]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commitdiff
mm: throttle on IO only when there are too many dirty and writeback pages
authorMichal Hocko <mhocko@suse.com>
Fri, 20 May 2016 23:57:03 +0000 (16:57 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sat, 21 May 2016 00:58:30 +0000 (17:58 -0700)
wait_iff_congested has been used to throttle allocator before it retried
another round of direct reclaim to allow the writeback to make some
progress and prevent reclaim from looping over dirty/writeback pages
without making any progress.

We used to do congestion_wait before commit 0e093d99763e ("writeback: do
not sleep on the congestion queue if there are no congested BDIs or if
significant congestion is not being encountered in the current zone")
but that led to undesirable stalls and sleeping for the full timeout
even when the BDI wasn't congested.  Hence wait_iff_congested was used
instead.

But it seems that even wait_iff_congested doesn't work as expected.  We
might have a small file LRU list with all pages dirty/writeback and yet
the bdi is not congested so this is just a cond_resched in the end and
can end up triggering pre mature OOM.

This patch replaces the unconditional wait_iff_congested by
congestion_wait which is executed only if we _know_ that the last round
of direct reclaim didn't make any progress and dirty+writeback pages are
more than a half of the reclaimable pages on the zone which might be
usable for our target allocation.  This shouldn't reintroduce stalls
fixed by 0e093d99763e because congestion_wait is called only when we are
getting hopeless when sleeping is a better choice than OOM with many
pages under IO.

We have to preserve logic introduced by commit 373ccbe59270 ("mm,
vmstat: allow WQ concurrency to discover memory reclaim doesn't make any
progress") into the __alloc_pages_slowpath now that wait_iff_congested
is not used anymore.  As the only remaining user of wait_iff_congested
is shrink_inactive_list we can remove the WQ specific short sleep from
wait_iff_congested because the sleep is needed to be done only once in
the allocation retry cycle.

[mhocko@suse.com: high_zoneidx->ac_classzone_idx to evaluate memory reserves properly]
Link: http://lkml.kernel.org/r/1463051677-29418-2-git-send-email-mhocko@kernel.org
Signed-off-by: Michal Hocko <mhocko@suse.com>
Acked-by: Hillf Danton <hillf.zj@alibaba-inc.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <js1304@gmail.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: Vladimir Davydov <vdavydov@virtuozzo.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/backing-dev.c
mm/page_alloc.c

index 0c6317b7db38a086191a043b5ec480e4f5264e27..ed173b8ae8f24a317e11da29735f0a5faa9e361e 100644 (file)
@@ -957,9 +957,8 @@ EXPORT_SYMBOL(congestion_wait);
  * jiffies for either a BDI to exit congestion of the given @sync queue
  * or a write to complete.
  *
- * In the absence of zone congestion, a short sleep or a cond_resched is
- * performed to yield the processor and to allow other subsystems to make
- * a forward progress.
+ * In the absence of zone congestion, cond_resched() is called to yield
+ * the processor if necessary but otherwise does not sleep.
  *
  * The return value is 0 if the sleep is for the full timeout. Otherwise,
  * it is the number of jiffies that were still remaining when the function
@@ -979,20 +978,7 @@ long wait_iff_congested(struct zone *zone, int sync, long timeout)
         */
        if (atomic_read(&nr_wb_congested[sync]) == 0 ||
            !test_bit(ZONE_CONGESTED, &zone->flags)) {
-
-               /*
-                * Memory allocation/reclaim might be called from a WQ
-                * context and the current implementation of the WQ
-                * concurrency control doesn't recognize that a particular
-                * WQ is congested if the worker thread is looping without
-                * ever sleeping. Therefore we have to do a short sleep
-                * here rather than calling cond_resched().
-                */
-               if (current->flags & PF_WQ_WORKER)
-                       schedule_timeout_uninterruptible(1);
-               else
-                       cond_resched();
-
+               cond_resched();
                /* In case we scheduled, work out time remaining */
                ret = timeout - (jiffies - start);
                if (ret < 0)
index fa39efc3a692e7040924af6781678e6c1dda036f..f51c302126a19232bf17c925a9ac09e420725841 100644 (file)
@@ -3436,8 +3436,9 @@ should_reclaim_retry(gfp_t gfp_mask, unsigned order,
        for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
                                        ac->nodemask) {
                unsigned long available;
+               unsigned long reclaimable;
 
-               available = zone_reclaimable_pages(zone);
+               available = reclaimable = zone_reclaimable_pages(zone);
                available -= DIV_ROUND_UP(no_progress_loops * available,
                                          MAX_RECLAIM_RETRIES);
                available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
@@ -3447,9 +3448,41 @@ should_reclaim_retry(gfp_t gfp_mask, unsigned order,
                 * available?
                 */
                if (__zone_watermark_ok(zone, order, min_wmark_pages(zone),
-                               ac->high_zoneidx, alloc_flags, available)) {
-                       /* Wait for some write requests to complete then retry */
-                       wait_iff_congested(zone, BLK_RW_ASYNC, HZ/50);
+                               ac_classzone_idx(ac), alloc_flags, available)) {
+                       /*
+                        * If we didn't make any progress and have a lot of
+                        * dirty + writeback pages then we should wait for
+                        * an IO to complete to slow down the reclaim and
+                        * prevent from pre mature OOM
+                        */
+                       if (!did_some_progress) {
+                               unsigned long writeback;
+                               unsigned long dirty;
+
+                               writeback = zone_page_state_snapshot(zone,
+                                                                    NR_WRITEBACK);
+                               dirty = zone_page_state_snapshot(zone, NR_FILE_DIRTY);
+
+                               if (2*(writeback + dirty) > reclaimable) {
+                                       congestion_wait(BLK_RW_ASYNC, HZ/10);
+                                       return true;
+                               }
+                       }
+
+                       /*
+                        * Memory allocation/reclaim might be called from a WQ
+                        * context and the current implementation of the WQ
+                        * concurrency control doesn't recognize that
+                        * a particular WQ is congested if the worker thread is
+                        * looping without ever sleeping. Therefore we have to
+                        * do a short sleep here rather than calling
+                        * cond_resched().
+                        */
+                       if (current->flags & PF_WQ_WORKER)
+                               schedule_timeout_uninterruptible(1);
+                       else
+                               cond_resched();
+
                        return true;
                }
        }