net/rds: fix possible double free on sock tear down

I got a report of a double free happening at RDS slab cache. One
suspicion was that may be somewhere we were doing a sock_hold/sock_put
on an already freed sock. Thus after providing a kernel with the
following change:

 static inline void sock_hold(struct sock *sk)
 {
-       atomic_inc(&sk->sk_refcnt);
+       if (!atomic_inc_not_zero(&sk->sk_refcnt))
+               WARN(1, "Trying to hold sock already gone: %p (family: %hd)\n",
+                       sk, sk->sk_family);
 }

The warning successfuly triggered:

Trying to hold sock already gone: ffff81f6dda61280 (family: 21)
WARNING: at include/net/sock.h:350 sock_hold()
Call Trace:
<IRQ>  [<ffffffff8adac135>] :rds:rds_send_remove_from_sock+0xf0/0x21b
[<ffffffff8adad35c>] :rds:rds_send_drop_acked+0xbf/0xcf
[<ffffffff8addf546>] :rds_rdma:rds_ib_recv_tasklet_fn+0x256/0x2dc
[<ffffffff8009899a>] tasklet_action+0x8f/0x12b
[<ffffffff800125a2>] __do_softirq+0x89/0x133
[<ffffffff8005f30c>] call_softirq+0x1c/0x28
[<ffffffff8006e644>] do_softirq+0x2c/0x7d
[<ffffffff8006e4d4>] do_IRQ+0xee/0xf7
[<ffffffff8005e625>] ret_from_intr+0x0/0xa
<EOI>

Looking at the call chain above, the only way I think this would be
possible is if somewhere we already released the same socket->sock which
is assigned to the rds_message at rds_send_remove_from_sock. Which seems
only possible to happen after the tear down done on rds_release.

rds_release properly calls rds_send_drop_to to drop the socket from any
rds_message, and some proper synchronization is in place to avoid race
with rds_send_drop_acked/rds_send_remove_from_sock. However, I still see
a very narrow window where it may be possible we touch a sock already
released: when rds_release races with rds_send_drop_acked, we check
RDS_MSG_ON_CONN to avoid cleanup on the same rds_message, but in this
specific case we don't clear rm->m_rs. In this case, it seems we could
then go on at rds_send_drop_to and after it returns, the sock is freed
by last sock_put on rds_release, with concurrently we being at
rds_send_remove_from_sock; then at some point in the loop at
rds_send_remove_from_sock we process an rds_message which didn't have
rm->m_rs unset for a freed sock, and a possible sock_hold on an sock
already gone at rds_release happens.

This hopefully address the described condition above and avoids a double
free on "second last" sock_put. In addition, I removed the comment about
socket destruction on top of rds_send_drop_acked: we call rds_send_drop_to
in rds_release and we should have things properly serialized there, thus
I can't see the comment being accurate there.

Signed-off-by: Herton R. Krzesinski <herton@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit is contained in:
Herton R. Krzesinski 2014-10-01 18:49:54 -03:00 committed by David S. Miller
parent eb74cc97b8
commit 593cbb3ec6

View File

@ -593,8 +593,11 @@ static void rds_send_remove_from_sock(struct list_head *messages, int status)
sock_put(rds_rs_to_sk(rs));
}
rs = rm->m_rs;
sock_hold(rds_rs_to_sk(rs));
if (rs)
sock_hold(rds_rs_to_sk(rs));
}
if (!rs)
goto unlock_and_drop;
spin_lock(&rs->rs_lock);
if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) {
@ -638,9 +641,6 @@ unlock_and_drop:
* queue. This means that in the TCP case, the message may not have been
* assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked
* checks the RDS_MSG_HAS_ACK_SEQ bit.
*
* XXX It's not clear to me how this is safely serialized with socket
* destruction. Maybe it should bail if it sees SOCK_DEAD.
*/
void rds_send_drop_acked(struct rds_connection *conn, u64 ack,
is_acked_func is_acked)
@ -711,6 +711,9 @@ void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
*/
if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) {
spin_unlock_irqrestore(&conn->c_lock, flags);
spin_lock_irqsave(&rm->m_rs_lock, flags);
rm->m_rs = NULL;
spin_unlock_irqrestore(&rm->m_rs_lock, flags);
continue;
}
list_del_init(&rm->m_conn_item);