Re: [RFC PATCH] mm: memcg: Do not count memory.low reclaim if it does not happen

[Roman Gushchin]

On Tue, Mar 22, 2022 at 07:22:48PM +0100, Michal Koutny wrote:
> This was observed with memcontrol selftest/new LTP test but can be also
> reproduced in simplified setup of two siblings:
> 
> 	`parent .low=50M
> 	  ` s1	.low=50M  .current=50M+ε
> 	  ` s2  .low=0M   .current=50M
> 
> The expectation is that s2/memory.events:low will be zero under outer
> reclaimer since no protection should be given to cgroup s2 (even with
> memory_recursiveprot).
> 
> However, this does not happen. The apparent reason is that when s1 is
> considered for (proportional) reclaim the scanned proportion is rounded
> up to SWAP_CLUSTER_MAX and slightly over-proportional amount is
> reclaimed. Consequently, when the effective low value of s2 is
> calculated, it observes unclaimed parent's protection from s1
> (ε-SWAP_CLUSTER_MAX in theory) and effectively appropriates it.
> The effect is slightly regularized protection (workload dependent)
> between siblings and misreported MEMCG_LOW event when reclaiming s2 with
> this protection.
> 
> Fix the behavior by not reporting breached memory.low in such
> situations. (This affects also setups where all siblings have
> memory.low=0, parent's memory.events:low will still be non-zero when
> parent's memory.low is breached but it will be reduced by the events
> originated in children.)
> 
> Fixes: 8a931f801340 ("mm: memcontrol: recursive memory.low protection")
> Reported-by: Richard Palethorpe <rpalethorpe@xxxxxxxx>
> Link: https://lore.kernel.org/all/20220321101429.3703-1-rpalethorpe@xxxxxxxx/
> Signed-off-by: Michal Koutný <mkoutny@xxxxxxxx>

Hi Michal!

Does it mean that in the following configuration:
	`parent .low=50M
	  ` s1	.low=0M   .current=50M
	  ` s2  .low=0M   .current=50M
there will be no memory.events::low at all? (assuming the recursive thing is on)

Thanks!