Re: [patch -mm 3/4] mm, memcg: replace memory.oom_group with policy tunable

[Michal Hocko <mhocko@kernel.org>]

On Wed 17-01-18 14:18:33, David Rientjes wrote:
> On Wed, 17 Jan 2018, Michal Hocko wrote:
> 
> > Absolutely agreed! And moreover, there are not all that many ways what
> > to do as an action. You just kill a logical entity - be it a process or
> > a logical group of processes. But you have way too many policies how
> > to select that entity. Do you want to chose the youngest process/group
> > because all the older ones have been computing real stuff and you would
> > lose days of your cpu time? Or should those who pay more should be
> > protected (aka give them static priorities), or you name it...
> > 
> 
> That's an argument for making the interface extensible, yes.

And there is no interface to control the selection yet so we can develop
one on top.
 
> > I am sorry, I still didn't grasp the full semantic of the proposed
> > soluton but the mere fact it is starting by conflating selection and the
> > action is a no go and a wrong API. This is why I've said that what you
> > (David) outlined yesterday is probably going to suffer from a much
> > longer discussion and most likely to be not acceptable. Your patchset
> > proves me correct...
> 
> I'm very happy to change the API if there are better suggestions.  That 
> may end up just being an memory.oom_policy file, as this implements, and 
> separating out a new memory.oom_action that isn't a boolean value to 
> either do a full group kill or only a single process.  Or it could be what 
> I suggested in my mail to Tejun, such as "hierarchy killall" written to
> memory.oom_policy, which would specify a single policy and then an 
> optional mechanism.  With my proposed patchset, there would then be three 
> policies: "none", "cgroup", and "tree" and one possible optional 
> mechanism: "killall".

You haven't convinced me at all. This all sounds more like "what if"
than a really thought through interface. I've tried to point out that
having a real policy driven victim selection is a _hard_ thing to do
_right_.

On the other hand oom_group makes semantic sense. It controls the
killable entity and there are usecases which want to consider the full
memcg as a single killable entity. No matter what selection policy we
chose on top. It is just a natural API.

Now you keep arguing about the victim selection and different strategies
to implement it. We will not move forward as long as you keep conflating
the two things, I am afraid.
-- 
Michal Hocko
SUSE Labs

--
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>