On Mon, Nov 30, 2015 at 02:36:28PM +0300, Vladimir Davydov wrote:

Suppose we have the following cgroup configuration.

A __ B
  \_ C

A is empty (which is natural for the unified hierarchy AFAIU). B has
some workload running in it, and C generates socket pressure. Due to the
socket pressure coming from C we start reclaim in A, which results in
thrashing of B, but we might not put sockets under pressure in A or C,
because vmpressure does not account pages scanned/reclaimed in B when
generating a vmpressure event for A or C. This might result in
aggressive reclaim and thrashing in B w/o generating a signal for C to
stop growing socket buffers.

Do you think such a situation is possible? If so, would it make sense to
switch to post-order walk in shrink_zone and pass sub-tree
scanned/reclaimed stats to vmpressure for each scanned memcg?

In that case the LRU pages in C would experience pressure as well,
which would then reign in the sockets in C. There must be some LRU
pages in there, otherwise who is creating socket pressure?

The same applies to shrinkers. All secondary reclaim is driven by LRU
reclaim results.

I can see that there is some unfairness in distributing memcg reclaim
pressure purely based on LRU size, because there are scenarios where
the auxiliary objects (incl. sockets, but mostly shrinker pools)
amount to a significant portion of the group's memory footprint. But
substitute group for NUMA node and we've had this behavior for
years. I'm not sure it's actually a problem in practice.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>