On Thu 19-12-13 13:29:59, Vladimir Davydov wrote: > On 12/19/2013 01:21 PM, Michal Hocko wrote: > > On Thu 19-12-13 13:16:01, Vladimir Davydov wrote: > >> On 12/19/2013 01:10 PM, Michal Hocko wrote: > >>> On Thu 19-12-13 10:37:27, Vladimir Davydov wrote: > >>>> On 12/18/2013 09:14 PM, Michal Hocko wrote: > >>>>> On Wed 18-12-13 17:16:54, Vladimir Davydov wrote: > >>>>>> First, in memcg_create_kmem_cache() we should issue the write barrier > >>>>>> after the kmem_cache is initialized, but before storing the pointer to > >>>>>> it in its parent's memcg_params. > >>>>>> > >>>>>> Second, we should always issue the read barrier after > >>>>>> cache_from_memcg_idx() to conform with the write barrier. > >>>>>> > >>>>>> Third, its better to use smp_* versions of barriers, because we don't > >>>>>> need them on UP systems. > >>>>> Please be (much) more verbose on Why. Barriers are tricky and should be > >>>>> documented accordingly. So if you say that we should issue a barrier > >>>>> always be specific why we should do it. > >>>> In short, we have kmem_cache::memcg_params::memcg_caches is an array of > >>>> pointers to per-memcg caches. We access it lock-free so we should use > >>>> memory barriers during initialization. Obviously we should place a write > >>>> barrier just before we set the pointer in order to make sure nobody will > >>>> see a partially initialized structure. Besides there must be a read > >>>> barrier between reading the pointer and accessing the structure, to > >>>> conform with the write barrier. It's all that similar to rcu_assign and > >>>> rcu_deref. Currently the barrier usage looks rather strange: > >>>> > >>>> memcg_create_kmem_cache: > >>>> initialize kmem > >>>> set the pointer in memcg_caches > >>>> wmb() // ??? > >>>> > >>>> __memcg_kmem_get_cache: > >>>> <...> > >>>> read_barrier_depends() // ??? > >>>> cachep = root_cache->memcg_params->memcg_caches[memcg_id] > >>>> <...> > >>> Why do we need explicit memory barriers when we can use RCU? > >>> __memcg_kmem_get_cache already dereferences within rcu_read_lock. > >> Because it's not RCU, IMO. RCU implies freeing the old version after a > >> grace period, while kmem_caches are freed immediately. We simply want to > >> be sure the kmem_cache is fully initialized. And we do not require > >> calling this in an RCU critical section. > > And you can use rcu_dereference and rcu_assign for that as well. > > rcu_dereference() will complain if called outside an RCU critical > section, while cache_from_memcg_idx() is called w/o RCU protection from > some places. Does anything prevents us from using RCU from those callers as well? > > It hides all the juicy details about memory barriers. > > IMO, a memory barrier with a good comment looks better than an > rcu_dereference() without RCU protection :-) OK, let's wait for a good comment then ;) > > Besides that nothing prevents us from freeing from rcu callback. Or? > > It's an overhead we can live without there. The point is that we can > access a cache only if it is active. I mean no allocation can go from a > cache that has already been destroyed. It would be a bug. So there is no > point in introducing RCU-protection for kmem_caches there. It would only > confuse, IMO. My point was that the current state is a disaster. Implicit assumptions on different locking with memory barriers to make it even more juicy. This should be cleaned up really. Replacing explicit memory barriers by RCU sounds like a straightforward and much easier to follow for many people (unlike memory barriers). I do not insist on RCU but please make this code comprehensible. My head is spinning anytime I look down there and try to find out which locks are actually held and whether that is safe. > > Thanks. -- Michal Hocko SUSE Labs -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>