Re: [PATCH v8 04/23] slab: add sheaf support for batching kfree_rcu() operations

[Harry Yoo <harry.yoo@oracle.com>]

On Wed, Sep 24, 2025 at 09:35:05PM -0700, Suren Baghdasaryan wrote:
> On Thu, Sep 18, 2025 at 1:09 AM Vlastimil Babka <vbabka@suse.cz> wrote:
> >
> > On 9/17/25 16:14, Vlastimil Babka wrote:
> > > On 9/17/25 15:34, Harry Yoo wrote:
> > >> On Wed, Sep 17, 2025 at 03:21:31PM +0200, Vlastimil Babka wrote:
> > >>> On 9/17/25 15:07, Harry Yoo wrote:
> > >>> > On Wed, Sep 17, 2025 at 02:05:49PM +0200, Vlastimil Babka wrote:
> > >>> >> On 9/17/25 13:32, Harry Yoo wrote:
> > >>> >> > On Wed, Sep 17, 2025 at 11:55:10AM +0200, Vlastimil Babka wrote:
> > >>> >> >> On 9/17/25 10:30, Harry Yoo wrote:
> > >>> >> >> > On Wed, Sep 10, 2025 at 10:01:06AM +0200, Vlastimil Babka wrote:
> > >>> >> >> >> +                          sfw->skip = true;
> > >>> >> >> >> +                          continue;
> > >>> >> >> >> +                  }
> > >>> >> >> >>
> > >>> >> >> >> +                  INIT_WORK(&sfw->work, flush_rcu_sheaf);
> > >>> >> >> >> +                  sfw->skip = false;
> > >>> >> >> >> +                  sfw->s = s;
> > >>> >> >> >> +                  queue_work_on(cpu, flushwq, &sfw->work);
> > >>> >> >> >> +                  flushed = true;
> > >>> >> >> >> +          }
> > >>> >> >> >> +
> > >>> >> >> >> +          for_each_online_cpu(cpu) {
> > >>> >> >> >> +                  sfw = &per_cpu(slub_flush, cpu);
> > >>> >> >> >> +                  if (sfw->skip)
> > >>> >> >> >> +                          continue;
> > >>> >> >> >> +                  flush_work(&sfw->work);
> > >>> >> >> >> +          }
> > >>> >> >> >> +
> > >>> >> >> >> +          mutex_unlock(&flush_lock);
> > >>> >> >> >> +  }
> > >>> >> >> >> +
> > >>> >> >> >> +  mutex_unlock(&slab_mutex);
> > >>> >> >> >> +  cpus_read_unlock();
> > >>> >> >> >> +
> > >>> >> >> >> +  if (flushed)
> > >>> >> >> >> +          rcu_barrier();
> > >>> >> >> >
> > >>> >> >> > I think we need to call rcu_barrier() even if flushed == false?
> > >>> >> >> >
> > >>> >> >> > Maybe a kvfree_rcu()'d object was already waiting for the rcu callback to
> > >>> >> >> > be processed before flush_all_rcu_sheaves() is called, and
> > >>> >> >> > in flush_all_rcu_sheaves() we skipped all (cache, cpu) pairs,
> > >>> >> >> > so flushed == false but the rcu callback isn't processed yet
> > >>> >> >> > by the end of the function?
> > >>> >> >> >
> > >>> >> >> > That sounds like a very unlikely to happen in a realistic scenario,
> > >>> >> >> > but still possible...
> > >>> >> >>
> > >>> >> >> Yes also good point, will flush unconditionally.
> > >>> >> >>
> > >>> >> >> Maybe in __kfree_rcu_sheaf() I should also move the call_rcu(...) before
> > >>> >> >> local_unlock().
> > >>> >> >>
> > >>> >> >> So we don't end up seeing a NULL pcs->rcu_free in
> > >>> >> >> flush_all_rcu_sheaves() because __kfree_rcu_sheaf() already set it to NULL,
> > >>> >> >> but didn't yet do the call_rcu() as it got preempted after local_unlock().
> > >>> >> >
> > >>> >> > Makes sense to me.
> > >>> >
> > >>> > Wait, I'm confused.
> > >>> >
> > >>> > I think the caller of kvfree_rcu_barrier() should make sure that it's invoked
> > >>> > only after a kvfree_rcu(X, rhs) call has returned, if the caller expects
> > >>> > the object X to be freed before kvfree_rcu_barrier() returns?
> > >>>
> > >>> Hmm, the caller of kvfree_rcu(X, rhs) might have returned without filling up
> > >>> the rcu_sheaf fully and thus without submitting it to call_rcu(), then
> > >>> migrated to another cpu. Then it calls kvfree_rcu_barrier() while another
> > >>> unrelated kvfree_rcu(X, rhs) call on the previous cpu is for the same
> > >>> kmem_cache (kvfree_rcu_barrier() is not only for cache destruction), fills
> > >>> up the rcu_sheaf fully and is about to call_rcu() on it. And since that
> > >>> sheaf also contains the object X, we should make sure that is flushed.
> > >>
> > >> I was going to say "but we queue and wait for the flushing work to
> > >> complete, so the sheaf containing object X should be flushed?"
> > >>
> > >> But nah, that's true only if we see pcs->rcu_free != NULL in
> > >> flush_all_rcu_sheaves().
> > >>
> > >> You are right...
> > >>
> > >> Hmm, maybe it's simpler to fix this by never skipping queueing the work
> > >> even when pcs->rcu_sheaf == NULL?
> > >
> > > I guess it's simpler, yeah.
> >
> > So what about this? The unconditional queueing should cover all races with
> > __kfree_rcu_sheaf() so there's just unconditional rcu_barrier() in the end.
> >
> > From 0722b29fa1625b31c05d659d1d988ec882247b38 Mon Sep 17 00:00:00 2001
> > From: Vlastimil Babka <vbabka@suse.cz>
> > Date: Wed, 3 Sep 2025 14:59:46 +0200
> > Subject: [PATCH] slab: add sheaf support for batching kfree_rcu() operations
> >
> > Extend the sheaf infrastructure for more efficient kfree_rcu() handling.
> > For caches with sheaves, on each cpu maintain a rcu_free sheaf in
> > addition to main and spare sheaves.
> >
> > kfree_rcu() operations will try to put objects on this sheaf. Once full,
> > the sheaf is detached and submitted to call_rcu() with a handler that
> > will try to put it in the barn, or flush to slab pages using bulk free,
> > when the barn is full. Then a new empty sheaf must be obtained to put
> > more objects there.
> >
> > It's possible that no free sheaves are available to use for a new
> > rcu_free sheaf, and the allocation in kfree_rcu() context can only use
> > GFP_NOWAIT and thus may fail. In that case, fall back to the existing
> > kfree_rcu() implementation.
> >
> > Expected advantages:
> > - batching the kfree_rcu() operations, that could eventually replace the
> >   existing batching
> > - sheaves can be reused for allocations via barn instead of being
> >   flushed to slabs, which is more efficient
> >   - this includes cases where only some cpus are allowed to process rcu
> >     callbacks (Android)
> 
> nit: I would say it's more CONFIG_RCU_NOCB_CPU related. Android is
> just an instance of that.
> 
> >
> > Possible disadvantage:
> > - objects might be waiting for more than their grace period (it is
> >   determined by the last object freed into the sheaf), increasing memory
> >   usage - but the existing batching does that too.
> >
> > Only implement this for CONFIG_KVFREE_RCU_BATCHED as the tiny
> > implementation favors smaller memory footprint over performance.
> >
> > Also for now skip the usage of rcu sheaf for CONFIG_PREEMPT_RT as the
> > contexts where kfree_rcu() is called might not be compatible with taking
> > a barn spinlock or a GFP_NOWAIT allocation of a new sheaf taking a
> > spinlock - the current kfree_rcu() implementation avoids doing that.
> >
> > Teach kvfree_rcu_barrier() to flush all rcu_free sheaves from all caches
> > that have them. This is not a cheap operation, but the barrier usage is
> > rare - currently kmem_cache_destroy() or on module unload.
> >
> > Add CONFIG_SLUB_STATS counters free_rcu_sheaf and free_rcu_sheaf_fail to
> > count how many kfree_rcu() used the rcu_free sheaf successfully and how
> > many had to fall back to the existing implementation.
> >
> > Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
> > ---
> >  mm/slab.h        |   3 +
> >  mm/slab_common.c |  26 +++++
> >  mm/slub.c        | 267 ++++++++++++++++++++++++++++++++++++++++++++++-
> >  3 files changed, 294 insertions(+), 2 deletions(-)
> >
> > diff --git a/mm/slub.c b/mm/slub.c
> > index cba188b7e04d..171273f90efd 100644
> > --- a/mm/slub.c
> > +++ b/mm/slub.c

[...snip...]

> > @@ -3840,6 +3895,77 @@ static void flush_all(struct kmem_cache *s)
> >         cpus_read_unlock();
> >  }
> >
> > +static void flush_rcu_sheaf(struct work_struct *w)
> > +{
> > +       struct slub_percpu_sheaves *pcs;
> > +       struct slab_sheaf *rcu_free;
> > +       struct slub_flush_work *sfw;
> > +       struct kmem_cache *s;
> > +
> > +       sfw = container_of(w, struct slub_flush_work, work);
> > +       s = sfw->s;
> > +
> > +       local_lock(&s->cpu_sheaves->lock);
> > +       pcs = this_cpu_ptr(s->cpu_sheaves);
> > +
> > +       rcu_free = pcs->rcu_free;
> > +       pcs->rcu_free = NULL;
> > +
> > +       local_unlock(&s->cpu_sheaves->lock);
> > +
> > +       if (rcu_free)
> > +               call_rcu(&rcu_free->rcu_head, rcu_free_sheaf_nobarn);
> > +}
> > +
> > +
> > +/* needed for kvfree_rcu_barrier() */
> > +void flush_all_rcu_sheaves(void)
> > +{
> > +       struct slub_flush_work *sfw;
> > +       struct kmem_cache *s;
> > +       unsigned int cpu;
> > +
> > +       cpus_read_lock();
> > +       mutex_lock(&slab_mutex);
> > +
> > +       list_for_each_entry(s, &slab_caches, list) {
> > +               if (!s->cpu_sheaves)
> > +                       continue;
> > +
> > +               mutex_lock(&flush_lock);
> > +
> > +               for_each_online_cpu(cpu) {
> > +                       sfw = &per_cpu(slub_flush, cpu);
> > +
> > +                       /*
> > +                        * we don't check if rcu_free sheaf exists - racing
> > +                        * __kfree_rcu_sheaf() might have just removed it.
> > +                        * by executing flush_rcu_sheaf() on the cpu we make
> > +                        * sure the __kfree_rcu_sheaf() finished its call_rcu()
> > +                        */
> > +
> > +                       INIT_WORK(&sfw->work, flush_rcu_sheaf);
> > +                       sfw->skip = false;
> 
> I think you don't need this sfw->skip flag since you never skip anymore, right?

Yes, at least in flush_all_rcu_sheaves().
I'm fine with or without sfw->skip in this function.

> > +                       sfw->s = s;
> > +                       queue_work_on(cpu, flushwq, &sfw->work);
> > +               }
> > +
> > +               for_each_online_cpu(cpu) {
> > +                       sfw = &per_cpu(slub_flush, cpu);
> > +                       if (sfw->skip)
> > +                               continue;
> > +                       flush_work(&sfw->work);
> 
> I'm sure I'm missing something but why can't we execute call_rcu()
> from here instead of queuing the work which does call_rcu() and then
> flushing all the queued work? I'm sure you have a good reason which
> I'm missing.

Because a local lock cannot be held by other CPUs, you can't take off the
rcu_free sheaf remotely and call call_rcu(). That's why the work is
queued on each CPU, ensuring the rcu_free sheaf is flushed by its local CPU.

> > +               }
> > +
> > +               mutex_unlock(&flush_lock);
> > +       }
> > +
> > +       mutex_unlock(&slab_mutex);
> > +       cpus_read_unlock();
> > +
> > +       rcu_barrier();
> > +}
> > +
> >  /*
> >   * Use the cpu notifier to insure that the cpu slabs are flushed when
> >   * necessary.
> > +bool __kfree_rcu_sheaf(struct kmem_cache *s, void *obj)
> > +{
> > +       struct slub_percpu_sheaves *pcs;
> > +       struct slab_sheaf *rcu_sheaf;
> > +
> > +       if (!local_trylock(&s->cpu_sheaves->lock))
> > +               goto fail;
> > +
> > +       pcs = this_cpu_ptr(s->cpu_sheaves);
> > +
> > +       if (unlikely(!pcs->rcu_free)) {
> > +
> > +               struct slab_sheaf *empty;
> > +               struct node_barn *barn;
> > +
> > +               if (pcs->spare && pcs->spare->size == 0) {
> > +                       pcs->rcu_free = pcs->spare;
> > +                       pcs->spare = NULL;
> > +                       goto do_free;
> > +               }
> > +
> > +               barn = get_barn(s);
> > +
> > +               empty = barn_get_empty_sheaf(barn);
> > +
> > +               if (empty) {
> > +                       pcs->rcu_free = empty;
> > +                       goto do_free;
> > +               }
> > +
> > +               local_unlock(&s->cpu_sheaves->lock);
> > +
> > +               empty = alloc_empty_sheaf(s, GFP_NOWAIT);
> > +
> > +               if (!empty)
> > +                       goto fail;
> > +
> > +               if (!local_trylock(&s->cpu_sheaves->lock)) {
> > +                       barn_put_empty_sheaf(barn, empty);
> > +                       goto fail;
> > +               }
> > +
> > +               pcs = this_cpu_ptr(s->cpu_sheaves);
> > +
> > +               if (unlikely(pcs->rcu_free))
> > +                       barn_put_empty_sheaf(barn, empty);
> > +               else
> > +                       pcs->rcu_free = empty;
> > +       }
> > +
> > +do_free:
> > +
> > +       rcu_sheaf = pcs->rcu_free;
> > +
> > +       rcu_sheaf->objects[rcu_sheaf->size++] = obj;
> 
> nit: The above would result in OOB write if we ever reached here with
> a full rcu_sheaf (rcu_sheaf->size == rcu_sheaf->sheaf_capacity) but I
> think it's impossible. You always start with an empty rcu_sheaf and
> objects are added only here with a following check for a full
> rcu_sheaf. I think a short comment clarifying that would be nice.

Sounds good to me.

-- 
Cheers,
Harry / Hyeonggon

> > +
> > +       if (likely(rcu_sheaf->size < s->sheaf_capacity))
> > +               rcu_sheaf = NULL;
> > +       else
> > +               pcs->rcu_free = NULL;
> > +
> > +       /*
> > +        * we flush before local_unlock to make sure a racing
> > +        * flush_all_rcu_sheaves() doesn't miss this sheaf
> > +        */
> > +       if (rcu_sheaf)
> > +               call_rcu(&rcu_sheaf->rcu_head, rcu_free_sheaf);
> > +
> > +       local_unlock(&s->cpu_sheaves->lock);
> > +
> > +       stat(s, FREE_RCU_SHEAF);
> > +       return true;
> > +
> > +fail:
> > +       stat(s, FREE_RCU_SHEAF_FAIL);
> > +       return false;
> > +}