net/rds/tcp_listen.c

0001 /*
0002  * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved.
0003  *
0004  * This software is available to you under a choice of one of two
0005  * licenses.  You may choose to be licensed under the terms of the GNU
0006  * General Public License (GPL) Version 2, available from the file
0007  * COPYING in the main directory of this source tree, or the
0008  * OpenIB.org BSD license below:
0009  *
0010  *     Redistribution and use in source and binary forms, with or
0011  *     without modification, are permitted provided that the following
0012  *     conditions are met:
0013  *
0014  *      - Redistributions of source code must retain the above
0015  *        copyright notice, this list of conditions and the following
0016  *        disclaimer.
0017  *
0018  *      - Redistributions in binary form must reproduce the above
0019  *        copyright notice, this list of conditions and the following
0020  *        disclaimer in the documentation and/or other materials
0021  *        provided with the distribution.
0022  *
0023  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
0024  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
0025  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
0026  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
0027  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
0028  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
0029  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
0030  * SOFTWARE.
0031  *
0032  */
0033 #include <linux/kernel.h>
0034 #include <linux/gfp.h>
0035 #include <linux/in.h>
0036 #include <net/tcp.h>
0037
0038 #include "rds.h"
0039 #include "tcp.h"
0040
0041 void rds_tcp_keepalive(struct socket *sock)
0042 {
0043     /* values below based on xs_udp_default_timeout */
0044     int keepidle = 5; /* send a probe 'keepidle' secs after last data */
0045     int keepcnt = 5; /* number of unack'ed probes before declaring dead */
0046
0047     sock_set_keepalive(sock->sk);
0048     tcp_sock_set_keepcnt(sock->sk, keepcnt);
0049     tcp_sock_set_keepidle(sock->sk, keepidle);
0050     /* KEEPINTVL is the interval between successive probes. We follow
0051      * the model in xs_tcp_finish_connecting() and re-use keepidle.
0052      */
0053     tcp_sock_set_keepintvl(sock->sk, keepidle);
0054 }
0055
0056 /* rds_tcp_accept_one_path(): if accepting on cp_index > 0, make sure the
0057  * client's ipaddr < server's ipaddr. Otherwise, close the accepted
0058  * socket and force a reconneect from smaller -> larger ip addr. The reason
0059  * we special case cp_index 0 is to allow the rds probe ping itself to itself
0060  * get through efficiently.
0061  * Since reconnects are only initiated from the node with the numerically
0062  * smaller ip address, we recycle conns in RDS_CONN_ERROR on the passive side
0063  * by moving them to CONNECTING in this function.
0064  */
0065 static
0066 struct rds_tcp_connection *rds_tcp_accept_one_path(struct rds_connection *conn)
0067 {
0068     int i;
0069     int npaths = max_t(int, 1, conn->c_npaths);
0070
0071     /* for mprds, all paths MUST be initiated by the peer
0072      * with the smaller address.
0073      */
0074     if (rds_addr_cmp(&conn->c_faddr, &conn->c_laddr) >= 0) {
0075         /* Make sure we initiate at least one path if this
0076          * has not already been done; rds_start_mprds() will
0077          * take care of additional paths, if necessary.
0078          */
0079         if (npaths == 1)
0080             rds_conn_path_connect_if_down(&conn->c_path[0]);
0081         return NULL;
0082     }
0083
0084     for (i = 0; i < npaths; i++) {
0085         struct rds_conn_path *cp = &conn->c_path[i];
0086
0087         if (rds_conn_path_transition(cp, RDS_CONN_DOWN,
0088                          RDS_CONN_CONNECTING) ||
0089             rds_conn_path_transition(cp, RDS_CONN_ERROR,
0090                          RDS_CONN_CONNECTING)) {
0091             return cp->cp_transport_data;
0092         }
0093     }
0094     return NULL;
0095 }
0096
0097 int rds_tcp_accept_one(struct socket *sock)
0098 {
0099     struct socket *new_sock = NULL;
0100     struct rds_connection *conn;
0101     int ret;
0102     struct inet_sock *inet;
0103     struct rds_tcp_connection *rs_tcp = NULL;
0104     int conn_state;
0105     struct rds_conn_path *cp;
0106     struct in6_addr *my_addr, *peer_addr;
0107 #if !IS_ENABLED(CONFIG_IPV6)
0108     struct in6_addr saddr, daddr;
0109 #endif
0110     int dev_if = 0;
0111
0112     if (!sock) /* module unload or netns delete in progress */
0113         return -ENETUNREACH;
0114
0115     ret = sock_create_lite(sock->sk->sk_family,
0116                    sock->sk->sk_type, sock->sk->sk_protocol,
0117                    &new_sock);
0118     if (ret)
0119         goto out;
0120
0121     ret = sock->ops->accept(sock, new_sock, O_NONBLOCK, true);
0122     if (ret < 0)
0123         goto out;
0124
0125     /* sock_create_lite() does not get a hold on the owner module so we
0126      * need to do it here.  Note that sock_release() uses sock->ops to
0127      * determine if it needs to decrement the reference count.  So set
0128      * sock->ops after calling accept() in case that fails.  And there's
0129      * no need to do try_module_get() as the listener should have a hold
0130      * already.
0131      */
0132     new_sock->ops = sock->ops;
0133     __module_get(new_sock->ops->owner);
0134
0135     rds_tcp_keepalive(new_sock);
0136     if (!rds_tcp_tune(new_sock)) {
0137         ret = -EINVAL;
0138         goto out;
0139     }
0140
0141     inet = inet_sk(new_sock->sk);
0142
0143 #if IS_ENABLED(CONFIG_IPV6)
0144     my_addr = &new_sock->sk->sk_v6_rcv_saddr;
0145     peer_addr = &new_sock->sk->sk_v6_daddr;
0146 #else
0147     ipv6_addr_set_v4mapped(inet->inet_saddr, &saddr);
0148     ipv6_addr_set_v4mapped(inet->inet_daddr, &daddr);
0149     my_addr = &saddr;
0150     peer_addr = &daddr;
0151 #endif
0152     rdsdebug("accepted family %d tcp %pI6c:%u -> %pI6c:%u\n",
0153          sock->sk->sk_family,
0154          my_addr, ntohs(inet->inet_sport),
0155          peer_addr, ntohs(inet->inet_dport));
0156
0157 #if IS_ENABLED(CONFIG_IPV6)
0158     /* sk_bound_dev_if is not set if the peer address is not link local
0159      * address.  In this case, it happens that mcast_oif is set.  So
0160      * just use it.
0161      */
0162     if ((ipv6_addr_type(my_addr) & IPV6_ADDR_LINKLOCAL) &&
0163         !(ipv6_addr_type(peer_addr) & IPV6_ADDR_LINKLOCAL)) {
0164         struct ipv6_pinfo *inet6;
0165
0166         inet6 = inet6_sk(new_sock->sk);
0167         dev_if = inet6->mcast_oif;
0168     } else {
0169         dev_if = new_sock->sk->sk_bound_dev_if;
0170     }
0171 #endif
0172
0173     if (!rds_tcp_laddr_check(sock_net(sock->sk), peer_addr, dev_if)) {
0174         /* local address connection is only allowed via loopback */
0175         ret = -EOPNOTSUPP;
0176         goto out;
0177     }
0178
0179     conn = rds_conn_create(sock_net(sock->sk),
0180                    my_addr, peer_addr,
0181                    &rds_tcp_transport, 0, GFP_KERNEL, dev_if);
0182
0183     if (IS_ERR(conn)) {
0184         ret = PTR_ERR(conn);
0185         goto out;
0186     }
0187     /* An incoming SYN request came in, and TCP just accepted it.
0188      *
0189      * If the client reboots, this conn will need to be cleaned up.
0190      * rds_tcp_state_change() will do that cleanup
0191      */
0192     rs_tcp = rds_tcp_accept_one_path(conn);
0193     if (!rs_tcp)
0194         goto rst_nsk;
0195     mutex_lock(&rs_tcp->t_conn_path_lock);
0196     cp = rs_tcp->t_cpath;
0197     conn_state = rds_conn_path_state(cp);
0198     WARN_ON(conn_state == RDS_CONN_UP);
0199     if (conn_state != RDS_CONN_CONNECTING && conn_state != RDS_CONN_ERROR)
0200         goto rst_nsk;
0201     if (rs_tcp->t_sock) {
0202         /* Duelling SYN has been handled in rds_tcp_accept_one() */
0203         rds_tcp_reset_callbacks(new_sock, cp);
0204         /* rds_connect_path_complete() marks RDS_CONN_UP */
0205         rds_connect_path_complete(cp, RDS_CONN_RESETTING);
0206     } else {
0207         rds_tcp_set_callbacks(new_sock, cp);
0208         rds_connect_path_complete(cp, RDS_CONN_CONNECTING);
0209     }
0210     new_sock = NULL;
0211     ret = 0;
0212     if (conn->c_npaths == 0)
0213         rds_send_ping(cp->cp_conn, cp->cp_index);
0214     goto out;
0215 rst_nsk:
0216     /* reset the newly returned accept sock and bail.
0217      * It is safe to set linger on new_sock because the RDS connection
0218      * has not been brought up on new_sock, so no RDS-level data could
0219      * be pending on it. By setting linger, we achieve the side-effect
0220      * of avoiding TIME_WAIT state on new_sock.
0221      */
0222     sock_no_linger(new_sock->sk);
0223     kernel_sock_shutdown(new_sock, SHUT_RDWR);
0224     ret = 0;
0225 out:
0226     if (rs_tcp)
0227         mutex_unlock(&rs_tcp->t_conn_path_lock);
0228     if (new_sock)
0229         sock_release(new_sock);
0230     return ret;
0231 }
0232
0233 void rds_tcp_listen_data_ready(struct sock *sk)
0234 {
0235     void (*ready)(struct sock *sk);
0236
0237     rdsdebug("listen data ready sk %p\n", sk);
0238
0239     read_lock_bh(&sk->sk_callback_lock);
0240     ready = sk->sk_user_data;
0241     if (!ready) { /* check for teardown race */
0242         ready = sk->sk_data_ready;
0243         goto out;
0244     }
0245
0246     /*
0247      * ->sk_data_ready is also called for a newly established child socket
0248      * before it has been accepted and the accepter has set up their
0249      * data_ready.. we only want to queue listen work for our listening
0250      * socket
0251      *
0252      * (*ready)() may be null if we are racing with netns delete, and
0253      * the listen socket is being torn down.
0254      */
0255     if (sk->sk_state == TCP_LISTEN)
0256         rds_tcp_accept_work(sk);
0257     else
0258         ready = rds_tcp_listen_sock_def_readable(sock_net(sk));
0259
0260 out:
0261     read_unlock_bh(&sk->sk_callback_lock);
0262     if (ready)
0263         ready(sk);
0264 }
0265
0266 struct socket *rds_tcp_listen_init(struct net *net, bool isv6)
0267 {
0268     struct socket *sock = NULL;
0269     struct sockaddr_storage ss;
0270     struct sockaddr_in6 *sin6;
0271     struct sockaddr_in *sin;
0272     int addr_len;
0273     int ret;
0274
0275     ret = sock_create_kern(net, isv6 ? PF_INET6 : PF_INET, SOCK_STREAM,
0276                    IPPROTO_TCP, &sock);
0277     if (ret < 0) {
0278         rdsdebug("could not create %s listener socket: %d\n",
0279              isv6 ? "IPv6" : "IPv4", ret);
0280         goto out;
0281     }
0282
0283     sock->sk->sk_reuse = SK_CAN_REUSE;
0284     tcp_sock_set_nodelay(sock->sk);
0285
0286     write_lock_bh(&sock->sk->sk_callback_lock);
0287     sock->sk->sk_user_data = sock->sk->sk_data_ready;
0288     sock->sk->sk_data_ready = rds_tcp_listen_data_ready;
0289     write_unlock_bh(&sock->sk->sk_callback_lock);
0290
0291     if (isv6) {
0292         sin6 = (struct sockaddr_in6 *)&ss;
0293         sin6->sin6_family = PF_INET6;
0294         sin6->sin6_addr = in6addr_any;
0295         sin6->sin6_port = (__force u16)htons(RDS_TCP_PORT);
0296         sin6->sin6_scope_id = 0;
0297         sin6->sin6_flowinfo = 0;
0298         addr_len = sizeof(*sin6);
0299     } else {
0300         sin = (struct sockaddr_in *)&ss;
0301         sin->sin_family = PF_INET;
0302         sin->sin_addr.s_addr = INADDR_ANY;
0303         sin->sin_port = (__force u16)htons(RDS_TCP_PORT);
0304         addr_len = sizeof(*sin);
0305     }
0306
0307     ret = sock->ops->bind(sock, (struct sockaddr *)&ss, addr_len);
0308     if (ret < 0) {
0309         rdsdebug("could not bind %s listener socket: %d\n",
0310              isv6 ? "IPv6" : "IPv4", ret);
0311         goto out;
0312     }
0313
0314     ret = sock->ops->listen(sock, 64);
0315     if (ret < 0)
0316         goto out;
0317
0318     return sock;
0319 out:
0320     if (sock)
0321         sock_release(sock);
0322     return NULL;
0323 }
0324
0325 void rds_tcp_listen_stop(struct socket *sock, struct work_struct *acceptor)
0326 {
0327     struct sock *sk;
0328
0329     if (!sock)
0330         return;
0331
0332     sk = sock->sk;
0333
0334     /* serialize with and prevent further callbacks */
0335     lock_sock(sk);
0336     write_lock_bh(&sk->sk_callback_lock);
0337     if (sk->sk_user_data) {
0338         sk->sk_data_ready = sk->sk_user_data;
0339         sk->sk_user_data = NULL;
0340     }
0341     write_unlock_bh(&sk->sk_callback_lock);
0342     release_sock(sk);
0343
0344     /* wait for accepts to stop and close the socket */
0345     flush_workqueue(rds_wq);
0346     flush_work(acceptor);
0347     sock_release(sock);
0348 }