On Mon, Aug 25, 2025 at 1:51 PM Stanislav Fomichev <stfomichev@xxxxxxxxx> wrote:
>
> On 08/25, Kuniyuki Iwashima wrote:
> > We will store a flag in sk->sk_memcg by bpf_setsockopt().
> >
> > For a new child socket, memcg is not allocated until accept(),
> > and the child's sk_memcg is not always the parent's one.
> >
> > For details, see commit e876ecc67db8 ("cgroup: memcg: net: do not
> > associate sock with unrelated cgroup") and commit d752a4986532
> > ("net: memcg: late association of sock to memcg").
> >
> > Let's add a new hook for BPF_PROG_TYPE_CGROUP_SOCK in
> > __inet_accept().
> >
> > This hook does not fail by not supporting bpf_set_retval().
> >
> > Signed-off-by: Kuniyuki Iwashima <kuniyu@xxxxxxxxxx>
>
> And similarly to [0], doing it in sock_ops's BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB
> is not an option because you want to run in the process context instead
> of softirq?
Yes, I considered the hook but ended up adding a new one
in accept(), only when we know sk_memcg is the intended one.
>
> 0: https://lore.kernel.org/netdev/daa73a77-3366-45b4-a770-fde87d4f50d8@xxxxxxxxx/
