On Thu 03-01-19 12:52:54, Shakeel Butt wrote:
> On Mon, Dec 31, 2018 at 2:12 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> >
> > On Sun 30-12-18 19:59:53, Shakeel Butt wrote:
> > > On Sun, Dec 30, 2018 at 12:00 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > > >
> > > > On Sun 30-12-18 08:45:13, Michal Hocko wrote:
> > > > > On Sat 29-12-18 11:34:29, Shakeel Butt wrote:
> > > > > > On Sat, Dec 29, 2018 at 2:06 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On Sat 29-12-18 10:52:15, Florian Westphal wrote:
> > > > > > > > Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> > > > > > > > > On Fri 28-12-18 17:55:24, Shakeel Butt wrote:
> > > > > > > > > > The [ip,ip6,arp]_tables use x_tables_info internally and the underlying
> > > > > > > > > > memory is already accounted to kmemcg. Do the same for ebtables. The
> > > > > > > > > > syzbot, by using setsockopt(EBT_SO_SET_ENTRIES), was able to OOM the
> > > > > > > > > > whole system from a restricted memcg, a potential DoS.
> > > > > > > > >
> > > > > > > > > What is the lifetime of these objects? Are they bound to any process?
> > > > > > > >
> > > > > > > > No, they are not.
> > > > > > > > They are free'd only when userspace requests it or the netns is
> > > > > > > > destroyed.
> > > > > > >
> > > > > > > Then this is problematic, because the oom killer is not able to
> > > > > > > guarantee the hard limit and so the excessive memory consumption cannot
> > > > > > > be really contained. As a result the memcg will be basically useless
> > > > > > > until somebody tears down the charged objects by other means. The memcg
> > > > > > > oom killer will surely kill all the existing tasks in the cgroup and
> > > > > > > this could somehow reduce the problem. Maybe this is sufficient for
> > > > > > > some usecases but that should be properly analyzed and described in the
> > > > > > > changelog.
> > > > > > >
> > > > > >
> > > > > > Can you explain why you think the memcg hard limit will not be
> > > > > > enforced? From what I understand, the memcg oom-killer will kill the
> > > > > > allocating processes as you have mentioned. We do force charging for
> > > > > > very limited conditions but here the memcg oom-killer will take care
> > > > > > of
> > > > >
> > > > > I was talking about the force charge part. Depending on a specific
> > > > > allocation and its life time this can gradually get us over hard limit
> > > > > without any bound theoretically.
> > > >
> > > > Forgot to mention. Since b8c8a338f75e ("Revert "vmalloc: back off when
> > > > the current task is killed"") there is no way to bail out from the
> > > > vmalloc allocation loop so if the request is really large then the memcg
> > > > oom will not help. Is that a problem here?
> > > >
> > >
> > > Yes, I think it will be an issue here.
> > >
> > > > Maybe it is time to revisit fatal_signal_pending check.
> > >
> > > Yes, we will need something to handle the memcg OOM. I will think more
> > > on that front or if you have any ideas, please do propose.
> >
> > I can see three options here:
> > - do not force charge on memcg oom or introduce a limited charge
> > overflow (reserves basically).
> > - revert the revert and reintroduce the fatal_signal_pending
> > check into vmalloc
> > - be more specific and check tsk_is_oom_victim in vmalloc and
> > fail
> >
>
> I think for the long term solution we might need something similar to
> memcg oom reserves (1) but for quick fix I think we can do the
> combination of (2) and (3).
Johannes argued that fatal_signal_pending is too general check for
vmalloc. I would argue that we already break out of some operations on
fatal signals. tsk_is_oom_victim is more subtle but much more targeted
on the other hand.
I do not have any strong preference to be honest but I agree that some
limited reserves would be the best solution long term. I just do not
have any idea how to scale those reserves to be meaningful.
--
Michal Hocko
SUSE Labs
