Re: mm: memcontrol: rewrite uncharge API: problems

From: Johannes Weiner
Date: Mon Jul 07 2014 - 09:45:14 EST


On Fri, Jul 04, 2014 at 07:12:04PM -0700, Hugh Dickins wrote:
> On Thu, 3 Jul 2014, Johannes Weiner wrote:
> > On Thu, Jul 03, 2014 at 12:54:36PM -0700, Hugh Dickins wrote:
> > > On Wed, 2 Jul 2014, Hugh Dickins wrote:
> > > > On Wed, 2 Jul 2014, Johannes Weiner wrote:
> > > > >
> > > > > Could you give the following patch a spin? I put it in the mmots
> > > > > stack on top of mm-memcontrol-rewrite-charge-api-fix-shmem_unuse-fix.
> > > >
> > > > I'm just with the laptop until this evening. I slapped it on top of
> > > > my 3.16-rc2-mm1 plus fixes (but obviously minus my memcg_batch one
> > > > - which incidentally continues to run without crashing on the G5),
> > > > and it quickly gave me this lockdep splat, which doesn't look very
> > > > different from the one before.
> > > >
> > > > I see there's now an -rc3-mm1, I'll try it out on that in half an
> > > > hour... but unless I send word otherwise, assume that's the same.
> > >
> > > Yes, I get that lockdep report each time on -rc3-mm1 + your patch.
> >
> > There are two instances where I missed to make &rtpz->lock IRQ-safe:
> >
> > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > index 91b621846e10..bbaa3f4cf4db 100644
> > --- a/mm/memcontrol.c
> > +++ b/mm/memcontrol.c
> > @@ -3919,7 +3919,7 @@ unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
> > gfp_mask, &nr_scanned);
> > nr_reclaimed += reclaimed;
> > *total_scanned += nr_scanned;
> > - spin_lock(&mctz->lock);
> > + spin_lock_irq(&mctz->lock);
> >
> > /*
> > * If we failed to reclaim anything from this memory cgroup
> > @@ -3959,7 +3959,7 @@ unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
> > */
> > /* If excess == 0, no tree ops */
> > __mem_cgroup_insert_exceeded(mz, mctz, excess);
> > - spin_unlock(&mctz->lock);
> > + spin_unlock_irq(&mctz->lock);
> > css_put(&mz->memcg->css);
> > loop++;
> > /*
>
> Thanks, that fixes my lockdep reports.
>
> >
> > That should make it complete - but the IRQ toggling costs are fairly
> > high so I'm rewriting the batching code to use the page lists that
> > most uncharges have anyway, and then batch the no-IRQ sections.
>
> Sounds good.
>
> >
> > > I also twice got a flurry of res_counter.c:28 underflow warnings.
> > > Hmm, 62 of them each time (I was checking for a number near 512,
> > > which would suggest a THP/4k confusion, but no). The majority
> > > of them coming from mem_cgroup_reparent_charges.
> >
> > I haven't seen these yet. But the location makes sense: if there are
> > any imbalances they'll be noticed during a group's final uncharges.
>
> I haven't seen any since adding your patch above, though I don't see
> how it could affect them. Of course I'll let you know if they reappear.
>
> >
> > > But the laptop stayed up fine (for two hours before I had to stop
> > > it), and the G5 has run fine with that load for 16 hours now, no
> > > problems with release_pages, and not even a res_counter.c:28 (but
> > > I don't use lockdep on it).
> >
> > Great!
> >
> > > The x86 workstation ran fine for 4.5 hours, then hit some deadlock
> > > which I doubt had any connection to your changes: looked more like
> > > a jbd2 transaction was failing to complete (which, with me trying
> > > ext4 on loop on tmpfs, might be more my problem than anyone else's).
> > >
> > > Oh, but nearly forgot, I did an earlier run on the laptop last night,
> > > which crashed within minutes on
> > >
> > > VM_BUG_ON_PAGE(!(pc->flags & PCG_MEM))
> > > mm/memcontrol.c:6680!
> > > page had count 1 mapcount 0 mapping anon index 0x196
> > > flags locked uptodate reclaim swapbacked, pcflags 1, memcg not root
> > > mem_cgroup_migrate < move_to_new_page < migrate_pages < compact_zone <
> > > compact_zone_order < try_to_compact_pages < __alloc_pages_direct_compact <
> > > __alloc_pages_nodemask < alloc_pages_vma < do_huge_pmd_anonymous_page <
> > > handle_mm_fault < __do_page_fault
>
> I got it again on the laptop, after 7 hours.
>
> >
> > Haven't seen that one yet, either. The only way I can see this happen
> > is when the same page gets selected for migration twice in a row.
> > Maybe a race with putback, where it gets added to the LRU but isolated
> > by compaction before putback drops the refcount - will verify that.
>
> Yes. This is one of those cases where I read a mail too quickly,
> misunderstand it, set it aside, plough through the source files,
> pace around thinking, finally come up with a hypothesis, go back to
> answer the mail, and find I've arrived at the same conclusion as you.
>
> Not verified in any way, but yes, mem_cgroup_migrate() looks anomalous
> to me, in clearing PCG_MEM and PGC_MEMSW but leaving PCG_USED. Once
> that old page is put back on LRU for freeing, it could get isolated
> by another migrator, who discovers the anomalous state in its own
> mem_cgroup_migrate().
>
> mem_cgroup_migrate() should just set pc->flags = 0, shouldn't it?
>
> But is there any point to PCG_USED now? Couldn't PageCgroupUsed
> (or better, PageCgroupCharged) just test PCG_MEM and PCG_MEMSW?
> Which should be low bits of pc->mem_cgroup, halving the array.

This is a great idea, actually.

You are right that by the point mem_cgroup_migrate() takes away the
charges, we might be able to uncharge the page entirely, but I have to
review the page's context there to make sure this is still safe.

mem_cgroup_swapout() similarly moves charges away from the page right
before it's freed. If we were to uncharge the page in there too, we
could remove the flags entirely: pc->mem_cgroup is only set when the
page is charged, and whether the charge includes memory+swap depends
on do_swap_account, which is a runtime constant.

I'll get back to this once the bugs are ironed out :)

Thanks, Hugh!

---