On Thu 11-01-18 18:23:57, Andrey Ryabinin wrote: > On 01/11/2018 03:46 PM, Michal Hocko wrote: > > On Thu 11-01-18 15:21:33, Andrey Ryabinin wrote: > >> > >> > >> On 01/11/2018 01:42 PM, Michal Hocko wrote: > >>> On Wed 10-01-18 15:43:17, Andrey Ryabinin wrote: > >>> [...] > >>>> @@ -2506,15 +2480,13 @@ static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, > >>>> if (!ret) > >>>> break; > >>>> > >>>> - try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, !memsw); > >>>> - > >>>> - curusage = page_counter_read(counter); > >>>> - /* Usage is reduced ? */ > >>>> - if (curusage >= oldusage) > >>>> - retry_count--; > >>>> - else > >>>> - oldusage = curusage; > >>>> - } while (retry_count); > >>>> + usage = page_counter_read(counter); > >>>> + if (!try_to_free_mem_cgroup_pages(memcg, usage - limit, > >>>> + GFP_KERNEL, !memsw)) { > >>> > >>> If the usage drops below limit in the meantime then you get underflow > >>> and reclaim the whole memcg. I do not think this is a good idea. This > >>> can also lead to over reclaim. Why don't you simply stick with the > >>> original SWAP_CLUSTER_MAX (aka 1 for try_to_free_mem_cgroup_pages)? > >>> > >> > >> Because, if new limit is gigabytes bellow the current usage, retrying to set > >> new limit after reclaiming only 32 pages seems unreasonable. > > > > Who would do insanity like that? > > > > What's insane about that? I haven't seen this being done in practice. Why would you want to reclaim GBs of memory from a cgroup? Anyway, if you believe this is really needed then simply do it in a separate patch. > >> @@ -2487,8 +2487,8 @@ static int mem_cgroup_resize_limit(struct mem_cgroup *memcg, > >> if (!ret) > >> break; > >> > >> - usage = page_counter_read(counter); > >> - if (!try_to_free_mem_cgroup_pages(memcg, usage - limit, > >> + nr_pages = max_t(long, 1, page_counter_read(counter) - limit); > >> + if (!try_to_free_mem_cgroup_pages(memcg, nr_pages, > >> GFP_KERNEL, !memsw)) { > >> ret = -EBUSY; > >> break; > > > > How does this address the over reclaim concern? > > It protects from over reclaim due to underflow. 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. -- Michal Hocko SUSE Labs -- 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