On Wed, Aug 13, 2025 at 6:00 AM Johannes Weiner [off-list ref] wrote:

On Tue, Aug 12, 2025 at 05:58:30PM +0000, Kuniyuki Iwashima wrote:

quoted

If all workloads were guaranteed to be controlled under memcg, the issue
could be worked around by setting tcp_mem[0~2] to UINT_MAX.

In reality, this assumption does not always hold, and processes that
belong to the root cgroup or opt out of memcg can consume memory up to
the global limit, becoming a noisy neighbour.

As per the last thread, this is not a supported usecase. Opting out of
memcg coverage for individual cgroups is a self-inflicted problem and
misconfiguration. There is *no* memory isolation *at all* on such
containers.

I think the commit message needs to be improved, but could
you read throughout the patch again ?  I think you have the
same misunderstanding that Shakeel had and corrected here.
https://lore.kernel.org/netdev/jmbszz4m7xkw7fzolpusjesbreaczmr4i64kynbs3zcoehrkpj@lwso5soc4dh3/ (local)

---8<---
Initially, I thought the series introduced multiple modes, including an
option to exclude network memory from memcg accounting. However, if I
understand correctly, that is not the case—the opt-out applies only to
the global TCP/UDP accounting. That’s a relief, and I apologize for the
misunderstanding.
---8<---

This patch does NOT change how memcg is applied to sockets
but changes how _another_ memory accounting in the networking
layer is applied to sockets.

Currently, memcg AND the other mem accounting are applied
to socket buffers.

With/without this patch, memcg is _always_ applied to socket
buffers.

Also, there is _no_ behavioural change for _uncontrolled
containers_ that have been subject to the two memory
accounting.  This behaviour hasn't been changed since
you added memcg support for the networking stack in
e805605c72102, and we want to _preserve_ this behaviour.

This change stop double-charging by opting out of _the
networking layer one_ because it interferes with memcg
and complicates configuration of memory.max and the
global networking limit.

Maybe their socket buffers is the only thing that happens
to matter to *you*, but this is in no way a generic, universal,
upstreamable solution. Knob or auto-detection is not the issue.

Nacked-by: Johannes Weiner [off-list ref]

Please let me know if this nack still applies with the
explanation above.