* Yang Shi:
> On Tue, Jan 30, 2024 at 11:53 PM Florian Weimer <fweimer@xxxxxxxxxx> wrote:
>>
>> * Yang Shi:
>>
>> > From: Yang Shi <yang@xxxxxxxxxxxxxxxxxxxxxx>
>> >
>> > The commit efa7df3e3bb5 ("mm: align larger anonymous mappings on THP
>> > boundaries") incured regression for stress-ng pthread benchmark [1].
>> > It is because THP get allocated to pthread's stack area much more possible
>> > than before. Pthread's stack area is allocated by mmap without VM_GROWSDOWN
>> > or VM_GROWSUP flag, so kernel can't tell whether it is a stack area or not.
>> >
>> > The MAP_STACK flag is used to mark the stack area, but it is a no-op on
>> > Linux. Mapping MAP_STACK to VM_NOHUGEPAGE to prevent from allocating
>> > THP for such stack area.
>>
>> Doesn't this introduce a regression in the other direction, where
>> workloads expect to use a hugepage TLB entry for the stack?
>
> Maybe, it is theoretically possible. But AFAICT, the real life
> workloads performance usually gets hurt if THP is used for stack.
> Willy has an example:
>
> https://lore.kernel.org/linux-mm/ZYPDwCcAjX+r+g6s@xxxxxxxxxxxxxxxxxxxx/#t
>
> And avoiding THP on stack is not new, VM_GROWSDOWN | VM_GROWSUP areas
> have been applied before, this patch just extends this to MAP_STACK.
If it's *always* beneficial then we should help it along in glibc as
well. We've started to offer a tunable in response to this observation
(also paper over in OpenJDK):
Make thread stacks not use huge pages
<https://bugs.openjdk.org/browse/JDK-8303215>
But this is specifically about RSS usage, and not directly about
reducing TLB misses etc.
Thanks,
Florian
