RDS: TCP: Synchronize accept() and connect() paths on t_conn_lock.
authorSowmini Varadhan <sowmini.varadhan@oracle.com>
Mon, 2 May 2016 18:24:52 +0000 (11:24 -0700)
committerDavid S. Miller <davem@davemloft.net>
Tue, 3 May 2016 20:03:44 +0000 (16:03 -0400)
An arbitration scheme for duelling SYNs is implemented as part of
commit 241b271952eb ("RDS-TCP: Reset tcp callbacks if re-using an
outgoing socket in rds_tcp_accept_one()") which ensures that both nodes
involved will arrive at the same arbitration decision. However, this
needs to be synchronized with an outgoing SYN to be generated by
rds_tcp_conn_connect(). This commit achieves the synchronization
through the t_conn_lock mutex in struct rds_tcp_connection.

The rds_conn_state is checked in rds_tcp_conn_connect() after acquiring
the t_conn_lock mutex.  A SYN is sent out only if the RDS connection is
not already UP (an UP would indicate that rds_tcp_accept_one() has
completed 3WH, so no SYN needs to be generated).

Similarly, the rds_conn_state is checked in rds_tcp_accept_one() after
acquiring the t_conn_lock mutex. The only acceptable states (to
allow continuation of the arbitration logic) are UP (i.e., outgoing SYN
was SYN-ACKed by peer after it sent us the SYN) or CONNECTING (we sent
outgoing SYN before we saw incoming SYN).

Signed-off-by: Sowmini Varadhan <sowmini.varadhan@oracle.com>
Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/rds/tcp.c
net/rds/tcp.h
net/rds/tcp_connect.c
net/rds/tcp_listen.c

index 9134544941c22fc000ca7e6702c7d9ec0aab9e5a..86187dad14403100ff6199ee79beed5262de818f 100644 (file)
@@ -216,6 +216,7 @@ static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
        if (!tc)
                return -ENOMEM;
 
+       mutex_init(&tc->t_conn_lock);
        tc->t_sock = NULL;
        tc->t_tinc = NULL;
        tc->t_tinc_hdr_rem = sizeof(struct rds_header);
index 64f873c0c6b6d15574ec9771506d04e617a6aaac..41c228300525c029df02472b609ada1a71cea98e 100644 (file)
@@ -12,6 +12,10 @@ struct rds_tcp_connection {
 
        struct list_head        t_tcp_node;
        struct rds_connection   *conn;
+       /* t_conn_lock synchronizes the connection establishment between
+        * rds_tcp_accept_one and rds_tcp_conn_connect
+        */
+       struct mutex            t_conn_lock;
        struct socket           *t_sock;
        void                    *t_orig_write_space;
        void                    *t_orig_data_ready;
index 5cb16875c4603dba71c733de6600ada40e39cffc..49a3fcfed360edfb146416976c700b24e4520864 100644 (file)
@@ -78,7 +78,14 @@ int rds_tcp_conn_connect(struct rds_connection *conn)
        struct socket *sock = NULL;
        struct sockaddr_in src, dest;
        int ret;
+       struct rds_tcp_connection *tc = conn->c_transport_data;
+
+       mutex_lock(&tc->t_conn_lock);
 
+       if (rds_conn_up(conn)) {
+               mutex_unlock(&tc->t_conn_lock);
+               return 0;
+       }
        ret = sock_create_kern(rds_conn_net(conn), PF_INET,
                               SOCK_STREAM, IPPROTO_TCP, &sock);
        if (ret < 0)
@@ -120,6 +127,7 @@ int rds_tcp_conn_connect(struct rds_connection *conn)
        }
 
 out:
+       mutex_unlock(&tc->t_conn_lock);
        if (sock)
                sock_release(sock);
        return ret;
index 0896187243d68c0ad1d3a324e7b396be367fa24e..be263cdf268bae5b59a4746d4011d1042ede58cd 100644 (file)
@@ -76,7 +76,9 @@ int rds_tcp_accept_one(struct socket *sock)
        struct rds_connection *conn;
        int ret;
        struct inet_sock *inet;
-       struct rds_tcp_connection *rs_tcp;
+       struct rds_tcp_connection *rs_tcp = NULL;
+       int conn_state;
+       struct sock *nsk;
 
        ret = sock_create_kern(sock_net(sock->sk), sock->sk->sk_family,
                               sock->sk->sk_type, sock->sk->sk_protocol,
@@ -116,6 +118,10 @@ int rds_tcp_accept_one(struct socket *sock)
         */
        rs_tcp = (struct rds_tcp_connection *)conn->c_transport_data;
        rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING);
+       mutex_lock(&rs_tcp->t_conn_lock);
+       conn_state = rds_conn_state(conn);
+       if (conn_state != RDS_CONN_CONNECTING && conn_state != RDS_CONN_UP)
+               goto rst_nsk;
        if (rs_tcp->t_sock) {
                /* Need to resolve a duelling SYN between peers.
                 * We have an outstanding SYN to this peer, which may
@@ -126,14 +132,7 @@ int rds_tcp_accept_one(struct socket *sock)
                wait_event(conn->c_waitq,
                           !test_bit(RDS_IN_XMIT, &conn->c_flags));
                if (ntohl(inet->inet_saddr) < ntohl(inet->inet_daddr)) {
-                       struct sock *nsk = new_sock->sk;
-
-                       nsk->sk_user_data = NULL;
-                       nsk->sk_prot->disconnect(nsk, 0);
-                       tcp_done(nsk);
-                       new_sock = NULL;
-                       ret = 0;
-                       goto out;
+                       goto rst_nsk;
                } else if (rs_tcp->t_sock) {
                        rds_tcp_restore_callbacks(rs_tcp->t_sock, rs_tcp);
                        conn->c_outgoing = 0;
@@ -143,8 +142,19 @@ int rds_tcp_accept_one(struct socket *sock)
        rds_connect_complete(conn); /* marks RDS_CONN_UP */
        new_sock = NULL;
        ret = 0;
-
+       goto out;
+rst_nsk:
+       /* reset the newly returned accept sock and bail */
+       nsk = new_sock->sk;
+       rds_tcp_stats_inc(s_tcp_listen_closed_stale);
+       nsk->sk_user_data = NULL;
+       nsk->sk_prot->disconnect(nsk, 0);
+       tcp_done(nsk);
+       new_sock = NULL;
+       ret = 0;
 out:
+       if (rs_tcp)
+               mutex_unlock(&rs_tcp->t_conn_lock);
        if (new_sock)
                sock_release(new_sock);
        return ret;
This page took 0.02791 seconds and 5 git commands to generate.