Re: [PATCH] mm: Optimize TLB flushes during page reclaim

[SeongJae Park <sj@kernel.org>]

On Thu, 23 Jan 2025 11:11:13 -0600 Vinay Banakar <vny@google.com> wrote:

> On Wed, Jan 22, 2025 at 2:05 PM SeongJae Park <sj@kernel.org> wrote:
> > damon_pa_pageout() from mm/damon/paddr.c also calls shrink_folio_list() similar
> > to madvise.c, but it doesn't aware such batching behavior.  Have you checked
> > that path?
> 
> Thanks for catching this path. In damon_pa_pageout(),
> shrink_folio_list() processes all pages from a single NUMA node that
> were collected (filtered) from a single DAMON region (r->ar.start to
> r->ar.end). This means it could be processing anywhere from 1 page up
> to ULONG_MAX pages from a single node at once.

Thank you Vinay.  That's same to my understanding, except that it is not
limited to a single NUMA node.  A region can have any start and end physical
addresses, so it could cover memory of different NUMA nodes.

> With the patch, we'll
> send a single IPI for TLB flush for the entire region, reducing IPIs
> by a factor equal to the number of pages being reclaimed by DAMON at
> once (decided by damon_reclaim_quota).

I guess the fact that the pages could belong to differnt NUMA nodes doesn't
make difference here?

> 
> My only concern here would be the overhead of maintaining the
> temporary pageout_list for batching. However, during BIO submission,
> the patch checks if the folio was reactivated, so submitting to BIO in
> bulk should be safe.
> 
> Another option would be to modify shrink_folio_list() to force batch
> flushes for up to N pages (512) at a time, rather than relying on
> callers to do the batching via folio_list.

Both sounds good to me :)


Thanks,
SJ

> 
> Thanks!
> Vinay