On Thu, Apr 05, 2018 at 07:59:20PM +0100, Roman Gushchin wrote:
> If memcg's usage is equal to the memory.low value, avoid reclaiming
> from this cgroup while there is a surplus of reclaimable memory.
>
> This sounds more logical and also matches memory.high and memory.max
> behavior: both are inclusive.
I was trying to figure out why we did it this way in the first place
and found this patch:
commit 4e54dede38b45052a941bcf709f7d29f2e18174d
Author: Michal Hocko <mhocko@xxxxxxx>
Date: Fri Feb 27 15:51:46 2015 -0800
memcg: fix low limit calculation
A memcg is considered low limited even when the current usage is equal to
the low limit. This leads to interesting side effects e.g.
groups/hierarchies with no memory accounted are considered protected and
so the reclaim will emit MEMCG_LOW event when encountering them.
Another and much bigger issue was reported by Joonsoo Kim. He has hit a
NULL ptr dereference with the legacy cgroup API which even doesn't have
low limit exposed. The limit is 0 by default but the initial check fails
for memcg with 0 consumption and parent_mem_cgroup() would return NULL if
use_hierarchy is 0 and so page_counter_read would try to dereference NULL.
I suppose that the current implementation is just an overlook because the
documentation in Documentation/cgroups/unified-hierarchy.txt says:
"The memory.low boundary on the other hand is a top-down allocated
reserve. A cgroup enjoys reclaim protection when it and all its
ancestors are below their low boundaries"
Fix the usage and the low limit comparision in mem_cgroup_low accordingly.
> @@ -5709,7 +5709,7 @@ bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
> elow = min(elow, parent_elow * low_usage / siblings_low_usage);
> exit:
> memcg->memory.elow = elow;
> - return usage < elow;
> + return usage <= elow;
So I think this needs to be usage && usage <= elow to not emit
MEMCG_LOW events in case usage == elow == 0.
--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
