Re: [PATCH 3/8] mm/vmscan: Throttle reclaim when no progress is being made

[Mel Gorman]

On Tue, Nov 23, 2021 at 05:49:14PM -0800, Darrick J. Wong wrote:
> > Ever since Christoph broke swapfiles, I've been carrying around a little
> > fstest in my dev tree[1] that tries to exercise paging things in and out
> > of a swapfile.  Sadly I've been trapped in about three dozen customer
> > escalations for over a month, which means I haven't been able to do much
> > upstream in weeks.  Like submit this test upstream. :(
> > 
> > Now that I've finally gotten around to trying out a 5.16-rc2 build, I
> > notice that the runtime of this test has gone from ~5s to 2 hours.
> > Among other things that it does, the test sets up a cgroup with a memory
> > controller limiting the memory usage to 25MB, then runs a program that
> > tries to dirty 50MB of memory.  There's 2GB of memory in the VM, so
> > we're not running reclaim globally, but the cgroup gets throttled very
> > severely.
> > 
> > AFAICT the system is mostly idle, but it's difficult to tell because ps
> > and top also get stuck waiting for this cgroup for whatever reason.  My
> > uninformed spculation is that usemem_and_swapoff takes a page fault
> > while dirtying the 50MB memory buffer, prepares to pull a page in from
> > swap, tries to evict another page to stay under the memcg limit, but
> > that decides that it's making no progress and calls
> > reclaim_throttle(..., VMSCAN_THROTTLE_NOPROGRESS).
> > 
> > The sleep is uninterruptible, so I can't even kill -9 fstests to shut it
> > down.  Eventually we either finish the test or (for the mlock part) the
> > OOM killer actually kills the process, but this takes a very long time.
> > 
> > Any thoughts?  For now I can just hack around this by skipping
> > reclaim_throttle if cgroup_reclaim() == true, but that's probably not
> > the correct fix. :)
> 
> Update: after adding timing information to usemem_and_swapoff, it looks
> like dirtying the 50MB buffer takes ~22s (up from 0.06s on 5.15).  The
> mlock call stalls for ~280s until the OOM killer kills it (up from
> nearly instantaneous on 5.15), and the swapon/swapoff variant takes
> 20 minutes to hours depending on the run.
> 

Can you try the patch below please? I think I'm running the test
correctly and it finishes for me in 16 seconds with this applied

diff --git a/mm/vmscan.c b/mm/vmscan.c
index 07db03883062..d9166e94eb95 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -1057,7 +1057,17 @@ void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason)
 
 		break;
 	case VMSCAN_THROTTLE_NOPROGRESS:
-		timeout = HZ/2;
+		timeout = 1;
+
+		/*
+		 * If kswapd is disabled, reschedule if necessary but do not
+		 * throttle as the system is likely near OOM.
+		 */
+		if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) {
+			cond_resched();
+			return;
+		}
+
 		break;
 	case VMSCAN_THROTTLE_ISOLATED:
 		timeout = HZ/50;
@@ -3395,7 +3405,7 @@ static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc)
 		return;
 
 	/* Throttle if making no progress at high prioities. */
-	if (sc->priority < DEF_PRIORITY - 2)
+	if (sc->priority < DEF_PRIORITY - 2 && !sc->nr_reclaimed)
 		reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS);
 }
 
@@ -3415,6 +3425,7 @@ static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
 	unsigned long nr_soft_scanned;
 	gfp_t orig_mask;
 	pg_data_t *last_pgdat = NULL;
+	pg_data_t *first_pgdat = NULL;
 
 	/*
 	 * If the number of buffer_heads in the machine exceeds the maximum
@@ -3478,14 +3489,18 @@ static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
 			/* need some check for avoid more shrink_zone() */
 		}
 
+		if (!first_pgdat)
+			first_pgdat = zone->zone_pgdat;
+
 		/* See comment about same check for global reclaim above */
 		if (zone->zone_pgdat == last_pgdat)
 			continue;
 		last_pgdat = zone->zone_pgdat;
 		shrink_node(zone->zone_pgdat, sc);
-		consider_reclaim_throttle(zone->zone_pgdat, sc);
 	}
 
+	consider_reclaim_throttle(first_pgdat, sc);
+
 	/*
 	 * Restore to original mask to avoid the impact on the caller if we
 	 * promoted it to __GFP_HIGHMEM.