[Intel-gfx] [PATCH] mm: Reward slab shrinkers that reclaim more than they were asked

[Andrew Morton]

On Sat, 12 Aug 2017 12:34:37 +0100 Chris Wilson <chris at chris-wilson.co.uk> wrote:

> Some shrinkers may only be able to free a bunch of objects at a time, and
> so free more than the requested nr_to_scan in one pass. Account for the
> extra freed objects against the total number of objects we intend to
> free, otherwise we may end up penalising the slab far more than intended.
> 
> ...
>
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -398,6 +398,7 @@ static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
>  			break;
>  		freed += ret;
>  
> +		nr_to_scan = max(nr_to_scan, ret);
>  		count_vm_events(SLABS_SCANNED, nr_to_scan);
>  		total_scan -= nr_to_scan;
>  		scanned += nr_to_scan;

Well...  kinda.  But what happens if the shrinker scanned more objects
than requested but failed to free many of them?  Of if the shrinker
scanned less than requested?

We really want to return nr_scanned from the shrinker invocation. 
Could we add a field to shrink_control for this?

--- a/mm/vmscan.c~a
+++ a/mm/vmscan.c
@@ -393,14 +393,15 @@ static unsigned long do_shrink_slab(stru
 		unsigned long nr_to_scan = min(batch_size, total_scan);
 
 		shrinkctl->nr_to_scan = nr_to_scan;
+		shrinkctl->nr_scanned = nr_to_scan;
 		ret = shrinker->scan_objects(shrinker, shrinkctl);
 		if (ret == SHRINK_STOP)
 			break;
 		freed += ret;
 
-		count_vm_events(SLABS_SCANNED, nr_to_scan);
-		total_scan -= nr_to_scan;
-		scanned += nr_to_scan;
+		count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned);
+		total_scan -= shrinkctl->nr_scanned;
+		scanned += shrinkctl->nr_scanned;
 
 		cond_resched();
 	}
_