On Thu, Nov 28, 2019 at 10:43:12AM +0100, Michal Hocko wrote: > On Wed 27-11-19 17:27:29, Roman Gushchin wrote: > > On Wed, Nov 27, 2019 at 01:32:25PM +0100, Michal Hocko wrote: > > > On Tue 26-11-19 18:41:41, Roman Gushchin wrote: > > > > On Tue, Nov 26, 2019 at 10:29:18AM +0100, Michal Hocko wrote: > > > > > On Mon 25-11-19 10:54:53, Roman Gushchin wrote: > > > [...] > > > > > > So in a rare case when not all children kmem_caches are destroyed > > > > > > at the moment when the root kmem_cache is about to be gone, we need > > > > > > to wait another rcu grace period before destroying the root > > > > > > kmem_cache. > > > > > > > > > > Could you explain how rare this really is please? > > > > > > > > It seems that we don't destroy root kmem_caches with enabled memcg > > > > accounting that often, but maybe I'm biased here. > > > > > > So this happens each time a root kmem_cache is destroyed? Which would > > > imply that only dynamically created ones? > > > > Yes, only dynamically created and only in those cases when destruction > > of the root cache happens immediately after the deactivation of the > > non-root cache. > > Tbh I can't imagine any other case except rmmod after > > removing the cgroup. > > Thanks for the confirmation! Could you please make this explicit in the > changelog please? Maybe it is obvious to you but it took me quite some > time to grasp what the hell is going on here. Both memcg and kmem_cache > destruction are quite complex and convoluted. Sure, will send v2 shorty. Thanks!
