Message-ID: <80b5f87e-c156-4ccc-98f0-96f1fd864273@arm.com>
Date: Wed, 24 Apr 2024 09:26:49 +0100
From: Ryan Roberts <ryan.roberts@....com>
To: Baolin Wang <baolin.wang@...ux.alibaba.com>, akpm@...ux-foundation.org,
 hughd@...gle.com
Cc: willy@...radead.org, david@...hat.com, wangkefeng.wang@...wei.com,
 21cnbao@...il.com, ying.huang@...el.com, shy828301@...il.com,
 ziy@...dia.com, linux-mm@...ck.org, linux-kernel@...r.kernel.org
Subject: Re: [RFC PATCH 0/5] add mTHP support for anonymous share pages

On 24/04/2024 07:55, Baolin Wang wrote:
> 
> 
> On 2024/4/23 18:41, Ryan Roberts wrote:
>> On 22/04/2024 08:02, Baolin Wang wrote:
>>> Anonymous pages have already been supported for multi-size (mTHP) allocation
>>> through commit 19eaf44954df, that can allow THP to be configured through the
>>> sysfs interface located at
>>> '/sys/kernel/mm/transparent_hugepage/hugepage-XXkb/enabled'.
>>>
>>> However, the anonymous shared pages will ignore the anonymous mTHP rule
>>> configured through the sysfs interface, and can only use the PMD-mapped
>>> THP, that is not reasonable. Many implement anonymous page sharing through
>>> mmap(MAP_SHARED | MAP_ANONYMOUS), especially in database usage scenarios,
>>> therefore, users expect to apply an unified mTHP strategy for anonymous pages,
>>> also including the anonymous shared pages, in order to enjoy the benefits of
>>> mTHP. For example, lower latency than PMD-mapped THP, smaller memory bloat
>>> than PMD-mapped THP, contiguous PTEs on ARM architecture to reduce TLB miss etc.
>>
>> This sounds like a very useful addition!
>>
>> Out of interest, can you point me at any workloads (and off-the-shelf benchmarks
>> for those workloads) that predominantly use shared anon memory?
> 
> As far as I know, some database related workloads make extensive use of shared
> anonymous page, such as PolarDB[1] in our Alibaba fleet, or MySQL likely also
> uses shared anonymous memory. And I still need to do some investigation to
> measure the performance.
> 
> [1] https://github.com/ApsaraDB/PolarDB-for-PostgreSQL

Thanks for the pointer!

> 
>>> The primary strategy is that, the use of huge pages for anonymous shared pages
>>> still follows the global control determined by the mount option "huge="
>>> parameter
>>> or the sysfs interface at '/sys/kernel/mm/transparent_hugepage/shmem_enabled'.
>>> The utilization of mTHP is allowed only when the global 'huge' switch is
>>> enabled.
>>> Subsequently, the mTHP sysfs interface
>>> (/sys/kernel/mm/transparent_hugepage/hugepage-XXkb/enabled)
>>> is checked to determine the mTHP size that can be used for large folio
>>> allocation
>>> for these anonymous shared pages.
>>
>> I'm not sure about this proposed control mechanism; won't it break
>> compatibility? I could be wrong, but I don't think shmem's use of THP used to
>> depend upon the value of /sys/kernel/mm/transparent_hugepage/enabled? So it
> 
> Yes, I realized this after more testing.
> 
>> doesn't make sense to me that we now depend upon the
>> /sys/kernel/mm/transparent_hugepage/hugepage-XXkb/enabled values (which by
>> default disables all sizes except 2M, which is set to "inherit" from
>> /sys/kernel/mm/transparent_hugepage/enabled).
>>
>> The other problem is that shmem_enabled has a different set of options
>> (always/never/within_size/advise/deny/force) to enabled (always/madvise/never)
>>
>> Perhaps it would be cleaner to do the same trick we did for enabled; Introduce
>> /mm/transparent_hugepage/hugepage-XXkb/shmem_enabled, which can have all the
>> same values as the top-level /sys/kernel/mm/transparent_hugepage/shmem_enabled,
>> plus the additional "inherit" option. By default all sizes will be set to
>> "never" except 2M, which is set to "inherit".
> 
> Sounds good to me. But I do not want to copy all same values from top-level
> '/sys/kernel/mm/transparent_hugepage/shmem_enabled':
> always within_size advise never deny force
> 
> For mTHP's shmem_enabled interface, we can just keep below values:
> always within_size advise never
> 
> Cause when checking if mTHP can be used for anon shmem, 'deny' is equal to
> 'never', and 'force' is equal to 'always'.

I'll admit it wasn't completely clear to me after reading the docs, but my rough
understanding is:

 - /sys/kernel/mm/transparent_hugepage/shmem_enabled controls
   mmap(SHARED|ANON) allocations (mostly; see rule 3)
 - huge=... controls tmpfs allocations
 - deny and force in shmem_enabled are equivalent to never and always for
   mmap(SHARED|ANON) but additionally override all tmpfs mounts so they act as
   if they were mounted with huge=never or huge=always

Is that correct? If so, then I think it still makes sense to support per-size
deny/force. Certainly if a per-size control is set to "inherit" and the
top-level control is set to deny or force, you would need that to mean something.

> 
>> Of course the huge= mount option would also need to take a per-size option in
>> this case. e.g. huge=2048kB:advise,64kB:always
> 
> IMO, I do not want to change the global 'huge=' mount option, which can control
> both anon shmem and tmpfs, but mTHP now is only applied for anon shmem. So let's

How does huge= control anon shmem? I thought it was only for mounted
filesystems; so tmpfs? Perhaps my mental model for how this works is broken...

> keep it be same with the global sysfs interface:
> /sys/kernel/mm/transparent_hugepage/shmem_enabled.
> 
> For tmpfs large folio strategy, I plan to address it later, and we may need more
> discussion to determine if it should follow the file large folio strategy or not
> (no investigation now).

OK. But until you get to tmpfs, you'll need an interim definition for what it
means if a per-size control is set to "inherit" and the top-level control is set
to deny/force.

> 
> Thanks for reviewing.

No problem! Thanks for doing the work!

> 
>>> TODO:
>>>   - More testing and provide some performance data.
>>>   - Need more discussion about the large folio allocation strategy for a
>>> 'regular
>>> file' operation created by memfd_create(), for example using ftruncate(fd) to
>>> specify
>>> the 'file' size, which need to follow the anonymous mTHP rule too?
>>>   - Do not split the large folio when share memory swap out.
>>>   - Can swap in a large folio for share memory.
>>>
>>> Baolin Wang (5):
>>>    mm: memory: extend finish_fault() to support large folio
>>>    mm: shmem: add an 'order' parameter for shmem_alloc_hugefolio()
>>>    mm: shmem: add THP validation for PMD-mapped THP related statistics
>>>    mm: shmem: add mTHP support for anonymous share pages
>>>    mm: shmem: add anonymous share mTHP counters
>>>
>>>   include/linux/huge_mm.h |   4 +-
>>>   mm/huge_memory.c        |   8 ++-
>>>   mm/memory.c             |  25 +++++++---
>>>   mm/shmem.c              | 107 ++++++++++++++++++++++++++++++----------
>>>   4 files changed, 108 insertions(+), 36 deletions(-)
>>>

Powered by blists - more mailing lists