before this patch series. It seems
> like an artifact of batching affecting performance negatively and
> a collection of micro-optimizations and "correct" placement of code/data
> to fix it. This seems to be very fragile.
I agree. This seems troubling.
What's the variance of your benchmark results, Kanchana? I have found
that kernel build tests can have quite a bit of variance, but is it as
big as the performance difference of these minor changes (i.e, can it
explain the "regression")?
>
> It is very possible that an incoming change will move the
> initializations or make other changes to the code, that will seemingly
> cause regressions when they really shouldn't.
>
> Additionally, what may seem optimal on the CPU you are testing on maybe
> be suboptimal for other CPUs. I think the big problem here is not where
> to initialize the entry or whether to precompute the store batch size,
> it's why the code has become extremely sensitive to these changes when
> it shouldn't be. zswap_store() is not a fast path by any means, and
> people will not expect such changes to cause meaningful regressions.
Agree. These micro-optimizations seem very fragile.
>
> >
> > >
> > > > + for (i = 0; i < nr_pages; ++i) {
> > > > + entries[i]->handle = (unsigned long)ERR_PTR(-EINVAL);
> > > > + entries[i]->pool = pool;
> > > > + entries[i]->swpentry = page_swap_entry(folio_page(folio,
> > > start + i));
> > > > + entries[i]->objcg = objcg;
> > > > + entries[i]->referenced = true;
> > > > + INIT_LIST_HEAD(&entries[i]->lru);
> > > > + }
> > > >
> > > > - old = xa_store(swap_zswap_tree(page_swpentry),
> > > > - swp_offset(page_swpentry),
> > > > - entry, GFP_KERNEL);
> > > > - if (xa_is_err(old)) {
> > > > - int err = xa_err(old);
> > > > + for (i = 0; i < nr_pages; ++i) {
> > > > + struct page *page = folio_page(folio, start + i);
> > > >
> > > > - WARN_ONCE(err != -ENOMEM, "unexpected xarray error:
> > > %d\n", err);
> > > > - zswap_reject_alloc_fail++;
> > > > - goto store_failed;
> > > > + if (!zswap_compress(page, entries[i], pool, folio_wb))
> > > > + goto store_pages_failed;
> > > > }
> > > >
> > > > - /*
> > > > - * We may have had an existing entry that became stale when
> > > > - * the folio was redirtied and now the new version is being
> > > > - * swapped out. Get rid of the old.
> > > > - */
> > > > - if (old)
> > > > - zswap_entry_free(old);
> > > > + for (i = 0; i < nr_pages; ++i) {
> > > > + struct zswap_entry *old, *entry = entries[i];
> > > >
> > > > - /*
> > > > - * The entry is successfully compressed and stored in the tree, there is
> > > > - * no further possibility of failure. Grab refs to the pool and objcg,
> > > > - * charge zswap memory, and increment zswap_stored_pages.
> > > > - * The opposite actions will be performed by zswap_entry_free()
> > > > - * when the entry is removed from the tree.
> > > > - */
> > > > - zswap_pool_get(pool);
> > > > - if (objcg) {
> > > > - obj_cgroup_get(objcg);
> > > > - obj_cgroup_charge_zswap(objcg, entry->length);
> > > > - }
> > > > - atomic_long_inc(&zswap_stored_pages);
> > > > - if (entry->length == PAGE_SIZE)
> > > > - atomic_long_inc(&zswap_stored_incompressible_pages);
> > > > + old = xa_store(swap_zswap_tree(entry->swpentry),
> > > > + swp_offset(entry->swpentry),
> > > > + entry, GFP_KERNEL);
> > > > + if (unlikely(xa_is_err(old))) {
> > > > + int err = xa_err(old);
> > > >
> > > > - /*
> > > > - * We finish initializing the entry while it's already in xarray.
> > > > - * This is safe because:
> > > > - *
> > > > - * 1. Concurrent stores and invalidations are excluded by folio lock.
> > > > - *
> > > > - * 2. Writeback is excluded by the entry not being on the LRU yet.
> > > > - * The publishing order matters to prevent writeback from seeing
> > > > - * an incoherent entry.
> > > > - */
> > > > - entry->pool = pool;
> > > > - entry->swpentry = page_swpentry;
> > > > - entry->objcg = objcg;
> > > > - entry->referenced = true;
> > > > - if (entry->length) {
> > > > - INIT_LIST_HEAD(&entry->lru);
> > > > - zswap_lru_add(&zswap_list_lru, entry);
> > > > + WARN_ONCE(err != -ENOMEM, "unexpected xarray
> > > error: %d\n", err);
> > > > + zswap_reject_alloc_fail++;
> > > > + /*
> > > > + * Entries up to this point have been stored in the
> > > > + * xarray. zswap_store() will erase them from the
> > > xarray
> > > > + * and call zswap_entry_free(). Local cleanup in
> > > > + * 'store_pages_failed' only needs to happen for
> > > > + * entries from [@i to @nr_pages).
> > > > + */
> > > > + store_fail_idx = i;
> > > > + goto store_pages_failed;
> > > > + }
> > > > +
> > > > + /*
> > > > + * We may have had an existing entry that became stale when
> > > > + * the folio was redirtied and now the new version is being
> > > > + * swapped out. Get rid of the old.
> > > > + */
> > > > + if (unlikely(old))
> > > > + zswap_entry_free(old);
> > > > +
> > > > + /*
> > > > + * The entry is successfully compressed and stored in the tree,
> > > there is
> > > > + * no further possibility of failure. Grab refs to the pool and
> > > objcg,
> > > > + * charge zswap memory, and increment
> > > zswap_stored_pages.
> > > > + * The opposite actions will be performed by
> > > zswap_entry_free()
> > > > + * when the entry is removed from the tree.
> > > > + */
> > >
> > > But there *is* further possibility of failure if a subsequent entry
> > > xa_store() fails, right?
> >
> > This comment is referring to this specific entry.
>
> I think it's currently misleading in its current form. Perhaps:
>
> The entry is successfully compressed and stored in the tree, and further
> failures will be cleaned up in zswap_store().
>
> >
> > >
> > > Seems like if xa_store() fails we do not cleanup previously charged
> > > objects, pool references, zswap_stored_pages, etc. Instead of rolling
> > > all this back on failure, can we do all the xarray stores first and only
> > > do the rest when we're at a point where no failure can happen? Would
> > > that cause a performance regression?
> >
> > It would make the code more complicated and thus cause a performance
> > regression. I have tried to balance code simplicity (which impacts performance)
> > with correctness here. The "store_failed_idx" ensures that all partial work done
> > in zswap_store_pages() for this batch, is cleaned up.
> >
> > The rest is handled in zswap_store() when it sees an error returned by
> > zswap_store_pages().
>
> Right, I missed the part about zswap_store() handling this, the comment
> above confused me.
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches.
2025-10-14 16:35 ` Nhat Pham
@ 2025-10-15 3:42 ` Sridhar, Kanchana P
2025-10-15 17:04 ` Nhat Pham
0 siblings, 1 reply; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-10-15 3:42 UTC (permalink / raw)
To: Nhat Pham, Yosry Ahmed
Cc: linux-kernel, linux-mm, hannes, chengming.zhou, usamaarif642,
ryan.roberts, 21cnbao, ying.huang, akpm, senozhatsky, sj, kasong,
linux-crypto, herbert, davem, clabbe, ardb, ebiggers, surenb,
Accardi, Kristen C, Gomes, Vinicius, Feghali, Wajdi K, Gopal,
Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Nhat Pham <nphamcs@gmail.com>
> Sent: Tuesday, October 14, 2025 9:35 AM
> To: Yosry Ahmed <yosry.ahmed@linux.dev>
> Cc: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>; linux-
> kernel@vger.kernel.org; linux-mm@kvack.org; hannes@cmpxchg.org;
> chengming.zhou@linux.dev; usamaarif642@gmail.com;
> ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>;
> Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> <vinodh.gopal@intel.com>
> Subject: Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a
> large folio in batches.
>
> On Tue, Oct 14, 2025 at 8:29 AM Yosry Ahmed <yosry.ahmed@linux.dev>
> wrote:
> >
> > [..]
> > > > > @@ -158,6 +161,8 @@ struct zswap_pool {
> > > > > struct work_struct release_work;
> > > > > struct hlist_node node;
> > > > > char tfm_name[CRYPTO_MAX_ALG_NAME];
> > > > > + u8 compr_batch_size;
> > > > > + u8 store_batch_size;
> > > >
> > > > I don't think we need to store store_batch_size, seems trivial to
> > > > calculate at store time (perhaps in a helper).
> > > >
> > > > Taking a step back, is there any benefit to limiting store_batch_size to
> > > > compr_batch_size? Is there a disadvantage to using
> > > > ZSWAP_MAX_BATCH_SIZE
> > > > even if it's higher than the HW compression batch size?
> > >
> > > Thanks Yosry, for the code review comments. I had a discussion with
> > > Barry earlier on these very same topics as follow up to his review
> comments
> > > for v11, starting with [1]. Can you please go through the rationale for
> > > these design choices, and let me know if you have any questions:
> > >
> > > [1]: https://patchwork.kernel.org/comment/26530319/
> >
> > I am surprised that calculating the value in zswap_store() causes a
> > regression, but I am fine with keeping the precalculation in this case.
> >
> > I think there's a bigger problem here tho, more below.
> >
> > > > > + */
> > > > > +static __always_inline int zswap_entries_cache_alloc_batch(void
> > > > **entries,
> > > > > + unsigned int
> > > > nr_entries,
> > > > > + gfp_t gfp)
> > > > > +{
> > > > > + return kmem_cache_alloc_bulk(zswap_entry_cache, gfp, nr_entries,
> > > > entries);
> > > >
> > > > We currently use kmem_cache_alloc_node() in
> zswap_entry_cache_alloc() to
> > > > allocate the entry on the same node as the compressed page. We use
> > > > entry_to_nid() to get the node for LRU operations.
> > > >
> > > > This breaks that assumption.
> > >
> > > You bring up a good point. I was looking at the code in slub.c and my
> > > understanding thus far is that both, bulk allocations and
> kmem_cache_alloc_node()
> > > allocations are made from a per-CPU "cpu_slab" that is allocated by SLUB.
> > >
> > > IIUC, the concern you are raising is in the mainline, the entry is allocated
> on
> > > the same node as the compressed page, and gets added to the LRU list of
> that
> > > node. IOW, the node to which the compressed page belongs is the one to
> whose
> > > LRU the entry will be added.
> > >
> > > With this patch, with kmem_cache_alloc_bulk(), the entry will be created
> on
> > > the per-CPU slab of the CPU on which zswap_store() is called and will be
> > > added to the LRU of that per-CPU slab's NUMA node. Hence, the end
> result
> > > could potentially be that the zswap_entry for a page could potentially be
> > > on a different NUMA node/memcg than the page's NUMA node.
>
> I think only the NUMA node is the problem, not the memcg.
>
> > >
> > > This is my thinking as to how this will impact the zswap shrinker:
> > >
> > > 1) memcg shrinker: if the memcg the entry ends up in is on the
> zswap_list_lru,
> > > the entry will be written back.
> > > 2) Global shrinker: will cycle through all memcg's that have pages in the
> > > zswap_list_lru, and the entry will be written back.
> > >
> > > Based on this, it is not clear to me if there is a problem, and would like to
> > > request you, Nhat and others to provide insights as well.
> > >
> > > Interestingly, most of the code in slub.c has unlikely(!node_match(slab,
> node)).
> > > Does this imply some higher level mm slab allocation requirements?
> > >
> > > I am Ok with just calling zswap_entry_cache_alloc() for "nr_pages" if we
> > > think this would be more correct.
> >
> > I saw your other response as well, but I think one thing is not clear
> > here. The zswap entry will get written back "eventually", sure, but
> > that's not the problem.
> >
> > If the zswap entry is on the wrong node lru, two things happen:
> > (a) When the "right" node is under memory pressure, we cannot free this
> > entry by writing it back since it's not available in the lru.
> > (b) When the "wrong" node is under memory pressure, it will potentially
> > writeback entries from other nodes AND report them as being freed
> > from this node.
> >
> > Both (a) and (b) cause less effective reclaim from the zswap shrinker.
> > Additionally (b) causes the shrinker to report the wrong amount of freed
> > memory from the node. While this may not be significant today, it's very
> > possible that more heuristics start relying on this number in the
> > future.
> >
> > I don't believe we should put zswap entries on the wrong LRU, but I will
> > defer to Nhat for the final verdict if he has a different opinion.
>
> Oh shoot. Yeah I missed that part.
>
> In the past, we sort of did not care - zswap was very poorly designed
> for NUMA architecture in general, and most of our test setups have
> been single-node, so these kinds of discrepancies did not show up in
> performance numbers.
>
> But we are getting more multi-node systems:
>
> 1. Bigger hosts (memory-wise) tend to also have more than one nodes.
> It scales better that way (especially because a lot of structures and
> locks protecting them are node-partitioned).
>
> 2. We have also seen different memory media that are often expressed
> to the kernel as nodes: CXL, GPU memory, etc.
>
> This will necessitate tightening memory placement. We recently had to
> fix one such issue:
>
> https://github.com/torvalds/linux/commit/56e5a103a721d0ef139bba7ff3d3a
> da6c8217d5b
>
> So I'm a bit nervous about this change, which will make us use the wrong
> LRU...
>
> Some work around:
>
> 1. Can we squeeze an extra int field anywhere in struct zswap_entry?
>
> 2. Can we pump nid all the way to zswap_lru_add()?
>
> This is still not 100% ideal - the metadata (struct zswap_entry) will
> still be allocated on the wrong node. But at least the data are
> properly managed, i.e on the right LRU.
Thanks, Nhat and Yosry for the discussion. Thank you Nhat, for the
zsmalloc change log reference and for the work arounds!
Following your suggestion in (2), can we pass in the folio's nid from
zswap_store_pages() to zswap_lru_add(), as follows:
diff --git a/mm/zswap.c b/mm/zswap.c
index 263bc6d7f5c6..44665deece80 100644
--- a/mm/zswap.c
+++ b/mm/zswap.c
@@ -694,9 +694,9 @@ static inline int entry_to_nid(struct zswap_entry *entry)
return page_to_nid(virt_to_page(entry));
}
-static void zswap_lru_add(struct list_lru *list_lru, struct zswap_entry *entry)
+static void zswap_lru_add(struct list_lru *list_lru, struct zswap_entry *entry,
+ int nid)
{
- int nid = entry_to_nid(entry);
struct mem_cgroup *memcg;
/*
@@ -1758,7 +1758,7 @@ static bool zswap_store_pages(struct folio *folio,
* an incoherent entry.
*/
if (likely(entry->length))
- zswap_lru_add(&zswap_list_lru, entry);
+ zswap_lru_add(&zswap_list_lru, entry, nid);
}
return true;
--
I believe this will add the entry to the LRU node of the folio being
compressed. If so, we may be able to avoid adding an int field to
struct zswap_entry?
>
> >
> > > >
> > > > Does it actually matter if we do the initializations here vs. right
> > > > before inserting to the LRU (current behavior)?
> > >
> > > Yes, it impacts batching performance with software quite a bit.
> >
> > Taking a step back, based on discussions in this thread and others, it
> > seems like the performance of zswap_store() is very sensitive to minor
> > changes like this one, calculating the store size, etc.
> >
> > I don't recall this being the case before this patch series. It seems
> > like an artifact of batching affecting performance negatively and
> > a collection of micro-optimizations and "correct" placement of code/data
> > to fix it. This seems to be very fragile.
>
> I agree. This seems troubling.
>
> What's the variance of your benchmark results, Kanchana? I have found
> that kernel build tests can have quite a bit of variance, but is it as
> big as the performance difference of these minor changes (i.e, can it
> explain the "regression")?
I tested the performance difference by not storing "u8 store_batch_size"
in struct zswap_pool and instead calculating it dynamically in zswap_store().
I tested it with usemem 30 processes, which is less prone to variance.
It appears that there is overall an improvement -- I must have been
confused or it could have been the case with an earlier mm-unstable, not
the one used for v12.
When I tested this change today, throughput is better for 64K/PMD,
and latency improves for PMD. There's a slight latency "regression" of
2% in sys time and 1% in elapsed time for 64K folios:
usemem30, 64K folios:
-----------------------------------------------------------------------
pool->store_batch_size store_batch_size
computed once and stored computed per call to
in zswap_pool zswap_store()
-----------------------------------------------------------------------
zswap compressor zstd zstd
-----------------------------------------------------------------------
Total throughput (KB/s) 6,106,539 6,147,419
Average throughput (KB/s) 203,551 204,913
elapsed time (sec) 100.12 101.30
sys time (sec) 2,458.79 2,511.54
-----------------------------------------------------------------------
usemem30, PMD folios:
-----------------------------------------------------------------------
pool->store_batch_size store_batch_size
computed once and stored computed per call to
in zswap_pool zswap_store()
-----------------------------------------------------------------------
zswap compressor zstd zstd
-----------------------------------------------------------------------
Total throughput (KB/s) 6,715,579 6,762,097
Average throughput (KB/s) 223,852 225,403
elapsed time (sec) 94.68 88.55
sys time (sec) 2,313.08 2,100.27
-----------------------------------------------------------------------
I can incorporate this change in v13 if it allays your concerns about
code fragility. Again, my apologies.
I believe the new batching implementation of zswap_compress() is
not relying on any other 'micro-optimizations and "correct" placement
of code/data'. It aims to directly represent the commonality of a
batching architecture for software/sequential and batching/parallel
compressors (hardware and software batching in future as our cover letter
states); borrowing the general code flow from the mainline per-page
implementation while incorporating Herbert's SG lists based
batching crypto interface suggestion. Are there any other specific
concerns you have that I might not be comprehending? If so, please
let me know and I am happy to factor this in as I work on v13.
The performance improvements data I posted in v12 for zstd and
deflate-iaa with this batched zswap_compress() version is also
something for us to consider, if we agree that eliminating the
zswap_pool's "store_batch_size" member will fix any concerns
about code fragility.
Thanks again for your helpful critique and suggestions!
Best regards,
Kanchana
>
> >
> > It is very possible that an incoming change will move the
> > initializations or make other changes to the code, that will seemingly
> > cause regressions when they really shouldn't.
> >
> > Additionally, what may seem optimal on the CPU you are testing on maybe
> > be suboptimal for other CPUs. I think the big problem here is not where
> > to initialize the entry or whether to precompute the store batch size,
> > it's why the code has become extremely sensitive to these changes when
> > it shouldn't be. zswap_store() is not a fast path by any means, and
> > people will not expect such changes to cause meaningful regressions.
>
> Agree. These micro-optimizations seem very fragile.
>
> >
> > >
> > > >
> > > > > + for (i = 0; i < nr_pages; ++i) {
> > > > > + entries[i]->handle = (unsigned long)ERR_PTR(-EINVAL);
> > > > > + entries[i]->pool = pool;
> > > > > + entries[i]->swpentry = page_swap_entry(folio_page(folio,
> > > > start + i));
> > > > > + entries[i]->objcg = objcg;
> > > > > + entries[i]->referenced = true;
> > > > > + INIT_LIST_HEAD(&entries[i]->lru);
> > > > > + }
> > > > >
> > > > > - old = xa_store(swap_zswap_tree(page_swpentry),
> > > > > - swp_offset(page_swpentry),
> > > > > - entry, GFP_KERNEL);
> > > > > - if (xa_is_err(old)) {
> > > > > - int err = xa_err(old);
> > > > > + for (i = 0; i < nr_pages; ++i) {
> > > > > + struct page *page = folio_page(folio, start + i);
> > > > >
> > > > > - WARN_ONCE(err != -ENOMEM, "unexpected xarray error:
> > > > %d\n", err);
> > > > > - zswap_reject_alloc_fail++;
> > > > > - goto store_failed;
> > > > > + if (!zswap_compress(page, entries[i], pool, folio_wb))
> > > > > + goto store_pages_failed;
> > > > > }
> > > > >
> > > > > - /*
> > > > > - * We may have had an existing entry that became stale when
> > > > > - * the folio was redirtied and now the new version is being
> > > > > - * swapped out. Get rid of the old.
> > > > > - */
> > > > > - if (old)
> > > > > - zswap_entry_free(old);
> > > > > + for (i = 0; i < nr_pages; ++i) {
> > > > > + struct zswap_entry *old, *entry = entries[i];
> > > > >
> > > > > - /*
> > > > > - * The entry is successfully compressed and stored in the tree, there
> is
> > > > > - * no further possibility of failure. Grab refs to the pool and objcg,
> > > > > - * charge zswap memory, and increment zswap_stored_pages.
> > > > > - * The opposite actions will be performed by zswap_entry_free()
> > > > > - * when the entry is removed from the tree.
> > > > > - */
> > > > > - zswap_pool_get(pool);
> > > > > - if (objcg) {
> > > > > - obj_cgroup_get(objcg);
> > > > > - obj_cgroup_charge_zswap(objcg, entry->length);
> > > > > - }
> > > > > - atomic_long_inc(&zswap_stored_pages);
> > > > > - if (entry->length == PAGE_SIZE)
> > > > > - atomic_long_inc(&zswap_stored_incompressible_pages);
> > > > > + old = xa_store(swap_zswap_tree(entry->swpentry),
> > > > > + swp_offset(entry->swpentry),
> > > > > + entry, GFP_KERNEL);
> > > > > + if (unlikely(xa_is_err(old))) {
> > > > > + int err = xa_err(old);
> > > > >
> > > > > - /*
> > > > > - * We finish initializing the entry while it's already in xarray.
> > > > > - * This is safe because:
> > > > > - *
> > > > > - * 1. Concurrent stores and invalidations are excluded by folio lock.
> > > > > - *
> > > > > - * 2. Writeback is excluded by the entry not being on the LRU yet.
> > > > > - * The publishing order matters to prevent writeback from seeing
> > > > > - * an incoherent entry.
> > > > > - */
> > > > > - entry->pool = pool;
> > > > > - entry->swpentry = page_swpentry;
> > > > > - entry->objcg = objcg;
> > > > > - entry->referenced = true;
> > > > > - if (entry->length) {
> > > > > - INIT_LIST_HEAD(&entry->lru);
> > > > > - zswap_lru_add(&zswap_list_lru, entry);
> > > > > + WARN_ONCE(err != -ENOMEM, "unexpected xarray
> > > > error: %d\n", err);
> > > > > + zswap_reject_alloc_fail++;
> > > > > + /*
> > > > > + * Entries up to this point have been stored in the
> > > > > + * xarray. zswap_store() will erase them from the
> > > > xarray
> > > > > + * and call zswap_entry_free(). Local cleanup in
> > > > > + * 'store_pages_failed' only needs to happen for
> > > > > + * entries from [@i to @nr_pages).
> > > > > + */
> > > > > + store_fail_idx = i;
> > > > > + goto store_pages_failed;
> > > > > + }
> > > > > +
> > > > > + /*
> > > > > + * We may have had an existing entry that became stale when
> > > > > + * the folio was redirtied and now the new version is being
> > > > > + * swapped out. Get rid of the old.
> > > > > + */
> > > > > + if (unlikely(old))
> > > > > + zswap_entry_free(old);
> > > > > +
> > > > > + /*
> > > > > + * The entry is successfully compressed and stored in the tree,
> > > > there is
> > > > > + * no further possibility of failure. Grab refs to the pool and
> > > > objcg,
> > > > > + * charge zswap memory, and increment
> > > > zswap_stored_pages.
> > > > > + * The opposite actions will be performed by
> > > > zswap_entry_free()
> > > > > + * when the entry is removed from the tree.
> > > > > + */
> > > >
> > > > But there *is* further possibility of failure if a subsequent entry
> > > > xa_store() fails, right?
> > >
> > > This comment is referring to this specific entry.
> >
> > I think it's currently misleading in its current form. Perhaps:
> >
> > The entry is successfully compressed and stored in the tree, and further
> > failures will be cleaned up in zswap_store().
> >
> > >
> > > >
> > > > Seems like if xa_store() fails we do not cleanup previously charged
> > > > objects, pool references, zswap_stored_pages, etc. Instead of rolling
> > > > all this back on failure, can we do all the xarray stores first and only
> > > > do the rest when we're at a point where no failure can happen? Would
> > > > that cause a performance regression?
> > >
> > > It would make the code more complicated and thus cause a performance
> > > regression. I have tried to balance code simplicity (which impacts
> performance)
> > > with correctness here. The "store_failed_idx" ensures that all partial work
> done
> > > in zswap_store_pages() for this batch, is cleaned up.
> > >
> > > The rest is handled in zswap_store() when it sees an error returned by
> > > zswap_store_pages().
> >
> > Right, I missed the part about zswap_store() handling this, the comment
> > above confused me.
^ permalink raw reply [flat|nested] 56+ messages in thread
* Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches.
2025-10-15 3:42 ` Sridhar, Kanchana P
@ 2025-10-15 17:04 ` Nhat Pham
2025-10-15 22:15 ` Sridhar, Kanchana P
0 siblings, 1 reply; 56+ messages in thread
From: Nhat Pham @ 2025-10-15 17:04 UTC (permalink / raw)
To: Sridhar, Kanchana P
Cc: Yosry Ahmed, linux-kernel, linux-mm, hannes, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius,
Feghali, Wajdi K, Gopal, Vinodh
On Tue, Oct 14, 2025 at 8:42 PM Sridhar, Kanchana P
<kanchana.p.sridhar@intel.com> wrote:
>
>
> > -----Original Message-----
> > From: Nhat Pham <nphamcs@gmail.com>
> > Sent: Tuesday, October 14, 2025 9:35 AM
> > To: Yosry Ahmed <yosry.ahmed@linux.dev>
> > Cc: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>; linux-
> > kernel@vger.kernel.org; linux-mm@kvack.org; hannes@cmpxchg.org;
> > chengming.zhou@linux.dev; usamaarif642@gmail.com;
> > ryan.roberts@arm.com; 21cnbao@gmail.com;
> > ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> > senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> > crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> > davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> > ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> > <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>;
> > Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> > <vinodh.gopal@intel.com>
> > Subject: Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a
> > large folio in batches.
> >
> > On Tue, Oct 14, 2025 at 8:29 AM Yosry Ahmed <yosry.ahmed@linux.dev>
> > wrote:
> > >
> > > [..]
> > > > > > @@ -158,6 +161,8 @@ struct zswap_pool {
> > > > > > struct work_struct release_work;
> > > > > > struct hlist_node node;
> > > > > > char tfm_name[CRYPTO_MAX_ALG_NAME];
> > > > > > + u8 compr_batch_size;
> > > > > > + u8 store_batch_size;
> > > > >
> > > > > I don't think we need to store store_batch_size, seems trivial to
> > > > > calculate at store time (perhaps in a helper).
> > > > >
> > > > > Taking a step back, is there any benefit to limiting store_batch_size to
> > > > > compr_batch_size? Is there a disadvantage to using
> > > > > ZSWAP_MAX_BATCH_SIZE
> > > > > even if it's higher than the HW compression batch size?
> > > >
> > > > Thanks Yosry, for the code review comments. I had a discussion with
> > > > Barry earlier on these very same topics as follow up to his review
> > comments
> > > > for v11, starting with [1]. Can you please go through the rationale for
> > > > these design choices, and let me know if you have any questions:
> > > >
> > > > [1]: https://patchwork.kernel.org/comment/26530319/
> > >
> > > I am surprised that calculating the value in zswap_store() causes a
> > > regression, but I am fine with keeping the precalculation in this case.
> > >
> > > I think there's a bigger problem here tho, more below.
> > >
> > > > > > + */
> > > > > > +static __always_inline int zswap_entries_cache_alloc_batch(void
> > > > > **entries,
> > > > > > + unsigned int
> > > > > nr_entries,
> > > > > > + gfp_t gfp)
> > > > > > +{
> > > > > > + return kmem_cache_alloc_bulk(zswap_entry_cache, gfp, nr_entries,
> > > > > entries);
> > > > >
> > > > > We currently use kmem_cache_alloc_node() in
> > zswap_entry_cache_alloc() to
> > > > > allocate the entry on the same node as the compressed page. We use
> > > > > entry_to_nid() to get the node for LRU operations.
> > > > >
> > > > > This breaks that assumption.
> > > >
> > > > You bring up a good point. I was looking at the code in slub.c and my
> > > > understanding thus far is that both, bulk allocations and
> > kmem_cache_alloc_node()
> > > > allocations are made from a per-CPU "cpu_slab" that is allocated by SLUB.
> > > >
> > > > IIUC, the concern you are raising is in the mainline, the entry is allocated
> > on
> > > > the same node as the compressed page, and gets added to the LRU list of
> > that
> > > > node. IOW, the node to which the compressed page belongs is the one to
> > whose
> > > > LRU the entry will be added.
> > > >
> > > > With this patch, with kmem_cache_alloc_bulk(), the entry will be created
> > on
> > > > the per-CPU slab of the CPU on which zswap_store() is called and will be
> > > > added to the LRU of that per-CPU slab's NUMA node. Hence, the end
> > result
> > > > could potentially be that the zswap_entry for a page could potentially be
> > > > on a different NUMA node/memcg than the page's NUMA node.
> >
> > I think only the NUMA node is the problem, not the memcg.
> >
> > > >
> > > > This is my thinking as to how this will impact the zswap shrinker:
> > > >
> > > > 1) memcg shrinker: if the memcg the entry ends up in is on the
> > zswap_list_lru,
> > > > the entry will be written back.
> > > > 2) Global shrinker: will cycle through all memcg's that have pages in the
> > > > zswap_list_lru, and the entry will be written back.
> > > >
> > > > Based on this, it is not clear to me if there is a problem, and would like to
> > > > request you, Nhat and others to provide insights as well.
> > > >
> > > > Interestingly, most of the code in slub.c has unlikely(!node_match(slab,
> > node)).
> > > > Does this imply some higher level mm slab allocation requirements?
> > > >
> > > > I am Ok with just calling zswap_entry_cache_alloc() for "nr_pages" if we
> > > > think this would be more correct.
> > >
> > > I saw your other response as well, but I think one thing is not clear
> > > here. The zswap entry will get written back "eventually", sure, but
> > > that's not the problem.
> > >
> > > If the zswap entry is on the wrong node lru, two things happen:
> > > (a) When the "right" node is under memory pressure, we cannot free this
> > > entry by writing it back since it's not available in the lru.
> > > (b) When the "wrong" node is under memory pressure, it will potentially
> > > writeback entries from other nodes AND report them as being freed
> > > from this node.
> > >
> > > Both (a) and (b) cause less effective reclaim from the zswap shrinker.
> > > Additionally (b) causes the shrinker to report the wrong amount of freed
> > > memory from the node. While this may not be significant today, it's very
> > > possible that more heuristics start relying on this number in the
> > > future.
> > >
> > > I don't believe we should put zswap entries on the wrong LRU, but I will
> > > defer to Nhat for the final verdict if he has a different opinion.
> >
> > Oh shoot. Yeah I missed that part.
> >
> > In the past, we sort of did not care - zswap was very poorly designed
> > for NUMA architecture in general, and most of our test setups have
> > been single-node, so these kinds of discrepancies did not show up in
> > performance numbers.
> >
> > But we are getting more multi-node systems:
> >
> > 1. Bigger hosts (memory-wise) tend to also have more than one nodes.
> > It scales better that way (especially because a lot of structures and
> > locks protecting them are node-partitioned).
> >
> > 2. We have also seen different memory media that are often expressed
> > to the kernel as nodes: CXL, GPU memory, etc.
> >
> > This will necessitate tightening memory placement. We recently had to
> > fix one such issue:
> >
> > https://github.com/torvalds/linux/commit/56e5a103a721d0ef139bba7ff3d3a
> > da6c8217d5b
> >
> > So I'm a bit nervous about this change, which will make us use the wrong
> > LRU...
> >
> > Some work around:
> >
> > 1. Can we squeeze an extra int field anywhere in struct zswap_entry?
> >
> > 2. Can we pump nid all the way to zswap_lru_add()?
> >
> > This is still not 100% ideal - the metadata (struct zswap_entry) will
> > still be allocated on the wrong node. But at least the data are
> > properly managed, i.e on the right LRU.
>
> Thanks, Nhat and Yosry for the discussion. Thank you Nhat, for the
> zsmalloc change log reference and for the work arounds!
>
> Following your suggestion in (2), can we pass in the folio's nid from
> zswap_store_pages() to zswap_lru_add(), as follows:
>
> diff --git a/mm/zswap.c b/mm/zswap.c
> index 263bc6d7f5c6..44665deece80 100644
> --- a/mm/zswap.c
> +++ b/mm/zswap.c
> @@ -694,9 +694,9 @@ static inline int entry_to_nid(struct zswap_entry *entry)
> return page_to_nid(virt_to_page(entry));
> }
>
> -static void zswap_lru_add(struct list_lru *list_lru, struct zswap_entry *entry)
> +static void zswap_lru_add(struct list_lru *list_lru, struct zswap_entry *entry,
> + int nid)
> {
> - int nid = entry_to_nid(entry);
> struct mem_cgroup *memcg;
>
> /*
> @@ -1758,7 +1758,7 @@ static bool zswap_store_pages(struct folio *folio,
> * an incoherent entry.
> */
> if (likely(entry->length))
> - zswap_lru_add(&zswap_list_lru, entry);
> + zswap_lru_add(&zswap_list_lru, entry, nid);
> }
>
> return true;
> --
>
> I believe this will add the entry to the LRU node of the folio being
> compressed. If so, we may be able to avoid adding an int field to
> struct zswap_entry?
Hmm that might not work for zswap_lru_del() :(
zswap_entry_free() might be called in context where we do not have
access to the node information (zswap_load()) for e.g.
Another alternative: can we instead determine the node from the
compressed object's storage? i.e store zswap_entry in the LRU
corresponding to the node that holds the compressed data?
You'll probably need the new zsmalloc API to get the node information.
And can zsmalloc migrate a backing page to a different node? This
seems complicated...
Taking a step back though, do we have to use the bulk allocation API
here? Calling the single-allocation version 512 times for a PMD-sized
page is no worse than the status quo, correct? We can leave this part
unoptimized for now - in the future, if the use case justifies it, we
can talk to slab allocator maintainers and ask for guidance on a
lock-optimized cross-node bulk allocation API.
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches.
2025-10-15 17:04 ` Nhat Pham
@ 2025-10-15 22:15 ` Sridhar, Kanchana P
2025-10-15 22:24 ` Yosry Ahmed
0 siblings, 1 reply; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-10-15 22:15 UTC (permalink / raw)
To: Nhat Pham
Cc: Yosry Ahmed, linux-kernel, linux-mm, hannes, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius,
Feghali, Wajdi K, Gopal, Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Nhat Pham <nphamcs@gmail.com>
> Sent: Wednesday, October 15, 2025 10:04 AM
> To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> Cc: Yosry Ahmed <yosry.ahmed@linux.dev>; linux-kernel@vger.kernel.org;
> linux-mm@kvack.org; hannes@cmpxchg.org; chengming.zhou@linux.dev;
> usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>;
> Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> <vinodh.gopal@intel.com>
> Subject: Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a
> large folio in batches.
>
> On Tue, Oct 14, 2025 at 8:42 PM Sridhar, Kanchana P
> <kanchana.p.sridhar@intel.com> wrote:
> >
> >
> > > -----Original Message-----
> > > From: Nhat Pham <nphamcs@gmail.com>
> > > Sent: Tuesday, October 14, 2025 9:35 AM
> > > To: Yosry Ahmed <yosry.ahmed@linux.dev>
> > > Cc: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>; linux-
> > > kernel@vger.kernel.org; linux-mm@kvack.org; hannes@cmpxchg.org;
> > > chengming.zhou@linux.dev; usamaarif642@gmail.com;
> > > ryan.roberts@arm.com; 21cnbao@gmail.com;
> > > ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> > > senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> > > crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> > > davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> > > ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> > > <kristen.c.accardi@intel.com>; Gomes, Vinicius
> <vinicius.gomes@intel.com>;
> > > Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> > > <vinodh.gopal@intel.com>
> > > Subject: Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a
> > > large folio in batches.
> > >
> > > On Tue, Oct 14, 2025 at 8:29 AM Yosry Ahmed <yosry.ahmed@linux.dev>
> > > wrote:
> > > >
> > > > [..]
> > > > > > > @@ -158,6 +161,8 @@ struct zswap_pool {
> > > > > > > struct work_struct release_work;
> > > > > > > struct hlist_node node;
> > > > > > > char tfm_name[CRYPTO_MAX_ALG_NAME];
> > > > > > > + u8 compr_batch_size;
> > > > > > > + u8 store_batch_size;
> > > > > >
> > > > > > I don't think we need to store store_batch_size, seems trivial to
> > > > > > calculate at store time (perhaps in a helper).
> > > > > >
> > > > > > Taking a step back, is there any benefit to limiting store_batch_size
> to
> > > > > > compr_batch_size? Is there a disadvantage to using
> > > > > > ZSWAP_MAX_BATCH_SIZE
> > > > > > even if it's higher than the HW compression batch size?
> > > > >
> > > > > Thanks Yosry, for the code review comments. I had a discussion with
> > > > > Barry earlier on these very same topics as follow up to his review
> > > comments
> > > > > for v11, starting with [1]. Can you please go through the rationale for
> > > > > these design choices, and let me know if you have any questions:
> > > > >
> > > > > [1]: https://patchwork.kernel.org/comment/26530319/
> > > >
> > > > I am surprised that calculating the value in zswap_store() causes a
> > > > regression, but I am fine with keeping the precalculation in this case.
> > > >
> > > > I think there's a bigger problem here tho, more below.
> > > >
> > > > > > > + */
> > > > > > > +static __always_inline int zswap_entries_cache_alloc_batch(void
> > > > > > **entries,
> > > > > > > + unsigned int
> > > > > > nr_entries,
> > > > > > > + gfp_t gfp)
> > > > > > > +{
> > > > > > > + return kmem_cache_alloc_bulk(zswap_entry_cache, gfp,
> nr_entries,
> > > > > > entries);
> > > > > >
> > > > > > We currently use kmem_cache_alloc_node() in
> > > zswap_entry_cache_alloc() to
> > > > > > allocate the entry on the same node as the compressed page. We
> use
> > > > > > entry_to_nid() to get the node for LRU operations.
> > > > > >
> > > > > > This breaks that assumption.
> > > > >
> > > > > You bring up a good point. I was looking at the code in slub.c and my
> > > > > understanding thus far is that both, bulk allocations and
> > > kmem_cache_alloc_node()
> > > > > allocations are made from a per-CPU "cpu_slab" that is allocated by
> SLUB.
> > > > >
> > > > > IIUC, the concern you are raising is in the mainline, the entry is
> allocated
> > > on
> > > > > the same node as the compressed page, and gets added to the LRU list
> of
> > > that
> > > > > node. IOW, the node to which the compressed page belongs is the one
> to
> > > whose
> > > > > LRU the entry will be added.
> > > > >
> > > > > With this patch, with kmem_cache_alloc_bulk(), the entry will be
> created
> > > on
> > > > > the per-CPU slab of the CPU on which zswap_store() is called and will
> be
> > > > > added to the LRU of that per-CPU slab's NUMA node. Hence, the end
> > > result
> > > > > could potentially be that the zswap_entry for a page could potentially
> be
> > > > > on a different NUMA node/memcg than the page's NUMA node.
> > >
> > > I think only the NUMA node is the problem, not the memcg.
> > >
> > > > >
> > > > > This is my thinking as to how this will impact the zswap shrinker:
> > > > >
> > > > > 1) memcg shrinker: if the memcg the entry ends up in is on the
> > > zswap_list_lru,
> > > > > the entry will be written back.
> > > > > 2) Global shrinker: will cycle through all memcg's that have pages in
> the
> > > > > zswap_list_lru, and the entry will be written back.
> > > > >
> > > > > Based on this, it is not clear to me if there is a problem, and would like
> to
> > > > > request you, Nhat and others to provide insights as well.
> > > > >
> > > > > Interestingly, most of the code in slub.c has unlikely(!node_match(slab,
> > > node)).
> > > > > Does this imply some higher level mm slab allocation requirements?
> > > > >
> > > > > I am Ok with just calling zswap_entry_cache_alloc() for "nr_pages" if
> we
> > > > > think this would be more correct.
> > > >
> > > > I saw your other response as well, but I think one thing is not clear
> > > > here. The zswap entry will get written back "eventually", sure, but
> > > > that's not the problem.
> > > >
> > > > If the zswap entry is on the wrong node lru, two things happen:
> > > > (a) When the "right" node is under memory pressure, we cannot free
> this
> > > > entry by writing it back since it's not available in the lru.
> > > > (b) When the "wrong" node is under memory pressure, it will potentially
> > > > writeback entries from other nodes AND report them as being freed
> > > > from this node.
> > > >
> > > > Both (a) and (b) cause less effective reclaim from the zswap shrinker.
> > > > Additionally (b) causes the shrinker to report the wrong amount of freed
> > > > memory from the node. While this may not be significant today, it's very
> > > > possible that more heuristics start relying on this number in the
> > > > future.
> > > >
> > > > I don't believe we should put zswap entries on the wrong LRU, but I will
> > > > defer to Nhat for the final verdict if he has a different opinion.
> > >
> > > Oh shoot. Yeah I missed that part.
> > >
> > > In the past, we sort of did not care - zswap was very poorly designed
> > > for NUMA architecture in general, and most of our test setups have
> > > been single-node, so these kinds of discrepancies did not show up in
> > > performance numbers.
> > >
> > > But we are getting more multi-node systems:
> > >
> > > 1. Bigger hosts (memory-wise) tend to also have more than one nodes.
> > > It scales better that way (especially because a lot of structures and
> > > locks protecting them are node-partitioned).
> > >
> > > 2. We have also seen different memory media that are often expressed
> > > to the kernel as nodes: CXL, GPU memory, etc.
> > >
> > > This will necessitate tightening memory placement. We recently had to
> > > fix one such issue:
> > >
> > >
> https://github.com/torvalds/linux/commit/56e5a103a721d0ef139bba7ff3d3a
> > > da6c8217d5b
> > >
> > > So I'm a bit nervous about this change, which will make us use the wrong
> > > LRU...
> > >
> > > Some work around:
> > >
> > > 1. Can we squeeze an extra int field anywhere in struct zswap_entry?
> > >
> > > 2. Can we pump nid all the way to zswap_lru_add()?
> > >
> > > This is still not 100% ideal - the metadata (struct zswap_entry) will
> > > still be allocated on the wrong node. But at least the data are
> > > properly managed, i.e on the right LRU.
> >
> > Thanks, Nhat and Yosry for the discussion. Thank you Nhat, for the
> > zsmalloc change log reference and for the work arounds!
> >
> > Following your suggestion in (2), can we pass in the folio's nid from
> > zswap_store_pages() to zswap_lru_add(), as follows:
> >
> > diff --git a/mm/zswap.c b/mm/zswap.c
> > index 263bc6d7f5c6..44665deece80 100644
> > --- a/mm/zswap.c
> > +++ b/mm/zswap.c
> > @@ -694,9 +694,9 @@ static inline int entry_to_nid(struct zswap_entry
> *entry)
> > return page_to_nid(virt_to_page(entry));
> > }
> >
> > -static void zswap_lru_add(struct list_lru *list_lru, struct zswap_entry
> *entry)
> > +static void zswap_lru_add(struct list_lru *list_lru, struct zswap_entry
> *entry,
> > + int nid)
> > {
> > - int nid = entry_to_nid(entry);
> > struct mem_cgroup *memcg;
> >
> > /*
> > @@ -1758,7 +1758,7 @@ static bool zswap_store_pages(struct folio *folio,
> > * an incoherent entry.
> > */
> > if (likely(entry->length))
> > - zswap_lru_add(&zswap_list_lru, entry);
> > + zswap_lru_add(&zswap_list_lru, entry, nid);
> > }
> >
> > return true;
> > --
> >
> > I believe this will add the entry to the LRU node of the folio being
> > compressed. If so, we may be able to avoid adding an int field to
> > struct zswap_entry?
>
> Hmm that might not work for zswap_lru_del() :(
>
> zswap_entry_free() might be called in context where we do not have
> access to the node information (zswap_load()) for e.g.
I was thinking that zswap_lru_del() would follow a similar approach,
i.e., zswap_load() would pass the folio's nid to zswap_lru_del(), but
you're right, this would not work if the process faulting in the page
is running on a different node than the one that stored the page.
>
> Another alternative: can we instead determine the node from the
> compressed object's storage? i.e store zswap_entry in the LRU
> corresponding to the node that holds the compressed data?
>
> You'll probably need the new zsmalloc API to get the node information.
> And can zsmalloc migrate a backing page to a different node? This
> seems complicated...
That's a great idea! It might be worth exploring if our goal is to maintain
parity with the current status for nodes/LRU/shrinker.
Good point about zsmalloc migrating a backing page to a different node:
although wouldn't this be a problem with the current status quo also?
To summarize my understanding: the current approach ensures that the
NUMA node in which the page was allocated, is the one that will hold the
compressed data (the zsmalloc commit log you shared), and is the node
which, under memory pressure, will cause the entry to be written back.
The entry being allocated on the same NUMA node as the page being
stored in zswap, imho is a "mechanism" to achieve the above. When the
page is faulted in, it is possible that the process has migrated to a different
node, and the folio is now assigned a different nid. IOW, there is no more
significance to the entry's nid than to facilitate the current approach, IIUC.
I think your suggestion (1) wherein we store the NUMA node as an int field
in the entry can accomplish the same thing. The entry doesn't have to be
allocated on the same node as the page being stored in zswap: we could
let the slab allocator decide this (potentially more optimal system-wide?).
The entry int field could also be more fail-safe than looking up the zsmalloc
node info (which could have migrated the compressed zspage [still need to
verify]).
I think the entry int field might also be cleaner with changes encapsulated
to zswap_lru_add()/del(). If we rely on zsmalloc node derivation, it might
require changes in zswap_decompress(). The downside is we add an int
member to the zswap_entry.
If its Ok with you, can I evaluate the feasibility of (1) and update shortly
after gathering data with usemem30 and kernel_compilation?
I am trying to avoid the latency penalty of not using the bulk allocation
API, and at the same time, ensure we don't change the NUMA node/LRU
lists/shrinker functionality. Based on the data I had gathered recently
in [1], reverting to use kmem_cache_alloc_node() for the batch in
zswap_store_pages() impacts latency considerably.
[1] https://patchwork.kernel.org/comment/26590517/
>
> Taking a step back though, do we have to use the bulk allocation API
> here? Calling the single-allocation version 512 times for a PMD-sized
> page is no worse than the status quo, correct? We can leave this part
> unoptimized for now - in the future, if the use case justifies it, we
> can talk to slab allocator maintainers and ask for guidance on a
> lock-optimized cross-node bulk allocation API.
Definitely. This is a sound fallback strategy if (1) doesn't work out, and
even if it does, we feel that adding an int field to the metadata is not
acceptable/needed. I will make sure to share before/after data with
usemem30 and kernel_compilation with the different options (the
int field in zswap_entry, bulk vs. single allocation).
Thanks,
Kanchana
^ permalink raw reply [flat|nested] 56+ messages in thread
* Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches.
2025-10-15 22:15 ` Sridhar, Kanchana P
@ 2025-10-15 22:24 ` Yosry Ahmed
2025-10-15 22:36 ` Nhat Pham
0 siblings, 1 reply; 56+ messages in thread
From: Yosry Ahmed @ 2025-10-15 22:24 UTC (permalink / raw)
To: Sridhar, Kanchana P
Cc: Nhat Pham, linux-kernel, linux-mm, hannes, chengming.zhou,
usamaarif642, ryan.roberts, 21cnbao, ying.huang, akpm,
senozhatsky, sj, kasong, linux-crypto, herbert, davem, clabbe,
ardb, ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius,
Feghali, Wajdi K, Gopal, Vinodh
On Wed, Oct 15, 2025 at 10:15:12PM +0000, Sridhar, Kanchana P wrote:
>
> > -----Original Message-----
> > From: Nhat Pham <nphamcs@gmail.com>
> > Sent: Wednesday, October 15, 2025 10:04 AM
> > To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> > Cc: Yosry Ahmed <yosry.ahmed@linux.dev>; linux-kernel@vger.kernel.org;
> > linux-mm@kvack.org; hannes@cmpxchg.org; chengming.zhou@linux.dev;
> > usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> > ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> > senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> > crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> > davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> > ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> > <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>;
> > Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> > <vinodh.gopal@intel.com>
> > Subject: Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a
> > large folio in batches.
> >
> > On Tue, Oct 14, 2025 at 8:42 PM Sridhar, Kanchana P
> > <kanchana.p.sridhar@intel.com> wrote:
> > >
> > >
> > > > -----Original Message-----
> > > > From: Nhat Pham <nphamcs@gmail.com>
> > > > Sent: Tuesday, October 14, 2025 9:35 AM
> > > > To: Yosry Ahmed <yosry.ahmed@linux.dev>
> > > > Cc: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>; linux-
> > > > kernel@vger.kernel.org; linux-mm@kvack.org; hannes@cmpxchg.org;
> > > > chengming.zhou@linux.dev; usamaarif642@gmail.com;
> > > > ryan.roberts@arm.com; 21cnbao@gmail.com;
> > > > ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> > > > senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> > > > crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> > > > davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> > > > ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> > > > <kristen.c.accardi@intel.com>; Gomes, Vinicius
> > <vinicius.gomes@intel.com>;
> > > > Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> > > > <vinodh.gopal@intel.com>
> > > > Subject: Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a
> > > > large folio in batches.
> > > >
> > > > On Tue, Oct 14, 2025 at 8:29 AM Yosry Ahmed <yosry.ahmed@linux.dev>
> > > > wrote:
> > > > >
> > > > > [..]
> > > > > > > > @@ -158,6 +161,8 @@ struct zswap_pool {
> > > > > > > > struct work_struct release_work;
> > > > > > > > struct hlist_node node;
> > > > > > > > char tfm_name[CRYPTO_MAX_ALG_NAME];
> > > > > > > > + u8 compr_batch_size;
> > > > > > > > + u8 store_batch_size;
> > > > > > >
> > > > > > > I don't think we need to store store_batch_size, seems trivial to
> > > > > > > calculate at store time (perhaps in a helper).
> > > > > > >
> > > > > > > Taking a step back, is there any benefit to limiting store_batch_size
> > to
> > > > > > > compr_batch_size? Is there a disadvantage to using
> > > > > > > ZSWAP_MAX_BATCH_SIZE
> > > > > > > even if it's higher than the HW compression batch size?
> > > > > >
> > > > > > Thanks Yosry, for the code review comments. I had a discussion with
> > > > > > Barry earlier on these very same topics as follow up to his review
> > > > comments
> > > > > > for v11, starting with [1]. Can you please go through the rationale for
> > > > > > these design choices, and let me know if you have any questions:
> > > > > >
> > > > > > [1]: https://patchwork.kernel.org/comment/26530319/
> > > > >
> > > > > I am surprised that calculating the value in zswap_store() causes a
> > > > > regression, but I am fine with keeping the precalculation in this case.
> > > > >
> > > > > I think there's a bigger problem here tho, more below.
> > > > >
> > > > > > > > + */
> > > > > > > > +static __always_inline int zswap_entries_cache_alloc_batch(void
> > > > > > > **entries,
> > > > > > > > + unsigned int
> > > > > > > nr_entries,
> > > > > > > > + gfp_t gfp)
> > > > > > > > +{
> > > > > > > > + return kmem_cache_alloc_bulk(zswap_entry_cache, gfp,
> > nr_entries,
> > > > > > > entries);
> > > > > > >
> > > > > > > We currently use kmem_cache_alloc_node() in
> > > > zswap_entry_cache_alloc() to
> > > > > > > allocate the entry on the same node as the compressed page. We
> > use
> > > > > > > entry_to_nid() to get the node for LRU operations.
> > > > > > >
> > > > > > > This breaks that assumption.
> > > > > >
> > > > > > You bring up a good point. I was looking at the code in slub.c and my
> > > > > > understanding thus far is that both, bulk allocations and
> > > > kmem_cache_alloc_node()
> > > > > > allocations are made from a per-CPU "cpu_slab" that is allocated by
> > SLUB.
> > > > > >
> > > > > > IIUC, the concern you are raising is in the mainline, the entry is
> > allocated
> > > > on
> > > > > > the same node as the compressed page, and gets added to the LRU list
> > of
> > > > that
> > > > > > node. IOW, the node to which the compressed page belongs is the one
> > to
> > > > whose
> > > > > > LRU the entry will be added.
> > > > > >
> > > > > > With this patch, with kmem_cache_alloc_bulk(), the entry will be
> > created
> > > > on
> > > > > > the per-CPU slab of the CPU on which zswap_store() is called and will
> > be
> > > > > > added to the LRU of that per-CPU slab's NUMA node. Hence, the end
> > > > result
> > > > > > could potentially be that the zswap_entry for a page could potentially
> > be
> > > > > > on a different NUMA node/memcg than the page's NUMA node.
> > > >
> > > > I think only the NUMA node is the problem, not the memcg.
> > > >
> > > > > >
> > > > > > This is my thinking as to how this will impact the zswap shrinker:
> > > > > >
> > > > > > 1) memcg shrinker: if the memcg the entry ends up in is on the
> > > > zswap_list_lru,
> > > > > > the entry will be written back.
> > > > > > 2) Global shrinker: will cycle through all memcg's that have pages in
> > the
> > > > > > zswap_list_lru, and the entry will be written back.
> > > > > >
> > > > > > Based on this, it is not clear to me if there is a problem, and would like
> > to
> > > > > > request you, Nhat and others to provide insights as well.
> > > > > >
> > > > > > Interestingly, most of the code in slub.c has unlikely(!node_match(slab,
> > > > node)).
> > > > > > Does this imply some higher level mm slab allocation requirements?
> > > > > >
> > > > > > I am Ok with just calling zswap_entry_cache_alloc() for "nr_pages" if
> > we
> > > > > > think this would be more correct.
> > > > >
> > > > > I saw your other response as well, but I think one thing is not clear
> > > > > here. The zswap entry will get written back "eventually", sure, but
> > > > > that's not the problem.
> > > > >
> > > > > If the zswap entry is on the wrong node lru, two things happen:
> > > > > (a) When the "right" node is under memory pressure, we cannot free
> > this
> > > > > entry by writing it back since it's not available in the lru.
> > > > > (b) When the "wrong" node is under memory pressure, it will potentially
> > > > > writeback entries from other nodes AND report them as being freed
> > > > > from this node.
> > > > >
> > > > > Both (a) and (b) cause less effective reclaim from the zswap shrinker.
> > > > > Additionally (b) causes the shrinker to report the wrong amount of freed
> > > > > memory from the node. While this may not be significant today, it's very
> > > > > possible that more heuristics start relying on this number in the
> > > > > future.
> > > > >
> > > > > I don't believe we should put zswap entries on the wrong LRU, but I will
> > > > > defer to Nhat for the final verdict if he has a different opinion.
> > > >
> > > > Oh shoot. Yeah I missed that part.
> > > >
> > > > In the past, we sort of did not care - zswap was very poorly designed
> > > > for NUMA architecture in general, and most of our test setups have
> > > > been single-node, so these kinds of discrepancies did not show up in
> > > > performance numbers.
> > > >
> > > > But we are getting more multi-node systems:
> > > >
> > > > 1. Bigger hosts (memory-wise) tend to also have more than one nodes.
> > > > It scales better that way (especially because a lot of structures and
> > > > locks protecting them are node-partitioned).
> > > >
> > > > 2. We have also seen different memory media that are often expressed
> > > > to the kernel as nodes: CXL, GPU memory, etc.
> > > >
> > > > This will necessitate tightening memory placement. We recently had to
> > > > fix one such issue:
> > > >
> > > >
> > https://github.com/torvalds/linux/commit/56e5a103a721d0ef139bba7ff3d3a
> > > > da6c8217d5b
> > > >
> > > > So I'm a bit nervous about this change, which will make us use the wrong
> > > > LRU...
> > > >
> > > > Some work around:
> > > >
> > > > 1. Can we squeeze an extra int field anywhere in struct zswap_entry?
> > > >
> > > > 2. Can we pump nid all the way to zswap_lru_add()?
> > > >
> > > > This is still not 100% ideal - the metadata (struct zswap_entry) will
> > > > still be allocated on the wrong node. But at least the data are
> > > > properly managed, i.e on the right LRU.
> > >
> > > Thanks, Nhat and Yosry for the discussion. Thank you Nhat, for the
> > > zsmalloc change log reference and for the work arounds!
> > >
> > > Following your suggestion in (2), can we pass in the folio's nid from
> > > zswap_store_pages() to zswap_lru_add(), as follows:
> > >
> > > diff --git a/mm/zswap.c b/mm/zswap.c
> > > index 263bc6d7f5c6..44665deece80 100644
> > > --- a/mm/zswap.c
> > > +++ b/mm/zswap.c
> > > @@ -694,9 +694,9 @@ static inline int entry_to_nid(struct zswap_entry
> > *entry)
> > > return page_to_nid(virt_to_page(entry));
> > > }
> > >
> > > -static void zswap_lru_add(struct list_lru *list_lru, struct zswap_entry
> > *entry)
> > > +static void zswap_lru_add(struct list_lru *list_lru, struct zswap_entry
> > *entry,
> > > + int nid)
> > > {
> > > - int nid = entry_to_nid(entry);
> > > struct mem_cgroup *memcg;
> > >
> > > /*
> > > @@ -1758,7 +1758,7 @@ static bool zswap_store_pages(struct folio *folio,
> > > * an incoherent entry.
> > > */
> > > if (likely(entry->length))
> > > - zswap_lru_add(&zswap_list_lru, entry);
> > > + zswap_lru_add(&zswap_list_lru, entry, nid);
> > > }
> > >
> > > return true;
> > > --
> > >
> > > I believe this will add the entry to the LRU node of the folio being
> > > compressed. If so, we may be able to avoid adding an int field to
> > > struct zswap_entry?
> >
> > Hmm that might not work for zswap_lru_del() :(
> >
> > zswap_entry_free() might be called in context where we do not have
> > access to the node information (zswap_load()) for e.g.
>
> I was thinking that zswap_lru_del() would follow a similar approach,
> i.e., zswap_load() would pass the folio's nid to zswap_lru_del(), but
> you're right, this would not work if the process faulting in the page
> is running on a different node than the one that stored the page.
>
> >
> > Another alternative: can we instead determine the node from the
> > compressed object's storage? i.e store zswap_entry in the LRU
> > corresponding to the node that holds the compressed data?
> >
> > You'll probably need the new zsmalloc API to get the node information.
> > And can zsmalloc migrate a backing page to a different node? This
> > seems complicated...
>
> That's a great idea! It might be worth exploring if our goal is to maintain
> parity with the current status for nodes/LRU/shrinker.
>
> Good point about zsmalloc migrating a backing page to a different node:
> although wouldn't this be a problem with the current status quo also?
>
> To summarize my understanding: the current approach ensures that the
> NUMA node in which the page was allocated, is the one that will hold the
> compressed data (the zsmalloc commit log you shared), and is the node
> which, under memory pressure, will cause the entry to be written back.
>
> The entry being allocated on the same NUMA node as the page being
> stored in zswap, imho is a "mechanism" to achieve the above. When the
> page is faulted in, it is possible that the process has migrated to a different
> node, and the folio is now assigned a different nid. IOW, there is no more
> significance to the entry's nid than to facilitate the current approach, IIUC.
>
> I think your suggestion (1) wherein we store the NUMA node as an int field
> in the entry can accomplish the same thing. The entry doesn't have to be
> allocated on the same node as the page being stored in zswap: we could
> let the slab allocator decide this (potentially more optimal system-wide?).
>
> The entry int field could also be more fail-safe than looking up the zsmalloc
> node info (which could have migrated the compressed zspage [still need to
> verify]).
>
> I think the entry int field might also be cleaner with changes encapsulated
> to zswap_lru_add()/del(). If we rely on zsmalloc node derivation, it might
> require changes in zswap_decompress(). The downside is we add an int
> member to the zswap_entry.
>
> If its Ok with you, can I evaluate the feasibility of (1) and update shortly
> after gathering data with usemem30 and kernel_compilation?
> I am trying to avoid the latency penalty of not using the bulk allocation
> API, and at the same time, ensure we don't change the NUMA node/LRU
> lists/shrinker functionality. Based on the data I had gathered recently
> in [1], reverting to use kmem_cache_alloc_node() for the batch in
> zswap_store_pages() impacts latency considerably.
>
> [1] https://patchwork.kernel.org/comment/26590517/
>
> >
> > Taking a step back though, do we have to use the bulk allocation API
> > here? Calling the single-allocation version 512 times for a PMD-sized
> > page is no worse than the status quo, correct? We can leave this part
> > unoptimized for now - in the future, if the use case justifies it, we
> > can talk to slab allocator maintainers and ask for guidance on a
> > lock-optimized cross-node bulk allocation API.
>
> Definitely. This is a sound fallback strategy if (1) doesn't work out, and
> even if it does, we feel that adding an int field to the metadata is not
> acceptable/needed. I will make sure to share before/after data with
> usemem30 and kernel_compilation with the different options (the
> int field in zswap_entry, bulk vs. single allocation).
I am against increasing the size of struct zswap_entry. On x86_64, there
is a 3 byte hole after 'referenced'. We can technically use that,
although the node id is usually an int, which is 4 bytes on x86_64.
In practice, I think 2 bytes (i.e. short) should be enough, but it will
be ugly to cast the node id to a short. We should at least WARN on
overflow.
Or we can take the simple route and drop the bulk allocation.
^ permalink raw reply [flat|nested] 56+ messages in thread
* Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches.
2025-10-15 22:24 ` Yosry Ahmed
@ 2025-10-15 22:36 ` Nhat Pham
2025-10-16 0:56 ` Sridhar, Kanchana P
0 siblings, 1 reply; 56+ messages in thread
From: Nhat Pham @ 2025-10-15 22:36 UTC (permalink / raw)
To: Yosry Ahmed
Cc: Sridhar, Kanchana P, linux-kernel, linux-mm, hannes,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, herbert, davem,
clabbe, ardb, ebiggers, surenb, Accardi, Kristen C, Gomes,
Vinicius, Feghali, Wajdi K, Gopal, Vinodh
On Wed, Oct 15, 2025 at 3:24 PM Yosry Ahmed <yosry.ahmed@linux.dev> wrote:
>
> On Wed, Oct 15, 2025 at 10:15:12PM +0000, Sridhar, Kanchana P wrote:
>
> I am against increasing the size of struct zswap_entry. On x86_64, there
> is a 3 byte hole after 'referenced'. We can technically use that,
> although the node id is usually an int, which is 4 bytes on x86_64.
>
> In practice, I think 2 bytes (i.e. short) should be enough, but it will
> be ugly to cast the node id to a short. We should at least WARN on
> overflow.
Can we pack length and referenced using bit fields? I assume
compressed length cannot exceed 2^31 - 1? :)
So, something along the line of:
struct {
unsigned int length:31;
bool referenced:1;
}
That way, we save up another 4 bytes hole, which can be repurposed for
the node's id.
Does that work? I'm not very experienced with this magickery - please
fact check me :)
>
> Or we can take the simple route and drop the bulk allocation.
Or this :)
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches.
2025-10-15 22:36 ` Nhat Pham
@ 2025-10-16 0:56 ` Sridhar, Kanchana P
2025-10-31 22:15 ` Sridhar, Kanchana P
0 siblings, 1 reply; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-10-16 0:56 UTC (permalink / raw)
To: Nhat Pham, Yosry Ahmed
Cc: linux-kernel, linux-mm, hannes, chengming.zhou, usamaarif642,
ryan.roberts, 21cnbao, ying.huang, akpm, senozhatsky, sj, kasong,
linux-crypto, herbert, davem, clabbe, ardb, ebiggers, surenb,
Accardi, Kristen C, Gomes, Vinicius, Feghali, Wajdi K, Gopal,
Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Nhat Pham <nphamcs@gmail.com>
> Sent: Wednesday, October 15, 2025 3:36 PM
> To: Yosry Ahmed <yosry.ahmed@linux.dev>
> Cc: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>; linux-
> kernel@vger.kernel.org; linux-mm@kvack.org; hannes@cmpxchg.org;
> chengming.zhou@linux.dev; usamaarif642@gmail.com;
> ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>;
> Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> <vinodh.gopal@intel.com>
> Subject: Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a
> large folio in batches.
>
> On Wed, Oct 15, 2025 at 3:24 PM Yosry Ahmed <yosry.ahmed@linux.dev>
> wrote:
> >
> > On Wed, Oct 15, 2025 at 10:15:12PM +0000, Sridhar, Kanchana P wrote:
> >
> > I am against increasing the size of struct zswap_entry. On x86_64, there
> > is a 3 byte hole after 'referenced'. We can technically use that,
> > although the node id is usually an int, which is 4 bytes on x86_64.
> >
> > In practice, I think 2 bytes (i.e. short) should be enough, but it will
> > be ugly to cast the node id to a short. We should at least WARN on
> > overflow.
>
> Can we pack length and referenced using bit fields? I assume
> compressed length cannot exceed 2^31 - 1? :)
>
> So, something along the line of:
>
> struct {
> unsigned int length:31;
> bool referenced:1;
> }
>
> That way, we save up another 4 bytes hole, which can be repurposed for
> the node's id.
>
> Does that work? I'm not very experienced with this magickery - please
> fact check me :)
>
>
> >
> > Or we can take the simple route and drop the bulk allocation.
>
> Or this :)
Thanks Nhat and Yosry, for sharing these constraints and ideas.
Sure, I will see if I can use the bit fields approach and will make sure
to not increase the size of struct zswap_entry. If this doesn't work
for some reason, we have the simple route.
Best regards,
Kanchana
^ permalink raw reply [flat|nested] 56+ messages in thread
* Re: [PATCH v12 19/23] crypto: acomp - Add crypto_acomp_batch_size() to get an algorithm's batch-size.
2025-09-26 3:34 ` [PATCH v12 19/23] crypto: acomp - Add crypto_acomp_batch_size() to get an algorithm's batch-size Kanchana P Sridhar
@ 2025-10-17 1:04 ` Herbert Xu
2025-10-17 4:02 ` Sridhar, Kanchana P
0 siblings, 1 reply; 56+ messages in thread
From: Herbert Xu @ 2025-10-17 1:04 UTC (permalink / raw)
To: Kanchana P Sridhar
Cc: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, davem, clabbe, ardb,
ebiggers, surenb, kristen.c.accardi, vinicius.gomes,
wajdi.k.feghali, vinodh.gopal
On Thu, Sep 25, 2025 at 08:34:58PM -0700, Kanchana P Sridhar wrote:
>
> diff --git a/include/crypto/acompress.h b/include/crypto/acompress.h
> index 0f1334168f1b..e94046529e46 100644
> --- a/include/crypto/acompress.h
> +++ b/include/crypto/acompress.h
> @@ -108,6 +108,8 @@ struct acomp_req {
> *
> * @compress: Function performs a compress operation
> * @decompress: Function performs a de-compress operation
> + * @batch_size: Maximum batch-size for batching compress/decompress
> + * operations.
> * @reqsize: Context size for (de)compression requests
> * @fb: Synchronous fallback tfm
> * @base: Common crypto API algorithm data structure
> @@ -115,6 +117,7 @@ struct acomp_req {
> struct crypto_acomp {
> int (*compress)(struct acomp_req *req);
> int (*decompress)(struct acomp_req *req);
> + unsigned int batch_size;
There is no need to store the batch_size in the tfm. Just return
the value from acomp_alg directly.
Thanks,
--
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
^ permalink raw reply [flat|nested] 56+ messages in thread
* Re: [PATCH v12 14/23] crypto: iaa - IAA Batching for parallel compressions/decompressions.
2025-09-26 3:34 ` [PATCH v12 14/23] crypto: iaa - IAA Batching for parallel compressions/decompressions Kanchana P Sridhar
@ 2025-10-17 1:09 ` Herbert Xu
2025-10-17 4:04 ` Sridhar, Kanchana P
0 siblings, 1 reply; 56+ messages in thread
From: Herbert Xu @ 2025-10-17 1:09 UTC (permalink / raw)
To: Kanchana P Sridhar
Cc: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, davem, clabbe, ardb,
ebiggers, surenb, kristen.c.accardi, vinicius.gomes,
wajdi.k.feghali, vinodh.gopal
On Thu, Sep 25, 2025 at 08:34:53PM -0700, Kanchana P Sridhar wrote:
>
> +__always_inline int iaa_comp_compress_batch(
> + enum iaa_mode mode,
> + struct iaa_req *parent_req,
> + unsigned int unit_size)
> +{
> + return iaa_comp_acompress_batch(iaa_ctx[mode], parent_req, unit_size);
> +}
> +EXPORT_SYMBOL_GPL(iaa_comp_compress_batch);
> +
> +__always_inline int iaa_comp_decompress_batch(
> + enum iaa_mode mode,
> + struct iaa_req *parent_req,
> + unsigned int unit_size)
> +{
> + return iaa_comp_adecompress_batch(iaa_ctx[mode], parent_req, unit_size);
> +}
> +EXPORT_SYMBOL_GPL(iaa_comp_decompress_batch);
What are these for?
Cheers,
--
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 19/23] crypto: acomp - Add crypto_acomp_batch_size() to get an algorithm's batch-size.
2025-10-17 1:04 ` Herbert Xu
@ 2025-10-17 4:02 ` Sridhar, Kanchana P
0 siblings, 0 replies; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-10-17 4:02 UTC (permalink / raw)
To: Herbert Xu
Cc: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, davem, clabbe, ardb,
ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius, Feghali,
Wajdi K, Gopal, Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Herbert Xu <herbert@gondor.apana.org.au>
> Sent: Thursday, October 16, 2025 6:04 PM
> To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> hannes@cmpxchg.org; yosry.ahmed@linux.dev; nphamcs@gmail.com;
> chengming.zhou@linux.dev; usamaarif642@gmail.com;
> ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; davem@davemloft.net; clabbe@baylibre.com;
> ardb@kernel.org; ebiggers@google.com; surenb@google.com; Accardi,
> Kristen C <kristen.c.accardi@intel.com>; Gomes, Vinicius
> <vinicius.gomes@intel.com>; Feghali, Wajdi K <wajdi.k.feghali@intel.com>;
> Gopal, Vinodh <vinodh.gopal@intel.com>
> Subject: Re: [PATCH v12 19/23] crypto: acomp - Add
> crypto_acomp_batch_size() to get an algorithm's batch-size.
>
> On Thu, Sep 25, 2025 at 08:34:58PM -0700, Kanchana P Sridhar wrote:
> >
> > diff --git a/include/crypto/acompress.h b/include/crypto/acompress.h
> > index 0f1334168f1b..e94046529e46 100644
> > --- a/include/crypto/acompress.h
> > +++ b/include/crypto/acompress.h
> > @@ -108,6 +108,8 @@ struct acomp_req {
> > *
> > * @compress: Function performs a compress operation
> > * @decompress: Function performs a de-compress operation
> > + * @batch_size: Maximum batch-size for batching
> compress/decompress
> > + * operations.
> > * @reqsize: Context size for (de)compression requests
> > * @fb: Synchronous fallback tfm
> > * @base: Common crypto API algorithm data structure
> > @@ -115,6 +117,7 @@ struct acomp_req {
> > struct crypto_acomp {
> > int (*compress)(struct acomp_req *req);
> > int (*decompress)(struct acomp_req *req);
> > + unsigned int batch_size;
>
> There is no need to store the batch_size in the tfm. Just return
> the value from acomp_alg directly.
I see. Ok, will do so.
Thanks,
Kanchana
>
> Thanks,
> --
> Email: Herbert Xu <herbert@gondor.apana.org.au>
> Home Page: http://gondor.apana.org.au/~herbert/
> PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 14/23] crypto: iaa - IAA Batching for parallel compressions/decompressions.
2025-10-17 1:09 ` Herbert Xu
@ 2025-10-17 4:04 ` Sridhar, Kanchana P
2025-10-17 4:09 ` Herbert Xu
2025-10-17 5:12 ` Sergey Senozhatsky
0 siblings, 2 replies; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-10-17 4:04 UTC (permalink / raw)
To: Herbert Xu
Cc: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, davem, clabbe, ardb,
ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius, Feghali,
Wajdi K, Gopal, Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Herbert Xu <herbert@gondor.apana.org.au>
> Sent: Thursday, October 16, 2025 6:10 PM
> To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> hannes@cmpxchg.org; yosry.ahmed@linux.dev; nphamcs@gmail.com;
> chengming.zhou@linux.dev; usamaarif642@gmail.com;
> ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; davem@davemloft.net; clabbe@baylibre.com;
> ardb@kernel.org; ebiggers@google.com; surenb@google.com; Accardi,
> Kristen C <kristen.c.accardi@intel.com>; Gomes, Vinicius
> <vinicius.gomes@intel.com>; Feghali, Wajdi K <wajdi.k.feghali@intel.com>;
> Gopal, Vinodh <vinodh.gopal@intel.com>
> Subject: Re: [PATCH v12 14/23] crypto: iaa - IAA Batching for parallel
> compressions/decompressions.
>
> On Thu, Sep 25, 2025 at 08:34:53PM -0700, Kanchana P Sridhar wrote:
> >
> > +__always_inline int iaa_comp_compress_batch(
> > + enum iaa_mode mode,
> > + struct iaa_req *parent_req,
> > + unsigned int unit_size)
> > +{
> > + return iaa_comp_acompress_batch(iaa_ctx[mode], parent_req,
> unit_size);
> > +}
> > +EXPORT_SYMBOL_GPL(iaa_comp_compress_batch);
> > +
> > +__always_inline int iaa_comp_decompress_batch(
> > + enum iaa_mode mode,
> > + struct iaa_req *parent_req,
> > + unsigned int unit_size)
> > +{
> > + return iaa_comp_adecompress_batch(iaa_ctx[mode], parent_req,
> unit_size);
> > +}
> > +EXPORT_SYMBOL_GPL(iaa_comp_decompress_batch);
>
> What are these for?
Hi Herbert,
These are for non-crypto users such as zram. I plan to submit an RFC for
zram that enables IAA as a zcomp backend with batching support. These
API will be used for this purpose.
Thanks,
Kanchana
>
> Cheers,
> --
> Email: Herbert Xu <herbert@gondor.apana.org.au>
> Home Page: http://gondor.apana.org.au/~herbert/
> PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
^ permalink raw reply [flat|nested] 56+ messages in thread
* Re: [PATCH v12 14/23] crypto: iaa - IAA Batching for parallel compressions/decompressions.
2025-10-17 4:04 ` Sridhar, Kanchana P
@ 2025-10-17 4:09 ` Herbert Xu
2025-10-17 5:12 ` Sergey Senozhatsky
1 sibling, 0 replies; 56+ messages in thread
From: Herbert Xu @ 2025-10-17 4:09 UTC (permalink / raw)
To: Sridhar, Kanchana P
Cc: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, davem, clabbe, ardb,
ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius, Feghali,
Wajdi K, Gopal, Vinodh
On Fri, Oct 17, 2025 at 04:04:56AM +0000, Sridhar, Kanchana P wrote:
>
> These are for non-crypto users such as zram. I plan to submit an RFC for
> zram that enables IAA as a zcomp backend with batching support. These
> API will be used for this purpose.
We should be converting zram to ahash instead of adding this
interface.
Thanks,
--
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
^ permalink raw reply [flat|nested] 56+ messages in thread
* Re: [PATCH v12 14/23] crypto: iaa - IAA Batching for parallel compressions/decompressions.
2025-10-17 4:04 ` Sridhar, Kanchana P
2025-10-17 4:09 ` Herbert Xu
@ 2025-10-17 5:12 ` Sergey Senozhatsky
2025-10-17 5:49 ` Sridhar, Kanchana P
1 sibling, 1 reply; 56+ messages in thread
From: Sergey Senozhatsky @ 2025-10-17 5:12 UTC (permalink / raw)
To: Sridhar, Kanchana P
Cc: Herbert Xu, linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, davem, clabbe, ardb,
ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius, Feghali,
Wajdi K, Gopal, Vinodh
On (25/10/17 04:04), Sridhar, Kanchana P wrote:
> > What are these for?
>
> Hi Herbert,
>
> These are for non-crypto users such as zram.
I agree with Herbert, I'd keep zram out of this discussion for now.
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 14/23] crypto: iaa - IAA Batching for parallel compressions/decompressions.
2025-10-17 5:12 ` Sergey Senozhatsky
@ 2025-10-17 5:49 ` Sridhar, Kanchana P
0 siblings, 0 replies; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-10-17 5:49 UTC (permalink / raw)
To: Sergey Senozhatsky
Cc: Herbert Xu, linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, sj, kasong, linux-crypto, davem, clabbe, ardb, ebiggers,
surenb, Accardi, Kristen C, Gomes, Vinicius, Feghali, Wajdi K,
Gopal, Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Sergey Senozhatsky <senozhatsky@chromium.org>
> Sent: Thursday, October 16, 2025 10:12 PM
> To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> Cc: Herbert Xu <herbert@gondor.apana.org.au>; linux-
> kernel@vger.kernel.org; linux-mm@kvack.org; hannes@cmpxchg.org;
> yosry.ahmed@linux.dev; nphamcs@gmail.com; chengming.zhou@linux.dev;
> usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; davem@davemloft.net; clabbe@baylibre.com;
> ardb@kernel.org; ebiggers@google.com; surenb@google.com; Accardi,
> Kristen C <kristen.c.accardi@intel.com>; Gomes, Vinicius
> <vinicius.gomes@intel.com>; Feghali, Wajdi K <wajdi.k.feghali@intel.com>;
> Gopal, Vinodh <vinodh.gopal@intel.com>
> Subject: Re: [PATCH v12 14/23] crypto: iaa - IAA Batching for parallel
> compressions/decompressions.
>
> On (25/10/17 04:04), Sridhar, Kanchana P wrote:
> > > What are these for?
> >
> > Hi Herbert,
> >
> > These are for non-crypto users such as zram.
>
> I agree with Herbert, I'd keep zram out of this discussion for now.
Sounds good, I will drop the IAA API intended for zram when I post v13
of this patch series.
Thanks Herbert and Sergey.
Best regards,
Kanchana
^ permalink raw reply [flat|nested] 56+ messages in thread
* Re: [PATCH v12 23/23] mm: zswap: Batched zswap_compress() with compress batching of large folios.
2025-09-26 3:35 ` [PATCH v12 23/23] mm: zswap: Batched zswap_compress() with compress batching of large folios Kanchana P Sridhar
@ 2025-10-22 5:18 ` Herbert Xu
2025-10-31 22:18 ` Sridhar, Kanchana P
0 siblings, 1 reply; 56+ messages in thread
From: Herbert Xu @ 2025-10-22 5:18 UTC (permalink / raw)
To: Kanchana P Sridhar
Cc: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, davem, clabbe, ardb,
ebiggers, surenb, kristen.c.accardi, vinicius.gomes,
wajdi.k.feghali, vinodh.gopal
On Thu, Sep 25, 2025 at 08:35:02PM -0700, Kanchana P Sridhar wrote:
>
> Another initialization level optimization that happens in
> zswap_cpu_comp_prepare() is, only for batching compressors, we allocate
> extra memory for "batch-size" int pointers in the
> @acomp_ctx->req->__ctx[] that are statically set to track the output SG
> lists' lengths. This optimization was necessary to avoid the latency
There is no need for this allocation. Just store the output lengths
in each destination scatterlist entry.
Before the call to acomp, the destination SG list should contain as
many elements as the number of units. On return, the dst lengths
should be stored in each destination SG entry.
Cheers,
--
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches.
2025-10-16 0:56 ` Sridhar, Kanchana P
@ 2025-10-31 22:15 ` Sridhar, Kanchana P
0 siblings, 0 replies; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-10-31 22:15 UTC (permalink / raw)
To: Nhat Pham, Yosry Ahmed
Cc: linux-kernel, linux-mm, hannes, chengming.zhou, usamaarif642,
ryan.roberts, 21cnbao, ying.huang, akpm, senozhatsky, sj, kasong,
linux-crypto, herbert, davem, clabbe, ardb, ebiggers, surenb,
Accardi, Kristen C, Gomes, Vinicius, Feghali, Wajdi K, Gopal,
Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> Sent: Wednesday, October 15, 2025 5:57 PM
> To: Nhat Pham <nphamcs@gmail.com>; Yosry Ahmed
> <yosry.ahmed@linux.dev>
> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> hannes@cmpxchg.org; chengming.zhou@linux.dev;
> usamaarif642@gmail.com; ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> <kristen.c.accardi@intel.com>; Gomes, Vinicius <vinicius.gomes@intel.com>;
> Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> <vinodh.gopal@intel.com>; Sridhar, Kanchana P
> <kanchana.p.sridhar@intel.com>
> Subject: RE: [PATCH v12 22/23] mm: zswap: zswap_store() will process a
> large folio in batches.
>
>
> > -----Original Message-----
> > From: Nhat Pham <nphamcs@gmail.com>
> > Sent: Wednesday, October 15, 2025 3:36 PM
> > To: Yosry Ahmed <yosry.ahmed@linux.dev>
> > Cc: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>; linux-
> > kernel@vger.kernel.org; linux-mm@kvack.org; hannes@cmpxchg.org;
> > chengming.zhou@linux.dev; usamaarif642@gmail.com;
> > ryan.roberts@arm.com; 21cnbao@gmail.com;
> > ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> > senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> > crypto@vger.kernel.org; herbert@gondor.apana.org.au;
> > davem@davemloft.net; clabbe@baylibre.com; ardb@kernel.org;
> > ebiggers@google.com; surenb@google.com; Accardi, Kristen C
> > <kristen.c.accardi@intel.com>; Gomes, Vinicius
> <vinicius.gomes@intel.com>;
> > Feghali, Wajdi K <wajdi.k.feghali@intel.com>; Gopal, Vinodh
> > <vinodh.gopal@intel.com>
> > Subject: Re: [PATCH v12 22/23] mm: zswap: zswap_store() will process a
> > large folio in batches.
> >
> > On Wed, Oct 15, 2025 at 3:24 PM Yosry Ahmed <yosry.ahmed@linux.dev>
> > wrote:
> > >
> > > On Wed, Oct 15, 2025 at 10:15:12PM +0000, Sridhar, Kanchana P wrote:
> > >
> > > I am against increasing the size of struct zswap_entry. On x86_64, there
> > > is a 3 byte hole after 'referenced'. We can technically use that,
> > > although the node id is usually an int, which is 4 bytes on x86_64.
> > >
> > > In practice, I think 2 bytes (i.e. short) should be enough, but it will
> > > be ugly to cast the node id to a short. We should at least WARN on
> > > overflow.
> >
> > Can we pack length and referenced using bit fields? I assume
> > compressed length cannot exceed 2^31 - 1? :)
> >
> > So, something along the line of:
> >
> > struct {
> > unsigned int length:31;
> > bool referenced:1;
> > }
> >
> > That way, we save up another 4 bytes hole, which can be repurposed for
> > the node's id.
> >
> > Does that work? I'm not very experienced with this magickery - please
> > fact check me :)
> >
> >
> > >
> > > Or we can take the simple route and drop the bulk allocation.
> >
> > Or this :)
>
> Thanks Nhat and Yosry, for sharing these constraints and ideas.
> Sure, I will see if I can use the bit fields approach and will make sure
> to not increase the size of struct zswap_entry. If this doesn't work
> for some reason, we have the simple route.
Hi Nhat and Yosry,
I have been able to get the bit fields approach working, as you suggested.
The size of "struct zswap_entry" is unchanged at 56 bytes with this change
and the addition of the "int nid" field.
zswap_lru_add()/del() use the entry->nid.
I am able to continue using kmem_cache_alloc_bulk() to benefit batching
performance.
I will include this in v13 that I expect to post within a couple of days.
Thanks!
Kanchana
>
> Best regards,
> Kanchana
^ permalink raw reply [flat|nested] 56+ messages in thread
* RE: [PATCH v12 23/23] mm: zswap: Batched zswap_compress() with compress batching of large folios.
2025-10-22 5:18 ` Herbert Xu
@ 2025-10-31 22:18 ` Sridhar, Kanchana P
0 siblings, 0 replies; 56+ messages in thread
From: Sridhar, Kanchana P @ 2025-10-31 22:18 UTC (permalink / raw)
To: Herbert Xu
Cc: linux-kernel, linux-mm, hannes, yosry.ahmed, nphamcs,
chengming.zhou, usamaarif642, ryan.roberts, 21cnbao, ying.huang,
akpm, senozhatsky, sj, kasong, linux-crypto, davem, clabbe, ardb,
ebiggers, surenb, Accardi, Kristen C, Gomes, Vinicius, Feghali,
Wajdi K, Gopal, Vinodh, Sridhar, Kanchana P
> -----Original Message-----
> From: Herbert Xu <herbert@gondor.apana.org.au>
> Sent: Tuesday, October 21, 2025 10:18 PM
> To: Sridhar, Kanchana P <kanchana.p.sridhar@intel.com>
> Cc: linux-kernel@vger.kernel.org; linux-mm@kvack.org;
> hannes@cmpxchg.org; yosry.ahmed@linux.dev; nphamcs@gmail.com;
> chengming.zhou@linux.dev; usamaarif642@gmail.com;
> ryan.roberts@arm.com; 21cnbao@gmail.com;
> ying.huang@linux.alibaba.com; akpm@linux-foundation.org;
> senozhatsky@chromium.org; sj@kernel.org; kasong@tencent.com; linux-
> crypto@vger.kernel.org; davem@davemloft.net; clabbe@baylibre.com;
> ardb@kernel.org; ebiggers@google.com; surenb@google.com; Accardi,
> Kristen C <kristen.c.accardi@intel.com>; Gomes, Vinicius
> <vinicius.gomes@intel.com>; Feghali, Wajdi K <wajdi.k.feghali@intel.com>;
> Gopal, Vinodh <vinodh.gopal@intel.com>
> Subject: Re: [PATCH v12 23/23] mm: zswap: Batched zswap_compress() with
> compress batching of large folios.
>
> On Thu, Sep 25, 2025 at 08:35:02PM -0700, Kanchana P Sridhar wrote:
> >
> > Another initialization level optimization that happens in
> > zswap_cpu_comp_prepare() is, only for batching compressors, we allocate
> > extra memory for "batch-size" int pointers in the
> > @acomp_ctx->req->__ctx[] that are statically set to track the output SG
> > lists' lengths. This optimization was necessary to avoid the latency
>
> There is no need for this allocation. Just store the output lengths
> in each destination scatterlist entry.
Thanks Herbert, for this suggestion. I have included this in the v13 code
that I am currently validating and expect to post an updated patch-set
within a couple of days.
>
> Before the call to acomp, the destination SG list should contain as
> many elements as the number of units. On return, the dst lengths
> should be stored in each destination SG entry.
Yes, confirming that this is the case.
Best regards,
Kanchana
>
> Cheers,
> --
> Email: Herbert Xu <herbert@gondor.apana.org.au>
> Home Page: http://gondor.apana.org.au/~herbert/
> PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
^ permalink raw reply [flat|nested] 56+ messages in thread
end of thread, other threads:[~2025-10-31 22:18 UTC | newest]
Thread overview: 56+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2025-09-26 3:34 [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 01/23] crypto: iaa - Reorganize the iaa_crypto driver code Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 02/23] crypto: iaa - New architecture for IAA device WQ comp/decomp usage & core mapping Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 03/23] crypto: iaa - Simplify, consistency of function parameters, minor stats bug fix Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 04/23] crypto: iaa - Descriptor allocation timeouts with mitigations Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 05/23] crypto: iaa - iaa_wq uses percpu_refs for get/put reference counting Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 06/23] crypto: iaa - Simplify the code flow in iaa_compress() and iaa_decompress() Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 07/23] crypto: iaa - Refactor hardware descriptor setup into separate procedures Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 08/23] crypto: iaa - Simplified, efficient job submissions for non-irq mode Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 09/23] crypto: iaa - Deprecate exporting add/remove IAA compression modes Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 10/23] crypto: iaa - Expect a single scatterlist for a [de]compress request's src/dst Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 11/23] crypto: iaa - Rearchitect the iaa_crypto driver to be usable by zswap and zram Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 12/23] crypto: iaa - Enablers for submitting descriptors then polling for completion Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 13/23] crypto: acomp - Define a unit_size in struct acomp_req to enable batching Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 14/23] crypto: iaa - IAA Batching for parallel compressions/decompressions Kanchana P Sridhar
2025-10-17 1:09 ` Herbert Xu
2025-10-17 4:04 ` Sridhar, Kanchana P
2025-10-17 4:09 ` Herbert Xu
2025-10-17 5:12 ` Sergey Senozhatsky
2025-10-17 5:49 ` Sridhar, Kanchana P
2025-09-26 3:34 ` [PATCH v12 15/23] crypto: iaa - Enable async mode and make it the default Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 16/23] crypto: iaa - Disable iaa_verify_compress by default Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 17/23] crypto: iaa - Submit the two largest source buffers first in decompress batching Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 18/23] crypto: iaa - Add deflate-iaa-dynamic compression mode Kanchana P Sridhar
2025-09-26 3:34 ` [PATCH v12 19/23] crypto: acomp - Add crypto_acomp_batch_size() to get an algorithm's batch-size Kanchana P Sridhar
2025-10-17 1:04 ` Herbert Xu
2025-10-17 4:02 ` Sridhar, Kanchana P
2025-09-26 3:34 ` [PATCH v12 20/23] mm: zswap: Per-CPU acomp_ctx resources exist from pool creation to deletion Kanchana P Sridhar
2025-09-30 15:49 ` Yosry Ahmed
2025-09-30 18:20 ` Sridhar, Kanchana P
2025-09-30 18:29 ` Yosry Ahmed
2025-09-30 21:00 ` Sridhar, Kanchana P
2025-09-30 21:20 ` Yosry Ahmed
2025-09-30 21:56 ` Sridhar, Kanchana P
2025-10-01 15:33 ` Yosry Ahmed
2025-10-01 17:37 ` Sridhar, Kanchana P
2025-09-26 3:35 ` [PATCH v12 21/23] mm: zswap: Consistently use IS_ERR_OR_NULL() to check acomp_ctx resources Kanchana P Sridhar
2025-09-26 3:35 ` [PATCH v12 22/23] mm: zswap: zswap_store() will process a large folio in batches Kanchana P Sridhar
2025-10-01 16:19 ` Yosry Ahmed
2025-10-01 21:20 ` Sridhar, Kanchana P
2025-10-03 19:10 ` Sridhar, Kanchana P
2025-10-13 17:47 ` Sridhar, Kanchana P
2025-10-13 17:58 ` Sridhar, Kanchana P
2025-10-14 15:29 ` Yosry Ahmed
2025-10-14 16:35 ` Nhat Pham
2025-10-15 3:42 ` Sridhar, Kanchana P
2025-10-15 17:04 ` Nhat Pham
2025-10-15 22:15 ` Sridhar, Kanchana P
2025-10-15 22:24 ` Yosry Ahmed
2025-10-15 22:36 ` Nhat Pham
2025-10-16 0:56 ` Sridhar, Kanchana P
2025-10-31 22:15 ` Sridhar, Kanchana P
2025-09-26 3:35 ` [PATCH v12 23/23] mm: zswap: Batched zswap_compress() with compress batching of large folios Kanchana P Sridhar
2025-10-22 5:18 ` Herbert Xu
2025-10-31 22:18 ` Sridhar, Kanchana P
2025-10-13 18:03 ` [PATCH v12 00/23] zswap compression batching with optimized iaa_crypto driver Sridhar, Kanchana P
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox