On Wed, Jul 05, 2017 at 01:27:04PM +0900, Minchan Kim wrote: > On Tue, Jul 04, 2017 at 08:33:38AM -0400, josef@xxxxxxxxxxxxxx wrote: > > From: Josef Bacik <jbacik@xxxxxx> > > > > Following patches will greatly increase our aggressiveness in slab > > reclaim, so we need checks in place to make sure we stop trying to > > reclaim slab once we've hit our reclaim target. > > > > Signed-off-by: Josef Bacik <jbacik@xxxxxx> > > --- > > v1->v2: > > - Don't bail out in shrink_slab() so that we always scan at least batch_size > > objects of every slab regardless of wether we've hit our target or not. > > It's no different with v1 for aging fairness POV. > > Imagine you have 3 shrinkers in shrinker_list and A has a lots of objects. > > HEAD-> A -> B -> C > > shrink_slab does scan/reclaims from A srhinker a lot until it meets > sc->nr_to_reclaim. Then, VM does aging B and C with batch_size which is > rather small. It breaks fairness. > > In next memory pressure, it shrinks A a lot again but B and C > a little bit. > Oh duh yeah I see what you are saying. I had a scheme previously to break up the scanning targets based on overall usage but it meant looping through the shrinkers twice, as we have to get a total count of objects first to determine individual ratios. I suppose since there's relatively low cost to getting object counts per shrinker and there don't tend to be a lot of shrinkers we could go with this to make it more fair. I'll write this up. Thanks, Josef -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
