Johannes Weiner wrote:

On Sat, Sep 16, 2017 at 12:23:53AM +0900, Tetsuo Handa wrote:

quoted

Johannes Weiner wrote:

quoted

How can we figure out if there is a bug here? Can we time the calls to
__alloc_pages_direct_reclaim() and __alloc_pages_direct_compact() and
drill down from there? Print out the number of times we have retried?
We're counting no_progress_loops, but we are also very much interested
in progress_loops that didn't result in a successful allocation. Too
many of those and I think we want to OOM kill as per above.

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index bec5e96f3b88..01736596389a 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c

@@ -3830,6 +3830,7 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
 			"page allocation stalls for %ums, order:%u",
 			jiffies_to_msecs(jiffies-alloc_start), order);
 		stall_timeout += 10 * HZ;
+		goto oom;
 	}
 
 	/* Avoid recursion of direct reclaim */

@@ -3882,6 +3883,7 @@ __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
 	if (read_mems_allowed_retry(cpuset_mems_cookie))
 		goto retry_cpuset;
 
+oom:
 	/* Reclaim has failed us, start killing things */
 	page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
 	if (page)

According to my stress tests, it is mutex_trylock() in __alloc_pages_may_oom()
that causes warn_alloc() to be called for so many times. The comment

	/*
	 * Acquire the oom lock.  If that fails, somebody else is
	 * making progress for us.
	 */

is true only if the owner of oom_lock can call out_of_memory() and is __GFP_FS
allocation. Consider a situation where there are 1 GFP_KERNEL allocating thread
and 99 GFP_NOFS/GFP_NOIO allocating threads contending the oom_lock. How likely
the OOM killer is invoked? It is very unlikely because GFP_KERNEL allocating thread
likely fails to grab oom_lock because GFP_NOFS/GFP_NOIO allocating threads is
grabing oom_lock. And GFP_KERNEL allocating thread yields CPU time for
GFP_NOFS/GFP_NOIO allocating threads to waste pointlessly.
s/!mutex_trylock(&oom_lock)/mutex_lock_killable()/ significantly improves
this situation for my stress tests. How is your case?

Interesting analysis, that definitely sounds plausible.

It just started happening to us in production and I haven't isolated
it yet. If you already have a reproducer, that's excellent.

Well, my reproducer is an artificial stressor. I think you want to test
using natural programs used in your production environment.

The synchronization has worked this way for a long time (trylock
failure assuming progress, but the order/NOFS/zone bailouts from
actually OOM-killing inside the locked section). We should really fix
*that* rather than serializing warn_alloc().

For GFP_NOFS, it seems to go back to 9879de7373fc ("mm: page_alloc:
embed OOM killing naturally into allocation slowpath"). Before that we
didn't use to call __alloc_pages_may_oom() for NOFS allocations. So I
still wonder why this only now appears to be causing problems.

In any case, converting that trylock to a sleeping lock in this case
makes sense to me. Nobody is blocking under this lock (except that one
schedule_timeout_killable(1) after dispatching a victim) and it's not
obvious to me why we'd need that level of concurrency under OOM.

You can try http://lkml.kernel.org/r/1500202791-5427-1-git-send-email-penguin-kernel@I-love.SAKURA.ne.jp
and http://lkml.kernel.org/r/1503577106-9196-2-git-send-email-penguin-kernel@I-love.SAKURA.ne.jp together.
Then, we can remove mutex_lock(&oom_lock) serialization from __oom_reap_task_mm()
which still exists because Andrea's patch was accepted instead of Michal's patch.

By the way, your environment is not using virtio, is it?
At least virtballoon_oom_notify() is blocking (i.e. might wait for memory
allocation) under oom_lock.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>