On Fri, Nov 02, 2018 at 05:51:47PM +0100, Michal Hocko wrote: > On Fri 02-11-18 16:22:41, Roman Gushchin wrote: > > On Fri, Nov 02, 2018 at 05:13:14PM +0100, Michal Hocko wrote: > > > On Fri 02-11-18 15:48:57, Roman Gushchin wrote: > > > > On Fri, Nov 02, 2018 at 09:03:55AM +0100, Michal Hocko wrote: > > > > > On Fri 02-11-18 02:45:42, Dexuan Cui wrote: > > > > > [...] > > > > > > I totally agree. I'm now just wondering if there is any temporary workaround, > > > > > > even if that means we have to run the kernel with some features disabled or > > > > > > with a suboptimal performance? > > > > > > > > > > One way would be to disable kmem accounting (cgroup.memory=nokmem kernel > > > > > option). That would reduce the memory isolation because quite a lot of > > > > > memory will not be accounted for but the primary source of in-flight and > > > > > hard to reclaim memory will be gone. > > > > > > > > In my experience disabling the kmem accounting doesn't really solve the issue > > > > (without patches), but can lower the rate of the leak. > > > > > > This is unexpected. 90cbc2508827e was introduced to address offline > > > memcgs to be reclaim even when they are small. But maybe you mean that > > > we still leak in an absence of the memory pressure. Or what does prevent > > > memcg from going down? > > > > There are 3 independent issues which are contributing to this leak: > > 1) Kernel stack accounting weirdness: processes can reuse stack accounted to > > different cgroups. So basically any running process can take a reference to any > > cgroup. > > yes, but kmem accounting should rule that out, right? If not then this > is a clear bug and easy to backport because that would mean to add a > missing memcg_kmem_enabled check. Yes, you're right, disabling kmem accounting should mitigate this problem. > > > 2) We do forget to scan the last page in the LRU list. So if we ended up with > > 1-page long LRU, it can stay there basically forever. > > Why > /* > * If the cgroup's already been deleted, make sure to > * scrape out the remaining cache. > */ > if (!scan && !mem_cgroup_online(memcg)) > scan = min(size, SWAP_CLUSTER_MAX); > > in get_scan_count doesn't work for that case? No, it doesn't. Let's look at the whole picture: size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); scan = size >> sc->priority; /* * If the cgroup's already been deleted, make sure to * scrape out the remaining cache. */ if (!scan && !mem_cgroup_online(memcg)) scan = min(size, SWAP_CLUSTER_MAX); If size == 1, scan == 0 => scan = min(1, 32) == 1. And after proportional adjustment we'll have 0. So, disabling kmem accounting mitigates 2 other issues, but not this one. Anyway, I'd prefer to wait a bit for test results, and backport the whole series as a whole. Thanks!
