Re: [PATCH v5] memcg: expose socket memory pressure in a cgroup

[Kuniyuki Iwashima <kuniyu@google.com>]

On Wed, Oct 15, 2025 at 11:39 AM Shakeel Butt <shakeel.butt@linux.dev> wrote:
>
> On Wed, Oct 15, 2025 at 11:21:17AM -0700, Kuniyuki Iwashima wrote:
> > On Tue, Oct 14, 2025 at 1:33 PM Shakeel Butt <shakeel.butt@linux.dev> wrote:
> > >
> > > On Mon, Oct 13, 2025 at 04:30:53PM +0200, Daniel Sedlak wrote:
> > > [...]
> > > > > > > > How about we track the actions taken by the callers of
> > > > > > > > mem_cgroup_sk_under_memory_pressure()? Basically if network stack
> > > > > > > > reduces the buffer size or whatever the other actions it may take when
> > > > > > > > mem_cgroup_sk_under_memory_pressure() returns, tracking those actions
> > > > > > > > is what I think is needed here, at least for the debugging use-case.
> > > >
> > > > I am not against it, but I feel that conveying those tracked actions (or how
> > > > to represent them) to the user will be much harder. Are there already
> > > > existing APIs to push this information to the user?
> > > >
> > >
> > > I discussed with Wei Wang and she suggested we should start tracking the
> > > calls to tcp_adjust_rcv_ssthresh() first. So, something like the
> > > following. I would like feedback frm networking folks as well:
> >
> > I think we could simply put memcg_memory_event() in
> > mem_cgroup_sk_under_memory_pressure() when it returns
> > true.
> >
> > Other than tcp_adjust_rcv_ssthresh(), if tcp_under_memory_pressure()
> > returns true, it indicates something bad will happen, failure to expand
> > rcvbuf and sndbuf, need to prune out-of-order queue more aggressively,
> > FIN deferred to a retransmitted packet.
> >
> > Also, we could cover mptcp and sctp too.
> >
>
> I wanted to start simple and focus on one specific action but I am open
> to other actins as well. Do we want a generic network throttled metric
> or do we want different metric for different action? At the moment I
> think for memcg, a single metric would be sufficient and then we can
> have tracepoints for more fine grained debugging.

I agree that a single metric would be enough if it can signal
something bad is happening as a first step, then we can take
further action with tracepoint, bpftrace, whatever.