[PATCH] rxrpc: fix bad unlock balance in rxrpc_do_sendmsg

[Khalid Masum]

On Sun, Aug 21, 2022 at 6:58 PM Hawkins Jiawei <yin31149 at gmail.com> wrote:
>
> Syzkaller reports bad unlock balance bug as follows:
> ------------[ cut here ]------------
> WARNING: bad unlock balance detected!
> syz-executor.0/4094 is trying to release lock (&call->user_mutex) at:
> [<ffffffff87c1d8d1>] rxrpc_do_sendmsg+0x851/0x1110 net/rxrpc/sendmsg.c:754
> but there are no more locks to release!
>
> other info that might help us debug this:
> no locks held by syz-executor.0/4094.
>
> stack backtrace:
> [...]
> Call Trace:
>  <TASK>
>  __dump_stack lib/dump_stack.c:88 [inline]
>  dump_stack_lvl+0x57/0x7d lib/dump_stack.c:106
>  print_unlock_imbalance_bug include/trace/events/lock.h:69 [inline]
>  __lock_release kernel/locking/lockdep.c:5333 [inline]
>  lock_release.cold+0x49/0x4e kernel/locking/lockdep.c:5686
>  __mutex_unlock_slowpath+0x99/0x5e0 kernel/locking/mutex.c:907
>  rxrpc_do_sendmsg+0x851/0x1110 net/rxrpc/sendmsg.c:754
>  sock_sendmsg_nosec net/socket.c:714 [inline]
>  sock_sendmsg+0xab/0xe0 net/socket.c:734
>  ____sys_sendmsg+0x5c2/0x7a0 net/socket.c:2485
>  ___sys_sendmsg+0xdb/0x160 net/socket.c:2539
>  __sys_sendmsg+0xc3/0x160 net/socket.c:2568
>  do_syscall_x64 arch/x86/entry/common.c:50 [inline]
>  do_syscall_64+0x35/0xb0 arch/x86/entry/common.c:80
>  entry_SYSCALL_64_after_hwframe+0x63/0xcd
>  [...]
>  </TASK>
> ------------------------------------
>
> When kernel wants to send a message through an RxRPC socket in
> rxrpc_do_sendmsg(), kernel should hold the call->user_mutex lock,
> or it will triggers bug when releasing this lock before returning
> from rxrpc_do_sendmsg().
>
> Yet the problem is that during rxrpc_do_sendmsg(), kernel may call
> rxrpc_wait_for_tx_window_intr() to wait for space to appear in the
> tx queue or a signal to occur. When kernel fails the
> mutex_lock_interruptible(), kernel will returns from the
> rxrpc_wait_for_tx_window_intr() without acquiring the mutex lock, then
> triggers bug when releasing the mutex lock in rxrpc_do_sendmsg().
>
> This patch solves it by acquiring the call->user_mutex lock, when
> kernel fails the mutex_lock_interruptible() before returning from
> the rxrpc_wait_for_tx_window_intr().
>
> Reported-and-tested-by: syzbot+7f0483225d0c94cb3441 at syzkaller.appspotmail.com
> Signed-off-by: Hawkins Jiawei <yin31149 at gmail.com>
> ---
>  net/rxrpc/sendmsg.c | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/net/rxrpc/sendmsg.c b/net/rxrpc/sendmsg.c
> index 1d38e279e2ef..e13043d357d5 100644
> --- a/net/rxrpc/sendmsg.c
> +++ b/net/rxrpc/sendmsg.c
> @@ -53,8 +53,10 @@ static int rxrpc_wait_for_tx_window_intr(struct rxrpc_sock *rx,
>                 trace_rxrpc_transmit(call, rxrpc_transmit_wait);
>                 mutex_unlock(&call->user_mutex);
>                 *timeo = schedule_timeout(*timeo);
> -               if (mutex_lock_interruptible(&call->user_mutex) < 0)
> +               if (mutex_lock_interruptible(&call->user_mutex) < 0) {
> +                       mutex_lock(&call->user_mutex);

The interruptible version fails to acquire the lock. So why is it okay to
force it to acquire the mutex_lock since we are in the interrupt context?
>                         return sock_intr_errno(*timeo);
> +               }
>         }
>  }

thanks,
  -- Khalid Masum