On Wed, Dec 08, 2021 at 10:20:47PM -0800, Hugh Dickins wrote:

On Thu, 2 Dec 2021, Mel Gorman wrote:
...

quoted

--- a/mm/vmscan.c
+++ b/mm/vmscan.c

...

quoted

@@ -3478,14 +3520,18 @@ 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);

My tmpfs swapping load (tweaked to use huge pages more heavily than
in real life) is far from being a realistic load: but it was notably
slowed down by your throttling mods in 5.16-rc, and this patch makes
it well again - thanks.

But: it very quickly hit NULL pointer until I changed that last line to

	if (first_pgdat)
		consider_reclaim_throttle(first_pgdat, sc);

I've given no thought as to whether that is the correct fix,
or if first_pgdat should be set earlier in the loop above.

It's the right fix, first_pgdat may be NULL if compaction can run for
each zone in the zonelist which could be the case for a tmpfs swapping
load that is huge page intensive.

Thanks Hugh.

-- 
Mel Gorman
SUSE Labs