netlink: fix netlink_change_ngroups()

commit 6c04bb18dd (netlink: use call_rcu for netlink_change_ngroups)
used a somewhat convoluted and racy way to perform call_rcu().

The old block of memory is freed after a grace period, but the rcu_head
used to track it is located in new block.

This can clash if we call two times or more netlink_change_ngroups(),
and a block is freed before another. call_rcu() called on different cpus
makes no guarantee in order of callbacks.

Fix this using a more standard way of handling this : Each block of
memory contains its own rcu_head, so that no 'use after free' can
happens.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Johannes Berg <johannes@sipsolutions.net>
CC: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/netlink/af_netlink.c

@@ -83,9 +83,9 @@ struct netlink_sock {
 	struct module		*module;
 };
 
-struct listeners_rcu_head {
+struct listeners {
-	struct rcu_head rcu_head;
+	struct rcu_head		rcu;
-	void *ptr;
+	unsigned long		masks[0];
 };
 
 #define NETLINK_KERNEL_SOCKET	0x1
@@ -119,7 +119,7 @@ struct nl_pid_hash {
 struct netlink_table {
 	struct nl_pid_hash hash;
 	struct hlist_head mc_list;
-	unsigned long *listeners;
+	struct listeners __rcu *listeners;
 	unsigned int nl_nonroot;
 	unsigned int groups;
 	struct mutex *cb_mutex;
@@ -338,7 +338,7 @@ netlink_update_listeners(struct sock *sk)
 			if (i < NLGRPLONGS(nlk_sk(sk)->ngroups))
 				mask |= nlk_sk(sk)->groups[i];
 		}
-		tbl->listeners[i] = mask;
+		tbl->listeners->masks[i] = mask;
 	}
 	/* this function is only called with the netlink table "grabbed", which
 	 * makes sure updates are visible before bind or setsockopt return. */
@@ -936,7 +936,7 @@ EXPORT_SYMBOL(netlink_unicast);
 int netlink_has_listeners(struct sock *sk, unsigned int group)
 {
 	int res = 0;
-	unsigned long *listeners;
+	struct listeners *listeners;
 
 	BUG_ON(!netlink_is_kernel(sk));
 
@@ -944,7 +944,7 @@ int netlink_has_listeners(struct sock *sk, unsigned int group)
 	listeners = rcu_dereference(nl_table[sk->sk_protocol].listeners);
 
 	if (group - 1 < nl_table[sk->sk_protocol].groups)
-		res = test_bit(group - 1, listeners);
+		res = test_bit(group - 1, listeners->masks);
 
 	rcu_read_unlock();
 
@@ -1498,7 +1498,7 @@ netlink_kernel_create(struct net *net, int unit, unsigned int groups,
 	struct socket *sock;
 	struct sock *sk;
 	struct netlink_sock *nlk;
-	unsigned long *listeners = NULL;
+	struct listeners *listeners = NULL;
 
 	BUG_ON(!nl_table);
 
@@ -1523,8 +1523,7 @@ netlink_kernel_create(struct net *net, int unit, unsigned int groups,
 	if (groups < 32)
 		groups = 32;
 
-	listeners = kzalloc(NLGRPSZ(groups) + sizeof(struct listeners_rcu_head),
+	listeners = kzalloc(sizeof(*listeners) + NLGRPSZ(groups), GFP_KERNEL);
-			    GFP_KERNEL);
 	if (!listeners)
 		goto out_sock_release;
 
@@ -1541,7 +1540,7 @@ netlink_kernel_create(struct net *net, int unit, unsigned int groups,
 	netlink_table_grab();
 	if (!nl_table[unit].registered) {
 		nl_table[unit].groups = groups;
-		nl_table[unit].listeners = listeners;
+		rcu_assign_pointer(nl_table[unit].listeners, listeners);
 		nl_table[unit].cb_mutex = cb_mutex;
 		nl_table[unit].module = module;
 		nl_table[unit].registered = 1;
@@ -1572,43 +1571,28 @@ netlink_kernel_release(struct sock *sk)
 EXPORT_SYMBOL(netlink_kernel_release);
 
 
-static void netlink_free_old_listeners(struct rcu_head *rcu_head)
+static void listeners_free_rcu(struct rcu_head *head)
 {
-	struct listeners_rcu_head *lrh;
+	kfree(container_of(head, struct listeners, rcu));
-
-	lrh = container_of(rcu_head, struct listeners_rcu_head, rcu_head);
-	kfree(lrh->ptr);
 }
 
 int __netlink_change_ngroups(struct sock *sk, unsigned int groups)
 {
-	unsigned long *listeners, *old = NULL;
+	struct listeners *new, *old;
-	struct listeners_rcu_head *old_rcu_head;
 	struct netlink_table *tbl = &nl_table[sk->sk_protocol];
 
 	if (groups < 32)
 		groups = 32;
 
 	if (NLGRPSZ(tbl->groups) < NLGRPSZ(groups)) {
-		listeners = kzalloc(NLGRPSZ(groups) +
+		new = kzalloc(sizeof(*new) + NLGRPSZ(groups), GFP_ATOMIC);
-				    sizeof(struct listeners_rcu_head),
+		if (!new)
-				    GFP_ATOMIC);
-		if (!listeners)
 			return -ENOMEM;
-		old = tbl->listeners;
+		old = rcu_dereference_raw(tbl->listeners);
-		memcpy(listeners, old, NLGRPSZ(tbl->groups));
+		memcpy(new->masks, old->masks, NLGRPSZ(tbl->groups));
-		rcu_assign_pointer(tbl->listeners, listeners);
+		rcu_assign_pointer(tbl->listeners, new);
-		/*
+
-		 * Free the old memory after an RCU grace period so we
+		call_rcu(&old->rcu, listeners_free_rcu);
-		 * don't leak it. We use call_rcu() here in order to be
-		 * able to call this function from atomic contexts. The
-		 * allocation of this memory will have reserved enough
-		 * space for struct listeners_rcu_head at the end.
-		 */
-		old_rcu_head = (void *)(tbl->listeners +
-					NLGRPLONGS(tbl->groups));
-		old_rcu_head->ptr = old;
-		call_rcu(&old_rcu_head->rcu_head, netlink_free_old_listeners);
 	}
 	tbl->groups = groups;
 
@@ -2104,18 +2088,17 @@ static void __net_exit netlink_net_exit(struct net *net)
 
 static void __init netlink_add_usersock_entry(void)
 {
-	unsigned long *listeners;
+	struct listeners *listeners;
 	int groups = 32;
 
-	listeners = kzalloc(NLGRPSZ(groups) + sizeof(struct listeners_rcu_head),
+	listeners = kzalloc(sizeof(*listeners) + NLGRPSZ(groups), GFP_KERNEL);
-			    GFP_KERNEL);
 	if (!listeners)
-		panic("netlink_add_usersock_entry: Cannot allocate listneres\n");
+		panic("netlink_add_usersock_entry: Cannot allocate listeners\n");
 
 	netlink_table_grab();
 
 	nl_table[NETLINK_USERSOCK].groups = groups;
-	nl_table[NETLINK_USERSOCK].listeners = listeners;
+	rcu_assign_pointer(nl_table[NETLINK_USERSOCK].listeners, listeners);
 	nl_table[NETLINK_USERSOCK].module = THIS_MODULE;
 	nl_table[NETLINK_USERSOCK].registered = 1;