Re: general protection fault in raw_seq_start

[Kuniyuki Iwashima]

From:   Eric Dumazet <edumazet@xxxxxxxxxx>
Date:   Fri, 31 Mar 2023 09:04:47 +0200
> On Thu, Mar 30, 2023 at 11:55 PM Kuniyuki Iwashima <kuniyu@xxxxxxxxxx> wrote:
> 
> > Thanks for reporting the issue.
> >
> > It seems we need to use RCU variant in raw_get_first().
> > I'll post a patch.
> >
> > ---
> > diff --git a/net/ipv4/raw.c b/net/ipv4/raw.c
> > index 3cf68695b40d..fe0d1ad20b35 100644
> > --- a/net/ipv4/raw.c
> > +++ b/net/ipv4/raw.c
> > @@ -957,7 +957,7 @@ static struct sock *raw_get_first(struct seq_file *seq, int bucket)
> >         for (state->bucket = bucket; state->bucket < RAW_HTABLE_SIZE;
> >                         ++state->bucket) {
> >                 hlist = &h->ht[state->bucket];
> > -               sk_nulls_for_each(sk, hnode, hlist) {
> > +               sk_nulls_for_each_rcu(sk, hnode, hlist) {
> >                         if (sock_net(sk) == seq_file_net(seq))
> >                                 return sk;
> >
> 
> No, we do not want this.
> You missed that sk_nulls_for_each_rcu() needs a specific protocol
> (see Documentation/RCU/rculist_nulls.rst for details)

Ah, exactly SOCK_RAW does not have SLAB_TYPESAFE_BY_RCU.
Thank you for pointing this out!

And I found this seems wrong.

c25b7a7a565e ("inet: ping: use hlist_nulls rcu iterator during lookup")

> 
> RCU is needed in the data path, not for this control path.
> 
> My patch went too far in the RCU conversion. I did not think about
> syzbot harassing /proc files :)
> 
> We need raw_seq_start and friends to go back to use the lock.

Ok, then I'll change /proc/net/{raw, icmp} to use spinlock :)