Re: [PATCH 00/14] replace call_rcu by kfree_rcu for simple kmem_cache_free callback

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 6/19/24 11:51 AM, Uladzislau Rezki wrote:
> On Tue, Jun 18, 2024 at 09:48:49AM -0700, Paul E. McKenney wrote:
>> On Tue, Jun 18, 2024 at 11:31:00AM +0200, Uladzislau Rezki wrote:
>> > > On 6/17/24 8:42 PM, Uladzislau Rezki wrote:
>> > > >> +
>> > > >> +	s = container_of(work, struct kmem_cache, async_destroy_work);
>> > > >> +
>> > > >> +	// XXX use the real kmem_cache_free_barrier() or similar thing here
>> > > > It implies that we need to introduce kfree_rcu_barrier(), a new API, which i
>> > > > wanted to avoid initially.
>> > > 
>> > > I wanted to avoid new API or flags for kfree_rcu() users and this would
>> > > be achieved. The barrier is used internally so I don't consider that an
>> > > API to avoid. How difficult is the implementation is another question,
>> > > depending on how the current batching works. Once (if) we have sheaves
>> > > proven to work and move kfree_rcu() fully into SLUB, the barrier might
>> > > also look different and hopefully easier. So maybe it's not worth to
>> > > invest too much into that barrier and just go for the potentially
>> > > longer, but easier to implement?
>> > > 
>> > Right. I agree here. If the cache is not empty, OK, we just defer the
>> > work, even we can use a big 21 seconds delay, after that we just "warn"
>> > if it is still not empty and leave it as it is, i.e. emit a warning and
>> > we are done.
>> > 
>> > Destroying the cache is not something that must happen right away. 
>> 
>> OK, I have to ask...
>> 
>> Suppose that the cache is created and destroyed by a module and
>> init/cleanup time, respectively.  Suppose that this module is rmmod'ed
>> then very quickly insmod'ed.
>> 
>> Do we need to fail the insmod if the kmem_cache has not yet been fully
>> cleaned up?  If not, do we have two versions of the same kmem_cache in
>> /proc during the overlap time?
>> 
> No fail :) If same cache is created several times, its s->refcount gets
> increased, so, it does not create two entries in the "slabinfo". But i
> agree that your point is good! We need to be carefully with removing and
> simultaneous creating.

Note that this merging may be disabled or not happen due to various flags on
the cache being incompatible with it. And I want to actually make sure it
never happens for caches being already destroyed as that would lead to
use-after-free (the workfn doesn't recheck the refcount in case a merge
would happen during the grace period)

--- a/mm/slab_common.c
+++ b/mm/slab_common.c
@@ -150,9 +150,10 @@ int slab_unmergeable(struct kmem_cache *s)
 #endif

        /*
-        * We may have set a slab to be unmergeable during bootstrap.
+        * We may have set a cache to be unmergeable during bootstrap.
+        * 0 is for cache being destroyed asynchronously
         */
-       if (s->refcount < 0)
+       if (s->refcount <= 0)
                return 1;

        return 0;

> From the first glance, there is a refcounter and a global "slab_mutex"
> which is used to protect a critical section. Destroying is almost fully
> protected(as noted above, by a global mutex) with one exception, it is:
> 
> static void kmem_cache_release(struct kmem_cache *s)
> {
> 	if (slab_state >= FULL) {
> 		sysfs_slab_unlink(s);
> 		sysfs_slab_release(s);
> 	} else {
> 		slab_kmem_cache_release(s);
> 	}
> }
> 
> this one can race, IMO.
> 
> --
> Uladzislau Rezki





[Index of Archives]     [Linux Crypto]     [Device Mapper Crypto]     [LARTC]     [Bugtraq]     [Yosemite Forum]

  Powered by Linux