On Sun, Aug 18, 2019 at 08:30:15AM +0800, Yafang Shao wrote: > On Sun, Aug 18, 2019 at 3:14 AM Roman Gushchin <guro@xxxxxx> wrote: > > > > On Sat, Aug 17, 2019 at 11:33:57AM +0800, Yafang Shao wrote: > > > On Sat, Aug 17, 2019 at 8:47 AM Roman Gushchin <guro@xxxxxx> wrote: > > > > > > > > Commit 766a4c19d880 ("mm/memcontrol.c: keep local VM counters in sync > > > > with the hierarchical ones") effectively decreased the precision of > > > > per-memcg vmstats_local and per-memcg-per-node lruvec percpu counters. > > > > > > > > That's good for displaying in memory.stat, but brings a serious regression > > > > into the reclaim process. > > > > > > > > One issue I've discovered and debugged is the following: > > > > lruvec_lru_size() can return 0 instead of the actual number of pages > > > > in the lru list, preventing the kernel to reclaim last remaining > > > > pages. Result is yet another dying memory cgroups flooding. > > > > The opposite is also happening: scanning an empty lru list > > > > is the waste of cpu time. > > > > > > > > Also, inactive_list_is_low() can return incorrect values, preventing > > > > the active lru from being scanned and freed. It can fail both because > > > > the size of active and inactive lists are inaccurate, and because > > > > the number of workingset refaults isn't precise. In other words, > > > > the result is pretty random. > > > > > > > > I'm not sure, if using the approximate number of slab pages in > > > > count_shadow_number() is acceptable, but issues described above > > > > are enough to partially revert the patch. > > > > > > > > Let's keep per-memcg vmstat_local batched (they are only used for > > > > displaying stats to the userspace), but keep lruvec stats precise. > > > > This change fixes the dead memcg flooding on my setup. > > > > > > > > > > That will make some misunderstanding if the local counters are not in > > > sync with the hierarchical ones > > > (someone may doubt whether there're something leaked.). > > > > Sure, but the actual leakage is a much more serious issue. > > > > > If we have to do it like this, I think we should better document this behavior. > > > > Lru size calculations can be done using per-zone counters, which is > > actually cheaper, because the number of zones is usually smaller than > > the number of cpus. I'll send a corresponding patch on Monday. > > > > Looks like a good idea. > > > Maybe other use cases can also be converted? > > We'd better keep the behavior the same across counters. I think you > can have a try. As I said, consistency of counters is important, but not nearly as important as the real behavior of the system. Especially because we talk about per-node memcg statistics, which I believe is mostly used for debugging. So for now I think the right thing to do is to revert the change to fix the memory reclaim process. And then we can discuss how to get counters right. Thanks!