On Tue, May 05, 2020 at 08:03:17AM -0700, Shakeel Butt wrote: > On Tue, May 5, 2020 at 12:13 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote: > > So instead of potentially removing a useful information can we focus on > > the remote charging side of the problem and deal with it in a sensible > > way? That would make memory.high usable for your usecase and I still > > believe that this is what you should be using in the first place. > > We talked about this at LSFMM'19 and I think the decision was to not > fix high reclaim for remote memcg until it will be an actual issue. I > suppose now we can treat it as an actual issue. > > There are a couple of open questions: > 1) Should the remote chargers be throttled and do the high reclaim? I would say it depends on the caller. We may need both: a centralized charger like a kthread that does work on behalf of many cgroups should not be throttled, or we risk priority inversions - maybe not even held up for reclaim, although that would need some thinking about memory.max as well - whether remote charges should be allowed to overrun memory.max temporarily, and we punt all reclaim from such contexts to a worker. (Preferrably, in the future, with all the cpu cycles of that worker accounted to the cgroup. We want that for kswapd-style background reclaim as well). We do similar things with IO for example, when offloading IO to a central kthread. The kthread is allowed to proceed past a cgroup's allowance in order to make forward progress, and the cgroup will be forced to pay up retroactively when it comes back for more of the resource.
