Re: [v4 2/4] mm, oom: cgroup-aware OOM killer

[Roman Gushchin <guro@fb.com>]

On Wed, Aug 02, 2017 at 09:29:01AM +0200, Michal Hocko wrote:
> On Tue 01-08-17 19:13:52, Roman Gushchin wrote:
> > On Tue, Aug 01, 2017 at 07:03:03PM +0200, Michal Hocko wrote:
> > > On Tue 01-08-17 16:25:48, Roman Gushchin wrote:
> > > > On Tue, Aug 01, 2017 at 04:54:35PM +0200, Michal Hocko wrote:
> > > [...]
> > > > > I would reap out the oom_kill_process into a separate patch.
> > > > 
> > > > It was a separate patch, I've merged it based on Vladimir's feedback.
> > > > No problems, I can divide it back.
> > > 
> > > It would make the review slightly more easier
> > > > 
> > > > > > -static void oom_kill_process(struct oom_control *oc, const char *message)
> > > > > > +static void __oom_kill_process(struct task_struct *victim)
> > > > > 
> > > > > To the rest of the patch. I have to say I do not quite like how it is
> > > > > implemented. I was hoping for something much simpler which would hook
> > > > > into oom_evaluate_task. If a task belongs to a memcg with kill-all flag
> > > > > then we would update the cumulative memcg badness (more specifically the
> > > > > badness of the topmost parent with kill-all flag). Memcg will then
> > > > > compete with existing self contained tasks (oom_badness will have to
> > > > > tell whether points belong to a task or a memcg to allow the caller to
> > > > > deal with it). But it shouldn't be much more complex than that.
> > > > 
> > > > I'm not sure, it will be any simpler. Basically I'm doing the same:
> > > > the difference is that you want to iterate over tasks and for each
> > > > task traverse the memcg tree, update per-cgroup oom score and find
> > > > the corresponding memcg(s) with the kill-all flag. I'm doing the opposite:
> > > > traverse the cgroup tree, and for each leaf cgroup iterate over processes.
> > > 
> > > Yeah but this doesn't fit very well to the existing scheme so we would
> > > need two different schemes which is not ideal from maint. point of view.
> > > We also do not have to duplicate all the tricky checks we already do in
> > > oom_evaluate_task. So I would prefer if we could try to hook there and
> > > do the special handling there.
> > 
> > I hope, that iterating over all tasks just to check if there are
> > in-flight OOM victims might be optimized at some point.
> > That means, we would be able to choose a victim much cheaper.
> > It's not easy, but it feels as a right direction to go.
> 
> You would have to count per each oom domain and that sounds quite
> unfeasible to me.

It's hard, but traversing the whole cgroup tree from bottom to top
for each task is just not scalable.

This is exactly why I've choosen a compromise right now: let's
iterate over all tasks, but do it by iterating over the cgroup tree.

>  
> > Also, adding new tricks to the oom_evaluate_task() will make the code
> > even more hairy. Some of the existing tricks are useless for memcg selection.
> 
> Not sure what you mean but oom_evaluate_task has been usable for both
> global and memcg oom paths so far. I do not see any reason why this
> shouldn't hold for a different oom killing strategy.

Yes, but in both cases we've evaluated tasks, not cgroups.

> 
> > > > Also, please note, that even without the kill-all flag the decision is made
> > > > on per-cgroup level (except tasks in the root cgroup).
> > > 
> > > Yeah and I am not sure this is a reasonable behavior. Why should we
> > > consider memcgs which are not kill-all as a single entity?
> > 
> > I think, it's reasonable to choose a cgroup/container to blow off based on
> > the cgroup oom_priority/size (including hierarchical settings), and then
> > kill one biggest or all tasks depending on cgroup settings.
> 
> But that doesn't mean you have to treat even !kill-all memcgs like a
> single entity. In fact we should compare killable entities which is
> either a task or the whole memcg if configured that way.

I believe it's absolutely valid user's intention to prioritize some
cgroups over other, even if only one task should be killed in case of OOM.

Thanks!

Roman