Re: [PATCH net] sctp: add mutual exclusion in proc_sctp_do_udp_port()

[David Laight <david.laight.linux@gmail.com>]

On Mon, 31 Mar 2025 18:11:38 +0200
Eric Dumazet <edumazet@google.com> wrote:

> On Mon, Mar 31, 2025 at 5:54 PM Xin Long <lucien.xin@gmail.com> wrote:
> >
> > On Mon, Mar 31, 2025 at 5:15 AM Eric Dumazet <edumazet@google.com> wrote:  
> > >
> > > We must serialize calls to sctp_udp_sock_stop() and sctp_udp_sock_start()
> > > or risk a crash as syzbot reported:
> > >
> > > Oops: general protection fault, probably for non-canonical address 0xdffffc000000000d: 0000 [#1] SMP KASAN PTI
> > > KASAN: null-ptr-deref in range [0x0000000000000068-0x000000000000006f]
> > > CPU: 1 UID: 0 PID: 6551 Comm: syz.1.44 Not tainted 6.14.0-syzkaller-g7f2ff7b62617 #0 PREEMPT(full)
> > > Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 02/12/2025
> > >  RIP: 0010:kernel_sock_shutdown+0x47/0x70 net/socket.c:3653
> > > Call Trace:
> > >  <TASK>
> > >   udp_tunnel_sock_release+0x68/0x80 net/ipv4/udp_tunnel_core.c:181
> > >   sctp_udp_sock_stop+0x71/0x160 net/sctp/protocol.c:930
> > >   proc_sctp_do_udp_port+0x264/0x450 net/sctp/sysctl.c:553
> > >   proc_sys_call_handler+0x3d0/0x5b0 fs/proc/proc_sysctl.c:601
> > >   iter_file_splice_write+0x91c/0x1150 fs/splice.c:738
> > >   do_splice_from fs/splice.c:935 [inline]
> > >   direct_splice_actor+0x18f/0x6c0 fs/splice.c:1158
> > >   splice_direct_to_actor+0x342/0xa30 fs/splice.c:1102
> > >   do_splice_direct_actor fs/splice.c:1201 [inline]
> > >   do_splice_direct+0x174/0x240 fs/splice.c:1227
> > >   do_sendfile+0xafd/0xe50 fs/read_write.c:1368
> > >   __do_sys_sendfile64 fs/read_write.c:1429 [inline]
> > >   __se_sys_sendfile64 fs/read_write.c:1415 [inline]
> > >   __x64_sys_sendfile64+0x1d8/0x220 fs/read_write.c:1415
> > >   do_syscall_x64 arch/x86/entry/syscall_64.c:63 [inline]
> > >
> > > Fixes: 046c052b475e ("sctp: enable udp tunneling socks")
> > > Reported-by: syzbot+fae49d997eb56fa7c74d@syzkaller.appspotmail.com
> > > Closes: https://lore.kernel.org/netdev/67ea5c01.050a0220.1547ec.012b.GAE@google.com/T/#u
> > > Signed-off-by: Eric Dumazet <edumazet@google.com>
> > > Cc: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
> > > Cc: Xin Long <lucien.xin@gmail.com>
> > > ---
> > >  net/sctp/sysctl.c | 4 ++++
> > >  1 file changed, 4 insertions(+)
> > >
> > > diff --git a/net/sctp/sysctl.c b/net/sctp/sysctl.c
> > > index 8e1e97be4df79f3245e2bbbeb0a75841abc67f58..ee3eac338a9deef064f273e29bb59b057835d3f1 100644
> > > --- a/net/sctp/sysctl.c
> > > +++ b/net/sctp/sysctl.c
> > > @@ -525,6 +525,8 @@ static int proc_sctp_do_auth(const struct ctl_table *ctl, int write,
> > >         return ret;
> > >  }
> > >
> > > +static DEFINE_MUTEX(sctp_sysctl_mutex);
> > > +
> > >  static int proc_sctp_do_udp_port(const struct ctl_table *ctl, int write,
> > >                                  void *buffer, size_t *lenp, loff_t *ppos)
> > >  {
> > > @@ -549,6 +551,7 @@ static int proc_sctp_do_udp_port(const struct ctl_table *ctl, int write,
> > >                 if (new_value > max || new_value < min)
> > >                         return -EINVAL;
> > >
> > > +               mutex_lock(&sctp_sysctl_mutex);
> > >                 net->sctp.udp_port = new_value;
> > >                 sctp_udp_sock_stop(net);
> > >                 if (new_value) {
> > > @@ -561,6 +564,7 @@ static int proc_sctp_do_udp_port(const struct ctl_table *ctl, int write,
> > >                 lock_sock(sk);
> > >                 sctp_sk(sk)->udp_port = htons(net->sctp.udp_port);
> > >                 release_sock(sk);
> > > +               mutex_unlock(&sctp_sysctl_mutex);
> > >         }
> > >
> > >         return ret;
> > > --
> > > 2.49.0.472.ge94155a9ec-goog
> > >  
> > Instead of introducing a new lock for this, wouldn't be better to just
> > move up `lock_sock(sk)` a little bit?  
> 
> It depends if calling synchronize_rcu() two times while holding the
> socket lock is ok or not ?
> 
> What is the issue about using a separate mutex ?
> 

Don't they need locking against a different path that is using the socket?
Not only against concurrent accesses to the sysctl?

Presuming the crash was because of the net->sctp.udp4_sock = NULL
assignment in sock_stop(), if 'min' is zero allowing 'new_value' zero
then the pointer is left NULL.

IIRC sctp_sk(sk) is fixed, so the sock_lock() doesn't do much apart
from stop some unlikely 'data tearing'.

	David