Re: [PATCH net] sctp: set newsk sk_socket before processing listening sk backlog

[Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>]

On Wed, Oct 02, 2019 at 04:23:52PM +0800, Xin Long wrote:
> On Wed, Oct 2, 2019 at 9:04 AM Marcelo Ricardo Leitner
> <marcelo.leitner@gmail.com> wrote:
> >
> > On Mon, Sep 30, 2019 at 09:10:18PM +0800, Xin Long wrote:
> > > This patch is to fix a NULL-ptr deref crash in selinux_sctp_bind_connect:
> > >
> > >   [...] kasan: GPF could be caused by NULL-ptr deref or user memory access
> > >   [...] RIP: 0010:selinux_sctp_bind_connect+0x16a/0x230
> > >   [...] Call Trace:
> > >   [...]  security_sctp_bind_connect+0x58/0x90
> > >   [...]  sctp_process_asconf+0xa52/0xfd0 [sctp]
> > >   [...]  sctp_sf_do_asconf+0x782/0x980 [sctp]
> > >   [...]  sctp_do_sm+0x139/0x520 [sctp]
> > >   [...]  sctp_assoc_bh_rcv+0x284/0x5c0 [sctp]
> > >   [...]  sctp_backlog_rcv+0x45f/0x880 [sctp]
> > >   [...]  __release_sock+0x120/0x370
> > >   [...]  release_sock+0x4f/0x180
> > >   [...]  sctp_accept+0x3f9/0x5a0 [sctp]
> > >   [...]  inet_accept+0xe7/0x6f0
> > >
> > > It was caused by that the 'newsk' sk_socket was not set before going to
> > > security sctp hook when doing accept() on a tcp-type socket:
> > >
> > >   inet_accept()->
> > >     sctp_accept():
> > >       lock_sock():
> > >           lock listening 'sk'
> > >                                           do_softirq():
> > >                                             sctp_rcv():  <-- [1]
> > >                                                 asconf chunk arrived and
> > >                                                 enqueued in 'sk' backlog
> > >       sctp_sock_migrate():
> > >           set asoc's sk to 'newsk'
> > >       release_sock():
> > >           sctp_backlog_rcv():
> > >             lock 'newsk'
> > >             sctp_process_asconf()  <-- [2]
> > >             unlock 'newsk'
> > >     sock_graft():
> > >         set sk_socket  <-- [3]
> > >
> > > As it shows, at [1] the asconf chunk would be put into the listening 'sk'
> > > backlog, as accept() was holding its sock lock. Then at [2] asconf would
> > > get processed with 'newsk' as asoc's sk had been set to 'newsk'. However,
> > > 'newsk' sk_socket is not set until [3], while selinux_sctp_bind_connect()
> > > would deref it, then kernel crashed.
> >
> > Note that sctp will migrate such incoming chunks from sk to newsk in
> > sctp_rcv() if they arrived after the mass-migration performed at
> > sctp_sock_migrate().
> >
> > That said, did you explore changing inet_accept() so that
> > sk1->sk_prot->accept() would return sk2 still/already locked?
> > That would be enough to block [2] from happening as then it would be
> > queued on newsk backlog this time and avoid nearly duplicating
> > inet_accept(). (too bad for this chunk, hit 2 backlogs..)
> We don't have to bother inet_accept() for it. I had this one below,
> and I was just thinking the locks order doesn't look nice. Do you
> think this is more acceptable?
> 
> @@ -4963,15 +4963,19 @@ static struct sock *sctp_accept(struct sock
> *sk, int flags, int *err, bool kern)
>          * asoc to the newsk.
>          */
>         error = sctp_sock_migrate(sk, newsk, asoc, SCTP_SOCKET_TCP);
> -       if (error) {
> -               sk_common_release(newsk);
> -               newsk = NULL;
> +       if (!error) {
> +               lock_sock_nested(newsk, SINGLE_DEPTH_NESTING);
> +               release_sock(sk);

Interesting. It fixes the backlog processing, ok. Question:

> +               release_sock(newsk);

As newsk is hashed already and unlocked here to be locked again later
on inet_accept(), it could receive a packet in between (thus before
sock_graft() could have a chance to run), no?

> +               *err = error;
> +
> +               return newsk;
>         }
> 
>  out:
>         release_sock(sk);
>         *err = error;
> -       return newsk;
> +       return NULL;
>  }
> 
> >
> > AFAICT TCP code would be fine with such change. Didn't check other
> > protocols.
> >
>