Re: [PATCH 3/3] mm: memcg: use non-unified stats flushing for userspace reads

[Michal Hocko <mhocko@suse.com>]

On Fri 25-08-23 11:21:16, Yosry Ahmed wrote:
> On Fri, Aug 25, 2023 at 11:17 AM Michal Hocko <mhocko@suse.com> wrote:
> >
> > On Fri 25-08-23 08:14:54, Yosry Ahmed wrote:
> > > On Fri, Aug 25, 2023 at 12:05 AM Michal Hocko <mhocko@suse.com> wrote:
> > [...]
> > > > I might be wrong but the whole discussion so far suggests that the
> > > > global rstat lock should be reconsidered. From my personal experience
> > > > global locks easily triggerable from the userspace are just a receip for
> > > > problems. Stats reading shouldn't be interfering with the system runtime
> > > > as much as possible and they should be deterministic wrt runtime as
> > > > well.
> > >
> > > The problem is that the global lock also serializes the global
> > > counters that we flush to. I will talk from the memcg flushing
> > > perspective as that's what I am familiar with. I am not sure how much
> > > this is transferable to other flushers.
> > >
> > > On the memcg side (see mem_cgroup_css_rstat_flush()), the global lock
> > > synchronizes access to multiple counters, for this discussion what's
> > > most important are:
> > > - The global stat counters of the memcg being flushed (e.g.
> > > memcg->vmstats->state).
> > > - The pending stat counters of the parent being flushed (e.g.
> > > parent->vmstats->state_pending).
> >
> > I haven't digested the rest of the email yet (Friday brain, sorry) but I
> > do not think you are adressing this particular part so let me ask before
> > I dive more into the following. I really do not follow the serialization
> > requirement here because the lock doesn't really serialize the flushing,
> > does it? At least not in a sense of a single caller to do the flushing
> > atomicaly from other flushers. It is possible that the current flusher
> > simply drops the lock midway and another one retakes the lock and
> > performs the operation again. So what additional flushing
> > synchronization does it provide and why cannot parallel flushers simply
> > compete over pcp spinlocks?
> >
> > So what am I missing?
> 
> Those counters are non-atomic. The lock makes sure we don't have two
> concurrent flushers updating the same counter locklessly and
> non-atomically, which would be possible if we flush the same cgroup on
> two different cpus in parallel.

pcp lock (cpu_lock) guarantees the very same, doesn't it?
-- 
Michal Hocko
SUSE Labs