On Mon, Aug 03, 2015 at 09:23:58AM -0400, Johannes Weiner wrote: > On Mon, Aug 03, 2015 at 03:04:22PM +0300, Vladimir Davydov wrote: > > @@ -179,8 +180,9 @@ static void unpack_shadow(void *shadow, > > eviction = entry; > > > > *zone = NODE_DATA(nid)->node_zones + zid; > > + *lruvec = mem_cgroup_page_lruvec(page, *zone); > > > > - refault = atomic_long_read(&(*zone)->inactive_age); > > + refault = atomic_long_read(&(*lruvec)->inactive_age); > > mask = ~0UL >> (NODES_SHIFT + ZONES_SHIFT + > > RADIX_TREE_EXCEPTIONAL_SHIFT); > > /* > > You can not compare an eviction shadow entry from one lruvec with the > inactive age of another lruvec. The inactive ages are not related and > might differ significantly: memcgs are created ad hoc, memory hotplug, > page allocator fairness drift. In those cases the result will be pure > noise. That's true. If a page is evicted in one cgroup and then refaulted in another, the activation will be random. However, is it a frequent event when a page used by and evicted from one cgroup is refaulted in another? If there is no active file sharing (is it common?), this should only happen to code pages, but those will most likely end up in the cgroup that has the greatest limit, so they shouldn't be evicted and refaulted frequently. So the question is can we tolerate some noise here? > > As much as I would like to see a simpler way, I am pessimistic that > there is a way around storing memcg ids in the shadow entries. On 32 bit there is too little space for storing memcg id. We can shift the distance so that it would fit and still contain something meaningful though, but that would take much more code, so I'm trying to try the simplest way first. Thanks, Vladimir -- 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>