Message-ID: <CAJD7tkad5NbqjXZ1qLaNx1g-FYsrv-BVLcNinycFStG_Bu0_zw@mail.gmail.com>
Date:   Thu, 23 Mar 2023 09:51:57 -0700
From:   Yosry Ahmed <yosryahmed@...gle.com>
To:     Shakeel Butt <shakeelb@...gle.com>
Cc:     Tejun Heo <tj@...nel.org>, Josef Bacik <josef@...icpanda.com>,
        Jens Axboe <axboe@...nel.dk>,
        Zefan Li <lizefan.x@...edance.com>,
        Johannes Weiner <hannes@...xchg.org>,
        Michal Hocko <mhocko@...nel.org>,
        Roman Gushchin <roman.gushchin@...ux.dev>,
        Muchun Song <muchun.song@...ux.dev>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Vasily Averin <vasily.averin@...ux.dev>,
        cgroups@...r.kernel.org, linux-block@...r.kernel.org,
        linux-kernel@...r.kernel.org, linux-mm@...ck.org,
        bpf@...r.kernel.org
Subject: Re: [RFC PATCH 1/7] cgroup: rstat: only disable interrupts for the
 percpu lock

On Thu, Mar 23, 2023 at 9:45 AM Shakeel Butt <shakeelb@...gle.com> wrote:
>
> On Thu, Mar 23, 2023 at 9:37 AM Yosry Ahmed <yosryahmed@...gle.com> wrote:
> >
> > On Thu, Mar 23, 2023 at 9:29 AM Shakeel Butt <shakeelb@...gle.com> wrote:
> > >
> > > On Thu, Mar 23, 2023 at 9:18 AM Yosry Ahmed <yosryahmed@...gle.com> wrote:
> > > >
> > > > On Thu, Mar 23, 2023 at 9:10 AM Shakeel Butt <shakeelb@...gle.com> wrote:
> > > > >
> > > > > On Thu, Mar 23, 2023 at 8:46 AM Shakeel Butt <shakeelb@...gle.com> wrote:
> > > > > >
> > > > > > On Thu, Mar 23, 2023 at 8:43 AM Yosry Ahmed <yosryahmed@...gle.com> wrote:
> > > > > > >
> > > > > > > On Thu, Mar 23, 2023 at 8:40 AM Shakeel Butt <shakeelb@...gle.com> wrote:
> > > > > > > >
> > > > > > > > On Thu, Mar 23, 2023 at 6:36 AM Yosry Ahmed <yosryahmed@...gle.com> wrote:
> > > > > > > > >
> > > > > > > > [...]
> > > > > > > > > > >
> > > > > > > > > > > > 2. Are we really calling rstat flush in irq context?
> > > > > > > > > > >
> > > > > > > > > > > I think it is possible through the charge/uncharge path:
> > > > > > > > > > > memcg_check_events()->mem_cgroup_threshold()->mem_cgroup_usage(). I
> > > > > > > > > > > added the protection against flushing in an interrupt context for
> > > > > > > > > > > future callers as well, as it may cause a deadlock if we don't disable
> > > > > > > > > > > interrupts when acquiring cgroup_rstat_lock.
> > > > > > > > > > >
> > > > > > > > > > > > 3. The mem_cgroup_flush_stats() call in mem_cgroup_usage() is only
> > > > > > > > > > > > done for root memcg. Why is mem_cgroup_threshold() interested in root
> > > > > > > > > > > > memcg usage? Why not ignore root memcg in mem_cgroup_threshold() ?
> > > > > > > > > > >
> > > > > > > > > > > I am not sure, but the code looks like event notifications may be set
> > > > > > > > > > > up on root memcg, which is why we need to check thresholds.
> > > > > > > > > >
> > > > > > > > > > This is something we should deprecate as root memcg's usage is ill defined.
> > > > > > > > >
> > > > > > > > > Right, but I think this would be orthogonal to this patch series.
> > > > > > > > >
> > > > > > > >
> > > > > > > > I don't think we can make cgroup_rstat_lock a non-irq-disabling lock
> > > > > > > > without either breaking a link between mem_cgroup_threshold and
> > > > > > > > cgroup_rstat_lock or make mem_cgroup_threshold work without disabling
> > > > > > > > irqs.
> > > > > > > >
> > > > > > > > So, this patch can not be applied before either of those two tasks are
> > > > > > > > done (and we may find more such scenarios).
> > > > > > >
> > > > > > >
> > > > > > > Could you elaborate why?
> > > > > > >
> > > > > > > My understanding is that with an in_task() check to make sure we only
> > > > > > > acquire cgroup_rstat_lock from non-irq context it should be fine to
> > > > > > > acquire cgroup_rstat_lock without disabling interrupts.
> > > > > >
> > > > > > From mem_cgroup_threshold() code path, cgroup_rstat_lock will be taken
> > > > > > with irq disabled while other code paths will take cgroup_rstat_lock
> > > > > > with irq enabled. This is a potential deadlock hazard unless
> > > > > > cgroup_rstat_lock is always taken with irq disabled.
> > > > >
> > > > > Oh you are making sure it is not taken in the irq context through
> > > > > should_skip_flush(). Hmm seems like a hack. Normally it is recommended
> > > > > to actually remove all such users instead of silently
> > > > > ignoring/bypassing the functionality.
> > > >
> > > > It is a workaround, we simply accept to read stale stats in irq
> > > > context instead of the expensive flush operation.
> > > >
> > > > >
> > > > > So, how about removing mem_cgroup_flush_stats() from
> > > > > mem_cgroup_usage(). It will break the known chain which is taking
> > > > > cgroup_rstat_lock with irq disabled and you can add
> > > > > WARN_ON_ONCE(!in_task()).
> > > >
> > > > This changes the behavior in a more obvious way because:
> > > > 1. The memcg_check_events()->mem_cgroup_threshold()->mem_cgroup_usage()
> > > > path is also exercised in a lot of paths outside irq context, this
> > > > will change the behavior for any event thresholds on the root memcg.
> > > > With proposed skipped flushing in irq context we only change the
> > > > behavior in a small subset of cases.
> > > >
> > > > I think we can skip flushing in irq context for now, and separately
> > > > deprecate threshold events for the root memcg. When that is done we
> > > > can come back and remove should_skip_flush() and add a VM_BUG_ON or
> > > > WARN_ON_ONCE instead. WDYT?
> > > >
> > > > 2. mem_cgroup_usage() is also used when reading usage from userspace.
> > > > This should be an easy workaround though.
> > >
> > > This is a cgroup v1 behavior and to me it is totally reasonable to get
> > > the 2 second stale root's usage. Even if you want to skip flushing in
> > > irq, do that in the memcg code and keep VM_BUG_ON/WARN_ON_ONCE in the
> > > rstat core code. This way we will know if other subsystems are doing
> > > the same or not.
> >
> > We can do that. Basically in mem_cgroup_usage() have:
> >
> > /* Some useful comment */
> > if (in_task())
> >     mem_cgroup_flush_stats();
> >
> > and in cgroup_rstat_flush() have:
> > WARN_ON_ONCE(!in_task());
> >
> > I am assuming VM_BUG_ON is not used outside mm code.
> >
> > The only thing that worries me is that if there is another unlikely
> > path somewhere that flushes stats in irq context we may run into a
> > deadlock. I am a little bit nervous about not skipping flushing if
> > !in_task() in cgroup_rstat_flush().
>
> I think it is a good thing. We will find such scenarios and fix those
> instead of hiding them forever or keeping the door open for new such
> scenarios.

Sure, I can do that in the next version. I will include a patch that
adds an in_task() check to mem_cgroup_usage() before this one. Since
BUG_ON() is discouraged and VM_BUG_ON() is mm specific, I guess we are
left with WARN_ON_ONCE() for the rstat core code, right?

Thanks Shakeel. Any other thoughts I should address for the next version?

Powered by blists - more mailing lists