[INET]: Fix inet_diag register vs rcv race
authorPavel Emelyanov <xemul@openvz.org>
Thu, 29 Nov 2007 13:08:14 +0000 (00:08 +1100)
committerHerbert Xu <herbert@gondor.apana.org.au>
Thu, 29 Nov 2007 13:08:14 +0000 (00:08 +1100)
The following race is possible when one cpu unregisters the handler
while other one is trying to receive a message and call this one:

CPU1:                                                 CPU2:
inet_diag_rcv()                                       inet_diag_unregister()
  mutex_lock(&inet_diag_mutex);
  netlink_rcv_skb(skb, &inet_diag_rcv_msg);
    if (inet_diag_table[nlh->nlmsg_type] ==
                               NULL) /* false handler is still registered */
    ...
    netlink_dump_start(idiagnl, skb, nlh,
                           inet_diag_dump, NULL);
           cb = kzalloc(sizeof(*cb), GFP_KERNEL);
                   /* sleep here freeing memory
                    * or preempt
                    * or sleep later on nlk->cb_mutex
                    */
                                                         spin_lock(&inet_diag_register_lock);
                                                         inet_diag_table[type] = NULL;
    ...                                                  spin_unlock(&inet_diag_register_lock);
                                                         synchronize_rcu();
                                                         /* CPU1 is sleeping - RCU quiescent
                                                          * state is passed
                                                          */
                                                         return;
    /* inet_diag_dump is finally called: */
    inet_diag_dump()
      handler = inet_diag_table[cb->nlh->nlmsg_type];
      BUG_ON(handler == NULL);
      /* OOPS! While we slept the unregister has set
       * handler to NULL :(
       */

Grep showed, that the register/unregister functions are called
from init/fini module callbacks for tcp_/dccp_diag, so it's OK
to use the inet_diag_mutex to synchronize manipulations with the
inet_diag_table and the access to it.

Besides, as Herbert pointed out, asynchronous dumps should hold
this mutex as well, and thus, we provide the mutex as cb_mutex one.

Signed-off-by: Pavel Emelyanov <xemul@openvz.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
net/ipv4/inet_diag.c

index b0170732b5e9e962ed1fca65cf930c3bf4cc98f0..6b3fffb554b6fa59f01abc5464934c28d6b38785 100644 (file)
@@ -853,8 +853,6 @@ static void inet_diag_rcv(struct sk_buff *skb)
        mutex_unlock(&inet_diag_mutex);
 }
 
-static DEFINE_SPINLOCK(inet_diag_register_lock);
-
 int inet_diag_register(const struct inet_diag_handler *h)
 {
        const __u16 type = h->idiag_type;
@@ -863,13 +861,13 @@ int inet_diag_register(const struct inet_diag_handler *h)
        if (type >= INET_DIAG_GETSOCK_MAX)
                goto out;
 
-       spin_lock(&inet_diag_register_lock);
+       mutex_lock(&inet_diag_mutex);
        err = -EEXIST;
        if (inet_diag_table[type] == NULL) {
                inet_diag_table[type] = h;
                err = 0;
        }
-       spin_unlock(&inet_diag_register_lock);
+       mutex_unlock(&inet_diag_mutex);
 out:
        return err;
 }
@@ -882,11 +880,9 @@ void inet_diag_unregister(const struct inet_diag_handler *h)
        if (type >= INET_DIAG_GETSOCK_MAX)
                return;
 
-       spin_lock(&inet_diag_register_lock);
+       mutex_lock(&inet_diag_mutex);
        inet_diag_table[type] = NULL;
-       spin_unlock(&inet_diag_register_lock);
-
-       synchronize_rcu();
+       mutex_unlock(&inet_diag_mutex);
 }
 EXPORT_SYMBOL_GPL(inet_diag_unregister);
 
@@ -901,7 +897,7 @@ static int __init inet_diag_init(void)
                goto out;
 
        idiagnl = netlink_kernel_create(&init_net, NETLINK_INET_DIAG, 0,
-                                       inet_diag_rcv, NULL, THIS_MODULE);
+                       inet_diag_rcv, &inet_diag_mutex, THIS_MODULE);
        if (idiagnl == NULL)
                goto out_free_table;
        err = 0;