Re: [PATCH] cgroup: rstat: use LOCK CMPXCHG in css_rstat_updated

[Tejun Heo]

Hello,

On Thu, Dec 04, 2025 at 06:24:37PM -0800, Shakeel Butt wrote:
...
> In Meta's fleet running the kernel with the commit 36df6e3dbd7e, we are
> observing on some machines the memcg stats are getting skewed by more
> than the actual memory on the system. On close inspection, we noticed
> that lockless node for a workload for specific CPU was in the bad state
> and thus all the updates on that CPU for that cgroup was being lost. At
> the moment, we are not sure if this CMPXCHG without LOCK is the cause of
> that but this needs to be fixed irrespective.

Is there a plausible theory of events that can explain the skew with the use
of this_cpu_cmpxchg()? lnode.next being set to self but this_cpu_cmpxchg()
returning something else? It may be useful to write a targeted repro for the
particular combination - this_cpu_cmpxchg() vs. remote NULL clearing and see
whether this_cpu_cmpxchg() can return a value that doesn't agree with what
gets written in the memory.

> @@ -113,9 +112,8 @@ __bpf_kfunc void css_rstat_updated(struct cgroup_subsys_state *css, int cpu)
>  	 * successful and the winner will eventually add the per-cpu lnode to
>  	 * the llist.
>  	 */
> -	self = &rstatc->lnode;
> -	rstatc_pcpu = css->rstat_cpu;
> -	if (this_cpu_cmpxchg(rstatc_pcpu->lnode.next, self, NULL) != self)
> +	expected = &rstatc->lnode;
> +	if (!try_cmpxchg(&rstatc->lnode.next, &expected, NULL))

Given that this is a relatively cold path, I don't see a problem with using
locked op here even if this wasn't necessarily the culprit; however, can you
please update the comment right above accordingly and explain why the locked
op is used? After this patch, the commend and code disagree.

Thanks.

-- 
tejun