On Sun, Jan 28, 2024 at 12:56:32PM -0800, Kuniyuki Iwashima wrote:
> From: Kent Overstreet <kent.overstreet@xxxxxxxxx>
> Date: Sun, 28 Jan 2024 14:38:02 -0500
> > On Sun, Jan 28, 2024 at 12:28:38AM -0800, Kuniyuki Iwashima wrote:
> > > From: Kent Overstreet <kent.overstreet@xxxxxxxxx>
> > > Date: Fri, 26 Jan 2024 21:08:31 -0500
> > > > Kill
> > > > - unix_state_lock_nested
> > > > - _nested usage for net->unx.table.locks[].
> > > >
> > > > replace both with lock_set_cmp_fn_ptr_order(&u->lock).
> > > >
> > > > The lock ordering in sk_diag_dump_icons() looks suspicious; this may
> > > > turn up a real issue.
> > >
> > > Yes, you cannot use lock_cmp_fn() for unix_state_lock_nested().
> > >
> > > The lock order in sk_diag_dump_icons() is
> > >
> > > listening socket -> child socket in the listener's queue
> > >
> > > , and the inverse order never happens. ptr comparison does not make
> > > sense in this case, and lockdep will complain about false positive.
> >
> > Is that a real lock ordering? Is this parent -> child relationship well
> > defined?
> >
> > If it is, we should be able to write a lock_cmp_fn for it, as long as
> > it's not some handwavy "this will never happen but _nested won't check
> > for it" like I saw elsewhere in the net code... :)
>
> The problem would be there's no handy way to detect the relationship
> except for iterating the queue again.
>
> ---8<---
> static int unix_state_lock_cmp_fn(const struct lockdep_map *_a,
> const struct lockdep_map *_b)
> {
> const struct unix_sock *a = container_of(_a, struct unix_sock, lock.dep_map);
> const struct unix_sock *b = container_of(_b, struct unix_sock, lock.dep_map);
>
> if (a->sk.sk_state == TCP_LISTEN && b->sk.sk_state == TCP_ESTABLISHED) {
> /* check if b is a's cihld */
> }
>
> /* otherwise, ptr comparison here. */
> }
> ---8<---
>
>
> This can be resolved by a patch like this, which is in my local tree
> for another series.
>
> So, after posting the series, I can revisit this and write lock_cmp_fn
> for u->lock.
Sounds good! Please CC me when you do.
