summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorMichal Hocko <mhocko@suse.cz>2012-08-01 01:45:55 +0200
committerLinus Torvalds <torvalds@linux-foundation.org>2012-08-01 03:42:49 +0200
commite62e384e9da8d9a0c599795464a7e76fd490931c (patch)
tree07d92bd50f42e5f919ddfb03b185a270daa2e1b1
parentmm: mmu_notifier: fix freed page still mapped in secondary MMU (diff)
downloadlinux-e62e384e9da8d9a0c599795464a7e76fd490931c.tar.xz
linux-e62e384e9da8d9a0c599795464a7e76fd490931c.zip
memcg: prevent OOM with too many dirty pages
The current implementation of dirty pages throttling is not memcg aware which makes it easy to have memcg LRUs full of dirty pages. Without throttling, these LRUs can be scanned faster than the rate of writeback, leading to memcg OOM conditions when the hard limit is small. This patch fixes the problem by throttling the allocating process (possibly a writer) during the hard limit reclaim by waiting on PageReclaim pages. We are waiting only for PageReclaim pages because those are the pages that made one full round over LRU and that means that the writeback is much slower than scanning. The solution is far from being ideal - long term solution is memcg aware dirty throttling - but it is meant to be a band aid until we have a real fix. We are seeing this happening during nightly backups which are placed into containers to prevent from eviction of the real working set. The change affects only memcg reclaim and only when we encounter PageReclaim pages which is a signal that the reclaim doesn't catch up on with the writers so somebody should be throttled. This could be potentially unfair because it could be somebody else from the group who gets throttled on behalf of the writer but as writers need to allocate as well and they allocate in higher rate the probability that only innocent processes would be penalized is not that high. I have tested this change by a simple dd copying /dev/zero to tmpfs or ext3 running under small memcg (1G copy under 5M, 60M, 300M and 2G containers) and dd got killed by OOM killer every time. With the patch I could run the dd with the same size under 5M controller without any OOM. The issue is more visible with slower devices for output. * With the patch ================ * tmpfs size=2G --------------- $ vim cgroup_cache_oom_test.sh $ ./cgroup_cache_oom_test.sh 5M using Limit 5M for group 1000+0 records in 1000+0 records out 1048576000 bytes (1.0 GB) copied, 30.4049 s, 34.5 MB/s $ ./cgroup_cache_oom_test.sh 60M using Limit 60M for group 1000+0 records in 1000+0 records out 1048576000 bytes (1.0 GB) copied, 31.4561 s, 33.3 MB/s $ ./cgroup_cache_oom_test.sh 300M using Limit 300M for group 1000+0 records in 1000+0 records out 1048576000 bytes (1.0 GB) copied, 20.4618 s, 51.2 MB/s $ ./cgroup_cache_oom_test.sh 2G using Limit 2G for group 1000+0 records in 1000+0 records out 1048576000 bytes (1.0 GB) copied, 1.42172 s, 738 MB/s * ext3 ------ $ ./cgroup_cache_oom_test.sh 5M using Limit 5M for group 1000+0 records in 1000+0 records out 1048576000 bytes (1.0 GB) copied, 27.9547 s, 37.5 MB/s $ ./cgroup_cache_oom_test.sh 60M using Limit 60M for group 1000+0 records in 1000+0 records out 1048576000 bytes (1.0 GB) copied, 30.3221 s, 34.6 MB/s $ ./cgroup_cache_oom_test.sh 300M using Limit 300M for group 1000+0 records in 1000+0 records out 1048576000 bytes (1.0 GB) copied, 24.5764 s, 42.7 MB/s $ ./cgroup_cache_oom_test.sh 2G using Limit 2G for group 1000+0 records in 1000+0 records out 1048576000 bytes (1.0 GB) copied, 3.35828 s, 312 MB/s * Without the patch =================== * tmpfs size=2G --------------- $ ./cgroup_cache_oom_test.sh 5M using Limit 5M for group ./cgroup_cache_oom_test.sh: line 46: 4668 Killed dd if=/dev/zero of=$OUT/zero bs=1M count=$count $ ./cgroup_cache_oom_test.sh 60M using Limit 60M for group 1000+0 records in 1000+0 records out 1048576000 bytes (1.0 GB) copied, 25.4989 s, 41.1 MB/s $ ./cgroup_cache_oom_test.sh 300M using Limit 300M for group 1000+0 records in 1000+0 records out 1048576000 bytes (1.0 GB) copied, 24.3928 s, 43.0 MB/s $ ./cgroup_cache_oom_test.sh 2G using Limit 2G for group 1000+0 records in 1000+0 records out 1048576000 bytes (1.0 GB) copied, 1.49797 s, 700 MB/s * ext3 ------ $ ./cgroup_cache_oom_test.sh 5M using Limit 5M for group ./cgroup_cache_oom_test.sh: line 46: 4689 Killed dd if=/dev/zero of=$OUT/zero bs=1M count=$count $ ./cgroup_cache_oom_test.sh 60M using Limit 60M for group ./cgroup_cache_oom_test.sh: line 46: 4692 Killed dd if=/dev/zero of=$OUT/zero bs=1M count=$count $ ./cgroup_cache_oom_test.sh 300M using Limit 300M for group 1000+0 records in 1000+0 records out 1048576000 bytes (1.0 GB) copied, 20.248 s, 51.8 MB/s $ ./cgroup_cache_oom_test.sh 2G using Limit 2G for group 1000+0 records in 1000+0 records out 1048576000 bytes (1.0 GB) copied, 2.85201 s, 368 MB/s [akpm@linux-foundation.org: tweak changelog, reordered the test to optimize for CONFIG_CGROUP_MEM_RES_CTLR=n] [hughd@google.com: fix deadlock with loop driver] Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujtisu.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Rik van Riel <riel@redhat.com> Cc: Ying Han <yinghan@google.com> Cc: Greg Thelen <gthelen@google.com> Cc: Hugh Dickins <hughd@google.com> Reviewed-by: Mel Gorman <mgorman@suse.de> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Reviewed-by: Fengguang Wu <fengguang.wu@intel.com> Signed-off-by: Michal Hocko <mhocko@suse.cz> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
-rw-r--r--mm/vmscan.c23
1 files changed, 20 insertions, 3 deletions
diff --git a/mm/vmscan.c b/mm/vmscan.c
index 88804017e7d6..ca43aa00ea0e 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -720,9 +720,26 @@ static unsigned long shrink_page_list(struct list_head *page_list,
(PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
if (PageWriteback(page)) {
- nr_writeback++;
- unlock_page(page);
- goto keep;
+ /*
+ * memcg doesn't have any dirty pages throttling so we
+ * could easily OOM just because too many pages are in
+ * writeback from reclaim and there is nothing else to
+ * reclaim.
+ *
+ * Check may_enter_fs, certainly because a loop driver
+ * thread might enter reclaim, and deadlock if it waits
+ * on a page for which it is needed to do the write
+ * (loop masks off __GFP_IO|__GFP_FS for this reason);
+ * but more thought would probably show more reasons.
+ */
+ if (!global_reclaim(sc) && PageReclaim(page) &&
+ may_enter_fs)
+ wait_on_page_writeback(page);
+ else {
+ nr_writeback++;
+ unlock_page(page);
+ goto keep;
+ }
}
references = page_check_references(page, sc);