inetpeer: fix race in unused_list manipulations
authorEric Dumazet <eric.dumazet@gmail.com>
Thu, 26 May 2011 17:27:11 +0000 (17:27 +0000)
committerDavid S. Miller <davem@davemloft.net>
Fri, 27 May 2011 17:39:11 +0000 (13:39 -0400)
Several crashes in cleanup_once() were reported in recent kernels.

Commit d6cc1d642de9 (inetpeer: various changes) added a race in
unlink_from_unused().

One way to avoid taking unused_peers.lock before doing the list_empty()
test is to catch 0->1 refcnt transitions, using full barrier atomic
operations variants (atomic_cmpxchg() and atomic_inc_return()) instead
of previous atomic_inc() and atomic_add_unless() variants.

We then call unlink_from_unused() only for the owner of the 0->1
transition.

Add a new atomic_add_unless_return() static helper

With help from Arun Sharma.

Refs: https://bugzilla.kernel.org/show_bug.cgi?id=32772

Reported-by: Arun Sharma <asharma@fb.com>
Reported-by: Maximilian Engelhardt <maxi@daemonizer.de>
Reported-by: Yann Dupont <Yann.Dupont@univ-nantes.fr>
Reported-by: Denys Fedoryshchenko <denys@visp.net.lb>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/inetpeer.c

index 9df4e635fb5fe5d4a0c92526ad68b845094aa39d..ce616d92cc544dd0a8e3ffb0974aff825a5939e9 100644 (file)
@@ -154,11 +154,9 @@ void __init inet_initpeers(void)
 /* Called with or without local BH being disabled. */
 static void unlink_from_unused(struct inet_peer *p)
 {
-       if (!list_empty(&p->unused)) {
-               spin_lock_bh(&unused_peers.lock);
-               list_del_init(&p->unused);
-               spin_unlock_bh(&unused_peers.lock);
-       }
+       spin_lock_bh(&unused_peers.lock);
+       list_del_init(&p->unused);
+       spin_unlock_bh(&unused_peers.lock);
 }
 
 static int addr_compare(const struct inetpeer_addr *a,
@@ -205,6 +203,20 @@ static int addr_compare(const struct inetpeer_addr *a,
        u;                                                      \
 })
 
+static bool atomic_add_unless_return(atomic_t *ptr, int a, int u, int *newv)
+{
+       int cur, old = atomic_read(ptr);
+
+       while (old != u) {
+               *newv = old + a;
+               cur = atomic_cmpxchg(ptr, old, *newv);
+               if (cur == old)
+                       return true;
+               old = cur;
+       }
+       return false;
+}
+
 /*
  * Called with rcu_read_lock()
  * Because we hold no lock against a writer, its quite possible we fall
@@ -213,7 +225,8 @@ static int addr_compare(const struct inetpeer_addr *a,
  * We exit from this function if number of links exceeds PEER_MAXDEPTH
  */
 static struct inet_peer *lookup_rcu(const struct inetpeer_addr *daddr,
-                                   struct inet_peer_base *base)
+                                   struct inet_peer_base *base,
+                                   int *newrefcnt)
 {
        struct inet_peer *u = rcu_dereference(base->root);
        int count = 0;
@@ -226,7 +239,7 @@ static struct inet_peer *lookup_rcu(const struct inetpeer_addr *daddr,
                         * distinction between an unused entry (refcnt=0) and
                         * a freed one.
                         */
-                       if (unlikely(!atomic_add_unless(&u->refcnt, 1, -1)))
+                       if (!atomic_add_unless_return(&u->refcnt, 1, -1, newrefcnt))
                                u = NULL;
                        return u;
                }
@@ -465,22 +478,23 @@ struct inet_peer *inet_getpeer(struct inetpeer_addr *daddr, int create)
        struct inet_peer_base *base = family_to_base(daddr->family);
        struct inet_peer *p;
        unsigned int sequence;
-       int invalidated;
+       int invalidated, newrefcnt = 0;
 
        /* Look up for the address quickly, lockless.
         * Because of a concurrent writer, we might not find an existing entry.
         */
        rcu_read_lock();
        sequence = read_seqbegin(&base->lock);
-       p = lookup_rcu(daddr, base);
+       p = lookup_rcu(daddr, base, &newrefcnt);
        invalidated = read_seqretry(&base->lock, sequence);
        rcu_read_unlock();
 
        if (p) {
-               /* The existing node has been found.
+found:         /* The existing node has been found.
                 * Remove the entry from unused list if it was there.
                 */
-               unlink_from_unused(p);
+               if (newrefcnt == 1)
+                       unlink_from_unused(p);
                return p;
        }
 
@@ -494,11 +508,9 @@ struct inet_peer *inet_getpeer(struct inetpeer_addr *daddr, int create)
        write_seqlock_bh(&base->lock);
        p = lookup(daddr, stack, base);
        if (p != peer_avl_empty) {
-               atomic_inc(&p->refcnt);
+               newrefcnt = atomic_inc_return(&p->refcnt);
                write_sequnlock_bh(&base->lock);
-               /* Remove the entry from unused list if it was there. */
-               unlink_from_unused(p);
-               return p;
+               goto found;
        }
        p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL;
        if (p) {