Re: net/sctp: list double add warning in sctp_endpoint_add_asoc

[Marcelo Ricardo Leitner]

On Wed, Apr 05, 2017 at 06:48:45PM +0800, Xin Long wrote:
> On Wed, Apr 5, 2017 at 5:14 AM, Marcelo Ricardo Leitner
> <marcelo.leitner@xxxxxxxxx> wrote:
> > On Wed, Apr 05, 2017 at 01:29:19AM +0800, Xin Long wrote:
> >> On Tue, Apr 4, 2017 at 9:28 PM, Andrey Konovalov <andreyknvl@xxxxxxxxxx> wrote:
> >> > Hi,
> >> >
> >> > I've got the following error report while fuzzing the kernel with syzkaller.
> >> >
> >> > On commit a71c9a1c779f2499fb2afc0553e543f18aff6edf (4.11-rc5).
> >> >
> >> > A reproducer and .config are attached.
> >> The script is pretty hard to reproduce the issue in my env.
> >
> > I didn't try running it but I also found the reproducer very complicated
> > to follow. Do you have any plans on having some PoC optimizer, so we can
> > have a more readable code?
> > strace is handy for filtering the noise, yes, but sometimes it doesn't
> > cut it.
> I got the script now:
> 1. create sk
> 2. set sk->sndbuf = x
> 3. sendmsg with size s1 (s1 < x)
> 4. sendmsg with size s2 (s1+s2 > x)
> 5. sendmsg with size s3 (wspace < 0), wait sndbuf, schedule out.
> 6. listen sk (abnormal operation on sctp client)
> 7. accept sk.
> 
> In step 6, sk->sk_state = listening, then step 7 could get the first asoc
> from ep->asoc_list and alloc a new sk2, attach the asoc to sk2.
> 
> after a while, sendmsg schedule in, but asoc->sk is sk2, !=sk.
> the same issue we fix for peeloff on commit dfcb9f4f99f1 ("sctp: deny
> peeloff operation on asocs with threads sleeping on it") happens.

Yes. That explains why the asoc isn't dead by when sendmsg comes back,
and avoid that dead check.

> 
> But we should not fix it by the same way as for peeloff. the real reason
> causes this issue is on step 6, it should disallow listen on the established sk.

Agreed.

> 
> The following fix should work for this, just similar with what
> inet_listen() did.
> 
> @@ -7174,6 +7175,9 @@ int sctp_inet_listen(struct socket *sock, int backlog)
>         if (sock->state != SS_UNCONNECTED)
>                 goto out;
> 
> +       if (!sctp_sstate(sk, LISTENING) && !sctp_sstate(sk,CLOSED))
> +               goto out;
> +
> 
> what do you think ?

Yes, agreed.
Thanks!

  Marcelo