Re: [RFC PATCH v3 0/4] Support large folios for tmpfs

[David Hildenbrand <david@redhat.com>]

On 22.10.24 05:41, Baolin Wang wrote:
> 
> 
> On 2024/10/21 21:34, Daniel Gomez wrote:
>> On Mon Oct 21, 2024 at 10:54 AM CEST, Kirill A. Shutemov wrote:
>>> On Mon, Oct 21, 2024 at 02:24:18PM +0800, Baolin Wang wrote:
>>>>
>>>>
>>>> On 2024/10/17 19:26, Kirill A. Shutemov wrote:
>>>>> On Thu, Oct 17, 2024 at 05:34:15PM +0800, Baolin Wang wrote:
>>>>>> + Kirill
>>>>>>
>>>>>> On 2024/10/16 22:06, Matthew Wilcox wrote:
>>>>>>> On Thu, Oct 10, 2024 at 05:58:10PM +0800, Baolin Wang wrote:
>>>>>>>> Considering that tmpfs already has the 'huge=' option to control the THP
>>>>>>>> allocation, it is necessary to maintain compatibility with the 'huge='
>>>>>>>> option, as well as considering the 'deny' and 'force' option controlled
>>>>>>>> by '/sys/kernel/mm/transparent_hugepage/shmem_enabled'.
>>>>>>>
>>>>>>> No, it's not.  No other filesystem honours these settings.  tmpfs would
>>>>>>> not have had these settings if it were written today.  It should simply
>>>>>>> ignore them, the way that NFS ignores the "intr" mount option now that
>>>>>>> we have a better solution to the original problem.
>>>>>>>
>>>>>>> To reiterate my position:
>>>>>>>
>>>>>>>      - When using tmpfs as a filesystem, it should behave like other
>>>>>>>        filesystems.
>>>>>>>      - When using tmpfs to implement MAP_ANONYMOUS | MAP_SHARED, it should
>>>>>>>        behave like anonymous memory.
>>>>>>
>>>>>> I do agree with your point to some extent, but the ‘huge=’ option has
>>>>>> existed for nearly 8 years, and the huge orders based on write size may not
>>>>>> achieve the performance of PMD-sized THP in some scenarios, such as when the
>>>>>> write length is consistently 4K. So, I am still concerned that ignoring the
>>>>>> 'huge' option could lead to compatibility issues.
>>>>>
>>>>> Yeah, I don't think we are there yet to ignore the mount option.
>>>>
>>>> OK.
>>>>
>>>>> Maybe we need to get a new generic interface to request the semantics
>>>>> tmpfs has with huge= on per-inode level on any fs. Like a set of FADV_*
>>>>> handles to make kernel allocate PMD-size folio on any allocation or on
>>>>> allocations within i_size. I think this behaviour is useful beyond tmpfs.
>>>>>
>>>>> Then huge= implementation for tmpfs can be re-defined to set these
>>>>> per-inode FADV_ flags by default. This way we can keep tmpfs compatible
>>>>> with current deployments and less special comparing to rest of
>>>>> filesystems on kernel side.
>>>>
>>>> I did a quick search, and I didn't find any other fs that require PMD-sized
>>>> huge pages, so I am not sure if FADV_* is useful for filesystems other than
>>>> tmpfs. Please correct me if I missed something.
>>>
>>> What do you mean by "require"? THPs are always opportunistic.
>>>
>>> IIUC, we don't have a way to hint kernel to use huge pages for a file on
>>> read from backing storage. Readahead is not always the right way.
>>>
>>>>> If huge= is not set, tmpfs would behave the same way as the rest of
>>>>> filesystems.
>>>>
>>>> So if 'huge=' is not set, tmpfs write()/fallocate() can still allocate large
>>>> folios based on the write size? If yes, that means it will change the
>>>> default huge behavior for tmpfs. Because previously having 'huge=' is not
>>>> set means the huge option is 'SHMEM_HUGE_NEVER', which is similar to what I
>>>> mentioned:
>>>> "Another possible choice is to make the huge pages allocation based on write
>>>> size as the *default* behavior for tmpfs, ..."
>>>
>>> I am more worried about breaking existing users of huge pages. So changing
>>> behaviour of users who don't specify huge is okay to me.
>>
>> I think moving tmpfs to allocate large folios opportunistically by
>> default (as it was proposed initially) doesn't necessary conflict with
>> the default behaviour (huge=never). We just need to clarify that in
>> the documentation.
>>
>> However, and IIRC, one of the requests from Hugh was to have a way to
>> disable large folios which is something other FS do not have control
>> of as of today. Ryan sent a proposal to actually control that globally
>> but I think it didn't move forward. So, what are we missing to go back
>> to implement large folios in tmpfs in the default case, as any other fs
>> leveraging large folios?
> 
> IMHO, as I discussed with Kirill, we still need maintain compatibility
> with the 'huge=' mount option. This means that if 'huge=never' is set
> for tmpfs, huge page allocation will still be prohibited (which can
> address Hugh's request?). However, if 'huge=' is not set, we can
> allocate large folios based on the write size.

I consider allocating large folios in shmem/tmpfs on the write path less 
controversial than allocating them on the page fault path -- especially 
as long as we stay within the size to-be-written.

I think in RHEL THP on shmem/tmpfs are disabled as default (e.g., 
shmem_enabled=never). Maybe because of some rather undesired 
side-effects (maybe some are historical?): I recall issues with VMs with 
THP+ memory ballooning, as we cannot reclaim pages of folios if 
splitting fails). I assume most of these problematic use cases don't use 
tmpfs as an ordinary file system (write()/read()), but mmap() the whole 
thing.

Sadly, I don't find any information about shmem/tmpfs + THP in the RHEL 
documentation; most documentation is only concerned about anon THP. 
Which makes me conclude that they are not suggested as of now.

I see more issues with allocating them on the page fault path and not 
having a way to disable it -- compared to allocating them on the write() 
path.

Getting Hugh's opinion in this would be very valuable.

-- 
Cheers,

David / dhildenb