lkml.org 
[lkml]   [2017]   [Oct]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [v10 3/6] mm, oom: cgroup-aware OOM killer
On Wed, 4 Oct 2017, Johannes Weiner wrote:

> > By only considering leaf memcgs, does this penalize users if their memcg
> > becomes oc->chosen_memcg purely because it has aggregated all of its
> > processes to be members of that memcg, which would otherwise be the
> > standard behavior?
> >
> > What prevents me from spreading my memcg with N processes attached over N
> > child memcgs instead so that memcg_oom_badness() becomes very small for
> > each child memcg specifically to avoid being oom killed?
>
> It's no different from forking out multiple mm to avoid being the
> biggest process.
>

It is, because it can quite clearly be a DoS, and was prevented with
Roman's earlier design of iterating usage up the hierarchy and comparing
siblings based on that criteria. I know exactly why he chose that
implementation detail early on, and it was to prevent cases such as this
and to not let userspace hide from the oom killer.

> It's up to the parent to enforce limits on that group and prevent you
> from being able to cause global OOM in the first place, in particular
> if you delegate to untrusted and potentially malicious users.
>

Let's resolve that global oom is a real condition and getting into that
situation is not a userspace problem. It's the result of overcommiting
the system, and is used in the enterprise to address business goals. If
the above is true, and its up to memcg to prevent global oom in the first
place, then this entire patchset is absolutely pointless. Limit userspace
to 95% of memory and when usage is approaching that limit, let userspace
attached to the root memcg iterate the hierarchy itself and kill from the
largest consumer.

This patchset exists because overcommit is real, exactly the same as
overcommit within memcg hierarchies is real. 99% of the time we don't run
into global oom because people aren't using their limits so it just works
out. 1% of the time we run into global oom and we need a decision to made
based for forward progress. Using Michal's earlier example of admins and
students, a student can easily use all of his limit and also, with v10 of
this patchset, 99% of the time avoid being oom killed just by forking N
processes over N cgroups. It's going to oom kill an admin every single
time.

I know exactly why earlier versions of this patchset iterated that usage
up the tree so you would pick from students, pick from this troublemaking
student, and then oom kill from his hierarchy. Roman has made that point
himself. My suggestion was to add userspace influence to it so that
enterprise users and users with business goals can actually define that we
really do want 80% of memory to be used by this process or this hierarchy,
it's in our best interest.

Earlier iterations of this patchset did this, and did it correctly.
Userspace influence over the decisionmaking makes it a very powerful
combination because you _can_ specify what your goals are or choose to
leave the priorities as default so you can compare based solely on usage.
It was a beautiful solution to the problem.

\
 
 \ /
  Last update: 2017-10-05 10:40    [W:0.096 / U:0.380 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site