On 01/13/2018 01:57 AM, Shakeel Butt wrote: > On Fri, Jan 12, 2018 at 4:24 AM, Michal Hocko <mhocko@xxxxxxxxxx> wrote: >> On Fri 12-01-18 00:59:38, Andrey Ryabinin wrote: >>> On 01/11/2018 07:29 PM, Michal Hocko wrote: >> [...] >>>> I do not think so. Consider that this reclaim races with other >>>> reclaimers. Now you are reclaiming a large chunk so you might end up >>>> reclaiming more than necessary. SWAP_CLUSTER_MAX would reduce the over >>>> reclaim to be negligible. >>>> >>> >>> I did consider this. And I think, I already explained that sort of race in previous email. >>> Whether "Task B" is really a task in cgroup or it's actually a bunch of reclaimers, >>> doesn't matter. That doesn't change anything. >> >> I would _really_ prefer two patches here. The first one removing the >> hard coded reclaim count. That thing is just dubious at best. If you >> _really_ think that the higher reclaim target is meaningfull then make >> it a separate patch. I am not conviced but I will not nack it it either. >> But it will make our life much easier if my over reclaim concern is >> right and we will need to revert it. Conceptually those two changes are >> independent anywa. >> > > Personally I feel that the cgroup-v2 semantics are much cleaner for > setting limit. There is no race with the allocators in the memcg, > though oom-killer can be triggered. For cgroup-v1, the user does not > expect OOM killer and EBUSY is expected on unsuccessful reclaim. How > about we do something similar here and make sure oom killer can not be > triggered for the given memcg? > > // pseudo code > disable_oom(memcg) > old = xchg(&memcg->memory.limit, requested_limit) > > reclaim memory until usage gets below new limit or retries are exhausted > > if (unsuccessful) { > reset_limit(memcg, old) > ret = EBUSY > } else > ret = 0; > enable_oom(memcg) > > This way there is no race with the allocators and oom killer will not > be triggered. The processes in the memcg can suffer but that should be > within the expectation of the user. One disclaimer though, disabling > oom for memcg needs more thought. That's might be worse. If limit is too low, all allocations (except __GFP_NOFAIL of course) will start failing. And the kernel not always careful enough in -ENOMEM handling. Also, it's not much different from oom killing everything, the end result is almost the same - nothing will work in that cgroup. > Shakeel > -- To unsubscribe from this list: send the line "unsubscribe cgroups" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html