Re: [PATCH v5 5/7] sched/numa: scan PID-inactive VMAs for promotion
From: Gregory Price (Meta)
Date: Fri Nov 13 2026 - 16:59:30 EST
On 2026-11-10 15:36 -0500, Gregory Price (Meta) wrote:
> On 2026-11-10 10:04 -0500, Gregory Price (Meta) wrote:
> ... snip ...
>
> Sashiko report feedback
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -4249,7 +4249,7 @@ static bool vma_needs_placement_scan(struct vm_area_struct *vma)
> > * threads can help scan this vma, force a vma scan.
> > */
> > if (READ_ONCE(mm->numa_scan_seq) >
> > - (vma->numab_state->prev_scan_seq + get_nr_threads(current)))
> > + (vma->numab_state->prev_placement_scan_seq + get_nr_threads(current)))
> > return true;
> ... snip ...
> > + if (placement_scan)
> > + vma->numab_state->prev_placement_scan_seq = mm->numa_scan_seq;
>
> Sashiko took issue with these skips, but its reported impact is
> overstated. It tries to point out a pre-existing condition:
>
> prev_scan_seq skips an update on a rare corner condition, but
> now the same skip applies to prev_placement_scan_seq.
>
> In modes 1 (placement) and 2 (tiering), the behavior is exactly the
> same as it was.
>
> In mode 3 (placement + tiering) the result is that the starvation
> counter isn't reset - but everything still works anyway because the
> VMA is still being scanned.
>
> I can follow up with *another* commit, but it's really low impact.
> Not sure it's worth the churn.
>
I will refresh this next week and just address it in another commit.
--
~Gregory