On Mon, Jun 17, 2024 at 06:38:52PM +0200, Vlastimil Babka wrote: > On 6/17/24 6:33 PM, Jason A. Donenfeld wrote: > > On Mon, Jun 17, 2024 at 6:30 PM Uladzislau Rezki <urezki@xxxxxxxxx> wrote: > >> Here if an "err" is less then "0" means there are still objects > >> whereas "is_destroyed" is set to "true" which is not correlated > >> with a comment: > >> > >> "Destruction happens when no objects" > > > > The comment is just poorly written. But the logic of the code is right. > > > >> > >> > out_unlock: > >> > mutex_unlock(&slab_mutex); > >> > cpus_read_unlock(); > >> > diff --git a/mm/slub.c b/mm/slub.c > >> > index 1373ac365a46..7db8fe90a323 100644 > >> > --- a/mm/slub.c > >> > +++ b/mm/slub.c > >> > @@ -4510,6 +4510,8 @@ void kmem_cache_free(struct kmem_cache *s, void *x) > >> > return; > >> > trace_kmem_cache_free(_RET_IP_, x, s); > >> > slab_free(s, virt_to_slab(x), x, _RET_IP_); > >> > + if (s->is_destroyed) > >> > + kmem_cache_destroy(s); > >> > } > >> > EXPORT_SYMBOL(kmem_cache_free); > >> > > >> > @@ -5342,9 +5344,6 @@ static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) > >> > if (!slab->inuse) { > >> > remove_partial(n, slab); > >> > list_add(&slab->slab_list, &discard); > >> > - } else { > >> > - list_slab_objects(s, slab, > >> > - "Objects remaining in %s on __kmem_cache_shutdown()"); > >> > } > >> > } > >> > spin_unlock_irq(&n->list_lock); > >> > > >> Anyway it looks like it was not welcome to do it in the kmem_cache_free() > >> function due to performance reason. > > > > "was not welcome" - Vlastimil mentioned *potential* performance > > concerns before I posted this. I suspect he might have a different > > view now, maybe? > > > > Vlastimil, this is just checking a boolean (which could be > > unlikely()'d), which should have pretty minimal overhead. Is that > > alright with you? > > Well I doubt we can just set and check it without any barriers? The > completion of the last pending kfree_rcu() might race with > kmem_cache_destroy() in a way that will leave the cache there forever, no? > And once we add barriers it becomes a perf issue? Hm, yea you might be right about barriers being required. But actually, might this point toward a larger problem with no matter what approach, polling or event, is chosen? If the current rule is that kmem_cache_free() must never race with kmem_cache_destroy(), because users have always made diligent use of call_rcu()/rcu_barrier() and such, but now we're going to let those race with each other - either by my thing above or by polling - so we're potentially going to get in trouble and need some barriers anyway. I think? Jason