Re: [PATCH] mm: ratelimit stat flush from workingset shrinker

[Shakeel Butt]

On Thu, Dec 28, 2023 at 07:13:23AM -0800, Yosry Ahmed wrote:
> On Wed, Dec 27, 2023 at 11:31 PM Shakeel Butt <shakeelb@xxxxxxxxxx> wrote:
> >
> > One of our internal workload regressed on newer upstream kernel and on
> > further investigation, it seems like the cause is the always synchronous
> > rstat flush in the count_shadow_nodes() added by the commit f82e6bf9bb9b
> > ("mm: memcg: use rstat for non-hierarchical stats"). On further
> > inspection it seems like we don't really need accurate stats in this
> > function as it was already approximating the amount of appropriate
> > shadow entried to keep for maintaining the refault information. Since
> 
> s/entried/entries
> 
> > there is already 2 sec periodic rstat flush, we don't need exact stats
> > here. Let's ratelimit the rstat flush in this code path.
> 
> Is the regression observed even with commit 7d7ef0a4686a ("mm: memcg:
> restore subtree stats flushing")? I think the answer is yes based on
> internal discussions, but this really surprises me.
> 

Yes, the regression was on latest mm-stable branch of Andrew's mm tree.

> Commit f82e6bf9bb9b removed the percpu loop in
> lruvec_page_state_local(), and added a flush call. With  7d7ef0a4686a,
> the flush call is only effective if there are pending updates in the
> cgroup subtree exceeding MEMCG_CHARGE_BATCH * num_online_cpus(). IOW,
> we should only be doing work when actually needed, whereas before we
> used to have multiple percpu loops in count_shadow_nodes() regardless
> of pending updates.
> 
> It seems like the cgroup subtree is very active that we continuously
> need to flush in count_shadow_nodes()? If that's the case, do we still
> think it's okay not to flush when we know there are pending updates? I
> don't have enough background about the workingset heuristics to judge
> this.

Not all updates might be related to the stats being read here. Also the
read value is further divided by 8 and manipulated more in
do_shrink_slab(). So, I don't think we need less than 2 seconds accuracy
for these stats here.