[NETLINK]: Synchronous message processing.
authorHerbert Xu <herbert@gondor.apana.org.au>
Tue, 3 May 2005 21:55:09 +0000 (14:55 -0700)
committerDavid S. Miller <davem@davemloft.net>
Tue, 3 May 2005 21:55:09 +0000 (14:55 -0700)
Let's recap the problem.  The current asynchronous netlink kernel
message processing is vulnerable to these attacks:

1) Hit and run: Attacker sends one or more messages and then exits
before they're processed.  This may confuse/disable the next netlink
user that gets the netlink address of the attacker since it may
receive the responses to the attacker's messages.

Proposed solutions:

a) Synchronous processing.
b) Stream mode socket.
c) Restrict/prohibit binding.

2) Starvation: Because various netlink rcv functions were written
to not return until all messages have been processed on a socket,
it is possible for these functions to execute for an arbitrarily
long period of time.  If this is successfully exploited it could
also be used to hold rtnl forever.

Proposed solutions:

a) Synchronous processing.
b) Stream mode socket.

Firstly let's cross off solution c).  It only solves the first
problem and it has user-visible impacts.  In particular, it'll
break user space applications that expect to bind or communicate
with specific netlink addresses (pid's).

So we're left with a choice of synchronous processing versus
SOCK_STREAM for netlink.

For the moment I'm sticking with the synchronous approach as
suggested by Alexey since it's simpler and I'd rather spend
my time working on other things.

However, it does have a number of deficiencies compared to the
stream mode solution:

1) User-space to user-space netlink communication is still vulnerable.

2) Inefficient use of resources.  This is especially true for rtnetlink
since the lock is shared with other users such as networking drivers.
The latter could hold the rtnl while communicating with hardware which
causes the rtnetlink user to wait when it could be doing other things.

3) It is still possible to DoS all netlink users by flooding the kernel
netlink receive queue.  The attacker simply fills the receive socket
with a single netlink message that fills up the entire queue.  The
attacker then continues to call sendmsg with the same message in a loop.

Point 3) can be countered by retransmissions in user-space code, however
it is pretty messy.

In light of these problems (in particular, point 3), we should implement
stream mode netlink at some point.  In the mean time, here is a patch
that implements synchronous processing.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
kernel/audit.c
net/core/rtnetlink.c
net/decnet/netfilter/dn_rtmsg.c
net/ipv4/netfilter/ip_queue.c
net/ipv4/tcp_diag.c
net/ipv6/netfilter/ip6_queue.c
net/xfrm/xfrm_user.c

index 0f84dd7af2c8d016a6992cf9388289399309b57f..ac26d4d960d3366d0473ede3533ab89446b75dc3 100644 (file)
@@ -427,7 +427,7 @@ static int audit_receive_msg(struct sk_buff *skb, struct nlmsghdr *nlh)
 /* Get message from skb (based on rtnetlink_rcv_skb).  Each message is
  * processed by audit_receive_msg.  Malformed skbs with wrong length are
  * discarded silently.  */
-static int audit_receive_skb(struct sk_buff *skb)
+static void audit_receive_skb(struct sk_buff *skb)
 {
        int             err;
        struct nlmsghdr *nlh;
@@ -436,7 +436,7 @@ static int audit_receive_skb(struct sk_buff *skb)
        while (skb->len >= NLMSG_SPACE(0)) {
                nlh = (struct nlmsghdr *)skb->data;
                if (nlh->nlmsg_len < sizeof(*nlh) || skb->len < nlh->nlmsg_len)
-                       return 0;
+                       return;
                rlen = NLMSG_ALIGN(nlh->nlmsg_len);
                if (rlen > skb->len)
                        rlen = skb->len;
@@ -446,23 +446,20 @@ static int audit_receive_skb(struct sk_buff *skb)
                        netlink_ack(skb, nlh, 0);
                skb_pull(skb, rlen);
        }
-       return 0;
 }
 
 /* Receive messages from netlink socket. */
 static void audit_receive(struct sock *sk, int length)
 {
        struct sk_buff  *skb;
+       unsigned int qlen;
 
-       if (down_trylock(&audit_netlink_sem))
-               return;
+       down(&audit_netlink_sem);
 
-                               /* FIXME: this must not cause starvation */
-       while ((skb = skb_dequeue(&sk->sk_receive_queue))) {
-               if (audit_receive_skb(skb) && skb->len)
-                       skb_queue_head(&sk->sk_receive_queue, skb);
-               else
-                       kfree_skb(skb);
+       for (qlen = skb_queue_len(&sk->sk_receive_queue); qlen; qlen--) {
+               skb = skb_dequeue(&sk->sk_receive_queue);
+               audit_receive_skb(skb);
+               kfree_skb(skb);
        }
        up(&audit_netlink_sem);
 }
index 5fb70cfa10850ae3d6abed0ad765ea76b7e80b4b..6e1ab1e34b2ec70630009aa4b4fd2f4ebac8a35c 100644 (file)
@@ -609,26 +609,31 @@ static inline int rtnetlink_rcv_skb(struct sk_buff *skb)
 
 /*
  *  rtnetlink input queue processing routine:
- *     - try to acquire shared lock. If it is failed, defer processing.
+ *     - process as much as there was in the queue upon entry.
  *     - feed skbs to rtnetlink_rcv_skb, until it refuse a message,
- *       that will occur, when a dump started and/or acquisition of
- *       exclusive lock failed.
+ *       that will occur, when a dump started.
  */
 
 static void rtnetlink_rcv(struct sock *sk, int len)
 {
+       unsigned int qlen = skb_queue_len(&sk->sk_receive_queue);
+
        do {
                struct sk_buff *skb;
 
-               if (rtnl_shlock_nowait())
-                       return;
+               rtnl_lock();
+
+               if (qlen > skb_queue_len(&sk->sk_receive_queue))
+                       qlen = skb_queue_len(&sk->sk_receive_queue);
 
-               while ((skb = skb_dequeue(&sk->sk_receive_queue)) != NULL) {
+               while (qlen--) {
+                       skb = skb_dequeue(&sk->sk_receive_queue);
                        if (rtnetlink_rcv_skb(skb)) {
-                               if (skb->len)
+                               if (skb->len) {
                                        skb_queue_head(&sk->sk_receive_queue,
                                                       skb);
-                               else
+                                       qlen++;
+                               } else
                                        kfree_skb(skb);
                                break;
                        }
@@ -638,7 +643,7 @@ static void rtnetlink_rcv(struct sock *sk, int len)
                up(&rtnl_sem);
 
                netdev_run_todo();
-       } while (rtnl && rtnl->sk_receive_queue.qlen);
+       } while (qlen);
 }
 
 static struct rtnetlink_link link_rtnetlink_table[RTM_NR_MSGTYPES] =
index f86a6259fd12171b1acc512713fff788089cb0f9..101ddef9ba9aa2aad76b4c26cb88a98551c7dbc5 100644 (file)
@@ -119,8 +119,9 @@ static inline void dnrmg_receive_user_skb(struct sk_buff *skb)
 static void dnrmg_receive_user_sk(struct sock *sk, int len)
 {
        struct sk_buff *skb;
+       unsigned int qlen = skb_queue_len(&sk->sk_receive_queue);
 
-       while((skb = skb_dequeue(&sk->sk_receive_queue)) != NULL) {
+       while (qlen-- && (skb = skb_dequeue(&sk->sk_receive_queue))) {
                dnrmg_receive_user_skb(skb);
                kfree_skb(skb);
        }
index 9e40dffc204f3438e0dbabc1ed89be0031f13449..e5746b674413464f6919d3b8880c30dc18feea42 100644 (file)
@@ -546,20 +546,18 @@ ipq_rcv_skb(struct sk_buff *skb)
 static void
 ipq_rcv_sk(struct sock *sk, int len)
 {
-       do {
-               struct sk_buff *skb;
+       struct sk_buff *skb;
+       unsigned int qlen;
 
-               if (down_trylock(&ipqnl_sem))
-                       return;
+       down(&ipqnl_sem);
                        
-               while ((skb = skb_dequeue(&sk->sk_receive_queue)) != NULL) {
-                       ipq_rcv_skb(skb);
-                       kfree_skb(skb);
-               }
+       for (qlen = skb_queue_len(&sk->sk_receive_queue); qlen; qlen--) {
+               skb = skb_dequeue(&sk->sk_receive_queue);
+               ipq_rcv_skb(skb);
+               kfree_skb(skb);
+       }
                
-               up(&ipqnl_sem);
-
-       } while (ipqnl && ipqnl->sk_receive_queue.qlen);
+       up(&ipqnl_sem);
 }
 
 static int
index 313c1408da33acbe328f8651ada04e3aa1261b68..8faa8948f75c2d67db9643dc07b24af7abc609a5 100644 (file)
@@ -777,8 +777,9 @@ static inline void tcpdiag_rcv_skb(struct sk_buff *skb)
 static void tcpdiag_rcv(struct sock *sk, int len)
 {
        struct sk_buff *skb;
+       unsigned int qlen = skb_queue_len(&sk->sk_receive_queue);
 
-       while ((skb = skb_dequeue(&sk->sk_receive_queue)) != NULL) {
+       while (qlen-- && (skb = skb_dequeue(&sk->sk_receive_queue))) {
                tcpdiag_rcv_skb(skb);
                kfree_skb(skb);
        }
index c54830b895939ed78f099978841d48a0d1d4a457..750943e2d34eed262fe1ccc5d9f276d2ed2fc832 100644 (file)
@@ -549,20 +549,18 @@ ipq_rcv_skb(struct sk_buff *skb)
 static void
 ipq_rcv_sk(struct sock *sk, int len)
 {
-       do {
-               struct sk_buff *skb;
+       struct sk_buff *skb;
+       unsigned int qlen;
 
-               if (down_trylock(&ipqnl_sem))
-                       return;
+       down(&ipqnl_sem);
                        
-               while ((skb = skb_dequeue(&sk->sk_receive_queue)) != NULL) {
-                       ipq_rcv_skb(skb);
-                       kfree_skb(skb);
-               }
+       for (qlen = skb_queue_len(&sk->sk_receive_queue); qlen; qlen--) {
+               skb = skb_dequeue(&sk->sk_receive_queue);
+               ipq_rcv_skb(skb);
+               kfree_skb(skb);
+       }
                
-               up(&ipqnl_sem);
-
-       } while (ipqnl && ipqnl->sk_receive_queue.qlen);
+       up(&ipqnl_sem);
 }
 
 static int
index 52b5843937c582f15ffca93f84440ff73d4255d2..dab112f1dd8a8c10eec2a044eb75852e2a64335a 100644 (file)
@@ -1008,17 +1008,24 @@ static int xfrm_user_rcv_skb(struct sk_buff *skb)
 
 static void xfrm_netlink_rcv(struct sock *sk, int len)
 {
+       unsigned int qlen = skb_queue_len(&sk->sk_receive_queue);
+
        do {
                struct sk_buff *skb;
 
                down(&xfrm_cfg_sem);
 
-               while ((skb = skb_dequeue(&sk->sk_receive_queue)) != NULL) {
+               if (qlen > skb_queue_len(&sk->sk_receive_queue))
+                       qlen = skb_queue_len(&sk->sk_receive_queue);
+
+               while (qlen--) {
+                       skb = skb_dequeue(&sk->sk_receive_queue);
                        if (xfrm_user_rcv_skb(skb)) {
-                               if (skb->len)
+                               if (skb->len) {
                                        skb_queue_head(&sk->sk_receive_queue,
                                                       skb);
-                               else
+                                       qlen++;
+                               } else
                                        kfree_skb(skb);
                                break;
                        }
@@ -1027,7 +1034,7 @@ static void xfrm_netlink_rcv(struct sock *sk, int len)
 
                up(&xfrm_cfg_sem);
 
-       } while (xfrm_nl && xfrm_nl->sk_receive_queue.qlen);
+       } while (qlen);
 }
 
 static int build_expire(struct sk_buff *skb, struct xfrm_state *x, int hard)