can: bcm: delay release of struct bcm_op after synchronize_rcu()

commit d5f9023 upstream.

can_rx_register() callbacks may be called concurrently to the call to
can_rx_unregister(). The callbacks and callback data, though, are
protected by RCU and the struct sock reference count.

So the callback data is really attached to the life of sk, meaning
that it should be released on sk_destruct. However, bcm_remove_op()
calls tasklet_kill(), and RCU callbacks may be called under RCU
softirq, so that cannot be used on kernels before the introduction of
HRTIMER_MODE_SOFT.

However, bcm_rx_handler() is called under RCU protection, so after
calling can_rx_unregister(), we may call synchronize_rcu() in order to
wait for any RCU read-side critical sections to finish. That is,
bcm_rx_handler() won't be called anymore for those ops. So, we only
free them, after we do that synchronize_rcu().

Fixes: ffd980f ("[CAN]: Add broadcast manager (bcm) protocol")
Link: https://lore.kernel.org/r/20210619161813.2098382-1-cascardo@canonical.com
Cc: linux-stable <stable@vger.kernel.org>
Reported-by: syzbot+0f7e7e5e2f4f40fa89c0@syzkaller.appspotmail.com
Reported-by: Norbert Slusarek <nslusarek@gmx.net>
Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@canonical.com>
Acked-by: Oliver Hartkopp <socketcan@hartkopp.net>
Signed-off-by: Marc Kleine-Budde <mkl@pengutronix.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

net/can/bcm.c

@@ -785,6 +785,7 @@ static int bcm_delete_rx_op(struct list_head *ops, struct bcm_msg_head *mh,
 						  bcm_rx_handler, op);
 
 			list_del(&op->list);
+			synchronize_rcu();
 			bcm_remove_op(op);
 			return 1; /* done */
 		}
@@ -1533,9 +1534,13 @@ static int bcm_release(struct socket *sock)
 					  REGMASK(op->can_id),
 					  bcm_rx_handler, op);
 
-		bcm_remove_op(op);
 	}
 
+	synchronize_rcu();
+
+	list_for_each_entry_safe(op, next, &bo->rx_ops, list)
+		bcm_remove_op(op);
+
 #if IS_ENABLED(CONFIG_PROC_FS)
 	/* remove procfs entry */
 	if (net->can.bcmproc_dir && bo->bcm_proc_read)