From: Gerd Rausch Delegate fan-out to a background worker in order to allow kernel_getpeername() to acquire a lock on the socket. This has become necessary since the introduction of commit "9dfc685e0262d ("inet: remove races in inet{6}_getname()") The socket is already locked in the context that "kernel_getpeername" used to get called by either rds_tcp_recv_path" or "tcp_v{4,6}_rcv", and therefore causing a deadlock. Luckily, the fan-out need not happen in-context nor fast, so we can easily just do the same in a background worker. Also, while we're doing this, we get rid of the unused struct members "t_conn_w", "t_send_w", "t_down_w" & "t_recv_w". Signed-off-by: Gerd Rausch Signed-off-by: Allison Henderson --- net/rds/tcp.c | 3 +++ net/rds/tcp.h | 7 ++---- net/rds/tcp_connect.c | 2 ++ net/rds/tcp_listen.c | 54 +++++++++++++++++++++++++++++++------------ 4 files changed, 46 insertions(+), 20 deletions(-) diff --git a/net/rds/tcp.c b/net/rds/tcp.c index 45484a93d75f..02f8f928c20b 100644 --- a/net/rds/tcp.c +++ b/net/rds/tcp.c @@ -358,6 +358,8 @@ static void rds_tcp_conn_free(void *arg) rdsdebug("freeing tc %p\n", tc); + cancel_work_sync(&tc->t_fan_out_w); + spin_lock_irqsave(&rds_tcp_conn_lock, flags); if (!tc->t_tcp_node_detached) list_del(&tc->t_tcp_node); @@ -384,6 +386,7 @@ static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp) tc->t_tinc = NULL; tc->t_tinc_hdr_rem = sizeof(struct rds_header); tc->t_tinc_data_rem = 0; + INIT_WORK(&tc->t_fan_out_w, rds_tcp_fan_out_w); init_waitqueue_head(&tc->t_recv_done_waitq); conn->c_path[i].cp_transport_data = tc; diff --git a/net/rds/tcp.h b/net/rds/tcp.h index 39c86347188c..9ecb0b6b658a 100644 --- a/net/rds/tcp.h +++ b/net/rds/tcp.h @@ -44,11 +44,7 @@ struct rds_tcp_connection { size_t t_tinc_hdr_rem; size_t t_tinc_data_rem; - /* XXX error report? */ - struct work_struct t_conn_w; - struct work_struct t_send_w; - struct work_struct t_down_w; - struct work_struct t_recv_w; + struct work_struct t_fan_out_w; /* for info exporting only */ struct list_head t_list_item; @@ -90,6 +86,7 @@ void rds_tcp_state_change(struct sock *sk); struct socket *rds_tcp_listen_init(struct net *net, bool isv6); void rds_tcp_listen_stop(struct socket *sock, struct work_struct *acceptor); void rds_tcp_listen_data_ready(struct sock *sk); +void rds_tcp_fan_out_w(struct work_struct *work); void rds_tcp_conn_slots_available(struct rds_connection *conn, bool fan_out); int rds_tcp_accept_one(struct rds_tcp_net *rtn); void rds_tcp_keepalive(struct socket *sock); diff --git a/net/rds/tcp_connect.c b/net/rds/tcp_connect.c index b77c88ffb199..6954b8c479f1 100644 --- a/net/rds/tcp_connect.c +++ b/net/rds/tcp_connect.c @@ -115,6 +115,8 @@ int rds_tcp_conn_path_connect(struct rds_conn_path *cp) if (cp->cp_index > 0 && cp->cp_conn->c_npaths < 2) return -EAGAIN; + cancel_work_sync(&tc->t_fan_out_w); + mutex_lock(&tc->t_conn_path_lock); if (rds_conn_path_up(cp)) { diff --git a/net/rds/tcp_listen.c b/net/rds/tcp_listen.c index 6fb5c928b8fd..8fb8f7d26683 100644 --- a/net/rds/tcp_listen.c +++ b/net/rds/tcp_listen.c @@ -123,27 +123,20 @@ rds_tcp_accept_one_path(struct rds_connection *conn, struct socket *sock) return NULL; } -void rds_tcp_conn_slots_available(struct rds_connection *conn, bool fan_out) +void rds_tcp_fan_out_w(struct work_struct *work) { - struct rds_tcp_connection *tc; - struct rds_tcp_net *rtn; - struct socket *sock; + struct rds_tcp_connection *tc = container_of(work, + struct rds_tcp_connection, + t_fan_out_w); + struct rds_connection *conn = tc->t_cpath->cp_conn; + struct rds_tcp_net *rtn = tc->t_rtn; + struct socket *sock = tc->t_sock; int sport, npaths; - if (rds_destroy_pending(conn)) - return; - - tc = conn->c_path->cp_transport_data; - rtn = tc->t_rtn; - if (!rtn) - return; - - sock = tc->t_sock; - /* During fan-out, check that the connection we already * accepted in slot#0 carried the proper source port modulo. */ - if (fan_out && conn->c_with_sport_idx && sock && + if (conn->c_with_sport_idx && sock && rds_addr_cmp(&conn->c_laddr, &conn->c_faddr) > 0) { /* cp->cp_index is encoded in lowest bits of source-port */ sport = rds_tcp_get_peer_sport(sock); @@ -167,6 +160,37 @@ void rds_tcp_conn_slots_available(struct rds_connection *conn, bool fan_out) rds_tcp_accept_work(rtn); } +void rds_tcp_conn_slots_available(struct rds_connection *conn, bool fan_out) +{ + struct rds_conn_path *cp0; + struct rds_tcp_connection *tc; + struct rds_tcp_net *rtn; + + if (rds_destroy_pending(conn)) + return; + + cp0 = conn->c_path; + tc = cp0->cp_transport_data; + rtn = tc->t_rtn; + if (!rtn) + return; + + if (fan_out) + /* Delegate fan-out to a background worker in order + * to allow "kernel_getpeername" to acquire a lock + * on the socket. + * The socket is already locked in this context + * by either "rds_tcp_recv_path" or "tcp_v{4,6}_rcv", + * depending on the origin of the dequeue-request. + */ + queue_work(cp0->cp_wq, &tc->t_fan_out_w); + else + /* Fan-out either already happened or is unnecessary. + * Just go ahead and attempt to accept more connections + */ + rds_tcp_accept_work(rtn); +} + int rds_tcp_accept_one(struct rds_tcp_net *rtn) { struct socket *listen_sock = rtn->rds_tcp_listen_sock; -- 2.43.0 From: Greg Jumper The function rds_tcp_get_peer_sport() should return the peer port of a socket, even when the socket is not currently connected, so that RDS can reliably determine the MPRDS "lane" corresponding to the port. rds_tcp_get_peer_sport() calls kernel_getpeername() to get the port number; however, when paths between endpoints frequently drop and reconnect, kernel_getpeername() can return -ENOTCONN, causing rds_tcp_get_peer_sport() to return an error, and ultimately causing RDS to use the wrong lane for a port when reconnecting to a peer. This patch modifies rds_tcp_get_peer_sport() to directly call the socket-specific get-name function (inet_getname() in this case) that kernel_getpeername() also calls. The socket-specific function offers an additional argument which, when set to a value greater than 1, causes the function to return the socket's peer name even when the socket is not connected, which in turn allows rds_tcp_get_peer_sport() to return the correct port number. Signed-off-by: Greg Jumper Signed-off-by: Allison Henderson --- net/rds/tcp_listen.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/net/rds/tcp_listen.c b/net/rds/tcp_listen.c index 8fb8f7d26683..db4938fd1672 100644 --- a/net/rds/tcp_listen.c +++ b/net/rds/tcp_listen.c @@ -67,7 +67,14 @@ rds_tcp_get_peer_sport(struct socket *sock) } saddr; int sport; - if (kernel_getpeername(sock, &saddr.addr) >= 0) { + /* Call the socket's getname() function (inet_getname() in this case) + * with a final argument greater than 1 to get the peer's port + * regardless of whether the socket is currently connected. + * Using peer=2 will get the peer port even during reconnection states + * (TCPF_CLOSE, TCPF_SYN_SENT). This avoids -ENOTCONN while + * inet_dport still contains the correct peer port. + */ + if (sock->ops->getname(sock, &saddr.addr, 2) >= 0) { switch (saddr.addr.sa_family) { case AF_INET: sport = ntohs(saddr.sin.sin_port); -- 2.43.0 From: Allison Henderson Commit 3db6e0d172c9 ("rds: use RCU to synchronize work-enqueue with connection teardown") modifies rds_sendmsg to avoid enqueueing work while a tear down is in progress. However, it also changed the return value of rds_sendmsg to that of rds_send_xmit instead of the payload_len. This means the user may incorrectly receive errno values when it should have simply received a payload of 0 while the peer attempts a reconnections. So this patch corrects the teardown handling code to only use the out error path in that case, thus restoring the original payload_len return value. Signed-off-by: Allison Henderson --- net/rds/send.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/net/rds/send.c b/net/rds/send.c index 6e96f108473e..a1039e422a38 100644 --- a/net/rds/send.c +++ b/net/rds/send.c @@ -1431,9 +1431,11 @@ int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len) else queue_delayed_work(cpath->cp_wq, &cpath->cp_send_w, 1); rcu_read_unlock(); + + if (ret) + goto out; } - if (ret) - goto out; + rds_message_put(rm); for (ind = 0; ind < vct.indx; ind++) -- 2.43.0