Re: [PATCH 2/2] mm: Consider subtrees in

From: Tejun Heo
Date: Thu Jan 24 2019 - 10:21:30 EST

Hello, Michal.

On Thu, Jan 24, 2019 at 09:22:52AM +0100, Michal Hocko wrote:
> I do not think we can do that for two reasons. It breaks the existing
> semantic userspace might depend on and more importantly this is not a
> correct behavior IMO.

This is a valid concern but I'll come back to this later.

> You have to realize that stats are hierarchical because that is how we
> account. Events represent a way to inform that something has happened at
> the specific level of the tree though. If you do not setup low/high/max

This isn't true. e.g.'s populated event is
hierarchical. Everything in cgroup should be hierarchical by default.

> limit then you simply cannot expect to be informed those get hit because
> they cannot by definition. Or put it other way, if you are waiting for
> those events you really want to know the (sub)tree they happened and if
> you propagate the event up the hierarchy you have hard time to tell that
> (you would basically have to exclude all but the lowest one and that is
> an awkward semantic at best.

I don't think it's a good idea to argue this for each piece of
information. Again, everything should be hierarchical unless there
are clear and strong reasons against; otherwise, we end up with random
mix of hierarchical and flat behaviors, something that we want to
avoid the most - remember .use_hierarchy?.

> Maybe we want to document this better but I do not see we are going to
> change the behavior.

I beg you to reconsider. This was a clear oversight and the cgroup2
usage is still relatively limited. We sure can add local-specific
counters if needed but must not mix local and hierarchical counters
without a clear way to tell what's what.