Message-ID: <aPtIqm3LaRfCVQ8L@hyeyoo>
Date: Fri, 24 Oct 2025 18:36:42 +0900
From: Harry Yoo <harry.yoo@...cle.com>
To: Vlastimil Babka <vbabka@...e.cz>
Cc: Alexei Starovoitov <alexei.starovoitov@...il.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Christoph Lameter <cl@...two.org>,
        David Rientjes <rientjes@...gle.com>,
        Roman Gushchin <roman.gushchin@...ux.dev>,
        Alexei Starovoitov <ast@...nel.org>, linux-mm <linux-mm@...ck.org>,
        LKML <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH v2] slab: fix slab accounting imbalance due to
 defer_deactivate_slab()

On Fri, Oct 24, 2025 at 10:55:20AM +0200, Vlastimil Babka wrote:
> On 10/24/25 04:03, Harry Yoo wrote:
> > On Thu, Oct 23, 2025 at 06:17:19PM -0700, Alexei Starovoitov wrote:
> >> On Thu, Oct 23, 2025 at 5:00 PM Harry Yoo <harry.yoo@...cle.com> wrote:
> >> >
> >> > On Thu, Oct 23, 2025 at 04:13:37PM -0700, Alexei Starovoitov wrote:
> >> > > On Thu, Oct 23, 2025 at 5:01 AM Vlastimil Babka <vbabka@...e.cz> wrote:
> >> > > >
> >> > > > Since commit af92793e52c3 ("slab: Introduce kmalloc_nolock() and
> >> > > > kfree_nolock().") there's a possibility in alloc_single_from_new_slab()
> >> > > > that we discard the newly allocated slab if we can't spin and we fail to
> >> > > > trylock. As a result we don't perform inc_slabs_node() later in the
> >> > > > function. Instead we perform a deferred deactivate_slab() which can
> >> > > > either put the unacounted slab on partial list, or discard it
> >> > > > immediately while performing dec_slabs_node(). Either way will cause an
> >> > > > accounting imbalance.
> >> > > >
> >> > > > Fix this by not marking the slab as frozen, and using free_slab()
> >> > > > instead of deactivate_slab() for non-frozen slabs in
> >> > > > free_deferred_objects(). For CONFIG_SLUB_TINY, that's the only possible
> >> > > > case. By not using discard_slab() we avoid dec_slabs_node().
> >> > > >
> >> > > > Fixes: af92793e52c3 ("slab: Introduce kmalloc_nolock() and kfree_nolock().")
> >> > > > Signed-off-by: Vlastimil Babka <vbabka@...e.cz>
> >> > > > ---
> >> > > > Changes in v2:
> >> > > > - Fix the problem differently. Harry pointed out that we can't move
> >> > > >   inc_slabs_node() outside of list_lock protected regions as that would
> >> > > >   reintroduce issues fixed by commit c7323a5ad078
> >> > > > - Link to v1: https://patch.msgid.link/20251022-fix-slab-accounting-v1-1-27870ec363ce@suse.cz
> >> > > > ---
> >> > > >  mm/slub.c | 8 +++++---
> >> > > >  1 file changed, 5 insertions(+), 3 deletions(-)
> >> > > >
> >> > > > diff --git a/mm/slub.c b/mm/slub.c
> >> > > > index 23d8f54e9486..87a1d2f9de0d 100644
> >> > > > --- a/mm/slub.c
> >> > > > +++ b/mm/slub.c
> >> > > > @@ -3422,7 +3422,6 @@ static void *alloc_single_from_new_slab(struct kmem_cache *s, struct slab *slab,
> >> > > >
> >> > > >         if (!allow_spin && !spin_trylock_irqsave(&n->list_lock, flags)) {
> >> > > >                 /* Unlucky, discard newly allocated slab */
> >> > > > -               slab->frozen = 1;
> >> > > >                 defer_deactivate_slab(slab, NULL);
> >> > > >                 return NULL;
> >> > > >         }
> >> > > > @@ -6471,9 +6470,12 @@ static void free_deferred_objects(struct irq_work *work)
> >> > > >                 struct slab *slab = container_of(pos, struct slab, llnode);
> >> > > >
> >> > > >  #ifdef CONFIG_SLUB_TINY
> >> > > > -               discard_slab(slab->slab_cache, slab);
> >> > > > +               free_slab(slab->slab_cache, slab);
> >> > > >  #else
> >> > > > -               deactivate_slab(slab->slab_cache, slab, slab->flush_freelist);
> >> > > > +               if (slab->frozen)
> >> > > > +                       deactivate_slab(slab->slab_cache, slab, slab->flush_freelist);
> >> > > > +               else
> >> > > > +                       free_slab(slab->slab_cache, slab);
> >> > >
> >> > > A bit odd to use 'frozen' flag as such a signal.
> >> > > I guess I'm worried that truly !frozen slab can come here
> >> > > via ___slab_alloc() -> retry_load_slab: -> defer_deactivate_slab().
> >> > > And things will be much worse than just accounting.
> >> >
> >> > But the cpu slab must have been frozen before it's attached to
> >> > c->slab?
> 
> Note that deactivate_slab() contains VM_BUG_ON(!old.frozen);
> we would have seen this triggered if we were passing unfrozen slabs to
> (defer_)deactivate_slab(). I assume it's also why the "unlucky, discard"
> code marks it frozen before calling defer_deactivate_slab() (and this patch
> removes that).
> 
> >> Is it?
> >> the path is
> >> c = slub_get_cpu_ptr(s->cpu_slab);
> >> if (unlikely(c->slab)) {
> >>    struct slab *flush_slab = c->slab;
> >>    defer_deactivate_slab(flush_slab, ...);
> >> 
> >> I don't see why it would be frozen.
> 
> c->slab is always frozen, that's an invariant
> 
> > 
> > Oh god. I was going to say the cpu slab is always frozen. It has been
> > true for very long time, but it seems it's not true after commit 90b1e56641
> > ("mm/slub: directly load freelist from cpu partial slab in the likely case").
> 
> It's still true. That commit only removes VM_BUG_ON(!new.frozen); where
> "new" is in fact the old state - when slab is on cpu partial list it's not
> yet frozen. get_freelist() then sets new.frozen = freelist != NULL;

(scratching cheek in embarrassment) You're right.

I was thinking that changing from calling freeze_slab() & get_freelist()
to calling just get_freelist() would make it non-frozen, but actually
get_freelist() freezes it! My mistake.

> and we know that freelist cant't be NULL for a slab on the cpu partial list.
> The commit even added VM_BUG_ON(!freelist); on the get_freelist() result for
> this case.

Yes, as long as it's in percpu partial slab list, it cannot be NULL.

> So I think we're fine?

Yes.

> > So I think you're right that a non-frozen slab can go through
> > free_slab() in free_deferred_objects()...
> > 
> > But fixing this should be simple. Add something like
> > freeze_and_get_freelist() and call it when SLUB take a slab from
> > per-cpu partial slab list?
> > 
> >> > > Maybe add
> >> > >   inc_slabs_node(s, nid, slab->objects);
> >> > > right before
> >> > >   defer_deactivate_slab(slab, NULL);
> >> > >   return NULL;
> >> > >
> >> > > I don't quite get why c7323a5ad078 is doing everything under n->list_lock.
> >> > > It's been 3 years since.
> >> >
> >> > When n->nr_slabs is inconsistent, validate_slab_node() might report an
> >> > error (false positive) when someone wrote '1' to
> >> > /sys/kernel/slab/<cache name>/validate
> >> 
> >> Ok. I see it now. It's the actual number of elements in n->full
> >> list needs to match n->nr_slabs.
> >> 
> >> But then how it's not broken already?
> >> I see that
> >> alloc_single_from_new_slab()
> >> unconditionally does inc_slabs_node(), but
> > 
> > It increments n->nr_slabs. It doesn't matter which list it's going to be
> > added to, because it's total number of slabs in that node.
> > 
> >> slab itself is added either to n->full or n->partial lists.
> > 
> > and then n->nr_partial is also incremented if it's added to n->partial.
> > 
> >> And validate_slab_node() should be complaining already.
> > 
> > The debug routine checks if:
> > - the number of slabs in n->partial == n->nr_partial
> > - the number of slabs in n->full + n->partial == n->nr_slabs
> > 
> > under n->list_lock. So it's not broken?
> > 
> >> Anyway, I'm not arguing. Just trying to understand.
> >> If you think the fix is fine, then go ahead.
> > 
> 

-- 
Cheers,
Harry / Hyeonggon

Powered by blists - more mailing lists