sctp: Fix a race between ICMP protocol unreachable and connect()
authorVlad Yasevich <vladislav.yasevich@hp.com>
Thu, 6 May 2010 07:56:07 +0000 (00:56 -0700)
committerDavid S. Miller <davem@davemloft.net>
Thu, 6 May 2010 07:56:07 +0000 (00:56 -0700)
ICMP protocol unreachable handling completely disregarded
the fact that the user may have locked the socket.  It proceeded
to destroy the association, even though the user may have
held the lock and had a ref on the association.  This resulted
in the following:

Attempt to release alive inet socket f6afcc00

=========================
[ BUG: held lock freed! ]
-------------------------
somenu/2672 is freeing memory f6afcc00-f6afcfff, with a lock still held
there!
 (sk_lock-AF_INET){+.+.+.}, at: [<c122098a>] sctp_connect+0x13/0x4c
1 lock held by somenu/2672:
 #0:  (sk_lock-AF_INET){+.+.+.}, at: [<c122098a>] sctp_connect+0x13/0x4c

stack backtrace:
Pid: 2672, comm: somenu Not tainted 2.6.32-telco #55
Call Trace:
 [<c1232266>] ? printk+0xf/0x11
 [<c1038553>] debug_check_no_locks_freed+0xce/0xff
 [<c10620b4>] kmem_cache_free+0x21/0x66
 [<c1185f25>] __sk_free+0x9d/0xab
 [<c1185f9c>] sk_free+0x1c/0x1e
 [<c1216e38>] sctp_association_put+0x32/0x89
 [<c1220865>] __sctp_connect+0x36d/0x3f4
 [<c122098a>] ? sctp_connect+0x13/0x4c
 [<c102d073>] ? autoremove_wake_function+0x0/0x33
 [<c12209a8>] sctp_connect+0x31/0x4c
 [<c11d1e80>] inet_dgram_connect+0x4b/0x55
 [<c11834fa>] sys_connect+0x54/0x71
 [<c103a3a2>] ? lock_release_non_nested+0x88/0x239
 [<c1054026>] ? might_fault+0x42/0x7c
 [<c1054026>] ? might_fault+0x42/0x7c
 [<c11847ab>] sys_socketcall+0x6d/0x178
 [<c10da994>] ? trace_hardirqs_on_thunk+0xc/0x10
 [<c1002959>] syscall_call+0x7/0xb

This was because the sctp_wait_for_connect() would aqcure the socket
lock and then proceed to release the last reference count on the
association, thus cause the fully destruction path to finish freeing
the socket.

The simplest solution is to start a very short timer in case the socket
is owned by user.  When the timer expires, we can do some verification
and be able to do the release properly.

Signed-off-by: Vlad Yasevich <vladislav.yasevich@hp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/sctp/sm.h
include/net/sctp/structs.h
net/sctp/input.c
net/sctp/sm_sideeffect.c
net/sctp/transport.c

index 851c813adb3ac2a144cbb0b98582331629a5c385..61d73e37d5436138d743372a48aecad3191526fe 100644 (file)
@@ -279,6 +279,7 @@ int sctp_do_sm(sctp_event_t event_type, sctp_subtype_t subtype,
 /* 2nd level prototypes */
 void sctp_generate_t3_rtx_event(unsigned long peer);
 void sctp_generate_heartbeat_event(unsigned long peer);
+void sctp_generate_proto_unreach_event(unsigned long peer);
 
 void sctp_ootb_pkt_free(struct sctp_packet *);
 
index 597f8e27aaf6dfccffa354ae4219c164a1bd078e..219043a67bf7ab6073eba985e80ebb8ca206f2c1 100644 (file)
@@ -1010,6 +1010,9 @@ struct sctp_transport {
        /* Heartbeat timer is per destination. */
        struct timer_list hb_timer;
 
+       /* Timer to handle ICMP proto unreachable envets */
+       struct timer_list proto_unreach_timer;
+
        /* Since we're using per-destination retransmission timers
         * (see above), we're also using per-destination "transmitted"
         * queues.  This probably ought to be a private struct
index 2a570184e5a9176c07291eb7ab6f8771c2dd95c0..ea2192444ce66413261d20ce8ec236279496d141 100644 (file)
@@ -440,11 +440,25 @@ void sctp_icmp_proto_unreachable(struct sock *sk,
 {
        SCTP_DEBUG_PRINTK("%s\n",  __func__);
 
-       sctp_do_sm(SCTP_EVENT_T_OTHER,
-                  SCTP_ST_OTHER(SCTP_EVENT_ICMP_PROTO_UNREACH),
-                  asoc->state, asoc->ep, asoc, t,
-                  GFP_ATOMIC);
+       if (sock_owned_by_user(sk)) {
+               if (timer_pending(&t->proto_unreach_timer))
+                       return;
+               else {
+                       if (!mod_timer(&t->proto_unreach_timer,
+                                               jiffies + (HZ/20)))
+                               sctp_association_hold(asoc);
+               }
+                       
+       } else {
+               if (timer_pending(&t->proto_unreach_timer) &&
+                   del_timer(&t->proto_unreach_timer))
+                       sctp_association_put(asoc);
 
+               sctp_do_sm(SCTP_EVENT_T_OTHER,
+                          SCTP_ST_OTHER(SCTP_EVENT_ICMP_PROTO_UNREACH),
+                          asoc->state, asoc->ep, asoc, t,
+                          GFP_ATOMIC);
+       }
 }
 
 /* Common lookup code for icmp/icmpv6 error handler. */
index d5ae450b6f02de3fa5bbe9e9ca58028c2bfe0576..eb1f42f45fdde604512f1eefcced4041843e51f8 100644 (file)
@@ -397,6 +397,41 @@ out_unlock:
        sctp_transport_put(transport);
 }
 
+/* Handle the timeout of the ICMP protocol unreachable timer.  Trigger
+ * the correct state machine transition that will close the association.
+ */
+void sctp_generate_proto_unreach_event(unsigned long data)
+{
+       struct sctp_transport *transport = (struct sctp_transport *) data;
+       struct sctp_association *asoc = transport->asoc;
+       
+       sctp_bh_lock_sock(asoc->base.sk);
+       if (sock_owned_by_user(asoc->base.sk)) {
+               SCTP_DEBUG_PRINTK("%s:Sock is busy.\n", __func__);
+
+               /* Try again later.  */
+               if (!mod_timer(&transport->proto_unreach_timer,
+                               jiffies + (HZ/20)))
+                       sctp_association_hold(asoc);
+               goto out_unlock;
+       }
+
+       /* Is this structure just waiting around for us to actually
+        * get destroyed?
+        */
+       if (asoc->base.dead)
+               goto out_unlock;
+
+       sctp_do_sm(SCTP_EVENT_T_OTHER,
+                  SCTP_ST_OTHER(SCTP_EVENT_ICMP_PROTO_UNREACH),
+                  asoc->state, asoc->ep, asoc, transport, GFP_ATOMIC);
+
+out_unlock:
+       sctp_bh_unlock_sock(asoc->base.sk);
+       sctp_association_put(asoc);
+}
+
+
 /* Inject a SACK Timeout event into the state machine.  */
 static void sctp_generate_sack_event(unsigned long data)
 {
index be4d63d5a5cc050d782a2355a3eb1b9ae61bda07..4a368038d46f77eb4370d6a65f636fa424470bf6 100644 (file)
@@ -108,6 +108,8 @@ static struct sctp_transport *sctp_transport_init(struct sctp_transport *peer,
                        (unsigned long)peer);
        setup_timer(&peer->hb_timer, sctp_generate_heartbeat_event,
                        (unsigned long)peer);
+       setup_timer(&peer->proto_unreach_timer,
+                   sctp_generate_proto_unreach_event, (unsigned long)peer);
 
        /* Initialize the 64-bit random nonce sent with heartbeat. */
        get_random_bytes(&peer->hb_nonce, sizeof(peer->hb_nonce));