Re: [PATCH v1] mm: Fix khugepaged activation policy

[Ryan Roberts <ryan.roberts@arm.com>]

On 02/07/2024 15:57, David Hildenbrand wrote:
> On 02.07.24 16:46, Ryan Roberts wrote:
>> Since the introduction of mTHP, the docuementation has stated that
>> khugepaged would be enabled when any mTHP size is enabled, and disabled
>> when all mTHP sizes are disabled. There are 2 problems with this; 1.
>> this is not what was implemented by the code and 2. this is not the
>> desirable behavior.
>>
>> Desirable behavior is for khugepaged to be enabled when any PMD-sized
>> THP is enabled, anon or file. (Note that file THP is still controlled by
>> the top-level control so we must always consider that, as well as the
>> PMD-size mTHP control for anon). khugepaged only supports collapsing to
>> PMD-sized THP so there is no value in enabling it when PMD-sized THP is
>> disabled. So let's change the code and documentation to reflect this
>> policy.
>>
>> Further, per-size enabled control modification events were not
>> previously forwarded to khugepaged to give it an opportunity to start or
>> stop. Consequently the following was resulting in khugepaged eroneously
>> not being activated:
>>
>>    echo never > /sys/kernel/mm/transparent_hugepage/enabled
>>    echo always > /sys/kernel/mm/transparent_hugepage/hugepages-2048kB/enabled
>>
>> Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
>> Fixes: 3485b88390b0 ("mm: thp: introduce multi-size THP sysfs interface")
>> Closes:
>> https://lore.kernel.org/linux-mm/7a0bbe69-1e3d-4263-b206-da007791a5c4@redhat.com/
>> Cc: stable@vger.kernel.org
>> ---
>>
>> Hi All,
>>
>> Applies on top of today's mm-unstable (9bb8753acdd8). No regressions observed in
>> mm selftests.
>>
>> When fixing this I also noticed that khugepaged doesn't get (and never has been)
>> activated/deactivated by `shmem_enabled=`. I'm not sure if khugepaged knows how
>> to collapse shmem - perhaps it should be activated in this case?
>>
> 
> Call me confused.
> 
> khugepaged_scan_mm_slot() and madvise_collapse() only all
> hpage_collapse_scan_file() with ... IS_ENABLED(CONFIG_SHMEM) ?

Looks like khugepaged_scan_mm_slot() was converted from:

  if (shmem_file(vma->vm_file)) {

to:

  if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) {

By 99cb0dbd47a15d395bf3faa78dc122bc5efe3fc0 which adds THP collapse support for
non-shmem files. Clearly that looks wrong, but I guess never spotted in practice
because noone disables shemem?

I guess madvise_collapse() was a copy/paste?

> 
> collapse_file() is only called by hpage_collapse_scan_file() ... and there we
> check "shmem_file(file)".
> 
> So why is the IS_ENABLED(CONFIG_SHMEM) check in there if collapse_file() seems
> to "collapse filemap/tmpfs/shmem pages into huge one".
> 
> Anyhow, we certainly can collapse shmem (that's how it all started IIUC).

Yes, thanks for pointing me at it. Should have just searched "shmem" in
khugepaged.c :-/

> 
> Besides that, khugepaged only seems to collapse !shmem with
>   VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem);

That makes sense. I guess I could use IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) to
tighen the (non-shmem) file THP check in hugepage_pmd_enabled() (currently I'm
unconditionally using the top-level enabled setting as a "is THP enabled for
files" check).

But back to my original question, I think hugepage_pmd_enabled() should also be
explicitly checking the appropriate shmem_enabled controls and ORing in the
result? Otherwise in a situation where only shmem is THP enabled (and file/anon
THP is disabled) khugepaged won't run.

> 
> The thp_vma_allowable_order() check tests if we are allowed to collapse a
> PMD_ORDER in that VMA.

I don't follow the relevance of this statement.