RDS: fix rds-ping deadlock over TCP transport
authorsantosh.shilimkar@oracle.com <santosh.shilimkar@oracle.com>
Sat, 17 Oct 2015 02:13:21 +0000 (22:13 -0400)
committerDavid S. Miller <davem@davemloft.net>
Mon, 19 Oct 2015 05:45:55 +0000 (22:45 -0700)
Sowmini found hang with rds-ping while testing RDS over TCP. Its
a corner case and doesn't happen always. The issue is not reproducible
with IB transport. Its clear from below dump why we see it with RDS TCP.

 [<ffffffff8153b7e5>] do_tcp_setsockopt+0xb5/0x740
 [<ffffffff8153bec4>] tcp_setsockopt+0x24/0x30
 [<ffffffff814d57d4>] sock_common_setsockopt+0x14/0x20
 [<ffffffffa096071d>] rds_tcp_xmit_prepare+0x5d/0x70 [rds_tcp]
 [<ffffffffa093b5f7>] rds_send_xmit+0xd7/0x740 [rds]
 [<ffffffffa093bda2>] rds_send_pong+0x142/0x180 [rds]
 [<ffffffffa0939d34>] rds_recv_incoming+0x274/0x330 [rds]
 [<ffffffff810815ae>] ? ttwu_queue+0x11e/0x130
 [<ffffffff814dcacd>] ? skb_copy_bits+0x6d/0x2c0
 [<ffffffffa0960350>] rds_tcp_data_recv+0x2f0/0x3d0 [rds_tcp]
 [<ffffffff8153d836>] tcp_read_sock+0x96/0x1c0
 [<ffffffffa0960060>] ? rds_tcp_recv_init+0x40/0x40 [rds_tcp]
 [<ffffffff814d6a90>] ? sock_def_write_space+0xa0/0xa0
 [<ffffffffa09604d1>] rds_tcp_data_ready+0xa1/0xf0 [rds_tcp]
 [<ffffffff81545249>] tcp_data_queue+0x379/0x5b0
 [<ffffffffa0960cdb>] ? rds_tcp_write_space+0xbb/0x110 [rds_tcp]
 [<ffffffff81547fd2>] tcp_rcv_established+0x2e2/0x6e0
 [<ffffffff81552602>] tcp_v4_do_rcv+0x122/0x220
 [<ffffffff81553627>] tcp_v4_rcv+0x867/0x880
 [<ffffffff8152e0b3>] ip_local_deliver_finish+0xa3/0x220

This happens because rds_send_xmit() chain wants to take
sock_lock which is already taken by tcp_v4_rcv() on its
way to rds_tcp_data_ready(). Commit db6526dcb51b ("RDS: use
rds_send_xmit() state instead of RDS_LL_SEND_FULL") which
was trying to opportunistically finish the send request
in same thread context.

But because of above recursive lock hang with RDS TCP,
the send work from rds_send_pong() needs to deferred to
worker to avoid lock up. Given RDS ping is more of connectivity
test than performance critical path, its should be ok even
for transport like IB.

Reported-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: Santosh Shilimkar <ssantosh@kernel.org>
Signed-off-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Acked-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/rds/send.c

index ee49c2556f4715ee7ad16cc4a4e376b9467af842..827155c2ead10376cb633c45c2f43917f5f5cd12 100644 (file)
@@ -1182,9 +1182,8 @@ rds_send_pong(struct rds_connection *conn, __be16 dport)
        rds_stats_inc(s_send_queued);
        rds_stats_inc(s_send_pong);
 
-       ret = rds_send_xmit(conn);
-       if (ret == -ENOMEM || ret == -EAGAIN)
-               queue_delayed_work(rds_wq, &conn->c_send_w, 1);
+       /* schedule the send work on rds_wq */
+       queue_delayed_work(rds_wq, &conn->c_send_w, 1);
 
        rds_message_put(rm);
        return 0;