Loading...
1/*
2 * Copyright (c) 2006, 2018 Oracle and/or its affiliates. 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#include <net/addrconf.h>
41
42#include "rds.h"
43#include "tcp.h"
44
45/* only for info exporting */
46static DEFINE_SPINLOCK(rds_tcp_tc_list_lock);
47static LIST_HEAD(rds_tcp_tc_list);
48
49/* rds_tcp_tc_count counts only IPv4 connections.
50 * rds6_tcp_tc_count counts both IPv4 and IPv6 connections.
51 */
52static unsigned int rds_tcp_tc_count;
53#if IS_ENABLED(CONFIG_IPV6)
54static unsigned int rds6_tcp_tc_count;
55#endif
56
57/* Track rds_tcp_connection structs so they can be cleaned up */
58static DEFINE_SPINLOCK(rds_tcp_conn_lock);
59static LIST_HEAD(rds_tcp_conn_list);
60static atomic_t rds_tcp_unloading = ATOMIC_INIT(0);
61
62static struct kmem_cache *rds_tcp_conn_slab;
63
64static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
65 void *buffer, size_t *lenp, loff_t *fpos);
66
67static int rds_tcp_min_sndbuf = SOCK_MIN_SNDBUF;
68static int rds_tcp_min_rcvbuf = SOCK_MIN_RCVBUF;
69
70static struct ctl_table rds_tcp_sysctl_table[] = {
71#define RDS_TCP_SNDBUF 0
72 {
73 .procname = "rds_tcp_sndbuf",
74 /* data is per-net pointer */
75 .maxlen = sizeof(int),
76 .mode = 0644,
77 .proc_handler = rds_tcp_skbuf_handler,
78 .extra1 = &rds_tcp_min_sndbuf,
79 },
80#define RDS_TCP_RCVBUF 1
81 {
82 .procname = "rds_tcp_rcvbuf",
83 /* data is per-net pointer */
84 .maxlen = sizeof(int),
85 .mode = 0644,
86 .proc_handler = rds_tcp_skbuf_handler,
87 .extra1 = &rds_tcp_min_rcvbuf,
88 },
89 { }
90};
91
92u32 rds_tcp_write_seq(struct rds_tcp_connection *tc)
93{
94 /* seq# of the last byte of data in tcp send buffer */
95 return tcp_sk(tc->t_sock->sk)->write_seq;
96}
97
98u32 rds_tcp_snd_una(struct rds_tcp_connection *tc)
99{
100 return tcp_sk(tc->t_sock->sk)->snd_una;
101}
102
103void rds_tcp_restore_callbacks(struct socket *sock,
104 struct rds_tcp_connection *tc)
105{
106 rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc);
107 write_lock_bh(&sock->sk->sk_callback_lock);
108
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);
112#if IS_ENABLED(CONFIG_IPV6)
113 rds6_tcp_tc_count--;
114#endif
115 if (!tc->t_cpath->cp_conn->c_isv6)
116 rds_tcp_tc_count--;
117 spin_unlock(&rds_tcp_tc_list_lock);
118
119 tc->t_sock = NULL;
120
121 sock->sk->sk_write_space = tc->t_orig_write_space;
122 sock->sk->sk_data_ready = tc->t_orig_data_ready;
123 sock->sk->sk_state_change = tc->t_orig_state_change;
124 sock->sk->sk_user_data = NULL;
125
126 write_unlock_bh(&sock->sk->sk_callback_lock);
127}
128
129/*
130 * rds_tcp_reset_callbacks() switches the to the new sock and
131 * returns the existing tc->t_sock.
132 *
133 * The only functions that set tc->t_sock are rds_tcp_set_callbacks
134 * and rds_tcp_reset_callbacks. Send and receive trust that
135 * it is set. The absence of RDS_CONN_UP bit protects those paths
136 * from being called while it isn't set.
137 */
138void rds_tcp_reset_callbacks(struct socket *sock,
139 struct rds_conn_path *cp)
140{
141 struct rds_tcp_connection *tc = cp->cp_transport_data;
142 struct socket *osock = tc->t_sock;
143
144 if (!osock)
145 goto newsock;
146
147 /* Need to resolve a duelling SYN between peers.
148 * We have an outstanding SYN to this peer, which may
149 * potentially have transitioned to the RDS_CONN_UP state,
150 * so we must quiesce any send threads before resetting
151 * cp_transport_data. We quiesce these threads by setting
152 * cp_state to something other than RDS_CONN_UP, and then
153 * waiting for any existing threads in rds_send_xmit to
154 * complete release_in_xmit(). (Subsequent threads entering
155 * rds_send_xmit() will bail on !rds_conn_up().
156 *
157 * However an incoming syn-ack at this point would end up
158 * marking the conn as RDS_CONN_UP, and would again permit
159 * rds_send_xmi() threads through, so ideally we would
160 * synchronize on RDS_CONN_UP after lock_sock(), but cannot
161 * do that: waiting on !RDS_IN_XMIT after lock_sock() may
162 * end up deadlocking with tcp_sendmsg(), and the RDS_IN_XMIT
163 * would not get set. As a result, we set c_state to
164 * RDS_CONN_RESETTTING, to ensure that rds_tcp_state_change
165 * cannot mark rds_conn_path_up() in the window before lock_sock()
166 */
167 atomic_set(&cp->cp_state, RDS_CONN_RESETTING);
168 wait_event(cp->cp_waitq, !test_bit(RDS_IN_XMIT, &cp->cp_flags));
169 lock_sock(osock->sk);
170 /* reset receive side state for rds_tcp_data_recv() for osock */
171 cancel_delayed_work_sync(&cp->cp_send_w);
172 cancel_delayed_work_sync(&cp->cp_recv_w);
173 if (tc->t_tinc) {
174 rds_inc_put(&tc->t_tinc->ti_inc);
175 tc->t_tinc = NULL;
176 }
177 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
178 tc->t_tinc_data_rem = 0;
179 rds_tcp_restore_callbacks(osock, tc);
180 release_sock(osock->sk);
181 sock_release(osock);
182newsock:
183 rds_send_path_reset(cp);
184 lock_sock(sock->sk);
185 rds_tcp_set_callbacks(sock, cp);
186 release_sock(sock->sk);
187}
188
189/* Add tc to rds_tcp_tc_list and set tc->t_sock. See comments
190 * above rds_tcp_reset_callbacks for notes about synchronization
191 * with data path
192 */
193void rds_tcp_set_callbacks(struct socket *sock, struct rds_conn_path *cp)
194{
195 struct rds_tcp_connection *tc = cp->cp_transport_data;
196
197 rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
198 write_lock_bh(&sock->sk->sk_callback_lock);
199
200 /* done under the callback_lock to serialize with write_space */
201 spin_lock(&rds_tcp_tc_list_lock);
202 list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
203#if IS_ENABLED(CONFIG_IPV6)
204 rds6_tcp_tc_count++;
205#endif
206 if (!tc->t_cpath->cp_conn->c_isv6)
207 rds_tcp_tc_count++;
208 spin_unlock(&rds_tcp_tc_list_lock);
209
210 /* accepted sockets need our listen data ready undone */
211 if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
212 sock->sk->sk_data_ready = sock->sk->sk_user_data;
213
214 tc->t_sock = sock;
215 tc->t_cpath = cp;
216 tc->t_orig_data_ready = sock->sk->sk_data_ready;
217 tc->t_orig_write_space = sock->sk->sk_write_space;
218 tc->t_orig_state_change = sock->sk->sk_state_change;
219
220 sock->sk->sk_user_data = cp;
221 sock->sk->sk_data_ready = rds_tcp_data_ready;
222 sock->sk->sk_write_space = rds_tcp_write_space;
223 sock->sk->sk_state_change = rds_tcp_state_change;
224
225 write_unlock_bh(&sock->sk->sk_callback_lock);
226}
227
228/* Handle RDS_INFO_TCP_SOCKETS socket option. It only returns IPv4
229 * connections for backward compatibility.
230 */
231static void rds_tcp_tc_info(struct socket *rds_sock, unsigned int len,
232 struct rds_info_iterator *iter,
233 struct rds_info_lengths *lens)
234{
235 struct rds_info_tcp_socket tsinfo;
236 struct rds_tcp_connection *tc;
237 unsigned long flags;
238
239 spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
240
241 if (len / sizeof(tsinfo) < rds_tcp_tc_count)
242 goto out;
243
244 list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
245 struct inet_sock *inet = inet_sk(tc->t_sock->sk);
246
247 if (tc->t_cpath->cp_conn->c_isv6)
248 continue;
249
250 tsinfo.local_addr = inet->inet_saddr;
251 tsinfo.local_port = inet->inet_sport;
252 tsinfo.peer_addr = inet->inet_daddr;
253 tsinfo.peer_port = inet->inet_dport;
254
255 tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
256 tsinfo.data_rem = tc->t_tinc_data_rem;
257 tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
258 tsinfo.last_expected_una = tc->t_last_expected_una;
259 tsinfo.last_seen_una = tc->t_last_seen_una;
260 tsinfo.tos = tc->t_cpath->cp_conn->c_tos;
261
262 rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
263 }
264
265out:
266 lens->nr = rds_tcp_tc_count;
267 lens->each = sizeof(tsinfo);
268
269 spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
270}
271
272#if IS_ENABLED(CONFIG_IPV6)
273/* Handle RDS6_INFO_TCP_SOCKETS socket option. It returns both IPv4 and
274 * IPv6 connections. IPv4 connection address is returned in an IPv4 mapped
275 * address.
276 */
277static void rds6_tcp_tc_info(struct socket *sock, unsigned int len,
278 struct rds_info_iterator *iter,
279 struct rds_info_lengths *lens)
280{
281 struct rds6_info_tcp_socket tsinfo6;
282 struct rds_tcp_connection *tc;
283 unsigned long flags;
284
285 spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
286
287 if (len / sizeof(tsinfo6) < rds6_tcp_tc_count)
288 goto out;
289
290 list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
291 struct sock *sk = tc->t_sock->sk;
292 struct inet_sock *inet = inet_sk(sk);
293
294 tsinfo6.local_addr = sk->sk_v6_rcv_saddr;
295 tsinfo6.local_port = inet->inet_sport;
296 tsinfo6.peer_addr = sk->sk_v6_daddr;
297 tsinfo6.peer_port = inet->inet_dport;
298
299 tsinfo6.hdr_rem = tc->t_tinc_hdr_rem;
300 tsinfo6.data_rem = tc->t_tinc_data_rem;
301 tsinfo6.last_sent_nxt = tc->t_last_sent_nxt;
302 tsinfo6.last_expected_una = tc->t_last_expected_una;
303 tsinfo6.last_seen_una = tc->t_last_seen_una;
304
305 rds_info_copy(iter, &tsinfo6, sizeof(tsinfo6));
306 }
307
308out:
309 lens->nr = rds6_tcp_tc_count;
310 lens->each = sizeof(tsinfo6);
311
312 spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
313}
314#endif
315
316static int rds_tcp_laddr_check(struct net *net, const struct in6_addr *addr,
317 __u32 scope_id)
318{
319 struct net_device *dev = NULL;
320#if IS_ENABLED(CONFIG_IPV6)
321 int ret;
322#endif
323
324 if (ipv6_addr_v4mapped(addr)) {
325 if (inet_addr_type(net, addr->s6_addr32[3]) == RTN_LOCAL)
326 return 0;
327 return -EADDRNOTAVAIL;
328 }
329
330 /* If the scope_id is specified, check only those addresses
331 * hosted on the specified interface.
332 */
333 if (scope_id != 0) {
334 rcu_read_lock();
335 dev = dev_get_by_index_rcu(net, scope_id);
336 /* scope_id is not valid... */
337 if (!dev) {
338 rcu_read_unlock();
339 return -EADDRNOTAVAIL;
340 }
341 rcu_read_unlock();
342 }
343#if IS_ENABLED(CONFIG_IPV6)
344 ret = ipv6_chk_addr(net, addr, dev, 0);
345 if (ret)
346 return 0;
347#endif
348 return -EADDRNOTAVAIL;
349}
350
351static void rds_tcp_conn_free(void *arg)
352{
353 struct rds_tcp_connection *tc = arg;
354 unsigned long flags;
355
356 rdsdebug("freeing tc %p\n", tc);
357
358 spin_lock_irqsave(&rds_tcp_conn_lock, flags);
359 if (!tc->t_tcp_node_detached)
360 list_del(&tc->t_tcp_node);
361 spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
362
363 kmem_cache_free(rds_tcp_conn_slab, tc);
364}
365
366static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
367{
368 struct rds_tcp_connection *tc;
369 int i, j;
370 int ret = 0;
371
372 for (i = 0; i < RDS_MPATH_WORKERS; i++) {
373 tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
374 if (!tc) {
375 ret = -ENOMEM;
376 goto fail;
377 }
378 mutex_init(&tc->t_conn_path_lock);
379 tc->t_sock = NULL;
380 tc->t_tinc = NULL;
381 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
382 tc->t_tinc_data_rem = 0;
383
384 conn->c_path[i].cp_transport_data = tc;
385 tc->t_cpath = &conn->c_path[i];
386 tc->t_tcp_node_detached = true;
387
388 rdsdebug("rds_conn_path [%d] tc %p\n", i,
389 conn->c_path[i].cp_transport_data);
390 }
391 spin_lock_irq(&rds_tcp_conn_lock);
392 for (i = 0; i < RDS_MPATH_WORKERS; i++) {
393 tc = conn->c_path[i].cp_transport_data;
394 tc->t_tcp_node_detached = false;
395 list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
396 }
397 spin_unlock_irq(&rds_tcp_conn_lock);
398fail:
399 if (ret) {
400 for (j = 0; j < i; j++)
401 rds_tcp_conn_free(conn->c_path[j].cp_transport_data);
402 }
403 return ret;
404}
405
406static bool list_has_conn(struct list_head *list, struct rds_connection *conn)
407{
408 struct rds_tcp_connection *tc, *_tc;
409
410 list_for_each_entry_safe(tc, _tc, list, t_tcp_node) {
411 if (tc->t_cpath->cp_conn == conn)
412 return true;
413 }
414 return false;
415}
416
417static void rds_tcp_set_unloading(void)
418{
419 atomic_set(&rds_tcp_unloading, 1);
420}
421
422static bool rds_tcp_is_unloading(struct rds_connection *conn)
423{
424 return atomic_read(&rds_tcp_unloading) != 0;
425}
426
427static void rds_tcp_destroy_conns(void)
428{
429 struct rds_tcp_connection *tc, *_tc;
430 LIST_HEAD(tmp_list);
431
432 /* avoid calling conn_destroy with irqs off */
433 spin_lock_irq(&rds_tcp_conn_lock);
434 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
435 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn))
436 list_move_tail(&tc->t_tcp_node, &tmp_list);
437 }
438 spin_unlock_irq(&rds_tcp_conn_lock);
439
440 list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node)
441 rds_conn_destroy(tc->t_cpath->cp_conn);
442}
443
444static void rds_tcp_exit(void);
445
446static u8 rds_tcp_get_tos_map(u8 tos)
447{
448 /* all user tos mapped to default 0 for TCP transport */
449 return 0;
450}
451
452struct rds_transport rds_tcp_transport = {
453 .laddr_check = rds_tcp_laddr_check,
454 .xmit_path_prepare = rds_tcp_xmit_path_prepare,
455 .xmit_path_complete = rds_tcp_xmit_path_complete,
456 .xmit = rds_tcp_xmit,
457 .recv_path = rds_tcp_recv_path,
458 .conn_alloc = rds_tcp_conn_alloc,
459 .conn_free = rds_tcp_conn_free,
460 .conn_path_connect = rds_tcp_conn_path_connect,
461 .conn_path_shutdown = rds_tcp_conn_path_shutdown,
462 .inc_copy_to_user = rds_tcp_inc_copy_to_user,
463 .inc_free = rds_tcp_inc_free,
464 .stats_info_copy = rds_tcp_stats_info_copy,
465 .exit = rds_tcp_exit,
466 .get_tos_map = rds_tcp_get_tos_map,
467 .t_owner = THIS_MODULE,
468 .t_name = "tcp",
469 .t_type = RDS_TRANS_TCP,
470 .t_prefer_loopback = 1,
471 .t_mp_capable = 1,
472 .t_unloading = rds_tcp_is_unloading,
473};
474
475static unsigned int rds_tcp_netid;
476
477/* per-network namespace private data for this module */
478struct rds_tcp_net {
479 struct socket *rds_tcp_listen_sock;
480 struct work_struct rds_tcp_accept_w;
481 struct ctl_table_header *rds_tcp_sysctl;
482 struct ctl_table *ctl_table;
483 int sndbuf_size;
484 int rcvbuf_size;
485};
486
487/* All module specific customizations to the RDS-TCP socket should be done in
488 * rds_tcp_tune() and applied after socket creation.
489 */
490void rds_tcp_tune(struct socket *sock)
491{
492 struct sock *sk = sock->sk;
493 struct net *net = sock_net(sk);
494 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
495
496 tcp_sock_set_nodelay(sock->sk);
497 lock_sock(sk);
498 if (rtn->sndbuf_size > 0) {
499 sk->sk_sndbuf = rtn->sndbuf_size;
500 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
501 }
502 if (rtn->rcvbuf_size > 0) {
503 sk->sk_sndbuf = rtn->rcvbuf_size;
504 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
505 }
506 release_sock(sk);
507}
508
509static void rds_tcp_accept_worker(struct work_struct *work)
510{
511 struct rds_tcp_net *rtn = container_of(work,
512 struct rds_tcp_net,
513 rds_tcp_accept_w);
514
515 while (rds_tcp_accept_one(rtn->rds_tcp_listen_sock) == 0)
516 cond_resched();
517}
518
519void rds_tcp_accept_work(struct sock *sk)
520{
521 struct net *net = sock_net(sk);
522 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
523
524 queue_work(rds_wq, &rtn->rds_tcp_accept_w);
525}
526
527static __net_init int rds_tcp_init_net(struct net *net)
528{
529 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
530 struct ctl_table *tbl;
531 int err = 0;
532
533 memset(rtn, 0, sizeof(*rtn));
534
535 /* {snd, rcv}buf_size default to 0, which implies we let the
536 * stack pick the value, and permit auto-tuning of buffer size.
537 */
538 if (net == &init_net) {
539 tbl = rds_tcp_sysctl_table;
540 } else {
541 tbl = kmemdup(rds_tcp_sysctl_table,
542 sizeof(rds_tcp_sysctl_table), GFP_KERNEL);
543 if (!tbl) {
544 pr_warn("could not set allocate sysctl table\n");
545 return -ENOMEM;
546 }
547 rtn->ctl_table = tbl;
548 }
549 tbl[RDS_TCP_SNDBUF].data = &rtn->sndbuf_size;
550 tbl[RDS_TCP_RCVBUF].data = &rtn->rcvbuf_size;
551 rtn->rds_tcp_sysctl = register_net_sysctl(net, "net/rds/tcp", tbl);
552 if (!rtn->rds_tcp_sysctl) {
553 pr_warn("could not register sysctl\n");
554 err = -ENOMEM;
555 goto fail;
556 }
557
558#if IS_ENABLED(CONFIG_IPV6)
559 rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net, true);
560#else
561 rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net, false);
562#endif
563 if (!rtn->rds_tcp_listen_sock) {
564 pr_warn("could not set up IPv6 listen sock\n");
565
566#if IS_ENABLED(CONFIG_IPV6)
567 /* Try IPv4 as some systems disable IPv6 */
568 rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net, false);
569 if (!rtn->rds_tcp_listen_sock) {
570#endif
571 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
572 rtn->rds_tcp_sysctl = NULL;
573 err = -EAFNOSUPPORT;
574 goto fail;
575#if IS_ENABLED(CONFIG_IPV6)
576 }
577#endif
578 }
579 INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
580 return 0;
581
582fail:
583 if (net != &init_net)
584 kfree(tbl);
585 return err;
586}
587
588static void rds_tcp_kill_sock(struct net *net)
589{
590 struct rds_tcp_connection *tc, *_tc;
591 LIST_HEAD(tmp_list);
592 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
593 struct socket *lsock = rtn->rds_tcp_listen_sock;
594
595 rtn->rds_tcp_listen_sock = NULL;
596 rds_tcp_listen_stop(lsock, &rtn->rds_tcp_accept_w);
597 spin_lock_irq(&rds_tcp_conn_lock);
598 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
599 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
600
601 if (net != c_net)
602 continue;
603 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn)) {
604 list_move_tail(&tc->t_tcp_node, &tmp_list);
605 } else {
606 list_del(&tc->t_tcp_node);
607 tc->t_tcp_node_detached = true;
608 }
609 }
610 spin_unlock_irq(&rds_tcp_conn_lock);
611 list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node)
612 rds_conn_destroy(tc->t_cpath->cp_conn);
613}
614
615static void __net_exit rds_tcp_exit_net(struct net *net)
616{
617 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
618
619 rds_tcp_kill_sock(net);
620
621 if (rtn->rds_tcp_sysctl)
622 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
623
624 if (net != &init_net)
625 kfree(rtn->ctl_table);
626}
627
628static struct pernet_operations rds_tcp_net_ops = {
629 .init = rds_tcp_init_net,
630 .exit = rds_tcp_exit_net,
631 .id = &rds_tcp_netid,
632 .size = sizeof(struct rds_tcp_net),
633};
634
635void *rds_tcp_listen_sock_def_readable(struct net *net)
636{
637 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
638 struct socket *lsock = rtn->rds_tcp_listen_sock;
639
640 if (!lsock)
641 return NULL;
642
643 return lsock->sk->sk_user_data;
644}
645
646/* when sysctl is used to modify some kernel socket parameters,this
647 * function resets the RDS connections in that netns so that we can
648 * restart with new parameters. The assumption is that such reset
649 * events are few and far-between.
650 */
651static void rds_tcp_sysctl_reset(struct net *net)
652{
653 struct rds_tcp_connection *tc, *_tc;
654
655 spin_lock_irq(&rds_tcp_conn_lock);
656 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
657 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
658
659 if (net != c_net || !tc->t_sock)
660 continue;
661
662 /* reconnect with new parameters */
663 rds_conn_path_drop(tc->t_cpath, false);
664 }
665 spin_unlock_irq(&rds_tcp_conn_lock);
666}
667
668static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
669 void *buffer, size_t *lenp, loff_t *fpos)
670{
671 struct net *net = current->nsproxy->net_ns;
672 int err;
673
674 err = proc_dointvec_minmax(ctl, write, buffer, lenp, fpos);
675 if (err < 0) {
676 pr_warn("Invalid input. Must be >= %d\n",
677 *(int *)(ctl->extra1));
678 return err;
679 }
680 if (write)
681 rds_tcp_sysctl_reset(net);
682 return 0;
683}
684
685static void rds_tcp_exit(void)
686{
687 rds_tcp_set_unloading();
688 synchronize_rcu();
689 rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
690#if IS_ENABLED(CONFIG_IPV6)
691 rds_info_deregister_func(RDS6_INFO_TCP_SOCKETS, rds6_tcp_tc_info);
692#endif
693 unregister_pernet_device(&rds_tcp_net_ops);
694 rds_tcp_destroy_conns();
695 rds_trans_unregister(&rds_tcp_transport);
696 rds_tcp_recv_exit();
697 kmem_cache_destroy(rds_tcp_conn_slab);
698}
699module_exit(rds_tcp_exit);
700
701static int rds_tcp_init(void)
702{
703 int ret;
704
705 rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
706 sizeof(struct rds_tcp_connection),
707 0, 0, NULL);
708 if (!rds_tcp_conn_slab) {
709 ret = -ENOMEM;
710 goto out;
711 }
712
713 ret = rds_tcp_recv_init();
714 if (ret)
715 goto out_slab;
716
717 ret = register_pernet_device(&rds_tcp_net_ops);
718 if (ret)
719 goto out_recv;
720
721 rds_trans_register(&rds_tcp_transport);
722
723 rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
724#if IS_ENABLED(CONFIG_IPV6)
725 rds_info_register_func(RDS6_INFO_TCP_SOCKETS, rds6_tcp_tc_info);
726#endif
727
728 goto out;
729out_recv:
730 rds_tcp_recv_exit();
731out_slab:
732 kmem_cache_destroy(rds_tcp_conn_slab);
733out:
734 return ret;
735}
736module_init(rds_tcp_init);
737
738MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
739MODULE_DESCRIPTION("RDS: TCP transport");
740MODULE_LICENSE("Dual BSD/GPL");
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 <net/tcp.h>
37
38#include "rds.h"
39#include "tcp.h"
40
41/* only for info exporting */
42static DEFINE_SPINLOCK(rds_tcp_tc_list_lock);
43static LIST_HEAD(rds_tcp_tc_list);
44static unsigned int rds_tcp_tc_count;
45
46/* Track rds_tcp_connection structs so they can be cleaned up */
47static DEFINE_SPINLOCK(rds_tcp_conn_lock);
48static LIST_HEAD(rds_tcp_conn_list);
49
50static struct kmem_cache *rds_tcp_conn_slab;
51
52#define RDS_TCP_DEFAULT_BUFSIZE (128 * 1024)
53
54/* doing it this way avoids calling tcp_sk() */
55void rds_tcp_nonagle(struct socket *sock)
56{
57 mm_segment_t oldfs = get_fs();
58 int val = 1;
59
60 set_fs(KERNEL_DS);
61 sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY, (char __user *)&val,
62 sizeof(val));
63 set_fs(oldfs);
64}
65
66void rds_tcp_tune(struct socket *sock)
67{
68 struct sock *sk = sock->sk;
69
70 rds_tcp_nonagle(sock);
71
72 /*
73 * We're trying to saturate gigabit with the default,
74 * see svc_sock_setbufsize().
75 */
76 lock_sock(sk);
77 sk->sk_sndbuf = RDS_TCP_DEFAULT_BUFSIZE;
78 sk->sk_rcvbuf = RDS_TCP_DEFAULT_BUFSIZE;
79 sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
80 release_sock(sk);
81}
82
83u32 rds_tcp_snd_nxt(struct rds_tcp_connection *tc)
84{
85 return tcp_sk(tc->t_sock->sk)->snd_nxt;
86}
87
88u32 rds_tcp_snd_una(struct rds_tcp_connection *tc)
89{
90 return tcp_sk(tc->t_sock->sk)->snd_una;
91}
92
93void rds_tcp_restore_callbacks(struct socket *sock,
94 struct rds_tcp_connection *tc)
95{
96 rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc);
97 write_lock_bh(&sock->sk->sk_callback_lock);
98
99 /* done under the callback_lock to serialize with write_space */
100 spin_lock(&rds_tcp_tc_list_lock);
101 list_del_init(&tc->t_list_item);
102 rds_tcp_tc_count--;
103 spin_unlock(&rds_tcp_tc_list_lock);
104
105 tc->t_sock = NULL;
106
107 sock->sk->sk_write_space = tc->t_orig_write_space;
108 sock->sk->sk_data_ready = tc->t_orig_data_ready;
109 sock->sk->sk_state_change = tc->t_orig_state_change;
110 sock->sk->sk_user_data = NULL;
111
112 write_unlock_bh(&sock->sk->sk_callback_lock);
113}
114
115/*
116 * This is the only path that sets tc->t_sock. Send and receive trust that
117 * it is set. The RDS_CONN_CONNECTED bit protects those paths from being
118 * called while it isn't set.
119 */
120void rds_tcp_set_callbacks(struct socket *sock, struct rds_connection *conn)
121{
122 struct rds_tcp_connection *tc = conn->c_transport_data;
123
124 rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
125 write_lock_bh(&sock->sk->sk_callback_lock);
126
127 /* done under the callback_lock to serialize with write_space */
128 spin_lock(&rds_tcp_tc_list_lock);
129 list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
130 rds_tcp_tc_count++;
131 spin_unlock(&rds_tcp_tc_list_lock);
132
133 /* accepted sockets need our listen data ready undone */
134 if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
135 sock->sk->sk_data_ready = sock->sk->sk_user_data;
136
137 tc->t_sock = sock;
138 tc->conn = conn;
139 tc->t_orig_data_ready = sock->sk->sk_data_ready;
140 tc->t_orig_write_space = sock->sk->sk_write_space;
141 tc->t_orig_state_change = sock->sk->sk_state_change;
142
143 sock->sk->sk_user_data = conn;
144 sock->sk->sk_data_ready = rds_tcp_data_ready;
145 sock->sk->sk_write_space = rds_tcp_write_space;
146 sock->sk->sk_state_change = rds_tcp_state_change;
147
148 write_unlock_bh(&sock->sk->sk_callback_lock);
149}
150
151static void rds_tcp_tc_info(struct socket *sock, unsigned int len,
152 struct rds_info_iterator *iter,
153 struct rds_info_lengths *lens)
154{
155 struct rds_info_tcp_socket tsinfo;
156 struct rds_tcp_connection *tc;
157 unsigned long flags;
158 struct sockaddr_in sin;
159 int sinlen;
160
161 spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
162
163 if (len / sizeof(tsinfo) < rds_tcp_tc_count)
164 goto out;
165
166 list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
167
168 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 0);
169 tsinfo.local_addr = sin.sin_addr.s_addr;
170 tsinfo.local_port = sin.sin_port;
171 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 1);
172 tsinfo.peer_addr = sin.sin_addr.s_addr;
173 tsinfo.peer_port = sin.sin_port;
174
175 tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
176 tsinfo.data_rem = tc->t_tinc_data_rem;
177 tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
178 tsinfo.last_expected_una = tc->t_last_expected_una;
179 tsinfo.last_seen_una = tc->t_last_seen_una;
180
181 rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
182 }
183
184out:
185 lens->nr = rds_tcp_tc_count;
186 lens->each = sizeof(tsinfo);
187
188 spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
189}
190
191static int rds_tcp_laddr_check(__be32 addr)
192{
193 if (inet_addr_type(&init_net, addr) == RTN_LOCAL)
194 return 0;
195 return -EADDRNOTAVAIL;
196}
197
198static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
199{
200 struct rds_tcp_connection *tc;
201
202 tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
203 if (!tc)
204 return -ENOMEM;
205
206 tc->t_sock = NULL;
207 tc->t_tinc = NULL;
208 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
209 tc->t_tinc_data_rem = 0;
210
211 conn->c_transport_data = tc;
212
213 spin_lock_irq(&rds_tcp_conn_lock);
214 list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
215 spin_unlock_irq(&rds_tcp_conn_lock);
216
217 rdsdebug("alloced tc %p\n", conn->c_transport_data);
218 return 0;
219}
220
221static void rds_tcp_conn_free(void *arg)
222{
223 struct rds_tcp_connection *tc = arg;
224 unsigned long flags;
225 rdsdebug("freeing tc %p\n", tc);
226
227 spin_lock_irqsave(&rds_tcp_conn_lock, flags);
228 list_del(&tc->t_tcp_node);
229 spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
230
231 kmem_cache_free(rds_tcp_conn_slab, tc);
232}
233
234static void rds_tcp_destroy_conns(void)
235{
236 struct rds_tcp_connection *tc, *_tc;
237 LIST_HEAD(tmp_list);
238
239 /* avoid calling conn_destroy with irqs off */
240 spin_lock_irq(&rds_tcp_conn_lock);
241 list_splice(&rds_tcp_conn_list, &tmp_list);
242 INIT_LIST_HEAD(&rds_tcp_conn_list);
243 spin_unlock_irq(&rds_tcp_conn_lock);
244
245 list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
246 if (tc->conn->c_passive)
247 rds_conn_destroy(tc->conn->c_passive);
248 rds_conn_destroy(tc->conn);
249 }
250}
251
252static void rds_tcp_exit(void)
253{
254 rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
255 rds_tcp_listen_stop();
256 rds_tcp_destroy_conns();
257 rds_trans_unregister(&rds_tcp_transport);
258 rds_tcp_recv_exit();
259 kmem_cache_destroy(rds_tcp_conn_slab);
260}
261module_exit(rds_tcp_exit);
262
263struct rds_transport rds_tcp_transport = {
264 .laddr_check = rds_tcp_laddr_check,
265 .xmit_prepare = rds_tcp_xmit_prepare,
266 .xmit_complete = rds_tcp_xmit_complete,
267 .xmit = rds_tcp_xmit,
268 .recv = rds_tcp_recv,
269 .conn_alloc = rds_tcp_conn_alloc,
270 .conn_free = rds_tcp_conn_free,
271 .conn_connect = rds_tcp_conn_connect,
272 .conn_shutdown = rds_tcp_conn_shutdown,
273 .inc_copy_to_user = rds_tcp_inc_copy_to_user,
274 .inc_free = rds_tcp_inc_free,
275 .stats_info_copy = rds_tcp_stats_info_copy,
276 .exit = rds_tcp_exit,
277 .t_owner = THIS_MODULE,
278 .t_name = "tcp",
279 .t_type = RDS_TRANS_TCP,
280 .t_prefer_loopback = 1,
281};
282
283static int rds_tcp_init(void)
284{
285 int ret;
286
287 rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
288 sizeof(struct rds_tcp_connection),
289 0, 0, NULL);
290 if (!rds_tcp_conn_slab) {
291 ret = -ENOMEM;
292 goto out;
293 }
294
295 ret = rds_tcp_recv_init();
296 if (ret)
297 goto out_slab;
298
299 ret = rds_trans_register(&rds_tcp_transport);
300 if (ret)
301 goto out_recv;
302
303 ret = rds_tcp_listen_init();
304 if (ret)
305 goto out_register;
306
307 rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
308
309 goto out;
310
311out_register:
312 rds_trans_unregister(&rds_tcp_transport);
313out_recv:
314 rds_tcp_recv_exit();
315out_slab:
316 kmem_cache_destroy(rds_tcp_conn_slab);
317out:
318 return ret;
319}
320module_init(rds_tcp_init);
321
322MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
323MODULE_DESCRIPTION("RDS: TCP transport");
324MODULE_LICENSE("Dual BSD/GPL");
325