2 * Copyright (c) 2006 Oracle. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include <linux/kernel.h>
34 #include <linux/slab.h>
36 #include <linux/module.h>
38 #include <net/net_namespace.h>
39 #include <net/netns/generic.h>
44 /* only for info exporting */
45 static DEFINE_SPINLOCK(rds_tcp_tc_list_lock);
46 static LIST_HEAD(rds_tcp_tc_list);
47 static unsigned int rds_tcp_tc_count;
49 /* Track rds_tcp_connection structs so they can be cleaned up */
50 static DEFINE_SPINLOCK(rds_tcp_conn_lock);
51 static LIST_HEAD(rds_tcp_conn_list);
53 static struct kmem_cache *rds_tcp_conn_slab;
55 static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
56 void __user *buffer, size_t *lenp,
59 static int rds_tcp_min_sndbuf = SOCK_MIN_SNDBUF;
60 static int rds_tcp_min_rcvbuf = SOCK_MIN_RCVBUF;
62 static struct ctl_table rds_tcp_sysctl_table[] = {
63 #define RDS_TCP_SNDBUF 0
65 .procname = "rds_tcp_sndbuf",
66 /* data is per-net pointer */
67 .maxlen = sizeof(int),
69 .proc_handler = rds_tcp_skbuf_handler,
70 .extra1 = &rds_tcp_min_sndbuf,
72 #define RDS_TCP_RCVBUF 1
74 .procname = "rds_tcp_rcvbuf",
75 /* data is per-net pointer */
76 .maxlen = sizeof(int),
78 .proc_handler = rds_tcp_skbuf_handler,
79 .extra1 = &rds_tcp_min_rcvbuf,
84 /* doing it this way avoids calling tcp_sk() */
85 void rds_tcp_nonagle(struct socket *sock)
89 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (void *)&val,
93 u32 rds_tcp_snd_nxt(struct rds_tcp_connection *tc)
95 return tcp_sk(tc->t_sock->sk)->snd_nxt;
98 u32 rds_tcp_snd_una(struct rds_tcp_connection *tc)
100 return tcp_sk(tc->t_sock->sk)->snd_una;
103 void rds_tcp_restore_callbacks(struct socket *sock,
104 struct rds_tcp_connection *tc)
106 rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc);
107 write_lock_bh(&sock->sk->sk_callback_lock);
109 /* done under the callback_lock to serialize with write_space */
110 spin_lock(&rds_tcp_tc_list_lock);
111 list_del_init(&tc->t_list_item);
113 spin_unlock(&rds_tcp_tc_list_lock);
117 sock->sk->sk_write_space = tc->t_orig_write_space;
118 sock->sk->sk_data_ready = tc->t_orig_data_ready;
119 sock->sk->sk_state_change = tc->t_orig_state_change;
120 sock->sk->sk_user_data = NULL;
122 write_unlock_bh(&sock->sk->sk_callback_lock);
126 * rds_tcp_reset_callbacks() switches the to the new sock and
127 * returns the existing tc->t_sock.
129 * The only functions that set tc->t_sock are rds_tcp_set_callbacks
130 * and rds_tcp_reset_callbacks. Send and receive trust that
131 * it is set. The absence of RDS_CONN_UP bit protects those paths
132 * from being called while it isn't set.
134 void rds_tcp_reset_callbacks(struct socket *sock,
135 struct rds_conn_path *cp)
137 struct rds_tcp_connection *tc = cp->cp_transport_data;
138 struct socket *osock = tc->t_sock;
143 /* Need to resolve a duelling SYN between peers.
144 * We have an outstanding SYN to this peer, which may
145 * potentially have transitioned to the RDS_CONN_UP state,
146 * so we must quiesce any send threads before resetting
147 * cp_transport_data. We quiesce these threads by setting
148 * cp_state to something other than RDS_CONN_UP, and then
149 * waiting for any existing threads in rds_send_xmit to
150 * complete release_in_xmit(). (Subsequent threads entering
151 * rds_send_xmit() will bail on !rds_conn_up().
153 * However an incoming syn-ack at this point would end up
154 * marking the conn as RDS_CONN_UP, and would again permit
155 * rds_send_xmi() threads through, so ideally we would
156 * synchronize on RDS_CONN_UP after lock_sock(), but cannot
157 * do that: waiting on !RDS_IN_XMIT after lock_sock() may
158 * end up deadlocking with tcp_sendmsg(), and the RDS_IN_XMIT
159 * would not get set. As a result, we set c_state to
160 * RDS_CONN_RESETTTING, to ensure that rds_tcp_state_change
161 * cannot mark rds_conn_path_up() in the window before lock_sock()
163 atomic_set(&cp->cp_state, RDS_CONN_RESETTING);
164 wait_event(cp->cp_waitq, !test_bit(RDS_IN_XMIT, &cp->cp_flags));
165 /* reset receive side state for rds_tcp_data_recv() for osock */
166 cancel_delayed_work_sync(&cp->cp_send_w);
167 cancel_delayed_work_sync(&cp->cp_recv_w);
168 lock_sock(osock->sk);
170 rds_inc_put(&tc->t_tinc->ti_inc);
173 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
174 tc->t_tinc_data_rem = 0;
175 rds_tcp_restore_callbacks(osock, tc);
176 release_sock(osock->sk);
179 rds_send_path_reset(cp);
181 rds_tcp_set_callbacks(sock, cp);
182 release_sock(sock->sk);
185 /* Add tc to rds_tcp_tc_list and set tc->t_sock. See comments
186 * above rds_tcp_reset_callbacks for notes about synchronization
189 void rds_tcp_set_callbacks(struct socket *sock, struct rds_conn_path *cp)
191 struct rds_tcp_connection *tc = cp->cp_transport_data;
193 rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
194 write_lock_bh(&sock->sk->sk_callback_lock);
196 /* done under the callback_lock to serialize with write_space */
197 spin_lock(&rds_tcp_tc_list_lock);
198 list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
200 spin_unlock(&rds_tcp_tc_list_lock);
202 /* accepted sockets need our listen data ready undone */
203 if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
204 sock->sk->sk_data_ready = sock->sk->sk_user_data;
208 tc->t_orig_data_ready = sock->sk->sk_data_ready;
209 tc->t_orig_write_space = sock->sk->sk_write_space;
210 tc->t_orig_state_change = sock->sk->sk_state_change;
212 sock->sk->sk_user_data = cp;
213 sock->sk->sk_data_ready = rds_tcp_data_ready;
214 sock->sk->sk_write_space = rds_tcp_write_space;
215 sock->sk->sk_state_change = rds_tcp_state_change;
217 write_unlock_bh(&sock->sk->sk_callback_lock);
220 static void rds_tcp_tc_info(struct socket *rds_sock, unsigned int len,
221 struct rds_info_iterator *iter,
222 struct rds_info_lengths *lens)
224 struct rds_info_tcp_socket tsinfo;
225 struct rds_tcp_connection *tc;
227 struct sockaddr_in sin;
231 spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
233 if (len / sizeof(tsinfo) < rds_tcp_tc_count)
236 list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
240 sock->ops->getname(sock, (struct sockaddr *)&sin,
242 tsinfo.local_addr = sin.sin_addr.s_addr;
243 tsinfo.local_port = sin.sin_port;
244 sock->ops->getname(sock, (struct sockaddr *)&sin,
246 tsinfo.peer_addr = sin.sin_addr.s_addr;
247 tsinfo.peer_port = sin.sin_port;
250 tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
251 tsinfo.data_rem = tc->t_tinc_data_rem;
252 tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
253 tsinfo.last_expected_una = tc->t_last_expected_una;
254 tsinfo.last_seen_una = tc->t_last_seen_una;
256 rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
260 lens->nr = rds_tcp_tc_count;
261 lens->each = sizeof(tsinfo);
263 spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
266 static int rds_tcp_laddr_check(struct net *net, __be32 addr)
268 if (inet_addr_type(net, addr) == RTN_LOCAL)
270 return -EADDRNOTAVAIL;
273 static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
275 struct rds_tcp_connection *tc;
278 for (i = 0; i < RDS_MPATH_WORKERS; i++) {
279 tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
283 mutex_init(&tc->t_conn_path_lock);
286 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
287 tc->t_tinc_data_rem = 0;
289 conn->c_path[i].cp_transport_data = tc;
290 tc->t_cpath = &conn->c_path[i];
292 spin_lock_irq(&rds_tcp_conn_lock);
293 list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
294 spin_unlock_irq(&rds_tcp_conn_lock);
295 rdsdebug("rds_conn_path [%d] tc %p\n", i,
296 conn->c_path[i].cp_transport_data);
302 static void rds_tcp_conn_free(void *arg)
304 struct rds_tcp_connection *tc = arg;
306 rdsdebug("freeing tc %p\n", tc);
308 spin_lock_irqsave(&rds_tcp_conn_lock, flags);
309 if (!tc->t_tcp_node_detached)
310 list_del(&tc->t_tcp_node);
311 spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
313 kmem_cache_free(rds_tcp_conn_slab, tc);
316 static bool list_has_conn(struct list_head *list, struct rds_connection *conn)
318 struct rds_tcp_connection *tc, *_tc;
320 list_for_each_entry_safe(tc, _tc, list, t_tcp_node) {
321 if (tc->t_cpath->cp_conn == conn)
327 static void rds_tcp_destroy_conns(void)
329 struct rds_tcp_connection *tc, *_tc;
332 /* avoid calling conn_destroy with irqs off */
333 spin_lock_irq(&rds_tcp_conn_lock);
334 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
335 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn))
336 list_move_tail(&tc->t_tcp_node, &tmp_list);
338 spin_unlock_irq(&rds_tcp_conn_lock);
340 list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node)
341 rds_conn_destroy(tc->t_cpath->cp_conn);
344 static void rds_tcp_exit(void);
346 struct rds_transport rds_tcp_transport = {
347 .laddr_check = rds_tcp_laddr_check,
348 .xmit_path_prepare = rds_tcp_xmit_path_prepare,
349 .xmit_path_complete = rds_tcp_xmit_path_complete,
350 .xmit = rds_tcp_xmit,
351 .recv_path = rds_tcp_recv_path,
352 .conn_alloc = rds_tcp_conn_alloc,
353 .conn_free = rds_tcp_conn_free,
354 .conn_path_connect = rds_tcp_conn_path_connect,
355 .conn_path_shutdown = rds_tcp_conn_path_shutdown,
356 .inc_copy_to_user = rds_tcp_inc_copy_to_user,
357 .inc_free = rds_tcp_inc_free,
358 .stats_info_copy = rds_tcp_stats_info_copy,
359 .exit = rds_tcp_exit,
360 .t_owner = THIS_MODULE,
362 .t_type = RDS_TRANS_TCP,
363 .t_prefer_loopback = 1,
367 static unsigned int rds_tcp_netid;
369 /* per-network namespace private data for this module */
371 struct socket *rds_tcp_listen_sock;
372 struct work_struct rds_tcp_accept_w;
373 struct ctl_table_header *rds_tcp_sysctl;
374 struct ctl_table *ctl_table;
379 /* All module specific customizations to the RDS-TCP socket should be done in
380 * rds_tcp_tune() and applied after socket creation.
382 void rds_tcp_tune(struct socket *sock)
384 struct sock *sk = sock->sk;
385 struct net *net = sock_net(sk);
386 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
388 rds_tcp_nonagle(sock);
390 if (rtn->sndbuf_size > 0) {
391 sk->sk_sndbuf = rtn->sndbuf_size;
392 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
394 if (rtn->rcvbuf_size > 0) {
395 sk->sk_rcvbuf = rtn->rcvbuf_size;
396 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
401 static void rds_tcp_accept_worker(struct work_struct *work)
403 struct rds_tcp_net *rtn = container_of(work,
407 while (rds_tcp_accept_one(rtn->rds_tcp_listen_sock) == 0)
411 void rds_tcp_accept_work(struct sock *sk)
413 struct net *net = sock_net(sk);
414 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
416 queue_work(rds_wq, &rtn->rds_tcp_accept_w);
419 static __net_init int rds_tcp_init_net(struct net *net)
421 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
422 struct ctl_table *tbl;
425 memset(rtn, 0, sizeof(*rtn));
427 /* {snd, rcv}buf_size default to 0, which implies we let the
428 * stack pick the value, and permit auto-tuning of buffer size.
430 if (net == &init_net) {
431 tbl = rds_tcp_sysctl_table;
433 tbl = kmemdup(rds_tcp_sysctl_table,
434 sizeof(rds_tcp_sysctl_table), GFP_KERNEL);
436 pr_warn("could not set allocate syctl table\n");
439 rtn->ctl_table = tbl;
441 tbl[RDS_TCP_SNDBUF].data = &rtn->sndbuf_size;
442 tbl[RDS_TCP_RCVBUF].data = &rtn->rcvbuf_size;
443 rtn->rds_tcp_sysctl = register_net_sysctl(net, "net/rds/tcp", tbl);
444 if (!rtn->rds_tcp_sysctl) {
445 pr_warn("could not register sysctl\n");
449 rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net);
450 if (!rtn->rds_tcp_listen_sock) {
451 pr_warn("could not set up listen sock\n");
452 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
453 rtn->rds_tcp_sysctl = NULL;
457 INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
461 if (net != &init_net)
466 static void __net_exit rds_tcp_exit_net(struct net *net)
468 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
470 if (rtn->rds_tcp_sysctl)
471 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
473 if (net != &init_net && rtn->ctl_table)
474 kfree(rtn->ctl_table);
476 /* If rds_tcp_exit_net() is called as a result of netns deletion,
477 * the rds_tcp_kill_sock() device notifier would already have cleaned
478 * up the listen socket, thus there is no work to do in this function.
480 * If rds_tcp_exit_net() is called as a result of module unload,
481 * i.e., due to rds_tcp_exit() -> unregister_pernet_subsys(), then
482 * we do need to clean up the listen socket here.
484 if (rtn->rds_tcp_listen_sock) {
485 struct socket *lsock = rtn->rds_tcp_listen_sock;
487 rtn->rds_tcp_listen_sock = NULL;
488 rds_tcp_listen_stop(lsock, &rtn->rds_tcp_accept_w);
492 static struct pernet_operations rds_tcp_net_ops = {
493 .init = rds_tcp_init_net,
494 .exit = rds_tcp_exit_net,
495 .id = &rds_tcp_netid,
496 .size = sizeof(struct rds_tcp_net),
499 /* explicitly send a RST on each socket, thereby releasing any socket refcnts
500 * that may otherwise hold up netns deletion.
502 static void rds_tcp_conn_paths_destroy(struct rds_connection *conn)
504 struct rds_conn_path *cp;
505 struct rds_tcp_connection *tc;
509 for (i = 0; i < RDS_MPATH_WORKERS; i++) {
510 cp = &conn->c_path[i];
511 tc = cp->cp_transport_data;
515 sk->sk_prot->disconnect(sk, 0);
520 static void rds_tcp_kill_sock(struct net *net)
522 struct rds_tcp_connection *tc, *_tc;
524 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
525 struct socket *lsock = rtn->rds_tcp_listen_sock;
527 rtn->rds_tcp_listen_sock = NULL;
528 rds_tcp_listen_stop(lsock, &rtn->rds_tcp_accept_w);
529 spin_lock_irq(&rds_tcp_conn_lock);
530 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
531 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
535 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn)) {
536 list_move_tail(&tc->t_tcp_node, &tmp_list);
538 list_del(&tc->t_tcp_node);
539 tc->t_tcp_node_detached = true;
542 spin_unlock_irq(&rds_tcp_conn_lock);
543 list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
544 rds_tcp_conn_paths_destroy(tc->t_cpath->cp_conn);
545 rds_conn_destroy(tc->t_cpath->cp_conn);
549 void *rds_tcp_listen_sock_def_readable(struct net *net)
551 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
552 struct socket *lsock = rtn->rds_tcp_listen_sock;
557 return lsock->sk->sk_user_data;
560 static int rds_tcp_dev_event(struct notifier_block *this,
561 unsigned long event, void *ptr)
563 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
565 /* rds-tcp registers as a pernet subys, so the ->exit will only
566 * get invoked after network acitivity has quiesced. We need to
567 * clean up all sockets to quiesce network activity, and use
568 * the unregistration of the per-net loopback device as a trigger
569 * to start that cleanup.
571 if (event == NETDEV_UNREGISTER_FINAL &&
572 dev->ifindex == LOOPBACK_IFINDEX)
573 rds_tcp_kill_sock(dev_net(dev));
578 static struct notifier_block rds_tcp_dev_notifier = {
579 .notifier_call = rds_tcp_dev_event,
580 .priority = -10, /* must be called after other network notifiers */
583 /* when sysctl is used to modify some kernel socket parameters,this
584 * function resets the RDS connections in that netns so that we can
585 * restart with new parameters. The assumption is that such reset
586 * events are few and far-between.
588 static void rds_tcp_sysctl_reset(struct net *net)
590 struct rds_tcp_connection *tc, *_tc;
592 spin_lock_irq(&rds_tcp_conn_lock);
593 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
594 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
596 if (net != c_net || !tc->t_sock)
599 /* reconnect with new parameters */
600 rds_conn_path_drop(tc->t_cpath, false);
602 spin_unlock_irq(&rds_tcp_conn_lock);
605 static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
606 void __user *buffer, size_t *lenp,
609 struct net *net = current->nsproxy->net_ns;
612 err = proc_dointvec_minmax(ctl, write, buffer, lenp, fpos);
614 pr_warn("Invalid input. Must be >= %d\n",
615 *(int *)(ctl->extra1));
619 rds_tcp_sysctl_reset(net);
623 static void rds_tcp_exit(void)
625 rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
626 unregister_pernet_subsys(&rds_tcp_net_ops);
627 if (unregister_netdevice_notifier(&rds_tcp_dev_notifier))
628 pr_warn("could not unregister rds_tcp_dev_notifier\n");
629 rds_tcp_destroy_conns();
630 rds_trans_unregister(&rds_tcp_transport);
632 kmem_cache_destroy(rds_tcp_conn_slab);
634 module_exit(rds_tcp_exit);
636 static int rds_tcp_init(void)
640 rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
641 sizeof(struct rds_tcp_connection),
643 if (!rds_tcp_conn_slab) {
648 ret = rds_tcp_recv_init();
652 ret = register_pernet_subsys(&rds_tcp_net_ops);
656 ret = register_netdevice_notifier(&rds_tcp_dev_notifier);
658 pr_warn("could not register rds_tcp_dev_notifier\n");
662 rds_trans_register(&rds_tcp_transport);
664 rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
669 unregister_pernet_subsys(&rds_tcp_net_ops);
673 kmem_cache_destroy(rds_tcp_conn_slab);
677 module_init(rds_tcp_init);
679 MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
680 MODULE_DESCRIPTION("RDS: TCP transport");
681 MODULE_LICENSE("Dual BSD/GPL");