On Sun, 29 May 2022, Hyeonggon Yoo wrote:
> > diff --git a/mm/slub.c b/mm/slub.c
> > index ed5c2c03a47a..310e56d99116 100644
> > --- a/mm/slub.c
> > +++ b/mm/slub.c
> > @@ -1374,15 +1374,12 @@ static noinline int free_debug_processing(
> > void *head, void *tail, int bulk_cnt,
> > unsigned long addr)
> > {
> > - struct kmem_cache_node *n = get_node(s, slab_nid(slab));
> > void *object = head;
> > int cnt = 0;
> > - unsigned long flags, flags2;
> > + unsigned long flags;
> > int ret = 0;
> >
> > - spin_lock_irqsave(&n->list_lock, flags);
> > - slab_lock(slab, &flags2);
> > -
> > + slab_lock(slab, &flags);
> > if (s->flags & SLAB_CONSISTENCY_CHECKS) {
> > if (!check_slab(s, slab))
> > goto out;
> > @@ -1414,8 +1411,7 @@ static noinline int free_debug_processing(
> > slab_err(s, slab, "Bulk freelist count(%d) invalid(%d)\n",
> > bulk_cnt, cnt);
> >
> > - slab_unlock(slab, &flags2);
> > - spin_unlock_irqrestore(&n->list_lock, flags);
> > + slab_unlock(slab, &flags);
> > if (!ret)
> > slab_fix(s, "Object at 0x%p not freed", object);
> > return ret;
> > @@ -3304,7 +3300,7 @@ static void __slab_free(struct kmem_cache *s, struct slab *slab,
> >
> > {
> > void *prior;
> > - int was_frozen;
> > + int was_frozen, to_take_off = 0;
> > struct slab new;
> > unsigned long counters;
> > struct kmem_cache_node *n = NULL;
> > @@ -3315,15 +3311,19 @@ static void __slab_free(struct kmem_cache *s, struct slab *slab,
> > if (kfence_free(head))
> > return;
> >
> > + n = get_node(s, slab_nid(slab));
> > + spin_lock_irqsave(&n->list_lock, flags);
> > +
>
> Oh please don't do this.
>
> SLUB free slowpath can be hit a lot depending on workload.
>
> __slab_free() try its best not to take n->list_lock. currently takes n->list_lock
> only when the slab need to be taken from list.
>
> Unconditionally taking n->list_lock will degrade performance.
>
This is a good point, it would be useful to gather some benchmarks for
workloads that are known to thrash some caches and would hit this path
such as netperf TCP_RR.
