Loading...
1/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
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:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
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.
22 *
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
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/slab.h>
35#include <linux/in.h>
36#include <linux/module.h>
37#include <net/tcp.h>
38#include <net/net_namespace.h>
39#include <net/netns/generic.h>
40
41#include "rds.h"
42#include "tcp.h"
43
44/* only for info exporting */
45static DEFINE_SPINLOCK(rds_tcp_tc_list_lock);
46static LIST_HEAD(rds_tcp_tc_list);
47static unsigned int rds_tcp_tc_count;
48
49/* Track rds_tcp_connection structs so they can be cleaned up */
50static DEFINE_SPINLOCK(rds_tcp_conn_lock);
51static LIST_HEAD(rds_tcp_conn_list);
52
53static struct kmem_cache *rds_tcp_conn_slab;
54
55static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
56 void __user *buffer, size_t *lenp,
57 loff_t *fpos);
58
59int rds_tcp_min_sndbuf = SOCK_MIN_SNDBUF;
60int rds_tcp_min_rcvbuf = SOCK_MIN_RCVBUF;
61
62static struct ctl_table rds_tcp_sysctl_table[] = {
63#define RDS_TCP_SNDBUF 0
64 {
65 .procname = "rds_tcp_sndbuf",
66 /* data is per-net pointer */
67 .maxlen = sizeof(int),
68 .mode = 0644,
69 .proc_handler = rds_tcp_skbuf_handler,
70 .extra1 = &rds_tcp_min_sndbuf,
71 },
72#define RDS_TCP_RCVBUF 1
73 {
74 .procname = "rds_tcp_rcvbuf",
75 /* data is per-net pointer */
76 .maxlen = sizeof(int),
77 .mode = 0644,
78 .proc_handler = rds_tcp_skbuf_handler,
79 .extra1 = &rds_tcp_min_rcvbuf,
80 },
81 { }
82};
83
84/* doing it this way avoids calling tcp_sk() */
85void rds_tcp_nonagle(struct socket *sock)
86{
87 mm_segment_t oldfs = get_fs();
88 int val = 1;
89
90 set_fs(KERNEL_DS);
91 sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY, (char __user *)&val,
92 sizeof(val));
93 set_fs(oldfs);
94}
95
96u32 rds_tcp_snd_nxt(struct rds_tcp_connection *tc)
97{
98 return tcp_sk(tc->t_sock->sk)->snd_nxt;
99}
100
101u32 rds_tcp_snd_una(struct rds_tcp_connection *tc)
102{
103 return tcp_sk(tc->t_sock->sk)->snd_una;
104}
105
106void rds_tcp_restore_callbacks(struct socket *sock,
107 struct rds_tcp_connection *tc)
108{
109 rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc);
110 write_lock_bh(&sock->sk->sk_callback_lock);
111
112 /* done under the callback_lock to serialize with write_space */
113 spin_lock(&rds_tcp_tc_list_lock);
114 list_del_init(&tc->t_list_item);
115 rds_tcp_tc_count--;
116 spin_unlock(&rds_tcp_tc_list_lock);
117
118 tc->t_sock = NULL;
119
120 sock->sk->sk_write_space = tc->t_orig_write_space;
121 sock->sk->sk_data_ready = tc->t_orig_data_ready;
122 sock->sk->sk_state_change = tc->t_orig_state_change;
123 sock->sk->sk_user_data = NULL;
124
125 write_unlock_bh(&sock->sk->sk_callback_lock);
126}
127
128/*
129 * This is the only path that sets tc->t_sock. Send and receive trust that
130 * it is set. The RDS_CONN_UP bit protects those paths from being
131 * called while it isn't set.
132 */
133void rds_tcp_set_callbacks(struct socket *sock, struct rds_connection *conn)
134{
135 struct rds_tcp_connection *tc = conn->c_transport_data;
136
137 rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
138 write_lock_bh(&sock->sk->sk_callback_lock);
139
140 /* done under the callback_lock to serialize with write_space */
141 spin_lock(&rds_tcp_tc_list_lock);
142 list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
143 rds_tcp_tc_count++;
144 spin_unlock(&rds_tcp_tc_list_lock);
145
146 /* accepted sockets need our listen data ready undone */
147 if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
148 sock->sk->sk_data_ready = sock->sk->sk_user_data;
149
150 tc->t_sock = sock;
151 tc->conn = conn;
152 tc->t_orig_data_ready = sock->sk->sk_data_ready;
153 tc->t_orig_write_space = sock->sk->sk_write_space;
154 tc->t_orig_state_change = sock->sk->sk_state_change;
155
156 sock->sk->sk_user_data = conn;
157 sock->sk->sk_data_ready = rds_tcp_data_ready;
158 sock->sk->sk_write_space = rds_tcp_write_space;
159 sock->sk->sk_state_change = rds_tcp_state_change;
160
161 write_unlock_bh(&sock->sk->sk_callback_lock);
162}
163
164static void rds_tcp_tc_info(struct socket *sock, unsigned int len,
165 struct rds_info_iterator *iter,
166 struct rds_info_lengths *lens)
167{
168 struct rds_info_tcp_socket tsinfo;
169 struct rds_tcp_connection *tc;
170 unsigned long flags;
171 struct sockaddr_in sin;
172 int sinlen;
173
174 spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
175
176 if (len / sizeof(tsinfo) < rds_tcp_tc_count)
177 goto out;
178
179 list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
180
181 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 0);
182 tsinfo.local_addr = sin.sin_addr.s_addr;
183 tsinfo.local_port = sin.sin_port;
184 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 1);
185 tsinfo.peer_addr = sin.sin_addr.s_addr;
186 tsinfo.peer_port = sin.sin_port;
187
188 tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
189 tsinfo.data_rem = tc->t_tinc_data_rem;
190 tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
191 tsinfo.last_expected_una = tc->t_last_expected_una;
192 tsinfo.last_seen_una = tc->t_last_seen_una;
193
194 rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
195 }
196
197out:
198 lens->nr = rds_tcp_tc_count;
199 lens->each = sizeof(tsinfo);
200
201 spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
202}
203
204static int rds_tcp_laddr_check(struct net *net, __be32 addr)
205{
206 if (inet_addr_type(net, addr) == RTN_LOCAL)
207 return 0;
208 return -EADDRNOTAVAIL;
209}
210
211static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
212{
213 struct rds_tcp_connection *tc;
214
215 tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
216 if (!tc)
217 return -ENOMEM;
218
219 mutex_init(&tc->t_conn_lock);
220 tc->t_sock = NULL;
221 tc->t_tinc = NULL;
222 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
223 tc->t_tinc_data_rem = 0;
224
225 conn->c_transport_data = tc;
226
227 spin_lock_irq(&rds_tcp_conn_lock);
228 list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
229 spin_unlock_irq(&rds_tcp_conn_lock);
230
231 rdsdebug("alloced tc %p\n", conn->c_transport_data);
232 return 0;
233}
234
235static void rds_tcp_conn_free(void *arg)
236{
237 struct rds_tcp_connection *tc = arg;
238 unsigned long flags;
239 rdsdebug("freeing tc %p\n", tc);
240
241 spin_lock_irqsave(&rds_tcp_conn_lock, flags);
242 list_del(&tc->t_tcp_node);
243 spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
244
245 kmem_cache_free(rds_tcp_conn_slab, tc);
246}
247
248static void rds_tcp_destroy_conns(void)
249{
250 struct rds_tcp_connection *tc, *_tc;
251 LIST_HEAD(tmp_list);
252
253 /* avoid calling conn_destroy with irqs off */
254 spin_lock_irq(&rds_tcp_conn_lock);
255 list_splice(&rds_tcp_conn_list, &tmp_list);
256 INIT_LIST_HEAD(&rds_tcp_conn_list);
257 spin_unlock_irq(&rds_tcp_conn_lock);
258
259 list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
260 if (tc->conn->c_passive)
261 rds_conn_destroy(tc->conn->c_passive);
262 rds_conn_destroy(tc->conn);
263 }
264}
265
266static void rds_tcp_exit(void);
267
268struct rds_transport rds_tcp_transport = {
269 .laddr_check = rds_tcp_laddr_check,
270 .xmit_prepare = rds_tcp_xmit_prepare,
271 .xmit_complete = rds_tcp_xmit_complete,
272 .xmit = rds_tcp_xmit,
273 .recv = rds_tcp_recv,
274 .conn_alloc = rds_tcp_conn_alloc,
275 .conn_free = rds_tcp_conn_free,
276 .conn_connect = rds_tcp_conn_connect,
277 .conn_shutdown = rds_tcp_conn_shutdown,
278 .inc_copy_to_user = rds_tcp_inc_copy_to_user,
279 .inc_free = rds_tcp_inc_free,
280 .stats_info_copy = rds_tcp_stats_info_copy,
281 .exit = rds_tcp_exit,
282 .t_owner = THIS_MODULE,
283 .t_name = "tcp",
284 .t_type = RDS_TRANS_TCP,
285 .t_prefer_loopback = 1,
286};
287
288static int rds_tcp_netid;
289
290/* per-network namespace private data for this module */
291struct rds_tcp_net {
292 struct socket *rds_tcp_listen_sock;
293 struct work_struct rds_tcp_accept_w;
294 struct ctl_table_header *rds_tcp_sysctl;
295 struct ctl_table *ctl_table;
296 int sndbuf_size;
297 int rcvbuf_size;
298};
299
300/* All module specific customizations to the RDS-TCP socket should be done in
301 * rds_tcp_tune() and applied after socket creation.
302 */
303void rds_tcp_tune(struct socket *sock)
304{
305 struct sock *sk = sock->sk;
306 struct net *net = sock_net(sk);
307 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
308
309 rds_tcp_nonagle(sock);
310 lock_sock(sk);
311 if (rtn->sndbuf_size > 0) {
312 sk->sk_sndbuf = rtn->sndbuf_size;
313 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
314 }
315 if (rtn->rcvbuf_size > 0) {
316 sk->sk_sndbuf = rtn->rcvbuf_size;
317 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
318 }
319 release_sock(sk);
320}
321
322static void rds_tcp_accept_worker(struct work_struct *work)
323{
324 struct rds_tcp_net *rtn = container_of(work,
325 struct rds_tcp_net,
326 rds_tcp_accept_w);
327
328 while (rds_tcp_accept_one(rtn->rds_tcp_listen_sock) == 0)
329 cond_resched();
330}
331
332void rds_tcp_accept_work(struct sock *sk)
333{
334 struct net *net = sock_net(sk);
335 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
336
337 queue_work(rds_wq, &rtn->rds_tcp_accept_w);
338}
339
340static __net_init int rds_tcp_init_net(struct net *net)
341{
342 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
343 struct ctl_table *tbl;
344 int err = 0;
345
346 memset(rtn, 0, sizeof(*rtn));
347
348 /* {snd, rcv}buf_size default to 0, which implies we let the
349 * stack pick the value, and permit auto-tuning of buffer size.
350 */
351 if (net == &init_net) {
352 tbl = rds_tcp_sysctl_table;
353 } else {
354 tbl = kmemdup(rds_tcp_sysctl_table,
355 sizeof(rds_tcp_sysctl_table), GFP_KERNEL);
356 if (!tbl) {
357 pr_warn("could not set allocate syctl table\n");
358 return -ENOMEM;
359 }
360 rtn->ctl_table = tbl;
361 }
362 tbl[RDS_TCP_SNDBUF].data = &rtn->sndbuf_size;
363 tbl[RDS_TCP_RCVBUF].data = &rtn->rcvbuf_size;
364 rtn->rds_tcp_sysctl = register_net_sysctl(net, "net/rds/tcp", tbl);
365 if (!rtn->rds_tcp_sysctl) {
366 pr_warn("could not register sysctl\n");
367 err = -ENOMEM;
368 goto fail;
369 }
370 rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net);
371 if (!rtn->rds_tcp_listen_sock) {
372 pr_warn("could not set up listen sock\n");
373 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
374 rtn->rds_tcp_sysctl = NULL;
375 err = -EAFNOSUPPORT;
376 goto fail;
377 }
378 INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
379 return 0;
380
381fail:
382 if (net != &init_net)
383 kfree(tbl);
384 return err;
385}
386
387static void __net_exit rds_tcp_exit_net(struct net *net)
388{
389 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
390
391 if (rtn->rds_tcp_sysctl)
392 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
393
394 if (net != &init_net && rtn->ctl_table)
395 kfree(rtn->ctl_table);
396
397 /* If rds_tcp_exit_net() is called as a result of netns deletion,
398 * the rds_tcp_kill_sock() device notifier would already have cleaned
399 * up the listen socket, thus there is no work to do in this function.
400 *
401 * If rds_tcp_exit_net() is called as a result of module unload,
402 * i.e., due to rds_tcp_exit() -> unregister_pernet_subsys(), then
403 * we do need to clean up the listen socket here.
404 */
405 if (rtn->rds_tcp_listen_sock) {
406 rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
407 rtn->rds_tcp_listen_sock = NULL;
408 flush_work(&rtn->rds_tcp_accept_w);
409 }
410}
411
412static struct pernet_operations rds_tcp_net_ops = {
413 .init = rds_tcp_init_net,
414 .exit = rds_tcp_exit_net,
415 .id = &rds_tcp_netid,
416 .size = sizeof(struct rds_tcp_net),
417};
418
419static void rds_tcp_kill_sock(struct net *net)
420{
421 struct rds_tcp_connection *tc, *_tc;
422 struct sock *sk;
423 LIST_HEAD(tmp_list);
424 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
425
426 rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
427 rtn->rds_tcp_listen_sock = NULL;
428 flush_work(&rtn->rds_tcp_accept_w);
429 spin_lock_irq(&rds_tcp_conn_lock);
430 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
431 struct net *c_net = read_pnet(&tc->conn->c_net);
432
433 if (net != c_net || !tc->t_sock)
434 continue;
435 list_move_tail(&tc->t_tcp_node, &tmp_list);
436 }
437 spin_unlock_irq(&rds_tcp_conn_lock);
438 list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
439 sk = tc->t_sock->sk;
440 sk->sk_prot->disconnect(sk, 0);
441 tcp_done(sk);
442 if (tc->conn->c_passive)
443 rds_conn_destroy(tc->conn->c_passive);
444 rds_conn_destroy(tc->conn);
445 }
446}
447
448static int rds_tcp_dev_event(struct notifier_block *this,
449 unsigned long event, void *ptr)
450{
451 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
452
453 /* rds-tcp registers as a pernet subys, so the ->exit will only
454 * get invoked after network acitivity has quiesced. We need to
455 * clean up all sockets to quiesce network activity, and use
456 * the unregistration of the per-net loopback device as a trigger
457 * to start that cleanup.
458 */
459 if (event == NETDEV_UNREGISTER_FINAL &&
460 dev->ifindex == LOOPBACK_IFINDEX)
461 rds_tcp_kill_sock(dev_net(dev));
462
463 return NOTIFY_DONE;
464}
465
466static struct notifier_block rds_tcp_dev_notifier = {
467 .notifier_call = rds_tcp_dev_event,
468 .priority = -10, /* must be called after other network notifiers */
469};
470
471/* when sysctl is used to modify some kernel socket parameters,this
472 * function resets the RDS connections in that netns so that we can
473 * restart with new parameters. The assumption is that such reset
474 * events are few and far-between.
475 */
476static void rds_tcp_sysctl_reset(struct net *net)
477{
478 struct rds_tcp_connection *tc, *_tc;
479
480 spin_lock_irq(&rds_tcp_conn_lock);
481 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
482 struct net *c_net = read_pnet(&tc->conn->c_net);
483
484 if (net != c_net || !tc->t_sock)
485 continue;
486
487 rds_conn_drop(tc->conn); /* reconnect with new parameters */
488 }
489 spin_unlock_irq(&rds_tcp_conn_lock);
490}
491
492static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
493 void __user *buffer, size_t *lenp,
494 loff_t *fpos)
495{
496 struct net *net = current->nsproxy->net_ns;
497 int err;
498
499 err = proc_dointvec_minmax(ctl, write, buffer, lenp, fpos);
500 if (err < 0) {
501 pr_warn("Invalid input. Must be >= %d\n",
502 *(int *)(ctl->extra1));
503 return err;
504 }
505 if (write)
506 rds_tcp_sysctl_reset(net);
507 return 0;
508}
509
510static void rds_tcp_exit(void)
511{
512 rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
513 unregister_pernet_subsys(&rds_tcp_net_ops);
514 if (unregister_netdevice_notifier(&rds_tcp_dev_notifier))
515 pr_warn("could not unregister rds_tcp_dev_notifier\n");
516 rds_tcp_destroy_conns();
517 rds_trans_unregister(&rds_tcp_transport);
518 rds_tcp_recv_exit();
519 kmem_cache_destroy(rds_tcp_conn_slab);
520}
521module_exit(rds_tcp_exit);
522
523static int rds_tcp_init(void)
524{
525 int ret;
526
527 rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
528 sizeof(struct rds_tcp_connection),
529 0, 0, NULL);
530 if (!rds_tcp_conn_slab) {
531 ret = -ENOMEM;
532 goto out;
533 }
534
535 ret = register_netdevice_notifier(&rds_tcp_dev_notifier);
536 if (ret) {
537 pr_warn("could not register rds_tcp_dev_notifier\n");
538 goto out;
539 }
540
541 ret = register_pernet_subsys(&rds_tcp_net_ops);
542 if (ret)
543 goto out_slab;
544
545 ret = rds_tcp_recv_init();
546 if (ret)
547 goto out_slab;
548
549 ret = rds_trans_register(&rds_tcp_transport);
550 if (ret)
551 goto out_recv;
552
553 rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
554
555 goto out;
556
557out_recv:
558 rds_tcp_recv_exit();
559out_slab:
560 unregister_pernet_subsys(&rds_tcp_net_ops);
561 kmem_cache_destroy(rds_tcp_conn_slab);
562out:
563 return ret;
564}
565module_init(rds_tcp_init);
566
567MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
568MODULE_DESCRIPTION("RDS: TCP transport");
569MODULE_LICENSE("Dual BSD/GPL");
570
1/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
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:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
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.
22 *
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
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/slab.h>
35#include <linux/in.h>
36#include <linux/module.h>
37#include <net/tcp.h>
38#include <net/net_namespace.h>
39#include <net/netns/generic.h>
40
41#include "rds.h"
42#include "tcp.h"
43
44/* only for info exporting */
45static DEFINE_SPINLOCK(rds_tcp_tc_list_lock);
46static LIST_HEAD(rds_tcp_tc_list);
47static unsigned int rds_tcp_tc_count;
48
49/* Track rds_tcp_connection structs so they can be cleaned up */
50static DEFINE_SPINLOCK(rds_tcp_conn_lock);
51static LIST_HEAD(rds_tcp_conn_list);
52static atomic_t rds_tcp_unloading = ATOMIC_INIT(0);
53
54static struct kmem_cache *rds_tcp_conn_slab;
55
56static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
57 void __user *buffer, size_t *lenp,
58 loff_t *fpos);
59
60static int rds_tcp_min_sndbuf = SOCK_MIN_SNDBUF;
61static int rds_tcp_min_rcvbuf = SOCK_MIN_RCVBUF;
62
63static struct ctl_table rds_tcp_sysctl_table[] = {
64#define RDS_TCP_SNDBUF 0
65 {
66 .procname = "rds_tcp_sndbuf",
67 /* data is per-net pointer */
68 .maxlen = sizeof(int),
69 .mode = 0644,
70 .proc_handler = rds_tcp_skbuf_handler,
71 .extra1 = &rds_tcp_min_sndbuf,
72 },
73#define RDS_TCP_RCVBUF 1
74 {
75 .procname = "rds_tcp_rcvbuf",
76 /* data is per-net pointer */
77 .maxlen = sizeof(int),
78 .mode = 0644,
79 .proc_handler = rds_tcp_skbuf_handler,
80 .extra1 = &rds_tcp_min_rcvbuf,
81 },
82 { }
83};
84
85/* doing it this way avoids calling tcp_sk() */
86void rds_tcp_nonagle(struct socket *sock)
87{
88 int val = 1;
89
90 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (void *)&val,
91 sizeof(val));
92}
93
94u32 rds_tcp_write_seq(struct rds_tcp_connection *tc)
95{
96 /* seq# of the last byte of data in tcp send buffer */
97 return tcp_sk(tc->t_sock->sk)->write_seq;
98}
99
100u32 rds_tcp_snd_una(struct rds_tcp_connection *tc)
101{
102 return tcp_sk(tc->t_sock->sk)->snd_una;
103}
104
105void rds_tcp_restore_callbacks(struct socket *sock,
106 struct rds_tcp_connection *tc)
107{
108 rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc);
109 write_lock_bh(&sock->sk->sk_callback_lock);
110
111 /* done under the callback_lock to serialize with write_space */
112 spin_lock(&rds_tcp_tc_list_lock);
113 list_del_init(&tc->t_list_item);
114 rds_tcp_tc_count--;
115 spin_unlock(&rds_tcp_tc_list_lock);
116
117 tc->t_sock = NULL;
118
119 sock->sk->sk_write_space = tc->t_orig_write_space;
120 sock->sk->sk_data_ready = tc->t_orig_data_ready;
121 sock->sk->sk_state_change = tc->t_orig_state_change;
122 sock->sk->sk_user_data = NULL;
123
124 write_unlock_bh(&sock->sk->sk_callback_lock);
125}
126
127/*
128 * rds_tcp_reset_callbacks() switches the to the new sock and
129 * returns the existing tc->t_sock.
130 *
131 * The only functions that set tc->t_sock are rds_tcp_set_callbacks
132 * and rds_tcp_reset_callbacks. Send and receive trust that
133 * it is set. The absence of RDS_CONN_UP bit protects those paths
134 * from being called while it isn't set.
135 */
136void rds_tcp_reset_callbacks(struct socket *sock,
137 struct rds_conn_path *cp)
138{
139 struct rds_tcp_connection *tc = cp->cp_transport_data;
140 struct socket *osock = tc->t_sock;
141
142 if (!osock)
143 goto newsock;
144
145 /* Need to resolve a duelling SYN between peers.
146 * We have an outstanding SYN to this peer, which may
147 * potentially have transitioned to the RDS_CONN_UP state,
148 * so we must quiesce any send threads before resetting
149 * cp_transport_data. We quiesce these threads by setting
150 * cp_state to something other than RDS_CONN_UP, and then
151 * waiting for any existing threads in rds_send_xmit to
152 * complete release_in_xmit(). (Subsequent threads entering
153 * rds_send_xmit() will bail on !rds_conn_up().
154 *
155 * However an incoming syn-ack at this point would end up
156 * marking the conn as RDS_CONN_UP, and would again permit
157 * rds_send_xmi() threads through, so ideally we would
158 * synchronize on RDS_CONN_UP after lock_sock(), but cannot
159 * do that: waiting on !RDS_IN_XMIT after lock_sock() may
160 * end up deadlocking with tcp_sendmsg(), and the RDS_IN_XMIT
161 * would not get set. As a result, we set c_state to
162 * RDS_CONN_RESETTTING, to ensure that rds_tcp_state_change
163 * cannot mark rds_conn_path_up() in the window before lock_sock()
164 */
165 atomic_set(&cp->cp_state, RDS_CONN_RESETTING);
166 wait_event(cp->cp_waitq, !test_bit(RDS_IN_XMIT, &cp->cp_flags));
167 lock_sock(osock->sk);
168 /* reset receive side state for rds_tcp_data_recv() for osock */
169 cancel_delayed_work_sync(&cp->cp_send_w);
170 cancel_delayed_work_sync(&cp->cp_recv_w);
171 if (tc->t_tinc) {
172 rds_inc_put(&tc->t_tinc->ti_inc);
173 tc->t_tinc = NULL;
174 }
175 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
176 tc->t_tinc_data_rem = 0;
177 rds_tcp_restore_callbacks(osock, tc);
178 release_sock(osock->sk);
179 sock_release(osock);
180newsock:
181 rds_send_path_reset(cp);
182 lock_sock(sock->sk);
183 rds_tcp_set_callbacks(sock, cp);
184 release_sock(sock->sk);
185}
186
187/* Add tc to rds_tcp_tc_list and set tc->t_sock. See comments
188 * above rds_tcp_reset_callbacks for notes about synchronization
189 * with data path
190 */
191void rds_tcp_set_callbacks(struct socket *sock, struct rds_conn_path *cp)
192{
193 struct rds_tcp_connection *tc = cp->cp_transport_data;
194
195 rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
196 write_lock_bh(&sock->sk->sk_callback_lock);
197
198 /* done under the callback_lock to serialize with write_space */
199 spin_lock(&rds_tcp_tc_list_lock);
200 list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
201 rds_tcp_tc_count++;
202 spin_unlock(&rds_tcp_tc_list_lock);
203
204 /* accepted sockets need our listen data ready undone */
205 if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
206 sock->sk->sk_data_ready = sock->sk->sk_user_data;
207
208 tc->t_sock = sock;
209 tc->t_cpath = cp;
210 tc->t_orig_data_ready = sock->sk->sk_data_ready;
211 tc->t_orig_write_space = sock->sk->sk_write_space;
212 tc->t_orig_state_change = sock->sk->sk_state_change;
213
214 sock->sk->sk_user_data = cp;
215 sock->sk->sk_data_ready = rds_tcp_data_ready;
216 sock->sk->sk_write_space = rds_tcp_write_space;
217 sock->sk->sk_state_change = rds_tcp_state_change;
218
219 write_unlock_bh(&sock->sk->sk_callback_lock);
220}
221
222static void rds_tcp_tc_info(struct socket *rds_sock, unsigned int len,
223 struct rds_info_iterator *iter,
224 struct rds_info_lengths *lens)
225{
226 struct rds_info_tcp_socket tsinfo;
227 struct rds_tcp_connection *tc;
228 unsigned long flags;
229 struct sockaddr_in sin;
230 struct socket *sock;
231
232 spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
233
234 if (len / sizeof(tsinfo) < rds_tcp_tc_count)
235 goto out;
236
237 list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
238
239 sock = tc->t_sock;
240 if (sock) {
241 sock->ops->getname(sock, (struct sockaddr *)&sin, 0);
242 tsinfo.local_addr = sin.sin_addr.s_addr;
243 tsinfo.local_port = sin.sin_port;
244 sock->ops->getname(sock, (struct sockaddr *)&sin, 1);
245 tsinfo.peer_addr = sin.sin_addr.s_addr;
246 tsinfo.peer_port = sin.sin_port;
247 }
248
249 tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
250 tsinfo.data_rem = tc->t_tinc_data_rem;
251 tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
252 tsinfo.last_expected_una = tc->t_last_expected_una;
253 tsinfo.last_seen_una = tc->t_last_seen_una;
254
255 rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
256 }
257
258out:
259 lens->nr = rds_tcp_tc_count;
260 lens->each = sizeof(tsinfo);
261
262 spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
263}
264
265static int rds_tcp_laddr_check(struct net *net, __be32 addr)
266{
267 if (inet_addr_type(net, addr) == RTN_LOCAL)
268 return 0;
269 return -EADDRNOTAVAIL;
270}
271
272static void rds_tcp_conn_free(void *arg)
273{
274 struct rds_tcp_connection *tc = arg;
275 unsigned long flags;
276
277 rdsdebug("freeing tc %p\n", tc);
278
279 spin_lock_irqsave(&rds_tcp_conn_lock, flags);
280 if (!tc->t_tcp_node_detached)
281 list_del(&tc->t_tcp_node);
282 spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
283
284 kmem_cache_free(rds_tcp_conn_slab, tc);
285}
286
287static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
288{
289 struct rds_tcp_connection *tc;
290 int i, j;
291 int ret = 0;
292
293 for (i = 0; i < RDS_MPATH_WORKERS; i++) {
294 tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
295 if (!tc) {
296 ret = -ENOMEM;
297 goto fail;
298 }
299 mutex_init(&tc->t_conn_path_lock);
300 tc->t_sock = NULL;
301 tc->t_tinc = NULL;
302 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
303 tc->t_tinc_data_rem = 0;
304
305 conn->c_path[i].cp_transport_data = tc;
306 tc->t_cpath = &conn->c_path[i];
307 tc->t_tcp_node_detached = true;
308
309 rdsdebug("rds_conn_path [%d] tc %p\n", i,
310 conn->c_path[i].cp_transport_data);
311 }
312 spin_lock_irq(&rds_tcp_conn_lock);
313 for (i = 0; i < RDS_MPATH_WORKERS; i++) {
314 tc = conn->c_path[i].cp_transport_data;
315 tc->t_tcp_node_detached = false;
316 list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
317 }
318 spin_unlock_irq(&rds_tcp_conn_lock);
319fail:
320 if (ret) {
321 for (j = 0; j < i; j++)
322 rds_tcp_conn_free(conn->c_path[j].cp_transport_data);
323 }
324 return ret;
325}
326
327static bool list_has_conn(struct list_head *list, struct rds_connection *conn)
328{
329 struct rds_tcp_connection *tc, *_tc;
330
331 list_for_each_entry_safe(tc, _tc, list, t_tcp_node) {
332 if (tc->t_cpath->cp_conn == conn)
333 return true;
334 }
335 return false;
336}
337
338static void rds_tcp_set_unloading(void)
339{
340 atomic_set(&rds_tcp_unloading, 1);
341}
342
343static bool rds_tcp_is_unloading(struct rds_connection *conn)
344{
345 return atomic_read(&rds_tcp_unloading) != 0;
346}
347
348static void rds_tcp_destroy_conns(void)
349{
350 struct rds_tcp_connection *tc, *_tc;
351 LIST_HEAD(tmp_list);
352
353 /* avoid calling conn_destroy with irqs off */
354 spin_lock_irq(&rds_tcp_conn_lock);
355 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
356 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn))
357 list_move_tail(&tc->t_tcp_node, &tmp_list);
358 }
359 spin_unlock_irq(&rds_tcp_conn_lock);
360
361 list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node)
362 rds_conn_destroy(tc->t_cpath->cp_conn);
363}
364
365static void rds_tcp_exit(void);
366
367struct rds_transport rds_tcp_transport = {
368 .laddr_check = rds_tcp_laddr_check,
369 .xmit_path_prepare = rds_tcp_xmit_path_prepare,
370 .xmit_path_complete = rds_tcp_xmit_path_complete,
371 .xmit = rds_tcp_xmit,
372 .recv_path = rds_tcp_recv_path,
373 .conn_alloc = rds_tcp_conn_alloc,
374 .conn_free = rds_tcp_conn_free,
375 .conn_path_connect = rds_tcp_conn_path_connect,
376 .conn_path_shutdown = rds_tcp_conn_path_shutdown,
377 .inc_copy_to_user = rds_tcp_inc_copy_to_user,
378 .inc_free = rds_tcp_inc_free,
379 .stats_info_copy = rds_tcp_stats_info_copy,
380 .exit = rds_tcp_exit,
381 .t_owner = THIS_MODULE,
382 .t_name = "tcp",
383 .t_type = RDS_TRANS_TCP,
384 .t_prefer_loopback = 1,
385 .t_mp_capable = 1,
386 .t_unloading = rds_tcp_is_unloading,
387};
388
389static unsigned int rds_tcp_netid;
390
391/* per-network namespace private data for this module */
392struct rds_tcp_net {
393 struct socket *rds_tcp_listen_sock;
394 struct work_struct rds_tcp_accept_w;
395 struct ctl_table_header *rds_tcp_sysctl;
396 struct ctl_table *ctl_table;
397 int sndbuf_size;
398 int rcvbuf_size;
399};
400
401/* All module specific customizations to the RDS-TCP socket should be done in
402 * rds_tcp_tune() and applied after socket creation.
403 */
404void rds_tcp_tune(struct socket *sock)
405{
406 struct sock *sk = sock->sk;
407 struct net *net = sock_net(sk);
408 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
409
410 rds_tcp_nonagle(sock);
411 lock_sock(sk);
412 if (rtn->sndbuf_size > 0) {
413 sk->sk_sndbuf = rtn->sndbuf_size;
414 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
415 }
416 if (rtn->rcvbuf_size > 0) {
417 sk->sk_sndbuf = rtn->rcvbuf_size;
418 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
419 }
420 release_sock(sk);
421}
422
423static void rds_tcp_accept_worker(struct work_struct *work)
424{
425 struct rds_tcp_net *rtn = container_of(work,
426 struct rds_tcp_net,
427 rds_tcp_accept_w);
428
429 while (rds_tcp_accept_one(rtn->rds_tcp_listen_sock) == 0)
430 cond_resched();
431}
432
433void rds_tcp_accept_work(struct sock *sk)
434{
435 struct net *net = sock_net(sk);
436 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
437
438 queue_work(rds_wq, &rtn->rds_tcp_accept_w);
439}
440
441static __net_init int rds_tcp_init_net(struct net *net)
442{
443 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
444 struct ctl_table *tbl;
445 int err = 0;
446
447 memset(rtn, 0, sizeof(*rtn));
448
449 /* {snd, rcv}buf_size default to 0, which implies we let the
450 * stack pick the value, and permit auto-tuning of buffer size.
451 */
452 if (net == &init_net) {
453 tbl = rds_tcp_sysctl_table;
454 } else {
455 tbl = kmemdup(rds_tcp_sysctl_table,
456 sizeof(rds_tcp_sysctl_table), GFP_KERNEL);
457 if (!tbl) {
458 pr_warn("could not set allocate syctl table\n");
459 return -ENOMEM;
460 }
461 rtn->ctl_table = tbl;
462 }
463 tbl[RDS_TCP_SNDBUF].data = &rtn->sndbuf_size;
464 tbl[RDS_TCP_RCVBUF].data = &rtn->rcvbuf_size;
465 rtn->rds_tcp_sysctl = register_net_sysctl(net, "net/rds/tcp", tbl);
466 if (!rtn->rds_tcp_sysctl) {
467 pr_warn("could not register sysctl\n");
468 err = -ENOMEM;
469 goto fail;
470 }
471 rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net);
472 if (!rtn->rds_tcp_listen_sock) {
473 pr_warn("could not set up listen sock\n");
474 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
475 rtn->rds_tcp_sysctl = NULL;
476 err = -EAFNOSUPPORT;
477 goto fail;
478 }
479 INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
480 return 0;
481
482fail:
483 if (net != &init_net)
484 kfree(tbl);
485 return err;
486}
487
488static void rds_tcp_kill_sock(struct net *net)
489{
490 struct rds_tcp_connection *tc, *_tc;
491 LIST_HEAD(tmp_list);
492 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
493 struct socket *lsock = rtn->rds_tcp_listen_sock;
494
495 rtn->rds_tcp_listen_sock = NULL;
496 rds_tcp_listen_stop(lsock, &rtn->rds_tcp_accept_w);
497 spin_lock_irq(&rds_tcp_conn_lock);
498 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
499 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
500
501 if (net != c_net || !tc->t_sock)
502 continue;
503 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn)) {
504 list_move_tail(&tc->t_tcp_node, &tmp_list);
505 } else {
506 list_del(&tc->t_tcp_node);
507 tc->t_tcp_node_detached = true;
508 }
509 }
510 spin_unlock_irq(&rds_tcp_conn_lock);
511 list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node)
512 rds_conn_destroy(tc->t_cpath->cp_conn);
513}
514
515static void __net_exit rds_tcp_exit_net(struct net *net)
516{
517 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
518
519 rds_tcp_kill_sock(net);
520
521 if (rtn->rds_tcp_sysctl)
522 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
523
524 if (net != &init_net && rtn->ctl_table)
525 kfree(rtn->ctl_table);
526}
527
528static struct pernet_operations rds_tcp_net_ops = {
529 .init = rds_tcp_init_net,
530 .exit = rds_tcp_exit_net,
531 .id = &rds_tcp_netid,
532 .size = sizeof(struct rds_tcp_net),
533};
534
535void *rds_tcp_listen_sock_def_readable(struct net *net)
536{
537 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
538 struct socket *lsock = rtn->rds_tcp_listen_sock;
539
540 if (!lsock)
541 return NULL;
542
543 return lsock->sk->sk_user_data;
544}
545
546/* when sysctl is used to modify some kernel socket parameters,this
547 * function resets the RDS connections in that netns so that we can
548 * restart with new parameters. The assumption is that such reset
549 * events are few and far-between.
550 */
551static void rds_tcp_sysctl_reset(struct net *net)
552{
553 struct rds_tcp_connection *tc, *_tc;
554
555 spin_lock_irq(&rds_tcp_conn_lock);
556 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
557 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
558
559 if (net != c_net || !tc->t_sock)
560 continue;
561
562 /* reconnect with new parameters */
563 rds_conn_path_drop(tc->t_cpath, false);
564 }
565 spin_unlock_irq(&rds_tcp_conn_lock);
566}
567
568static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
569 void __user *buffer, size_t *lenp,
570 loff_t *fpos)
571{
572 struct net *net = current->nsproxy->net_ns;
573 int err;
574
575 err = proc_dointvec_minmax(ctl, write, buffer, lenp, fpos);
576 if (err < 0) {
577 pr_warn("Invalid input. Must be >= %d\n",
578 *(int *)(ctl->extra1));
579 return err;
580 }
581 if (write)
582 rds_tcp_sysctl_reset(net);
583 return 0;
584}
585
586static void rds_tcp_exit(void)
587{
588 rds_tcp_set_unloading();
589 synchronize_rcu();
590 rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
591 unregister_pernet_device(&rds_tcp_net_ops);
592 rds_tcp_destroy_conns();
593 rds_trans_unregister(&rds_tcp_transport);
594 rds_tcp_recv_exit();
595 kmem_cache_destroy(rds_tcp_conn_slab);
596}
597module_exit(rds_tcp_exit);
598
599static int rds_tcp_init(void)
600{
601 int ret;
602
603 rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
604 sizeof(struct rds_tcp_connection),
605 0, 0, NULL);
606 if (!rds_tcp_conn_slab) {
607 ret = -ENOMEM;
608 goto out;
609 }
610
611 ret = rds_tcp_recv_init();
612 if (ret)
613 goto out_slab;
614
615 ret = register_pernet_device(&rds_tcp_net_ops);
616 if (ret)
617 goto out_recv;
618
619 rds_trans_register(&rds_tcp_transport);
620
621 rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
622
623 goto out;
624out_recv:
625 rds_tcp_recv_exit();
626out_slab:
627 kmem_cache_destroy(rds_tcp_conn_slab);
628out:
629 return ret;
630}
631module_init(rds_tcp_init);
632
633MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
634MODULE_DESCRIPTION("RDS: TCP transport");
635MODULE_LICENSE("Dual BSD/GPL");
636