[LSF/MM/BPF TOPIC] Per-process page size

[LSF/MM/BPF TOPIC] Per-process page size

[Dev Jain]

Hi everyone,

We propose per-process page size on arm64. Although the proposal is for
arm64, perhaps the concept can be extended to other arches, thus the
generic topic name.

-------------
INTRODUCTION
-------------
While mTHP has brought the performance of many workloads running on an arm64 4K
kernel closer to that of the performance on an arm64 64K kernel, a performance
gap still remains. This is attributed to a combination of greater number of
pgtable levels, less reach within the walk cache and higher data cache footprint
for pgtable memory. At the same time, 64K is not suitable for general
purpose environments due to it's significantly higher memory footprint.

To solve this, we have been experimenting with a concept called "per-process
page size". This breaks the historic assumption of a single page size for the
entire system: a process will now operate on a page size ABI that is greater
than or equal to the kernel's page size. This is enabled by a key architectural
feature on Arm: the separation of user and kernel page tables.

This can also lead to a future of a single kernel image instead of 4K, 16K
and 64K images.

--------------
CURRENT DESIGN
--------------
The design is based on one core idea; most of the kernel continues to believe
there is only one page size in use across the whole system. That page size is
the size selected at compile-time, as is done today. But every process (more
accurately mm_struct) has a page size ABI which is one of the 3 page sizes
(4K, 16K or 64K) as long as that page size is greater than or equal to the
kernel page size (kernel page size is the macro PAGE_SIZE).

Pagesize selection
------------------
A process' selected page size ABI comes into force at execve() time and
remains fixed until the process exits or until the next execve(). Any forked
processes inherit the page size of their parent.
The personality() mechanism already exists for similar cases, so we propose
to extend it to enable specifying the required page size.

There are 3 layers to the design. The first two are not arch-dependent,
and makes Linux support a per-process pagesize ABI. The last layer is
arch-specific.

1. ABI adapter
--------------
A translation layer is added at the syscall boundary to convert between the
process page size and the kernel page size. This effectively means enforcing
alignment requirements for addresses passed to syscalls and ensuring that
quantities passed as “number of pages” are interpreted relative to the process
page size and not the kernel page size. In this way the process has the illusion
that it is working in units of its page size, but the kernel is working in
units of the kernel page size.

2. Generic Linux MM enlightenment
---------------------------------
We enlighten the Linux MM code to always hand out memory in the granularity
of process pages. Most of this work is greatly simplified because of the
existing mTHP allocation paths, and the ongoing support for large folios
across different areas of the kernel. The process order will be used as the
hard minimum mTHP order to allocate.

File memory
-----------
For a growing list of compliant file systems, large folios can already be
stored in the page cache. There is even a mechanism, introduced to support
filesystems with block sizes larger than the system page size, to set a
hard-minimum size for folios on a per-address-space basis. This mechanism
will be reused and extended to service the per-process page size requirements.

One key reason that the 64K kernel currently consumes considerably more memory
than the 4K kernel is that Linux systems often have lots of small
configuration files which each require a page in the page cache. But these
small files are (likely) only used by certain processes. So, we prefer to
continue to cache those using a 4K page.
Therefore, if a process with a larger page size maps a file whose pagecache
contains smaller folios, we drop them and re-read the range with a folio
order at least that of the process order.

3. Translation from Linux pagetable to native pagetable
-------------------------------------------------------
Assume the case of a kernel pagesize of 4K and app pagesize of 64K.
Now that enlightenment is done, it is guaranteed that every single mapping
in the 4K pagetable (which we call the Linux pagetable) is of granularity
at least 64K. In the arm64 MM code, we maintain a "native" pagetable per
mm_struct, which is based off a 64K geometry. Because of the guarantee
aforementioned, any pagetable operation on the Linux pagetable
(set_ptes, clear_flush_ptes, modify_prot_start_ptes, etc) is going to happen
at a granularity of at least 16 PTEs - therefore we can translate this
operation to modify a single PTE entry in the native pagetable.
Given that enlightenment may miss corner cases, we insert a warning in the
architecture code - on being presented with an operation not translatable
into a native operation, we fallback to the Linux pagetable, thus losing
the benefits borne out of the pagetable geometry but keeping
the emulation intact.

-----------------------
What we want to discuss
-----------------------
 - Are there other arches which could benefit from this?
 - What level of compatibility we can achieve - is it even possible to
   contain userspace within the emulated ABI?
 - Rough edges of compatibility layer - pfnmaps, ksm, procfs, etc. For
   example, what happens when a 64K process opens a procfs file of
   a 4K process?
 - native pgtable implementation - perhaps inspiration can be taken
   from other arches with an involved pgtable logic (ppc, s390)?

-------------
Key Attendees
-------------
 - Ryan Roberts (co-presenter)
 - mm folks (David Hildenbrand, Matthew Wilcox, Liam Howlett, Lorenzo Stoakes,
             and many others)
 - arch folks

Re: [LSF/MM/BPF TOPIC] Per-process page size

[Matthew Wilcox]

On Tue, Feb 17, 2026 at 08:20:26PM +0530, Dev Jain wrote:
> 2. Generic Linux MM enlightenment
> ---------------------------------
> We enlighten the Linux MM code to always hand out memory in the granularity

Please don't use the term "enlighten".  Tht's used to describe something
something or other with hypervisors.  Come up with a new term or use one
that already exists.

> File memory
> -----------
> For a growing list of compliant file systems, large folios can already be
> stored in the page cache. There is even a mechanism, introduced to support
> filesystems with block sizes larger than the system page size, to set a
> hard-minimum size for folios on a per-address-space basis. This mechanism
> will be reused and extended to service the per-process page size requirements.
> 
> One key reason that the 64K kernel currently consumes considerably more memory
> than the 4K kernel is that Linux systems often have lots of small
> configuration files which each require a page in the page cache. But these
> small files are (likely) only used by certain processes. So, we prefer to
> continue to cache those using a 4K page.
> Therefore, if a process with a larger page size maps a file whose pagecache
> contains smaller folios, we drop them and re-read the range with a folio
> order at least that of the process order.

That's going to be messy.  I don't have a good idea for solving this
problem, but the page cache really isn't set up to change minimum folio
order while the inode is in use.

>  - Are there other arches which could benefit from this?

Some architectures walk the page tables entirely in software, but on the
other hand, those tend to be, er, "legacy" architectures these days and
it's doubtful that anybody would invest in adding support.

Sounds like a good question for Arnd ;-)

>  - What level of compatibility we can achieve - is it even possible to
>    contain userspace within the emulated ABI?
>  - Rough edges of compatibility layer - pfnmaps, ksm, procfs, etc. For
>    example, what happens when a 64K process opens a procfs file of
>    a 4K process?
>  - native pgtable implementation - perhaps inspiration can be taken
>    from other arches with an involved pgtable logic (ppc, s390)?

I question who decides what page size a particular process will use.
The programmer?  The sysadmin?  It seems too disruptive for the kernel
to monitor and decide for the app what page size it will use.

Re: [LSF/MM/BPF TOPIC] Per-process page size

[David Hildenbrand (Arm)]

On 2/17/26 16:22, Matthew Wilcox wrote:
> On Tue, Feb 17, 2026 at 08:20:26PM +0530, Dev Jain wrote:
>> 2. Generic Linux MM enlightenment
>> ---------------------------------
>> We enlighten the Linux MM code to always hand out memory in the granularity
> 
> Please don't use the term "enlighten".  Tht's used to describe something
> something or other with hypervisors.  Come up with a new term or use one
> that already exists.
> 
>> File memory
>> -----------
>> For a growing list of compliant file systems, large folios can already be
>> stored in the page cache. There is even a mechanism, introduced to support
>> filesystems with block sizes larger than the system page size, to set a
>> hard-minimum size for folios on a per-address-space basis. This mechanism
>> will be reused and extended to service the per-process page size requirements.
>>
>> One key reason that the 64K kernel currently consumes considerably more memory
>> than the 4K kernel is that Linux systems often have lots of small
>> configuration files which each require a page in the page cache. But these
>> small files are (likely) only used by certain processes. So, we prefer to
>> continue to cache those using a 4K page.
>> Therefore, if a process with a larger page size maps a file whose pagecache
>> contains smaller folios, we drop them and re-read the range with a folio
>> order at least that of the process order.
> 
> That's going to be messy.  I don't have a good idea for solving this
> problem, but the page cache really isn't set up to change minimum folio
> order while the inode is in use.

In a private conversation I also raised that some situations might make 
it impossible/hard to drop+re-read.

One example I cam up with if a folio is simply long-term R/O pinned. But 
I am also not quite sure how mlock might interfere here.

So yes, I think the page cache is likely the one of the most 
problematic/messy thing to handle.

-- 
Cheers,

David

Re: [LSF/MM/BPF TOPIC] Per-process page size

[Ryan Roberts]

On 17/02/2026 15:30, David Hildenbrand (Arm) wrote:
> On 2/17/26 16:22, Matthew Wilcox wrote:
>> On Tue, Feb 17, 2026 at 08:20:26PM +0530, Dev Jain wrote:
>>> 2. Generic Linux MM enlightenment
>>> ---------------------------------
>>> We enlighten the Linux MM code to always hand out memory in the granularity
>>
>> Please don't use the term "enlighten".  Tht's used to describe something
>> something or other with hypervisors.  Come up with a new term or use one
>> that already exists.
>>
>>> File memory
>>> -----------
>>> For a growing list of compliant file systems, large folios can already be
>>> stored in the page cache. There is even a mechanism, introduced to support
>>> filesystems with block sizes larger than the system page size, to set a
>>> hard-minimum size for folios on a per-address-space basis. This mechanism
>>> will be reused and extended to service the per-process page size requirements.
>>>
>>> One key reason that the 64K kernel currently consumes considerably more memory
>>> than the 4K kernel is that Linux systems often have lots of small
>>> configuration files which each require a page in the page cache. But these
>>> small files are (likely) only used by certain processes. So, we prefer to
>>> continue to cache those using a 4K page.
>>> Therefore, if a process with a larger page size maps a file whose pagecache
>>> contains smaller folios, we drop them and re-read the range with a folio
>>> order at least that of the process order.
>>
>> That's going to be messy.  I don't have a good idea for solving this
>> problem, but the page cache really isn't set up to change minimum folio
>> order while the inode is in use.

Dev has a prototype up and running, but based on your comments, I'm guessing
there is some horrible race that hasn't hit yet. Would be good to debug the gap
in understanding at some point!

> 
> In a private conversation I also raised that some situations might make it
> impossible/hard to drop+re-read.
> 
> One example I cam up with if a folio is simply long-term R/O pinned. But I am
> also not quite sure how mlock might interfere here.
> 
> So yes, I think the page cache is likely the one of the most problematic/messy
> thing to handle.
> 

I guess we could side step the problem for now, by initially requiring that the
minimum folio size always be the maximum supported process page size. That would
allow us to get something up and running at least. But then we lose the memory
saving benefits.

Of course, I'm conveniently ignoring that not all filesystems support large
folios, but perhaps we could do a generic fallback adapter with a bounce buffer
for that case?

Re: [LSF/MM/BPF TOPIC] Per-process page size

[Matthew Wilcox]

On Tue, Feb 17, 2026 at 04:30:59PM +0100, David Hildenbrand (Arm) wrote:
> In a private conversation I also raised that some situations might make it
> impossible/hard to drop+re-read.
> 
> One example I cam up with if a folio is simply long-term R/O pinned. But I
> am also not quite sure how mlock might interfere here.
> 
> So yes, I think the page cache is likely the one of the most
> problematic/messy thing to handle.

So what if we convert to max-supported-order the first time somebody
calls mmap on a given file?  Most files are never mmaped, so it won't
affect them.  And files that are mmaped are generally not written to.
So there should not be much in the page cache for the common case.
And if no pages from the file have been mmaped yet, they cannot be pinned
or mlocked.

Re: [LSF/MM/BPF TOPIC] Per-process page size

[David Hildenbrand (Arm)]

On 2/20/26 05:49, Matthew Wilcox wrote:
> On Tue, Feb 17, 2026 at 04:30:59PM +0100, David Hildenbrand (Arm) wrote:
>> In a private conversation I also raised that some situations might make it
>> impossible/hard to drop+re-read.
>>
>> One example I cam up with if a folio is simply long-term R/O pinned. But I
>> am also not quite sure how mlock might interfere here.
>>
>> So yes, I think the page cache is likely the one of the most
>> problematic/messy thing to handle.
> 
> So what if we convert to max-supported-order the first time somebody
> calls mmap on a given file?  Most files are never mmaped, so it won't
> affect them. 

Yes!

> And files that are mmaped are generally not written to.

Well, let's say many mmaped files are not written to. :)

> So there should not be much in the page cache for the common case.

You'd assume many files to either get mmaped or read/written, yes.

> And if no pages from the file have been mmaped yet, they cannot be pinned
> or mlocked.

Is there some other way for someone to block a page from getting evicted 
from the pagecache?

We have this memfd_pin_folios() thing, but I don't think we have 
something comparable for ordinary pagecache files.

... putting them into a pipe and never reading from the pipe maybe (I 
assume that's what splice() does, but not sure if it actually places the 
pages in there or whether it creates a copy first)?

But yes, doing the conversion on first mmap() should handle mlock+gup.

-- 
Cheers,

David

Re: [LSF/MM/BPF TOPIC] Per-process page size

[Dev Jain]

On 17/02/26 8:52 pm, Matthew Wilcox wrote:
> On Tue, Feb 17, 2026 at 08:20:26PM +0530, Dev Jain wrote:
>> 2. Generic Linux MM enlightenment
>> ---------------------------------
>> We enlighten the Linux MM code to always hand out memory in the granularity
> Please don't use the term "enlighten".  Tht's used to describe something
> something or other with hypervisors.  Come up with a new term or use one
> that already exists.

Sure.

>
>> File memory
>> -----------
>> For a growing list of compliant file systems, large folios can already be
>> stored in the page cache. There is even a mechanism, introduced to support
>> filesystems with block sizes larger than the system page size, to set a
>> hard-minimum size for folios on a per-address-space basis. This mechanism
>> will be reused and extended to service the per-process page size requirements.
>>
>> One key reason that the 64K kernel currently consumes considerably more memory
>> than the 4K kernel is that Linux systems often have lots of small
>> configuration files which each require a page in the page cache. But these
>> small files are (likely) only used by certain processes. So, we prefer to
>> continue to cache those using a 4K page.
>> Therefore, if a process with a larger page size maps a file whose pagecache
>> contains smaller folios, we drop them and re-read the range with a folio
>> order at least that of the process order.
> That's going to be messy.  I don't have a good idea for solving this
> problem, but the page cache really isn't set up to change minimum folio
> order while the inode is in use.

Holding mapping->invalidate_lock, bumping mapping->min_folio_order and
dropping-rereading the range suffers from a race - filemap_fault operating
on some other partially populated 64K range will observe in filemap_get_folio
that nothing is in the pagecache. Then, it will read the updated min_order
in __filemap_get_folio, then use filemap_add_folio to add a 64K folio, but since
the 64K range is partially populated, we get stuck in an infinite loop due to -EEXIST.

So I figured that deleting the entire pagecache is simpler. We will also bail
out early in __filemap_add_folio if the folio order asked by the caller to
create is less than mapping_min_folio_order. Eventually the caller is going
to read the correct min order. This algorithm avoids the race above, however...

my assumption here was that we are synchronized on mapping->invalidate_lock.
The kerneldoc above read_cache_folio() and some other comments convinced me
of that, but I just checked with a VM_WARN_ON(!is_rwsem_locked()) in
__filemap_add_folio and this doesn't seem to be the case for all code paths...
If the algorithm sounds reasonable, I wonder what is the correct synchronization
mechanism here.

>
>>  - Are there other arches which could benefit from this?
> Some architectures walk the page tables entirely in software, but on the
> other hand, those tend to be, er, "legacy" architectures these days and
> it's doubtful that anybody would invest in adding support.
>
> Sounds like a good question for Arnd ;-)
>
>>  - What level of compatibility we can achieve - is it even possible to
>>    contain userspace within the emulated ABI?
>>  - Rough edges of compatibility layer - pfnmaps, ksm, procfs, etc. For
>>    example, what happens when a 64K process opens a procfs file of
>>    a 4K process?
>>  - native pgtable implementation - perhaps inspiration can be taken
>>    from other arches with an involved pgtable logic (ppc, s390)?
> I question who decides what page size a particular process will use.
> The programmer?  The sysadmin?  It seems too disruptive for the kernel
> to monitor and decide for the app what page size it will use.

It's the sysadmin. The latter method you mention is similar to the problem
of the kernel choosing the correct mTHP order, which we don't have an
elegant idea for solving yet.

Re: [LSF/MM/BPF TOPIC] Per-process page size

[Dev Jain]

On 18/02/26 2:09 pm, Dev Jain wrote:
> On 17/02/26 8:52 pm, Matthew Wilcox wrote:
>> On Tue, Feb 17, 2026 at 08:20:26PM +0530, Dev Jain wrote:
>>> 2. Generic Linux MM enlightenment
>>> ---------------------------------
>>> We enlighten the Linux MM code to always hand out memory in the granularity
>> Please don't use the term "enlighten".  Tht's used to describe something
>> something or other with hypervisors.  Come up with a new term or use one
>> that already exists.
> Sure.
>
>>> File memory
>>> -----------
>>> For a growing list of compliant file systems, large folios can already be
>>> stored in the page cache. There is even a mechanism, introduced to support
>>> filesystems with block sizes larger than the system page size, to set a
>>> hard-minimum size for folios on a per-address-space basis. This mechanism
>>> will be reused and extended to service the per-process page size requirements.
>>>
>>> One key reason that the 64K kernel currently consumes considerably more memory
>>> than the 4K kernel is that Linux systems often have lots of small
>>> configuration files which each require a page in the page cache. But these
>>> small files are (likely) only used by certain processes. So, we prefer to
>>> continue to cache those using a 4K page.
>>> Therefore, if a process with a larger page size maps a file whose pagecache
>>> contains smaller folios, we drop them and re-read the range with a folio
>>> order at least that of the process order.
>> That's going to be messy.  I don't have a good idea for solving this
>> problem, but the page cache really isn't set up to change minimum folio
>> order while the inode is in use.
> Holding mapping->invalidate_lock, bumping mapping->min_folio_order and
> dropping-rereading the range suffers from a race - filemap_fault operating
> on some other partially populated 64K range will observe in filemap_get_folio
> that nothing is in the pagecache. Then, it will read the updated min_order
> in __filemap_get_folio, then use filemap_add_folio to add a 64K folio, but since
> the 64K range is partially populated, we get stuck in an infinite loop due to -EEXIST.
>
> So I figured that deleting the entire pagecache is simpler. We will also bail
> out early in __filemap_add_folio if the folio order asked by the caller to
> create is less than mapping_min_folio_order. Eventually the caller is going
> to read the correct min order. This algorithm avoids the race above, however...
>
> my assumption here was that we are synchronized on mapping->invalidate_lock.
> The kerneldoc above read_cache_folio() and some other comments convinced me
> of that, but I just checked with a VM_WARN_ON(!is_rwsem_locked()) in
> __filemap_add_folio and this doesn't seem to be the case for all code paths...
> If the algorithm sounds reasonable, I wonder what is the correct synchronization
> mechanism here.

I may have been vague here... to avoid the race I described above, we must
ensure that after all folios have been dropped from pagecache, and min order
is bumped up, no other code path remembers the old order and partially
populates a 64K range. For this we need synchronization.

>
>>>  - Are there other arches which could benefit from this?
>> Some architectures walk the page tables entirely in software, but on the
>> other hand, those tend to be, er, "legacy" architectures these days and
>> it's doubtful that anybody would invest in adding support.
>>
>> Sounds like a good question for Arnd ;-)
>>
>>>  - What level of compatibility we can achieve - is it even possible to
>>>    contain userspace within the emulated ABI?
>>>  - Rough edges of compatibility layer - pfnmaps, ksm, procfs, etc. For
>>>    example, what happens when a 64K process opens a procfs file of
>>>    a 4K process?
>>>  - native pgtable implementation - perhaps inspiration can be taken
>>>    from other arches with an involved pgtable logic (ppc, s390)?
>> I question who decides what page size a particular process will use.
>> The programmer?  The sysadmin?  It seems too disruptive for the kernel
>> to monitor and decide for the app what page size it will use.
> It's the sysadmin. The latter method you mention is similar to the problem
> of the kernel choosing the correct mTHP order, which we don't have an
> elegant idea for solving yet.
>
>

Re: [LSF/MM/BPF TOPIC] Per-process page size

[David Hildenbrand (Arm)]

On 2/18/26 09:58, Dev Jain wrote:
> 
> On 18/02/26 2:09 pm, Dev Jain wrote:
>> On 17/02/26 8:52 pm, Matthew Wilcox wrote:
>>> Please don't use the term "enlighten".  Tht's used to describe something
>>> something or other with hypervisors.  Come up with a new term or use one
>>> that already exists.
>> Sure.
>>
>>> That's going to be messy.  I don't have a good idea for solving this
>>> problem, but the page cache really isn't set up to change minimum folio
>>> order while the inode is in use.
>> Holding mapping->invalidate_lock, bumping mapping->min_folio_order and
>> dropping-rereading the range suffers from a race - filemap_fault operating
>> on some other partially populated 64K range will observe in filemap_get_folio
>> that nothing is in the pagecache. Then, it will read the updated min_order
>> in __filemap_get_folio, then use filemap_add_folio to add a 64K folio, but since
>> the 64K range is partially populated, we get stuck in an infinite loop due to -EEXIST.
>>
>> So I figured that deleting the entire pagecache is simpler. We will also bail
>> out early in __filemap_add_folio if the folio order asked by the caller to
>> create is less than mapping_min_folio_order. Eventually the caller is going
>> to read the correct min order. This algorithm avoids the race above, however...
>>
>> my assumption here was that we are synchronized on mapping->invalidate_lock.
>> The kerneldoc above read_cache_folio() and some other comments convinced me
>> of that, but I just checked with a VM_WARN_ON(!is_rwsem_locked()) in
>> __filemap_add_folio and this doesn't seem to be the case for all code paths...
>> If the algorithm sounds reasonable, I wonder what is the correct synchronization
>> mechanism here.
> 
> I may have been vague here... to avoid the race I described above, we must
> ensure that after all folios have been dropped from pagecache, and min order
> is bumped up, no other code path remembers the old order and partially
> populates a 64K range. For this we need synchronization.

And I don't think you can reliably do that when other processes might be 
using the files concurrently.

It's best to start like Ryan suggested: lifting min_order on these 
systems for now and leaving dynamically switching the min order as 
future work.

-- 
Cheers,

David

Re: [LSF/MM/BPF TOPIC] Per-process page size

[Arnd Bergmann]

On Tue, Feb 17, 2026, at 16:22, Matthew Wilcox wrote:
> On Tue, Feb 17, 2026 at 08:20:26PM +0530, Dev Jain wrote:
>>
>>  - Are there other arches which could benefit from this?
>
> Some architectures walk the page tables entirely in software, but on the
> other hand, those tend to be, er, "legacy" architectures these days and
> it's doubtful that anybody would invest in adding support.
>
> Sounds like a good question for Arnd ;-)

I think Loongarch and RISC-V are the candidates for doing whatever
Arm does here. MIPS and PowerPC64 could do it in theory, but it's
less clear that someone will spend the effort here.

>>  - Rough edges of compatibility layer - pfnmaps, ksm, procfs, etc. For
>>    example, what happens when a 64K process opens a procfs file of
>>    a 4K process?

This would also be my main concern. There are hundreds of device drivers
that implement a custom .mmap() file operation, and a few dozen file
systems, all of which need to be audited and likely changed to allow
mapping larger granules.

>>  - native pgtable implementation - perhaps inspiration can be taken
>>    from other arches with an involved pgtable logic (ppc, s390)?
>
> I question who decides what page size a particular process will use.
> The programmer?  The sysadmin?

I would expect this to be done by a combination of these two, it
seems simple enough to have a wrapper like numactl or setarch to
start an application one way or another.


Another concern I have is for the actual performance trade-offs here.
As I understand it, the idea is to have most of the memory size
advantages of a 4KB page kernel, and most of the performance
advantages of a 64KB page kernel for the special applications that
care about this. However, the same is true for 16KB page kernel,
which also aims for the same trade-off with a much simpler model
and a different set of compatibility problems.

Do we expect per-process page size kernels to actually be better
than fixed 16KB page kernels, and better enough that it's worth
the added complexity? In particular, this approach would likely
only get the advantages of the TLB but not the file systems using
larger pages, while also suffering from the extra overhead of
compacting smaller pages in order to map them.

   Arnd

Re: [LSF/MM/BPF TOPIC] Per-process page size

[Pedro Falcato]

On Tue, Feb 17, 2026 at 08:20:26PM +0530, Dev Jain wrote:
> Hi everyone,
> 
> We propose per-process page size on arm64. Although the proposal is for
> arm64, perhaps the concept can be extended to other arches, thus the
> generic topic name.
> 
> -------------
> INTRODUCTION
> -------------
> While mTHP has brought the performance of many workloads running on an arm64 4K
> kernel closer to that of the performance on an arm64 64K kernel, a performance
> gap still remains. This is attributed to a combination of greater number of
> pgtable levels, less reach within the walk cache and higher data cache footprint
> for pgtable memory. At the same time, 64K is not suitable for general
> purpose environments due to it's significantly higher memory footprint.

Could this perhaps be because of larger-page-size kernels being able to use
mTHP (and THP) more aggressively? It would be interesting to compare arm64
"4K" vs "4K with mTHP" vs "4K with _only_ mTHP" vs "64K" vs "64K with mTHP".

> 
> To solve this, we have been experimenting with a concept called "per-process
> page size". This breaks the historic assumption of a single page size for the
> entire system: a process will now operate on a page size ABI that is greater
> than or equal to the kernel's page size. This is enabled by a key architectural
> feature on Arm: the separation of user and kernel page tables.
> 
> This can also lead to a future of a single kernel image instead of 4K, 16K
> and 64K images.
> 
> --------------
> CURRENT DESIGN
> --------------
> The design is based on one core idea; most of the kernel continues to believe
> there is only one page size in use across the whole system. That page size is
> the size selected at compile-time, as is done today. But every process (more
> accurately mm_struct) has a page size ABI which is one of the 3 page sizes
> (4K, 16K or 64K) as long as that page size is greater than or equal to the
> kernel page size (kernel page size is the macro PAGE_SIZE).
> 
> Pagesize selection
> ------------------
> A process' selected page size ABI comes into force at execve() time and
> remains fixed until the process exits or until the next execve(). Any forked
> processes inherit the page size of their parent.
> The personality() mechanism already exists for similar cases, so we propose
> to extend it to enable specifying the required page size.
> 
> There are 3 layers to the design. The first two are not arch-dependent,
> and makes Linux support a per-process pagesize ABI. The last layer is
> arch-specific.
> 
> 1. ABI adapter
> --------------
> A translation layer is added at the syscall boundary to convert between the
> process page size and the kernel page size. This effectively means enforcing
> alignment requirements for addresses passed to syscalls and ensuring that
> quantities passed as “number of pages” are interpreted relative to the process
> page size and not the kernel page size. In this way the process has the illusion
> that it is working in units of its page size, but the kernel is working in
> units of the kernel page size.
> 
> 2. Generic Linux MM enlightenment
> ---------------------------------
> We enlighten the Linux MM code to always hand out memory in the granularity
> of process pages. Most of this work is greatly simplified because of the
> existing mTHP allocation paths, and the ongoing support for large folios
> across different areas of the kernel. The process order will be used as the
> hard minimum mTHP order to allocate.
> 
> File memory
> -----------
> For a growing list of compliant file systems, large folios can already be
> stored in the page cache. There is even a mechanism, introduced to support
> filesystems with block sizes larger than the system page size, to set a
> hard-minimum size for folios on a per-address-space basis. This mechanism
> will be reused and extended to service the per-process page size requirements.
> 
> One key reason that the 64K kernel currently consumes considerably more memory
> than the 4K kernel is that Linux systems often have lots of small
> configuration files which each require a page in the page cache. But these
> small files are (likely) only used by certain processes. So, we prefer to
> continue to cache those using a 4K page.
> Therefore, if a process with a larger page size maps a file whose pagecache
> contains smaller folios, we drop them and re-read the range with a folio
> order at least that of the process order.
> 
> 3. Translation from Linux pagetable to native pagetable
> -------------------------------------------------------
> Assume the case of a kernel pagesize of 4K and app pagesize of 64K.
> Now that enlightenment is done, it is guaranteed that every single mapping
> in the 4K pagetable (which we call the Linux pagetable) is of granularity
> at least 64K. In the arm64 MM code, we maintain a "native" pagetable per
> mm_struct, which is based off a 64K geometry. Because of the guarantee
> aforementioned, any pagetable operation on the Linux pagetable
> (set_ptes, clear_flush_ptes, modify_prot_start_ptes, etc) is going to happen
> at a granularity of at least 16 PTEs - therefore we can translate this
> operation to modify a single PTE entry in the native pagetable.
> Given that enlightenment may miss corner cases, we insert a warning in the
> architecture code - on being presented with an operation not translatable
> into a native operation, we fallback to the Linux pagetable, thus losing
> the benefits borne out of the pagetable geometry but keeping
> the emulation intact.

I don't understand. What exactly are you trying to do here? Maintain 2
different paging structures, one for core mm and the other for the arch? As
done in architectures with no radix tree paging structures?

If so, that's wildly inefficient, unless you're willing to go into reclaimable
page tables on the arm64 side. And that brings extra problems and extra fun :)

-- 
Pedro