Messages in this thread | | | Date | Thu, 19 Oct 2017 21:35:42 +0200 | From | Michal Hocko <> | Subject | Re: [PATCH] mm: mlock: remove lru_add_drain_all() |
| |
On Thu 19-10-17 12:19:26, Shakeel Butt wrote: > On Thu, Oct 19, 2017 at 5:32 AM, Michal Hocko <mhocko@kernel.org> wrote: > > On Wed 18-10-17 16:17:30, Shakeel Butt wrote: > >> Recently we have observed high latency in mlock() in our generic > >> library and noticed that users have started using tmpfs files even > >> without swap and the latency was due to expensive remote LRU cache > >> draining. > > > > some numbers would be really nice > > > > On a production workload, customers complained that single mlock() > call took around 10 seconds on mapped tmpfs files and the perf profile > showed lru_add_drain_all as culprit.
draining can take some time. I wouldn't expect orders of seconds so perf data would be definitely helpful in the changelog.
[...] > > Is this really true? lru_add_drain_all will flush the previously cached > > LRU pages. We are not flushing after the pages have been faulted in so > > this might not do anything wrt. mlocked pages, right? > > > > Sorry for the confusion. I wanted to say that if the pages which are > being mlocked are on caches of remote cpus then lru_add_drain_all will > move them to their corresponding LRUs and then remaining functionality > of mlock will move them again from their evictable LRUs to unevictable > LRU.
yes, but the point is that we are draining pages which might be not directly related to pages which _will_ be mlocked by the syscall. In fact those will stay on the cache. This is the primary reason why this draining doesn't make much sense.
Or am I still misunderstanding what you are saying here? -- Michal Hocko SUSE Labs
| |