Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S933251AbdGTG4a (ORCPT ); Thu, 20 Jul 2017 02:56:30 -0400 Received: from mx2.suse.de ([195.135.220.15]:34671 "EHLO mx1.suse.de" rhost-flags-OK-OK-OK-FAIL) by vger.kernel.org with ESMTP id S1751810AbdGTG43 (ORCPT ); Thu, 20 Jul 2017 02:56:29 -0400 Date: Thu, 20 Jul 2017 08:56:26 +0200 From: Michal Hocko To: Andrew Morton Cc: Mel Gorman , Tetsuo Handa , Rik van Riel , Johannes Weiner , Vlastimil Babka , linux-mm@kvack.org, LKML Subject: Re: [PATCH] mm, vmscan: do not loop on too_many_isolated for ever Message-ID: <20170720065625.GB9058@dhcp22.suse.cz> References: <20170710074842.23175-1-mhocko@kernel.org> <20170719152014.53a861c57bcb636d6cd9d002@linux-foundation.org> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20170719152014.53a861c57bcb636d6cd9d002@linux-foundation.org> User-Agent: Mutt/1.5.23 (2014-03-12) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 3329 Lines: 77 On Wed 19-07-17 15:20:14, Andrew Morton wrote: > On Mon, 10 Jul 2017 09:48:42 +0200 Michal Hocko wrote: > > > From: Michal Hocko > > > > Tetsuo Handa has reported [1][2][3]that direct reclaimers might get stuck > > in too_many_isolated loop basically for ever because the last few pages > > on the LRU lists are isolated by the kswapd which is stuck on fs locks > > when doing the pageout or slab reclaim. This in turn means that there is > > nobody to actually trigger the oom killer and the system is basically > > unusable. > > > > too_many_isolated has been introduced by 35cd78156c49 ("vmscan: throttle > > direct reclaim when too many pages are isolated already") to prevent > > from pre-mature oom killer invocations because back then no reclaim > > progress could indeed trigger the OOM killer too early. But since the > > oom detection rework 0a0337e0d1d1 ("mm, oom: rework oom detection") > > the allocation/reclaim retry loop considers all the reclaimable pages > > and throttles the allocation at that layer so we can loosen the direct > > reclaim throttling. > > > > Make shrink_inactive_list loop over too_many_isolated bounded and returns > > immediately when the situation hasn't resolved after the first sleep. > > Replace congestion_wait by a simple schedule_timeout_interruptible because > > we are not really waiting on the IO congestion in this path. > > > > Please note that this patch can theoretically cause the OOM killer to > > trigger earlier while there are many pages isolated for the reclaim > > which makes progress only very slowly. This would be obvious from the oom > > report as the number of isolated pages are printed there. If we ever hit > > this should_reclaim_retry should consider those numbers in the evaluation > > in one way or another. > > Need to figure out which kernels to patch. Maybe just 4.13-rc after a > week or two? I do not think we need to rush it and the next merge window should be just OK. > > --- a/mm/vmscan.c > > +++ b/mm/vmscan.c > > @@ -1713,9 +1713,15 @@ shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, > > int file = is_file_lru(lru); > > struct pglist_data *pgdat = lruvec_pgdat(lruvec); > > struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; > > + bool stalled = false; > > > > while (unlikely(too_many_isolated(pgdat, file, sc))) { > > - congestion_wait(BLK_RW_ASYNC, HZ/10); > > + if (stalled) > > + return 0; > > + > > + /* wait a bit for the reclaimer. */ > > + schedule_timeout_interruptible(HZ/10); > > a) if this task has signal_pending(), this falls straight through > and I suspect the code breaks? It will not break. It will return to the allocation path more quickly but no over-reclaim will happen and it will/should get throttled there. So nothing critical. > b) replacing congestion_wait() with schedule_timeout_interruptible() > means this task no longer contributes to load average here and it's > a (slightly) user-visible change. you are right. I am not sure it matters but it might be visible. > c) msleep_interruptible() is nicer > > d) IOW, methinks we should be using msleep() here? OK, I do not have objections. Are you going to squash this in or want a separate patch explaining all the above? -- Michal Hocko SUSE Labs