Message-ID: <CAJD7tkasteOdJ_waQ0r9WqbiwRz3OCzmx35v7-XMNOLb2WRddQ@mail.gmail.com>
Date:   Thu, 7 Sep 2023 18:11:22 -0700
From:   Yosry Ahmed <yosryahmed@...gle.com>
To:     Ivan Babrou <ivan@...udflare.com>
Cc:     Wei Xu <weixugc@...gle.com>, Michal Hocko <mhocko@...e.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Johannes Weiner <hannes@...xchg.org>,
        Roman Gushchin <roman.gushchin@...ux.dev>,
        Shakeel Butt <shakeelb@...gle.com>,
        Muchun Song <muchun.song@...ux.dev>, Tejun Heo <tj@...nel.org>,
        Michal Koutný <mkoutny@...e.com>,
        Waiman Long <longman@...hat.com>, linux-mm@...ck.org,
        cgroups@...r.kernel.org, linux-kernel@...r.kernel.org,
        Greg Thelen <gthelen@...gle.com>
Subject: Re: [PATCH v4 4/4] mm: memcg: use non-unified stats flushing for
 userspace reads

On Thu, Sep 7, 2023 at 6:03 PM Ivan Babrou <ivan@...udflare.com> wrote:
>
> On Thu, Sep 7, 2023 at 5:52 PM Wei Xu <weixugc@...gle.com> wrote:
> >
> > On Mon, Sep 4, 2023 at 8:15 AM Michal Hocko <mhocko@...e.com> wrote:
> > >
> > > On Thu 31-08-23 16:56:11, Yosry Ahmed wrote:
> > > > Unified flushing allows for great concurrency for paths that attempt to
> > > > flush the stats, at the expense of potential staleness and a single
> > > > flusher paying the extra cost of flushing the full tree.
> > > >
> > > > This tradeoff makes sense for in-kernel flushers that may observe high
> > > > concurrency (e.g. reclaim, refault). For userspace readers, stale stats
> > > > may be unexpected and problematic, especially when such stats are used
> > > > for critical paths such as userspace OOM handling. Additionally, a
> > > > userspace reader will occasionally pay the cost of flushing the entire
> > > > hierarchy, which also causes problems in some cases [1].
> > > >
> > > > Opt userspace reads out of unified flushing. This makes the cost of
> > > > reading the stats more predictable (proportional to the size of the
> > > > subtree), as well as the freshness of the stats. Userspace readers are
> > > > not expected to have similar concurrency to in-kernel flushers,
> > > > serializing them among themselves and among in-kernel flushers should be
> > > > okay. Nonetheless, for extra safety, introduce a mutex when flushing for
> > > > userspace readers to make sure only a single userspace reader can compete
> > > > with in-kernel flushers at a time. This takes away userspace ability to
> > > > directly influence or hurt in-kernel lock contention.
> > >
> > > I think it would be helpful to note that the primary reason this is a
> > > concern is that the spinlock is dropped during flushing under
> > > contention.
> > >
> > > > An alternative is to remove flushing from the stats reading path
> > > > completely, and rely on the periodic flusher. This should be accompanied
> > > > by making the periodic flushing period tunable, and providing an
> > > > interface for userspace to force a flush, following a similar model to
> > > > /proc/vmstat. However, such a change will be hard to reverse if the
> > > > implementation needs to be changed because:
> > > > - The cost of reading stats will be very cheap and we won't be able to
> > > >   take that back easily.
> > > > - There are user-visible interfaces involved.
> > > >
> > > > Hence, let's go with the change that's most reversible first and revisit
> > > > as needed.
> > > >
> > > > This was tested on a machine with 256 cpus by running a synthetic test
> > > > script [2] that creates 50 top-level cgroups, each with 5 children (250
> > > > leaf cgroups). Each leaf cgroup has 10 processes running that allocate
> > > > memory beyond the cgroup limit, invoking reclaim (which is an in-kernel
> > > > unified flusher). Concurrently, one thread is spawned per-cgroup to read
> > > > the stats every second (including root, top-level, and leaf cgroups --
> > > > so total 251 threads). No significant regressions were observed in the
> > > > total run time, which means that userspace readers are not significantly
> > > > affecting in-kernel flushers:
> > > >
> > > > Base (mm-unstable):
> > > >
> > > > real  0m22.500s
> > > > user  0m9.399s
> > > > sys   73m41.381s
> > > >
> > > > real  0m22.749s
> > > > user  0m15.648s
> > > > sys   73m13.113s
> > > >
> > > > real  0m22.466s
> > > > user  0m10.000s
> > > > sys   73m11.933s
> > > >
> > > > With this patch:
> > > >
> > > > real  0m23.092s
> > > > user  0m10.110s
> > > > sys   75m42.774s
> > > >
> > > > real  0m22.277s
> > > > user  0m10.443s
> > > > sys   72m7.182s
> > > >
> > > > real  0m24.127s
> > > > user  0m12.617s
> > > > sys   78m52.765s
> > > >
> > > > [1]https://lore.kernel.org/lkml/CABWYdi0c6__rh-K7dcM_pkf9BJdTRtAU08M43KO9ME4-dsgfoQ@mail.gmail.com/
> > > > [2]https://lore.kernel.org/lkml/CAJD7tka13M-zVZTyQJYL1iUAYvuQ1fcHbCjcOBZcz6POYTV-4g@mail.gmail.com/
> > > >
> > > > Signed-off-by: Yosry Ahmed <yosryahmed@...gle.com>
> > >
> > > OK, I can live with that but I still believe that locking involved in
> > > the user interface only begs for issues later on as there is no control
> > > over that lock contention other than the number of processes involved.
> > > As it seems that we cannot make a consensus on this concern now and this
> > > should be already helping existing workloads then let's just buy some
> > > more time ;)
> >
> > Indeed, even though the new global mutex protects the kernel from the
> > userspace contention on the rstats spinlock, its current
> > implementation doesn't have any protection for the lock contention
> > among the userspace threads and can cause significant delays to memcg
> > stats reads.
> >
> > I tested this patch on a machine with 384 CPUs using a microbenchmark
> > that spawns 10K threads, each reading its memory.stat every 100
> > milliseconds.  Most of memory.stat reads take 5ms-10ms in kernel, with
> > ~5% reads even exceeding 1 second. This is a significant regression.
> > In comparison, without contention, each memory.stat read only takes
> > 20us-50us in the kernel.  Almost all of the extra read time is spent
> > on waiting for the new mutex. The time to flush rstats only accounts
> > for 10us-50us (This test creates only 1K memory cgroups and doesn't
> > generate any loads other than these stat reader threads).
> >
> >  Here are some ideas to control the lock contention on the mutex and
> > reduce both the median and tail latencies of concurrent memcg stat
> > reads:


Thanks for the analysis, Wei!

I will update the patch series based on your ideas to limit the
contention on the userspace read mutex.

>
> >
> > - Bring back the stats_flush_threshold check in
> > mem_cgroup_try_flush_stats() to mem_cgroup_user_flush_stats().  This
> > check provides a reasonable bound on the stats staleness while being
> > able to filter out a large number of rstats flush requests, which
> > reduces the contention on the mutex.
> >
> > - When contended, upgrade the per-memcg rstats flush in
> > mem_cgroup_user_flush_stats() to a root memcg flush and coalesce these
> > contended flushes together.  We can wait for the ongoing flush to
> > complete and eliminate repeated flush requests.
>
> Full root memcg flush being slow is one of the issues that prompted this patch:
>
> * https://lore.kernel.org/lkml/CABWYdi0c6__rh-K7dcM_pkf9BJdTRtAU08M43KO9ME4-dsgfoQ@mail.gmail.com/
>
> I don't want us to regress in this regard.


It will only be a fallback if there is high concurrency among
userspace reads, which will cause high contention on the mutex. In
that case, the userspace reads will be slowed down by contention,
which can be even worse than flush slowness as it is theoretically
unbounded.

I am working on a v5 now to incorporate Wei's suggestions. Would you
be able to test that and verify that there are no regressions?

>
>
> > - Wait for the mutex and the ongoing flush with a timeout.  We should
> > not use busy-wait, though.  We can bail out to read the stats without
> > a flush after enough wait.  A long-stalled system call is much worse
> > than somewhat stale stats in the corner cases in my opinion.
> >
> > Wei Xu
> >

Powered by blists - more mailing lists