On Sun, Oct 08, 2017 at 03:03:32AM +0100, Al Viro wrote: > On Sun, Oct 08, 2017 at 01:56:08AM +0100, Al Viro wrote: > > > What's more, we need to be careful about resize vs. drain. Right now it's > > on list_lrus_mutex, but if we drop that around actual resize of an individual > > list_lru, we'll need something else. Would there be any problem if we > > took memcg_cache_ids_sem shared in memcg_offline_kmem()? > > > > The first problem is not fatal - we can e.g. use the sign of the field used > > to store the number of ->memcg_lrus elements (i.e. stashed value of > > memcg_nr_cache_ids at allocation or last resize) to indicate that actual > > freeing is left for resizer... > > Ugh. That spinlock would have to be held over too much work, or bounced back > and forth a lot on memcg shutdowns ;-/ Gets especially nasty if we want > list_lru_destroy() callable from rcu callbacks. Oh, well... > > I still suspect that locking there is too heavy, but it looks like I don't have > a better replacement. > > What are the realistic numbers of memcg on a big system? Several thousand. I guess we could turn list_lrus_mutex into a spin lock by making resize/drain procedures handle list_lru destruction as you suggested above, but list_lru_destroy() would still have to iterate over all elements of list_lru_node->memcg_lrus array to free per-memcg objects, which is too heavy to be performed under sb_lock IMHO. Thanks, Vladimir
