On Thu, Aug 16, 2012 at 10:23 PM, Glauber Costa <glommer@xxxxxxxxxxxxx> wrote: > On 08/17/2012 12:54 AM, Ying Han wrote: >> This patch makes target reclaim shrinks slabs in addition to userpages. >> >> Slab shrinkers determine the amount of pressure to put on slabs based on how >> many pages are on lru (inversely proportional relationship). Calculate the >> lru_pages correctly based on memcg lru lists instead of global lru lists. >> >> Signed-off-by: Ying Han <yinghan@xxxxxxxxxx> > > This seems fine from where I stand. > > So imagining for an instant we apply this patch, and this patch only. > The behavior we get is that when memcg gets pressure, it will shrink > globally, but it will at least shrink anything. > > It is needless to say this is not enough. But I wonder if this isn't > better than no shrinking at all ? Maybe this could be put ontop of the > slab series and be the temporary default while we sort out the whole > shrinkers problem? It is a balance between breaking isolation or risking the memcg to OOM. Today there is no shrink_slab under target reclaim, and I think that is bad after your slab accounting patch. But I do worry about the isolation bit where you might end up throwing memcg B's pages for memcg A's pressure. Not only for slab pages but also user pages, like the example I listed on the first patch. --Ying > > -- 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>