Re: [PATCH v2] mm: shmem: don't set large-order range for internal shmem mount

[Zi Yan <ziy@nvidia.com>]

On 15 Apr 2026, at 21:22, Baolin Wang wrote:

> On 4/16/26 9:11 AM, Zi Yan wrote:
>> On 15 Apr 2026, at 21:05, Baolin Wang wrote:
>>
>>> On 4/15/26 10:36 PM, David Hildenbrand (Arm) wrote:
>>>> On 4/15/26 12:05, Baolin Wang wrote:
>>>>>
>>>>>
>>>>> On 4/15/26 5:54 PM, David Hildenbrand (Arm) wrote:
>>>>>>>
>>>>>>> Yes, that makes sense.
>>>>>>>
>>>>>>> However, it’s also possible that the mapping does not support large
>>>>>>> folios, yet anonymous shmem can still allocate large folios via the
>>>>>>> sysfs interfaces. That doesn't make sense, right?
>>>>>>
>>>>>> That's what I am saying: if there could be large folios in there, then
>>>>>> let's tell the world.
>>>>>>
>>>>>> Getting in a scenario where the mapping claims to not support large
>>>>>> folios, but then we have large folios in there is inconsistent, not?
>>>>>>
>>>>>> [...]
>>>>>>
>>>>>>>
>>>>>>> For the current anonymous shmem (tmpfs is already clear, no questions),
>>>>>>> I don’t think there will be any "will never have/does never allow"
>>>>>>> cases, because it can be changed dynamically via the sysfs interfaces.
>>>>>>
>>>>>> Right. It's about non-anon shmem with huge=off.
>>>>>>
>>>>>>>
>>>>>>> If we still want that logic, then for anonymous shmem we can treat it as
>>>>>>> always "might have large folios".
>>>>>
>>>>> OK. To resolve the confusion about 1, the logic should be changed as
>>>>> follows. Does that make sense to you?
>>>>>
>>>>> if (sbinfo->huge || (sb->s_flags & SB_KERNMOUNT))
>>>>>       mapping_set_large_folios(inode->i_mapping);
>>>>
>>>> I think that's better.
>>>
>>> Thanks for your valuable input.
>>>
>>> But has Willy says, maybe we can just
>>>> unconditionally set it and have it even simpler.
>>>
>>> However, for tmpfs mounts, we should still respect the 'huge=' mount option. See commit 5a90c155defa ("tmpfs: don't enable large folios if not supported").
>>
>> Is it possible to get sbinfo->huge during tmpfs’s folio allocation time, so that
>> even if all tmpfs has mapping_set_large_folios() but sbinfo->huge can still
>> decide whether huge page will be allocated for a tmpfs?
>
> Yes, of course. However, the issue isn’t whether tmpfs allows allocating large folios.
>
> The problem commit 5a90c155defa tries to fix is that when tmpfs is mounted with the 'huge=never' option, we will not allocate large folios for it. Then when writing tmpfs files, generic_perform_write() will call mapping_max_folio_size() to get the chunk size and ends up with an order-9 size for writing tmpfs files. However, this tmpfs file is populated only with small folios, resulting in a performance regression.

IIUC, generic_perform_write() needs to use a small chunk if tmpfs denies huge.
It seems that Kefeng did that in the first try[1]. But willy suggested
the current fix.

I wonder if we should revisit Kefeng’s first version.

[1] https://lore.kernel.org/all/20240914140613.2334139-1-wangkefeng.wang@huawei.com/

Best Regards,
Yan, Zi