Re: [PATCH net] sctp: annotate data-races around sk->sk_wmem_queued

[Eric Dumazet <edumazet@google.com>]

On Tue, Aug 29, 2023 at 9:05 PM Xin Long <lucien.xin@gmail.com> wrote:
>
> On Tue, Aug 29, 2023 at 2:19 PM Eric Dumazet <edumazet@google.com> wrote:
> >
> > On Tue, Aug 29, 2023 at 8:14 PM Xin Long <lucien.xin@gmail.com> wrote:
> > >
> > > On Mon, Aug 28, 2023 at 8:44 AM Eric Dumazet <edumazet@google.com> wrote:
> > > >
> > > > sk->sk_wmem_queued can be read locklessly from sctp_poll()
> > > >
> > > >                 sk->sk_rcvbuf);
> > > Just wondering why sk->sk_sndbuf/sk_rcvbuf doesn't need READ_ONCE()
> > > while adding READ_ONCE for sk->sk_wmem_queued in here?
> > >
> >
> > Separate patches for sk_sndbuf, sk_rcvbuf, sk_err, sk_shutdown, and
> > many other socket fields.
> >
> > I prefer having small patches to reduce merge conflicts in backports.
> >
> > Note that I used  READ_ONCE(sk->sk_sndbuf) in sctp_writeable(),
> > (I assume this is why you asked)
> Yes.
>
> Not sure about tcp's seq_show, but as sctp_assocs_seq_show() is only

tcp seq_show uses requested spinlocks on lhash or ehash tables, but
not socket lock.

> under rcu_read_lock() and with a hold of transport/association/socket,
> does it mean all members of assoc should also use READ_ONCE()?

Probably...

> (Note I think we don't expect the seq show to be that accurate.)

Yeah, this is really best effort, inet_diag is probably what we want
to harden these days.