Message-ID: <756c359e-bb8f-481e-a33f-163c729afa31@redhat.com>
Date: Sat, 13 Jul 2024 03:08:50 +0200
From: David Hildenbrand <david@...hat.com>
To: Lance Yang <ioworker0@...il.com>,
 Baolin Wang <baolin.wang@...ux.alibaba.com>
Cc: Ryan Roberts <ryan.roberts@....com>,
 Andrew Morton <akpm@...ux-foundation.org>, Hugh Dickins <hughd@...gle.com>,
 Jonathan Corbet <corbet@....net>,
 "Matthew Wilcox (Oracle)" <willy@...radead.org>,
 Barry Song <baohua@...nel.org>, linux-kernel@...r.kernel.org,
 linux-mm@...ck.org
Subject: Re: [PATCH v1 2/2] mm: mTHP stats for pagecache folio allocations

On 12.07.24 14:22, Lance Yang wrote:
> On Fri, Jul 12, 2024 at 11:00 AM Baolin Wang
> <baolin.wang@...ux.alibaba.com> wrote:
>>
>>
>>
>> On 2024/7/11 15:29, Ryan Roberts wrote:
>>> Expose 3 new mTHP stats for file (pagecache) folio allocations:
>>>
>>>     /sys/kernel/mm/transparent_hugepage/hugepages-*kB/stats/file_alloc
>>>     /sys/kernel/mm/transparent_hugepage/hugepages-*kB/stats/file_fallback
>>>     /sys/kernel/mm/transparent_hugepage/hugepages-*kB/stats/file_fallback_charge
>>>
>>> This will provide some insight on the sizes of large folios being
>>> allocated for file-backed memory, and how often allocation is failing.
>>>
>>> All non-order-0 (and most order-0) folio allocations are currently done
>>> through filemap_alloc_folio(), and folios are charged in a subsequent
>>> call to filemap_add_folio(). So count file_fallback when allocation
>>> fails in filemap_alloc_folio() and count file_alloc or
>>> file_fallback_charge in filemap_add_folio(), based on whether charging
>>> succeeded or not. There are some users of filemap_add_folio() that
>>> allocate their own order-0 folio by other means, so we would not count
>>> an allocation failure in this case, but we also don't care about order-0
>>> allocations. This approach feels like it should be good enough and
>>> doesn't require any (impractically large) refactoring.
>>>
>>> The existing mTHP stats interface is reused to provide consistency to
>>> users. And because we are reusing the same interface, we can reuse the
>>> same infrastructure on the kernel side. The one small wrinkle is that
>>> the set of folio sizes supported by the pagecache are not identical to
>>> those supported by anon and shmem; pagecache supports order-1, unlike
>>> anon and shmem, and the max pagecache order may be less than PMD-size
>>> (see arm64 with 64K base pages), again unlike anon and shmem. So we now
>>> create a hugepages-*kB directory for the union of the sizes supported by
>>> all 3 memory types and populate it with the relevant stats and controls.
>>
>> Personally, I like the idea that can help analyze the allocation of
>> large folios for the page cache.
>>
>> However, I have a slight concern about the consistency of the interface.
>>
>> For 64K, the fields layout:
>> ├── hugepages-64kB
>> │   ├── enabled
>> │   ├── shmem_enabled
>> │   └── stats
>> │       ├── anon_fault_alloc
>> │       ├── anon_fault_fallback
>> │       ├── anon_fault_fallback_charge
>> │       ├── file_alloc
>> │       ├── file_fallback
>> │       ├── file_fallback_charge
>> │       ├── shmem_alloc
>> │       ├── shmem_fallback
>> │       ├── shmem_fallback_charge
>> │       ├── split
>> │       ├── split_deferred
>> │       ├── split_failed
>> │       ├── swpout
>> │       └── swpout_fallback
>>
>> But for 8K (for pagecache), you removed some fields (of course, I
>> understand why they are not supported).
>>
>> ├── hugepages-8kB
>> │   └── stats
>> │       ├── file_alloc
>> │       ├── file_fallback
>> │       └── file_fallback_charge
>>
>> This might not be user-friendly for some user-space parsing tools, as
>> they lack certain fields for the same pattern interfaces. Of course,
>> this might not be an issue if we have clear documentation describing the
>> differences here:)
>>
>> Another possible approach is to maintain the same field layout to keep
>> consistent, but prohibit writing to the fields that are not supported by
>> the pagecache, and any stats read from them would be 0.
> 
> I agree that maintaining a uniform field layout, especially at the stats
> level, might be necessary ;)
> 
> Keeping a consistent interface could future-proof the design. It allows
> for the possibility that features not currently supported for 8kB pages
> might be enabled in the future.

I'll just note that, with shmem/file effectively being disabled for 
order > 11, we'll also have entries there that are effectively unused.

Good question how we want to deal with that (stats are easy, but what 
about when we enable something? Maybe we should document that "enabled" 
is only effective when supported).

Hmmmmm

-- 
Cheers,

David / dhildenb

Powered by blists - more mailing lists