[PATCH RFC v5 0/3] block: enable RWF_DONTCACHE for block devices

[PATCH RFC v5 0/3] block: enable RWF_DONTCACHE for block devices

[Tal Zussman]

Add support for using RWF_DONTCACHE with block devices.

Dropbehind pruning needs to be done in non-IRQ context, but block
devices complete writeback in IRQ context.

To fix this, we can defer dropbehind invalidation to task context. We
introduce a new BIO_COMPLETE_IN_TASK flag that allows the bio submitter
to request task-context completion of bi_end_io. When bio_endio() sees
this flag in non-task context, it queues the bio to a per-CPU lockless
list and schedules a delayed work item to do bio completion.

Patch 1 adds the BIO_COMPLETE_IN_TASK infrastructure in the block
layer.

Patch 2 wires BIO_COMPLETE_IN_TASK into iomap writeback for dropbehind
folios, removes IOMAP_IOEND_DONTCACHE, and removes the DONTCACHE
workqueue deferral from XFS.

Patch 3 enables RWF_DONTCACHE for block devices, setting
BIO_COMPLETE_IN_TASK in submit_bh_wbc() for the CONFIG_BUFFER_HEAD
path.

This support is useful for databases that operate on raw block devices,
among other userspace applications.

I tested this (with CONFIG_BUFFER_HEAD=y) for reads and writes on a
single block device on a VM, so results may be noisy.

Reads were tested on the root partition with a 45GB range (~2x RAM).
Writes were tested on a disabled swap parition (~1GB) in a memcg of size
244MB to force reclaim pressure.

Results:

===== READS (/dev/nvme0n1p2) =====
 sec   normal MB/s  dontcache MB/s
----  ------------  --------------
   1        1098.6          1609.0
   2        1270.3          1506.6
   3        1093.3          1576.5
   4        1141.8          2393.9
   5        1365.3          2793.8
   6        1324.6          2065.9
   7         879.6          1920.7
   8        1434.1          1662.4
   9        1184.9          1857.9
  10        1166.4          1702.8
  11        1161.4          1653.4
  12        1086.9          1555.4
  13        1198.5          1718.9
  14        1111.9          1752.2
----  ------------  --------------
 avg        1173.7          1828.8  (+56%)

==== WRITES (/dev/nvme0n1p3) =====
 sec   normal MB/s  dontcache MB/s
----  ------------  --------------
   1         692.4          9297.7
   2        4810.8          9342.8
   3        5221.7          2955.2
   4         396.7          8488.3
   5        7249.2          9249.3
   6        6695.4          1376.2
   7         122.9          9125.8
   8        5486.5          9414.7
   9        6921.5          8743.5
  10          27.9          8997.8
----  ------------  --------------
 avg        3762.5          7699.1  (+105%)

---
Changes in v5:
- 1/3: Replace local_lock + bio_list with struct llist, per Dave.
- 1/3: Use delayed_work with 1-jiffie delay, per Dave.
- 1/3: Add dedicated workqueue to avoid deadlocks, per Christoph.
- 1/3: Restructure work function as do/while loop and only schedule work
  originally when the list was previously empty, per Jens.
- 2/3: Delete IOMAP_IOEND_DONTCACHE and its NOMERGE entry, per Matthew
  and Christoph.
- Link to v4: https://lore.kernel.org/r/20260325-blk-dontcache-v4-0-c4b56db43f64@columbia.edu

Changes in v4:
- 1/3: Move dropbehind deferral from folio-level to bio-level using
  BIO_COMPLETE_IN_TASK, per Matthew and Jan.
- 1/3: Work function yields on need_resched() to avoid hogging the CPU,
  per Jan.
- 2/3: New patch. Set BIO_COMPLETE_IN_TASK on iomap writeback bios for
  DONTCACHE folios, removing the need for XFS-specific workqueue
  deferral.
- 3/3: Set BIO_COMPLETE_IN_TASK in submit_bh_wbc() for buffer_head
  path.
- 3/3: Update commit message to mention CONFIG_BUFFER_HEAD=n path.
- Link to v3: https://lore.kernel.org/r/20260227-blk-dontcache-v3-0-cd309ccd5868@columbia.edu

Changes in v3:
- 1/2: Convert dropbehind deferral to per-CPU folio_batches protected by
  local_lock using per-CPU work items, to reduce contention, per Jens.
- 1/2: Call folio_end_dropbehind_irq() directly from
  folio_end_writeback(), per Jens.
- 1/2: Add CPU hotplug dead callback to drain the departing CPU's folio
  batch.
- 2/2: Introduce block_write_begin_iocb(), per Christoph.
- 2/2: Dropped R-b due to changes.
- Link to v2: https://lore.kernel.org/r/20260225-blk-dontcache-v2-0-70e7ac4f7108@columbia.edu

Changes in v2:
- Add R-b from Jan Kara for 2/2.
- Add patch to defer dropbehind completion from IRQ context via a work
  item (1/2).
- Add initial performance numbers to cover letter.
- Link to v1: https://lore.kernel.org/r/20260218-blk-dontcache-v1-1-fad6675ef71f@columbia.edu

---
Tal Zussman (3):
      block: add BIO_COMPLETE_IN_TASK for task-context completion
      iomap: use BIO_COMPLETE_IN_TASK for dropbehind writeback
      block: enable RWF_DONTCACHE for block devices

 block/bio.c                 | 83 ++++++++++++++++++++++++++++++++++++++++++++-
 block/fops.c                |  5 +--
 fs/buffer.c                 | 22 ++++++++++--
 fs/iomap/ioend.c            |  5 +--
 fs/xfs/xfs_aops.c           |  4 ---
 include/linux/blk_types.h   |  7 +++-
 include/linux/buffer_head.h |  3 ++
 include/linux/iomap.h       |  6 +---
 8 files changed, 117 insertions(+), 18 deletions(-)
---
base-commit: f384d0b7710d3edaab718c02bbae46a4d3fd09de
change-id: 20260218-blk-dontcache-338133dd045e

Best regards,
-- 
Tal Zussman <tz2294@columbia.edu>

[PATCH RFC v5 1/3] block: add BIO_COMPLETE_IN_TASK for task-context completion

[Tal Zussman]

Some bio completion handlers need to run in task context but bio_endio()
can be called from IRQ context (e.g. buffer_head writeback). Add a
BIO_COMPLETE_IN_TASK flag that bio submitters can set to request
task-context completion of their bi_end_io callback.

When bio_endio() sees this flag and is running in non-task context, it
queues the bio to a per-cpu lockless list and schedules a delayed work
item to call bi_end_io() from task context.  The delayed work uses a
1-jiffie delay to allow batches of completions to accumulate before
processing. A CPU hotplug dead callback drains any remaining bios from
the departing CPU's batch.

This will be used to enable RWF_DONTCACHE for block devices, and could
be used for other subsystems like fscrypt that need task-context bio
completion.

Suggested-by: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Tal Zussman <tz2294@columbia.edu>
---
 block/bio.c               | 83 ++++++++++++++++++++++++++++++++++++++++++++++-
 include/linux/blk_types.h |  7 +++-
 2 files changed, 88 insertions(+), 2 deletions(-)

diff --git a/block/bio.c b/block/bio.c
index 8203bb7455a9..21b403eb1c04 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -18,6 +18,7 @@
 #include <linux/highmem.h>
 #include <linux/blk-crypto.h>
 #include <linux/xarray.h>
+#include <linux/llist.h>
 
 #include <trace/events/block.h>
 #include "blk.h"
@@ -1714,6 +1715,51 @@ void bio_check_pages_dirty(struct bio *bio)
 }
 EXPORT_SYMBOL_GPL(bio_check_pages_dirty);
 
+struct bio_complete_batch {
+	struct llist_head list;
+	struct delayed_work work;
+	int cpu;
+};
+
+static DEFINE_PER_CPU(struct bio_complete_batch, bio_complete_batch);
+static struct workqueue_struct *bio_complete_wq;
+
+static void bio_complete_work_fn(struct work_struct *w)
+{
+	struct delayed_work *dw = to_delayed_work(w);
+	struct bio_complete_batch *batch =
+		container_of(dw, struct bio_complete_batch, work);
+	struct llist_node *node;
+	struct bio *bio, *next;
+
+	do {
+		node = llist_del_all(&batch->list);
+		if (!node)
+			break;
+
+		node = llist_reverse_order(node);
+		llist_for_each_entry_safe(bio, next, node, bi_llist)
+			bio->bi_end_io(bio);
+
+		if (need_resched()) {
+			if (!llist_empty(&batch->list))
+				mod_delayed_work_on(batch->cpu,
+						    bio_complete_wq,
+						    &batch->work, 0);
+			break;
+		}
+	} while (1);
+}
+
+static void bio_queue_completion(struct bio *bio)
+{
+	struct bio_complete_batch *batch = this_cpu_ptr(&bio_complete_batch);
+
+	if (llist_add(&bio->bi_llist, &batch->list))
+		mod_delayed_work_on(batch->cpu, bio_complete_wq,
+				    &batch->work, 1);
+}
+
 static inline bool bio_remaining_done(struct bio *bio)
 {
 	/*
@@ -1788,7 +1834,9 @@ void bio_endio(struct bio *bio)
 	}
 #endif
 
-	if (bio->bi_end_io)
+	if (!in_task() && bio_flagged(bio, BIO_COMPLETE_IN_TASK))
+		bio_queue_completion(bio);
+	else if (bio->bi_end_io)
 		bio->bi_end_io(bio);
 }
 EXPORT_SYMBOL(bio_endio);
@@ -1974,6 +2022,24 @@ int bioset_init(struct bio_set *bs,
 }
 EXPORT_SYMBOL(bioset_init);
 
+/*
+ * Drain a dead CPU's deferred bio completions.
+ */
+static int bio_complete_batch_cpu_dead(unsigned int cpu)
+{
+	struct bio_complete_batch *batch =
+		per_cpu_ptr(&bio_complete_batch, cpu);
+	struct llist_node *node;
+	struct bio *bio, *next;
+
+	node = llist_del_all(&batch->list);
+	node = llist_reverse_order(node);
+	llist_for_each_entry_safe(bio, next, node, bi_llist)
+		bio->bi_end_io(bio);
+
+	return 0;
+}
+
 static int __init init_bio(void)
 {
 	int i;
@@ -1988,6 +2054,21 @@ static int __init init_bio(void)
 				SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
 	}
 
+	for_each_possible_cpu(i) {
+		struct bio_complete_batch *batch =
+			per_cpu_ptr(&bio_complete_batch, i);
+
+		init_llist_head(&batch->list);
+		INIT_DELAYED_WORK(&batch->work, bio_complete_work_fn);
+		batch->cpu = i;
+	}
+
+	bio_complete_wq = alloc_workqueue("bio_complete", WQ_MEM_RECLAIM, 0);
+	if (!bio_complete_wq)
+		panic("bio: can't allocate bio_complete workqueue\n");
+
+	cpuhp_setup_state(CPUHP_BP_PREPARE_DYN, "block/bio:complete:dead",
+				NULL, bio_complete_batch_cpu_dead);
 	cpuhp_setup_state_multi(CPUHP_BIO_DEAD, "block/bio:dead", NULL,
 					bio_cpu_dead);
 
diff --git a/include/linux/blk_types.h b/include/linux/blk_types.h
index 8808ee76e73c..0b55159d110d 100644
--- a/include/linux/blk_types.h
+++ b/include/linux/blk_types.h
@@ -11,6 +11,7 @@
 #include <linux/device.h>
 #include <linux/ktime.h>
 #include <linux/rw_hint.h>
+#include <linux/llist.h>
 
 struct bio_set;
 struct bio;
@@ -208,7 +209,10 @@ typedef unsigned int blk_qc_t;
  * stacking drivers)
  */
 struct bio {
-	struct bio		*bi_next;	/* request queue link */
+	union {
+		struct bio	*bi_next;	/* request queue link */
+		struct llist_node bi_llist;	/* deferred completion */
+	};
 	struct block_device	*bi_bdev;
 	blk_opf_t		bi_opf;		/* bottom bits REQ_OP, top bits
 						 * req_flags.
@@ -322,6 +326,7 @@ enum {
 	BIO_REMAPPED,
 	BIO_ZONE_WRITE_PLUGGING, /* bio handled through zone write plugging */
 	BIO_EMULATES_ZONE_APPEND, /* bio emulates a zone append operation */
+	BIO_COMPLETE_IN_TASK, /* complete bi_end_io() in task context */
 	BIO_FLAG_LAST
 };
 

-- 
2.39.5

[PATCH RFC v5 2/3] iomap: use BIO_COMPLETE_IN_TASK for dropbehind writeback

[Tal Zussman]

Set BIO_COMPLETE_IN_TASK on iomap writeback bios when a dropbehind folio
is added. This ensures that bi_end_io runs in task context, where
folio_end_dropbehind() can safely invalidate folios.

With the bio layer now handling task-context deferral generically,
IOMAP_IOEND_DONTCACHE is no longer needed, as XFS no longer needs to
route DONTCACHE ioends through its completion workqueue. Remove the flag
and its NOMERGE entry.

Without the NOMERGE, regular I/Os that get merged with a dropbehind
folio will also have their completion deferred to task context.

Signed-off-by: Tal Zussman <tz2294@columbia.edu>
---
 fs/iomap/ioend.c      | 5 +++--
 fs/xfs/xfs_aops.c     | 4 ----
 include/linux/iomap.h | 6 +-----
 3 files changed, 4 insertions(+), 11 deletions(-)

diff --git a/fs/iomap/ioend.c b/fs/iomap/ioend.c
index e4d57cb969f1..fe2a4c3dae42 100644
--- a/fs/iomap/ioend.c
+++ b/fs/iomap/ioend.c
@@ -182,8 +182,6 @@ ssize_t iomap_add_to_ioend(struct iomap_writepage_ctx *wpc, struct folio *folio,
 		ioend_flags |= IOMAP_IOEND_UNWRITTEN;
 	if (wpc->iomap.flags & IOMAP_F_SHARED)
 		ioend_flags |= IOMAP_IOEND_SHARED;
-	if (folio_test_dropbehind(folio))
-		ioend_flags |= IOMAP_IOEND_DONTCACHE;
 	if (pos == wpc->iomap.offset && (wpc->iomap.flags & IOMAP_F_BOUNDARY))
 		ioend_flags |= IOMAP_IOEND_BOUNDARY;
 
@@ -200,6 +198,9 @@ ssize_t iomap_add_to_ioend(struct iomap_writepage_ctx *wpc, struct folio *folio,
 	if (!bio_add_folio(&ioend->io_bio, folio, map_len, poff))
 		goto new_ioend;
 
+	if (folio_test_dropbehind(folio))
+		bio_set_flag(&ioend->io_bio, BIO_COMPLETE_IN_TASK);
+
 	/*
 	 * Clamp io_offset and io_size to the incore EOF so that ondisk
 	 * file size updates in the ioend completion are byte-accurate.
diff --git a/fs/xfs/xfs_aops.c b/fs/xfs/xfs_aops.c
index 76678814f46f..0d469b91377d 100644
--- a/fs/xfs/xfs_aops.c
+++ b/fs/xfs/xfs_aops.c
@@ -510,10 +510,6 @@ xfs_ioend_needs_wq_completion(
 	if (ioend->io_flags & (IOMAP_IOEND_UNWRITTEN | IOMAP_IOEND_SHARED))
 		return true;
 
-	/* Page cache invalidation cannot be done in irq context. */
-	if (ioend->io_flags & IOMAP_IOEND_DONTCACHE)
-		return true;
-
 	return false;
 }
 
diff --git a/include/linux/iomap.h b/include/linux/iomap.h
index 99b7209dabd7..a5d6401ebd80 100644
--- a/include/linux/iomap.h
+++ b/include/linux/iomap.h
@@ -392,16 +392,12 @@ sector_t iomap_bmap(struct address_space *mapping, sector_t bno,
 #define IOMAP_IOEND_BOUNDARY		(1U << 2)
 /* is direct I/O */
 #define IOMAP_IOEND_DIRECT		(1U << 3)
-/* is DONTCACHE I/O */
-#define IOMAP_IOEND_DONTCACHE		(1U << 4)
-
 /*
  * Flags that if set on either ioend prevent the merge of two ioends.
  * (IOMAP_IOEND_BOUNDARY also prevents merges, but only one-way)
  */
 #define IOMAP_IOEND_NOMERGE_FLAGS \
-	(IOMAP_IOEND_SHARED | IOMAP_IOEND_UNWRITTEN | IOMAP_IOEND_DIRECT | \
-	 IOMAP_IOEND_DONTCACHE)
+	(IOMAP_IOEND_SHARED | IOMAP_IOEND_UNWRITTEN | IOMAP_IOEND_DIRECT)
 
 /*
  * Structure for writeback I/O completions.

-- 
2.39.5

[PATCH RFC v5 3/3] block: enable RWF_DONTCACHE for block devices

[Tal Zussman]

Block device buffered reads and writes already pass through
filemap_read() and iomap_file_buffered_write() respectively, both of
which handle IOCB_DONTCACHE. Enable RWF_DONTCACHE for block device files
by setting FOP_DONTCACHE in def_blk_fops.

For CONFIG_BUFFER_HEAD=y paths, add block_write_begin_iocb() which
threads the kiocb through so that buffer_head-based I/O can use
DONTCACHE behavior. The existing block_write_begin() is preserved as a
wrapper that passes a NULL iocb. Set BIO_COMPLETE_IN_TASK in
submit_bh_wbc() when the folio has dropbehind so that buffer_head
writeback completions get deferred to task context.

CONFIG_BUFFER_HEAD=n paths are handled by the previously added iomap
BIO_COMPLETE_IN_TASK support.

This support is useful for databases that operate on raw block devices,
among other userspace applications.

Signed-off-by: Tal Zussman <tz2294@columbia.edu>
---
 block/fops.c                |  5 +++--
 fs/buffer.c                 | 22 +++++++++++++++++++---
 include/linux/buffer_head.h |  3 +++
 3 files changed, 25 insertions(+), 5 deletions(-)

diff --git a/block/fops.c b/block/fops.c
index 4d32785b31d9..d8165f6ba71c 100644
--- a/block/fops.c
+++ b/block/fops.c
@@ -505,7 +505,8 @@ static int blkdev_write_begin(const struct kiocb *iocb,
 			      unsigned len, struct folio **foliop,
 			      void **fsdata)
 {
-	return block_write_begin(mapping, pos, len, foliop, blkdev_get_block);
+	return block_write_begin_iocb(iocb, mapping, pos, len, foliop,
+				     blkdev_get_block);
 }
 
 static int blkdev_write_end(const struct kiocb *iocb,
@@ -967,7 +968,7 @@ const struct file_operations def_blk_fops = {
 	.splice_write	= iter_file_splice_write,
 	.fallocate	= blkdev_fallocate,
 	.uring_cmd	= blkdev_uring_cmd,
-	.fop_flags	= FOP_BUFFER_RASYNC,
+	.fop_flags	= FOP_BUFFER_RASYNC | FOP_DONTCACHE,
 };
 
 static __init int blkdev_init(void)
diff --git a/fs/buffer.c b/fs/buffer.c
index ed724a902657..c60c0ad6cc35 100644
--- a/fs/buffer.c
+++ b/fs/buffer.c
@@ -2239,14 +2239,19 @@ EXPORT_SYMBOL(block_commit_write);
  *
  * The filesystem needs to handle block truncation upon failure.
  */
-int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
+int block_write_begin_iocb(const struct kiocb *iocb,
+		struct address_space *mapping, loff_t pos, unsigned len,
 		struct folio **foliop, get_block_t *get_block)
 {
 	pgoff_t index = pos >> PAGE_SHIFT;
+	fgf_t fgp_flags = FGP_WRITEBEGIN;
 	struct folio *folio;
 	int status;
 
-	folio = __filemap_get_folio(mapping, index, FGP_WRITEBEGIN,
+	if (iocb && iocb->ki_flags & IOCB_DONTCACHE)
+		fgp_flags |= FGP_DONTCACHE;
+
+	folio = __filemap_get_folio(mapping, index, fgp_flags,
 			mapping_gfp_mask(mapping));
 	if (IS_ERR(folio))
 		return PTR_ERR(folio);
@@ -2261,6 +2266,13 @@ int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
 	*foliop = folio;
 	return status;
 }
+
+int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
+		struct folio **foliop, get_block_t *get_block)
+{
+	return block_write_begin_iocb(NULL, mapping, pos, len, foliop,
+				      get_block);
+}
 EXPORT_SYMBOL(block_write_begin);
 
 int block_write_end(loff_t pos, unsigned len, unsigned copied,
@@ -2589,7 +2601,8 @@ int cont_write_begin(const struct kiocb *iocb, struct address_space *mapping,
 		(*bytes)++;
 	}
 
-	return block_write_begin(mapping, pos, len, foliop, get_block);
+	return block_write_begin_iocb(iocb, mapping, pos, len, foliop,
+				     get_block);
 }
 EXPORT_SYMBOL(cont_write_begin);
 
@@ -2801,6 +2814,9 @@ static void submit_bh_wbc(blk_opf_t opf, struct buffer_head *bh,
 
 	bio = bio_alloc(bh->b_bdev, 1, opf, GFP_NOIO);
 
+	if (folio_test_dropbehind(bh->b_folio))
+		bio_set_flag(bio, BIO_COMPLETE_IN_TASK);
+
 	fscrypt_set_bio_crypt_ctx_bh(bio, bh, GFP_NOIO);
 
 	bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
diff --git a/include/linux/buffer_head.h b/include/linux/buffer_head.h
index b16b88bfbc3e..ddf88ce290f2 100644
--- a/include/linux/buffer_head.h
+++ b/include/linux/buffer_head.h
@@ -260,6 +260,9 @@ int block_read_full_folio(struct folio *, get_block_t *);
 bool block_is_partially_uptodate(struct folio *, size_t from, size_t count);
 int block_write_begin(struct address_space *mapping, loff_t pos, unsigned len,
 		struct folio **foliop, get_block_t *get_block);
+int block_write_begin_iocb(const struct kiocb *iocb,
+		struct address_space *mapping, loff_t pos, unsigned len,
+		struct folio **foliop, get_block_t *get_block);
 int __block_write_begin(struct folio *folio, loff_t pos, unsigned len,
 		get_block_t *get_block);
 int block_write_end(loff_t pos, unsigned len, unsigned copied, struct folio *);

-- 
2.39.5

Re: [PATCH RFC v5 0/3] block: enable RWF_DONTCACHE for block devices

[Christoph Hellwig]

What tree is this against?  I tried Jens' for-next and for-7.1/block
trees, linux-next and current mainline and it does not apply to any
of them.

Re: [PATCH RFC v5 3/3] block: enable RWF_DONTCACHE for block devices

[Christoph Hellwig]

For the next round please split this into the buffer.c parts for
the infrastructure and the use in the block device file operations.

Re: [PATCH RFC v5 0/3] block: enable RWF_DONTCACHE for block devices

[Tal Zussman]

On 4/9/26 2:53 AM, Christoph Hellwig wrote:
> What tree is this against?  I tried Jens' for-next and for-7.1/block
> trees, linux-next and current mainline and it does not apply to any
> of them.

Sorry about that -- I had previously based this on an earlier rc and forgot
to rebase. For the next iteration, is there a preferred tree to rebase
onto? I see you rebased this on linux-next.

Re: [PATCH RFC v5 3/3] block: enable RWF_DONTCACHE for block devices

[Tal Zussman]

On 4/9/26 12:08 PM, Christoph Hellwig wrote:
> For the next round please split this into the buffer.c parts for
> the infrastructure and the use in the block device file operations.

Will do.