Re: [v11 3/6] mm, oom: cgroup-aware OOM killer

[Roman Gushchin]

On Thu, Oct 12, 2017 at 02:50:38PM -0700, David Rientjes wrote:
> On Wed, 11 Oct 2017, Roman Gushchin wrote:
> 
> Think about it in a different way: we currently compare per-process usage 
> and userspace has /proc/pid/oom_score_adj to adjust that usage depending 
> on priorities of that process and still oom kill if there's a memory leak.  
> Your heuristic compares per-cgroup usage, it's the cgroup-aware oom killer 
> after all.  We don't need a strict memory.oom_priority that outranks all 
> other sibling cgroups regardless of usage.  We need a memory.oom_score_adj 
> to adjust the per-cgroup usage.  The decisionmaking in your earlier 
> example would be under the control of C/memory.oom_score_adj and 
> D/memory.oom_score_adj.  Problem solved.
> 
> It also solves the problem of userspace being able to influence oom victim 
> selection so now they can protect important cgroups just like we can 
> protect important processes today.
> 
> And since this would be hierarchical usage, you can trivially infer root 
> mem cgroup usage by subtraction of top-level mem cgroup usage.
> 
> This is a powerful solution to the problem and gives userspace the control 
> they need so that it can work in all usecases, not a subset of usecases.

You're right that per-cgroup oom_score_adj may resolve the issue with
too strict semantics of oom_priorities. But I believe nobody likes
the existing per-process oom_score_adj interface, and there are reasons behind.
Especially in case of memcg-OOM, getting the idea how exactly oom_score_adj
will work is not trivial.
For example, earlier in this thread I've shown an example, when a decision
which of two processes should be killed depends on whether it's global or
memcg-wide oom, despite both belong to a single cgroup!

Of course, it's technically trivial to implement some analog of oom_score_adj
for cgroups (and early versions of this patchset did that).
But the right question is: is this an interface we want to support
for the next many years? I'm not sure.