Re: [stable-6.6.y] mm: khugepaged refuses to freeze

[Baolin Wang]

On 2/6/26 4:36 PM, David Hildenbrand (Arm) wrote:
> On 2/6/26 06:12, Baolin Wang wrote:
> > On 2/6/26 12:31 PM, Sergey Senozhatsky wrote:
> > > On (26/02/06 12:38), Sergey Senozhatsky wrote:
> > > [..]
> > > > Right, I thought about it but wasn't sure.  Could the inner loop (e.g.
> > > > collapse_file() in this particular case) loop long enough to fail 
> > > > suspend
> > > > w/o ever giving the outer loop (khugepaged_do_scan()) a chance to 
> > > > freeze?
> >
> > Yes, that’s possible. However, if we add a try_to_freeze() check in 
> > the inner loop, we need to consider various scenarios (such as 
> > anonymous folio swap-in and other potential cases?), which feels too 
> > hacky to me.
> >
> > > For inner loops I wondered if cond_resched() could be an indicator of
> > > where try_to_freeze() should be placed.  Those cond_resched() calls
> > > are there for a reason, after all.   E.g. something like:
> > >
> > > ---
> > >
> > > diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> > > index fa6a018b20a8..cee08466a069 100644
> > > --- a/mm/khugepaged.c
> > > +++ b/mm/khugepaged.c
> > > @@ -2431,6 +2431,9 @@ static unsigned int 
> > > khugepaged_scan_mm_slot(unsigned int pages, enum scan_result
> > >           unsigned long hstart, hend;
> > >           cond_resched();
> > > +        if (try_to_freeze())
> > > +            break;
> > > +
> > >           if (unlikely(hpage_collapse_test_exit_or_disable(mm))) {
> > >               progress++;
> > >               break;
> > > @@ -2453,6 +2456,9 @@ static unsigned int 
> > > khugepaged_scan_mm_slot(unsigned int pages, enum scan_result
> > >               bool mmap_locked = true;
> > >               cond_resched();
> > > +            if (try_to_freeze())
> > > +                goto breakouterloop;
> > > +
> > >               if (unlikely(hpage_collapse_test_exit_or_disable(mm)))
> > >                   goto breakouterloop;
> >
> > This looks better than the previous version. Let’s also wait to see if 
> > others have any better suggestions.
>
> What prevents other callpaths (faults, read(), write(), etc) from 
> similarly triggering swapin?

Usually it’s just a userspace process triggering one page fault to swap 
a page in, then will return to userspace. There aren’t other kernel 
threads like khugepaged continuously do swap-in in a loop.

> I recall that there is a notifier when the system is preparing to sleep 
> (pm notifier or something). Could we simply hook into that to tell 
> khugepaged to suspend+resume?

Do you mean “struct dev_pm_ops”, which is used to register PM callbacks 
for devices? However, I don’t know how to use it with a kernel thread.

Also look at how kswapd does it, kswapd also uses 
kthread_freezable_should_stop() to check the freeze state.


> Essentially, making hpage_collapse_test_exit_or_disable() break our for us.

Ah, yes, even better:)