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, 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 __user *buffer, size_t *lenp,
66 loff_t *fpos);
67
68static int rds_tcp_min_sndbuf = SOCK_MIN_SNDBUF;
69static int rds_tcp_min_rcvbuf = SOCK_MIN_RCVBUF;
70
71static struct ctl_table rds_tcp_sysctl_table[] = {
72#define RDS_TCP_SNDBUF 0
73 {
74 .procname = "rds_tcp_sndbuf",
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_sndbuf,
80 },
81#define RDS_TCP_RCVBUF 1
82 {
83 .procname = "rds_tcp_rcvbuf",
84 /* data is per-net pointer */
85 .maxlen = sizeof(int),
86 .mode = 0644,
87 .proc_handler = rds_tcp_skbuf_handler,
88 .extra1 = &rds_tcp_min_rcvbuf,
89 },
90 { }
91};
92
93/* doing it this way avoids calling tcp_sk() */
94void rds_tcp_nonagle(struct socket *sock)
95{
96 int val = 1;
97
98 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (void *)&val,
99 sizeof(val));
100}
101
102u32 rds_tcp_write_seq(struct rds_tcp_connection *tc)
103{
104 /* seq# of the last byte of data in tcp send buffer */
105 return tcp_sk(tc->t_sock->sk)->write_seq;
106}
107
108u32 rds_tcp_snd_una(struct rds_tcp_connection *tc)
109{
110 return tcp_sk(tc->t_sock->sk)->snd_una;
111}
112
113void rds_tcp_restore_callbacks(struct socket *sock,
114 struct rds_tcp_connection *tc)
115{
116 rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc);
117 write_lock_bh(&sock->sk->sk_callback_lock);
118
119 /* done under the callback_lock to serialize with write_space */
120 spin_lock(&rds_tcp_tc_list_lock);
121 list_del_init(&tc->t_list_item);
122#if IS_ENABLED(CONFIG_IPV6)
123 rds6_tcp_tc_count--;
124#endif
125 if (!tc->t_cpath->cp_conn->c_isv6)
126 rds_tcp_tc_count--;
127 spin_unlock(&rds_tcp_tc_list_lock);
128
129 tc->t_sock = NULL;
130
131 sock->sk->sk_write_space = tc->t_orig_write_space;
132 sock->sk->sk_data_ready = tc->t_orig_data_ready;
133 sock->sk->sk_state_change = tc->t_orig_state_change;
134 sock->sk->sk_user_data = NULL;
135
136 write_unlock_bh(&sock->sk->sk_callback_lock);
137}
138
139/*
140 * rds_tcp_reset_callbacks() switches the to the new sock and
141 * returns the existing tc->t_sock.
142 *
143 * The only functions that set tc->t_sock are rds_tcp_set_callbacks
144 * and rds_tcp_reset_callbacks. Send and receive trust that
145 * it is set. The absence of RDS_CONN_UP bit protects those paths
146 * from being called while it isn't set.
147 */
148void rds_tcp_reset_callbacks(struct socket *sock,
149 struct rds_conn_path *cp)
150{
151 struct rds_tcp_connection *tc = cp->cp_transport_data;
152 struct socket *osock = tc->t_sock;
153
154 if (!osock)
155 goto newsock;
156
157 /* Need to resolve a duelling SYN between peers.
158 * We have an outstanding SYN to this peer, which may
159 * potentially have transitioned to the RDS_CONN_UP state,
160 * so we must quiesce any send threads before resetting
161 * cp_transport_data. We quiesce these threads by setting
162 * cp_state to something other than RDS_CONN_UP, and then
163 * waiting for any existing threads in rds_send_xmit to
164 * complete release_in_xmit(). (Subsequent threads entering
165 * rds_send_xmit() will bail on !rds_conn_up().
166 *
167 * However an incoming syn-ack at this point would end up
168 * marking the conn as RDS_CONN_UP, and would again permit
169 * rds_send_xmi() threads through, so ideally we would
170 * synchronize on RDS_CONN_UP after lock_sock(), but cannot
171 * do that: waiting on !RDS_IN_XMIT after lock_sock() may
172 * end up deadlocking with tcp_sendmsg(), and the RDS_IN_XMIT
173 * would not get set. As a result, we set c_state to
174 * RDS_CONN_RESETTTING, to ensure that rds_tcp_state_change
175 * cannot mark rds_conn_path_up() in the window before lock_sock()
176 */
177 atomic_set(&cp->cp_state, RDS_CONN_RESETTING);
178 wait_event(cp->cp_waitq, !test_bit(RDS_IN_XMIT, &cp->cp_flags));
179 lock_sock(osock->sk);
180 /* reset receive side state for rds_tcp_data_recv() for osock */
181 cancel_delayed_work_sync(&cp->cp_send_w);
182 cancel_delayed_work_sync(&cp->cp_recv_w);
183 if (tc->t_tinc) {
184 rds_inc_put(&tc->t_tinc->ti_inc);
185 tc->t_tinc = NULL;
186 }
187 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
188 tc->t_tinc_data_rem = 0;
189 rds_tcp_restore_callbacks(osock, tc);
190 release_sock(osock->sk);
191 sock_release(osock);
192newsock:
193 rds_send_path_reset(cp);
194 lock_sock(sock->sk);
195 rds_tcp_set_callbacks(sock, cp);
196 release_sock(sock->sk);
197}
198
199/* Add tc to rds_tcp_tc_list and set tc->t_sock. See comments
200 * above rds_tcp_reset_callbacks for notes about synchronization
201 * with data path
202 */
203void rds_tcp_set_callbacks(struct socket *sock, struct rds_conn_path *cp)
204{
205 struct rds_tcp_connection *tc = cp->cp_transport_data;
206
207 rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
208 write_lock_bh(&sock->sk->sk_callback_lock);
209
210 /* done under the callback_lock to serialize with write_space */
211 spin_lock(&rds_tcp_tc_list_lock);
212 list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
213#if IS_ENABLED(CONFIG_IPV6)
214 rds6_tcp_tc_count++;
215#endif
216 if (!tc->t_cpath->cp_conn->c_isv6)
217 rds_tcp_tc_count++;
218 spin_unlock(&rds_tcp_tc_list_lock);
219
220 /* accepted sockets need our listen data ready undone */
221 if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
222 sock->sk->sk_data_ready = sock->sk->sk_user_data;
223
224 tc->t_sock = sock;
225 tc->t_cpath = cp;
226 tc->t_orig_data_ready = sock->sk->sk_data_ready;
227 tc->t_orig_write_space = sock->sk->sk_write_space;
228 tc->t_orig_state_change = sock->sk->sk_state_change;
229
230 sock->sk->sk_user_data = cp;
231 sock->sk->sk_data_ready = rds_tcp_data_ready;
232 sock->sk->sk_write_space = rds_tcp_write_space;
233 sock->sk->sk_state_change = rds_tcp_state_change;
234
235 write_unlock_bh(&sock->sk->sk_callback_lock);
236}
237
238/* Handle RDS_INFO_TCP_SOCKETS socket option. It only returns IPv4
239 * connections for backward compatibility.
240 */
241static void rds_tcp_tc_info(struct socket *rds_sock, unsigned int len,
242 struct rds_info_iterator *iter,
243 struct rds_info_lengths *lens)
244{
245 struct rds_info_tcp_socket tsinfo;
246 struct rds_tcp_connection *tc;
247 unsigned long flags;
248
249 spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
250
251 if (len / sizeof(tsinfo) < rds_tcp_tc_count)
252 goto out;
253
254 list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
255 struct inet_sock *inet = inet_sk(tc->t_sock->sk);
256
257 if (tc->t_cpath->cp_conn->c_isv6)
258 continue;
259
260 tsinfo.local_addr = inet->inet_saddr;
261 tsinfo.local_port = inet->inet_sport;
262 tsinfo.peer_addr = inet->inet_daddr;
263 tsinfo.peer_port = inet->inet_dport;
264
265 tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
266 tsinfo.data_rem = tc->t_tinc_data_rem;
267 tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
268 tsinfo.last_expected_una = tc->t_last_expected_una;
269 tsinfo.last_seen_una = tc->t_last_seen_una;
270 tsinfo.tos = tc->t_cpath->cp_conn->c_tos;
271
272 rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
273 }
274
275out:
276 lens->nr = rds_tcp_tc_count;
277 lens->each = sizeof(tsinfo);
278
279 spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
280}
281
282#if IS_ENABLED(CONFIG_IPV6)
283/* Handle RDS6_INFO_TCP_SOCKETS socket option. It returns both IPv4 and
284 * IPv6 connections. IPv4 connection address is returned in an IPv4 mapped
285 * address.
286 */
287static void rds6_tcp_tc_info(struct socket *sock, unsigned int len,
288 struct rds_info_iterator *iter,
289 struct rds_info_lengths *lens)
290{
291 struct rds6_info_tcp_socket tsinfo6;
292 struct rds_tcp_connection *tc;
293 unsigned long flags;
294
295 spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
296
297 if (len / sizeof(tsinfo6) < rds6_tcp_tc_count)
298 goto out;
299
300 list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
301 struct sock *sk = tc->t_sock->sk;
302 struct inet_sock *inet = inet_sk(sk);
303
304 tsinfo6.local_addr = sk->sk_v6_rcv_saddr;
305 tsinfo6.local_port = inet->inet_sport;
306 tsinfo6.peer_addr = sk->sk_v6_daddr;
307 tsinfo6.peer_port = inet->inet_dport;
308
309 tsinfo6.hdr_rem = tc->t_tinc_hdr_rem;
310 tsinfo6.data_rem = tc->t_tinc_data_rem;
311 tsinfo6.last_sent_nxt = tc->t_last_sent_nxt;
312 tsinfo6.last_expected_una = tc->t_last_expected_una;
313 tsinfo6.last_seen_una = tc->t_last_seen_una;
314
315 rds_info_copy(iter, &tsinfo6, sizeof(tsinfo6));
316 }
317
318out:
319 lens->nr = rds6_tcp_tc_count;
320 lens->each = sizeof(tsinfo6);
321
322 spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
323}
324#endif
325
326static int rds_tcp_laddr_check(struct net *net, const struct in6_addr *addr,
327 __u32 scope_id)
328{
329 struct net_device *dev = NULL;
330#if IS_ENABLED(CONFIG_IPV6)
331 int ret;
332#endif
333
334 if (ipv6_addr_v4mapped(addr)) {
335 if (inet_addr_type(net, addr->s6_addr32[3]) == RTN_LOCAL)
336 return 0;
337 return -EADDRNOTAVAIL;
338 }
339
340 /* If the scope_id is specified, check only those addresses
341 * hosted on the specified interface.
342 */
343 if (scope_id != 0) {
344 rcu_read_lock();
345 dev = dev_get_by_index_rcu(net, scope_id);
346 /* scope_id is not valid... */
347 if (!dev) {
348 rcu_read_unlock();
349 return -EADDRNOTAVAIL;
350 }
351 rcu_read_unlock();
352 }
353#if IS_ENABLED(CONFIG_IPV6)
354 ret = ipv6_chk_addr(net, addr, dev, 0);
355 if (ret)
356 return 0;
357#endif
358 return -EADDRNOTAVAIL;
359}
360
361static void rds_tcp_conn_free(void *arg)
362{
363 struct rds_tcp_connection *tc = arg;
364 unsigned long flags;
365
366 rdsdebug("freeing tc %p\n", tc);
367
368 spin_lock_irqsave(&rds_tcp_conn_lock, flags);
369 if (!tc->t_tcp_node_detached)
370 list_del(&tc->t_tcp_node);
371 spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
372
373 kmem_cache_free(rds_tcp_conn_slab, tc);
374}
375
376static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
377{
378 struct rds_tcp_connection *tc;
379 int i, j;
380 int ret = 0;
381
382 for (i = 0; i < RDS_MPATH_WORKERS; i++) {
383 tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
384 if (!tc) {
385 ret = -ENOMEM;
386 goto fail;
387 }
388 mutex_init(&tc->t_conn_path_lock);
389 tc->t_sock = NULL;
390 tc->t_tinc = NULL;
391 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
392 tc->t_tinc_data_rem = 0;
393
394 conn->c_path[i].cp_transport_data = tc;
395 tc->t_cpath = &conn->c_path[i];
396 tc->t_tcp_node_detached = true;
397
398 rdsdebug("rds_conn_path [%d] tc %p\n", i,
399 conn->c_path[i].cp_transport_data);
400 }
401 spin_lock_irq(&rds_tcp_conn_lock);
402 for (i = 0; i < RDS_MPATH_WORKERS; i++) {
403 tc = conn->c_path[i].cp_transport_data;
404 tc->t_tcp_node_detached = false;
405 list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
406 }
407 spin_unlock_irq(&rds_tcp_conn_lock);
408fail:
409 if (ret) {
410 for (j = 0; j < i; j++)
411 rds_tcp_conn_free(conn->c_path[j].cp_transport_data);
412 }
413 return ret;
414}
415
416static bool list_has_conn(struct list_head *list, struct rds_connection *conn)
417{
418 struct rds_tcp_connection *tc, *_tc;
419
420 list_for_each_entry_safe(tc, _tc, list, t_tcp_node) {
421 if (tc->t_cpath->cp_conn == conn)
422 return true;
423 }
424 return false;
425}
426
427static void rds_tcp_set_unloading(void)
428{
429 atomic_set(&rds_tcp_unloading, 1);
430}
431
432static bool rds_tcp_is_unloading(struct rds_connection *conn)
433{
434 return atomic_read(&rds_tcp_unloading) != 0;
435}
436
437static void rds_tcp_destroy_conns(void)
438{
439 struct rds_tcp_connection *tc, *_tc;
440 LIST_HEAD(tmp_list);
441
442 /* avoid calling conn_destroy with irqs off */
443 spin_lock_irq(&rds_tcp_conn_lock);
444 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
445 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn))
446 list_move_tail(&tc->t_tcp_node, &tmp_list);
447 }
448 spin_unlock_irq(&rds_tcp_conn_lock);
449
450 list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node)
451 rds_conn_destroy(tc->t_cpath->cp_conn);
452}
453
454static void rds_tcp_exit(void);
455
456static u8 rds_tcp_get_tos_map(u8 tos)
457{
458 /* all user tos mapped to default 0 for TCP transport */
459 return 0;
460}
461
462struct rds_transport rds_tcp_transport = {
463 .laddr_check = rds_tcp_laddr_check,
464 .xmit_path_prepare = rds_tcp_xmit_path_prepare,
465 .xmit_path_complete = rds_tcp_xmit_path_complete,
466 .xmit = rds_tcp_xmit,
467 .recv_path = rds_tcp_recv_path,
468 .conn_alloc = rds_tcp_conn_alloc,
469 .conn_free = rds_tcp_conn_free,
470 .conn_path_connect = rds_tcp_conn_path_connect,
471 .conn_path_shutdown = rds_tcp_conn_path_shutdown,
472 .inc_copy_to_user = rds_tcp_inc_copy_to_user,
473 .inc_free = rds_tcp_inc_free,
474 .stats_info_copy = rds_tcp_stats_info_copy,
475 .exit = rds_tcp_exit,
476 .get_tos_map = rds_tcp_get_tos_map,
477 .t_owner = THIS_MODULE,
478 .t_name = "tcp",
479 .t_type = RDS_TRANS_TCP,
480 .t_prefer_loopback = 1,
481 .t_mp_capable = 1,
482 .t_unloading = rds_tcp_is_unloading,
483};
484
485static unsigned int rds_tcp_netid;
486
487/* per-network namespace private data for this module */
488struct rds_tcp_net {
489 struct socket *rds_tcp_listen_sock;
490 struct work_struct rds_tcp_accept_w;
491 struct ctl_table_header *rds_tcp_sysctl;
492 struct ctl_table *ctl_table;
493 int sndbuf_size;
494 int rcvbuf_size;
495};
496
497/* All module specific customizations to the RDS-TCP socket should be done in
498 * rds_tcp_tune() and applied after socket creation.
499 */
500void rds_tcp_tune(struct socket *sock)
501{
502 struct sock *sk = sock->sk;
503 struct net *net = sock_net(sk);
504 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
505
506 rds_tcp_nonagle(sock);
507 lock_sock(sk);
508 if (rtn->sndbuf_size > 0) {
509 sk->sk_sndbuf = rtn->sndbuf_size;
510 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
511 }
512 if (rtn->rcvbuf_size > 0) {
513 sk->sk_sndbuf = rtn->rcvbuf_size;
514 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
515 }
516 release_sock(sk);
517}
518
519static void rds_tcp_accept_worker(struct work_struct *work)
520{
521 struct rds_tcp_net *rtn = container_of(work,
522 struct rds_tcp_net,
523 rds_tcp_accept_w);
524
525 while (rds_tcp_accept_one(rtn->rds_tcp_listen_sock) == 0)
526 cond_resched();
527}
528
529void rds_tcp_accept_work(struct sock *sk)
530{
531 struct net *net = sock_net(sk);
532 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
533
534 queue_work(rds_wq, &rtn->rds_tcp_accept_w);
535}
536
537static __net_init int rds_tcp_init_net(struct net *net)
538{
539 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
540 struct ctl_table *tbl;
541 int err = 0;
542
543 memset(rtn, 0, sizeof(*rtn));
544
545 /* {snd, rcv}buf_size default to 0, which implies we let the
546 * stack pick the value, and permit auto-tuning of buffer size.
547 */
548 if (net == &init_net) {
549 tbl = rds_tcp_sysctl_table;
550 } else {
551 tbl = kmemdup(rds_tcp_sysctl_table,
552 sizeof(rds_tcp_sysctl_table), GFP_KERNEL);
553 if (!tbl) {
554 pr_warn("could not set allocate sysctl table\n");
555 return -ENOMEM;
556 }
557 rtn->ctl_table = tbl;
558 }
559 tbl[RDS_TCP_SNDBUF].data = &rtn->sndbuf_size;
560 tbl[RDS_TCP_RCVBUF].data = &rtn->rcvbuf_size;
561 rtn->rds_tcp_sysctl = register_net_sysctl(net, "net/rds/tcp", tbl);
562 if (!rtn->rds_tcp_sysctl) {
563 pr_warn("could not register sysctl\n");
564 err = -ENOMEM;
565 goto fail;
566 }
567
568#if IS_ENABLED(CONFIG_IPV6)
569 rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net, true);
570#else
571 rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net, false);
572#endif
573 if (!rtn->rds_tcp_listen_sock) {
574 pr_warn("could not set up IPv6 listen sock\n");
575
576#if IS_ENABLED(CONFIG_IPV6)
577 /* Try IPv4 as some systems disable IPv6 */
578 rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net, false);
579 if (!rtn->rds_tcp_listen_sock) {
580#endif
581 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
582 rtn->rds_tcp_sysctl = NULL;
583 err = -EAFNOSUPPORT;
584 goto fail;
585#if IS_ENABLED(CONFIG_IPV6)
586 }
587#endif
588 }
589 INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
590 return 0;
591
592fail:
593 if (net != &init_net)
594 kfree(tbl);
595 return err;
596}
597
598static void rds_tcp_kill_sock(struct net *net)
599{
600 struct rds_tcp_connection *tc, *_tc;
601 LIST_HEAD(tmp_list);
602 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
603 struct socket *lsock = rtn->rds_tcp_listen_sock;
604
605 rtn->rds_tcp_listen_sock = NULL;
606 rds_tcp_listen_stop(lsock, &rtn->rds_tcp_accept_w);
607 spin_lock_irq(&rds_tcp_conn_lock);
608 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
609 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
610
611 if (net != c_net)
612 continue;
613 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn)) {
614 list_move_tail(&tc->t_tcp_node, &tmp_list);
615 } else {
616 list_del(&tc->t_tcp_node);
617 tc->t_tcp_node_detached = true;
618 }
619 }
620 spin_unlock_irq(&rds_tcp_conn_lock);
621 list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node)
622 rds_conn_destroy(tc->t_cpath->cp_conn);
623}
624
625static void __net_exit rds_tcp_exit_net(struct net *net)
626{
627 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
628
629 rds_tcp_kill_sock(net);
630
631 if (rtn->rds_tcp_sysctl)
632 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
633
634 if (net != &init_net)
635 kfree(rtn->ctl_table);
636}
637
638static struct pernet_operations rds_tcp_net_ops = {
639 .init = rds_tcp_init_net,
640 .exit = rds_tcp_exit_net,
641 .id = &rds_tcp_netid,
642 .size = sizeof(struct rds_tcp_net),
643};
644
645void *rds_tcp_listen_sock_def_readable(struct net *net)
646{
647 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
648 struct socket *lsock = rtn->rds_tcp_listen_sock;
649
650 if (!lsock)
651 return NULL;
652
653 return lsock->sk->sk_user_data;
654}
655
656/* when sysctl is used to modify some kernel socket parameters,this
657 * function resets the RDS connections in that netns so that we can
658 * restart with new parameters. The assumption is that such reset
659 * events are few and far-between.
660 */
661static void rds_tcp_sysctl_reset(struct net *net)
662{
663 struct rds_tcp_connection *tc, *_tc;
664
665 spin_lock_irq(&rds_tcp_conn_lock);
666 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
667 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
668
669 if (net != c_net || !tc->t_sock)
670 continue;
671
672 /* reconnect with new parameters */
673 rds_conn_path_drop(tc->t_cpath, false);
674 }
675 spin_unlock_irq(&rds_tcp_conn_lock);
676}
677
678static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
679 void __user *buffer, size_t *lenp,
680 loff_t *fpos)
681{
682 struct net *net = current->nsproxy->net_ns;
683 int err;
684
685 err = proc_dointvec_minmax(ctl, write, buffer, lenp, fpos);
686 if (err < 0) {
687 pr_warn("Invalid input. Must be >= %d\n",
688 *(int *)(ctl->extra1));
689 return err;
690 }
691 if (write)
692 rds_tcp_sysctl_reset(net);
693 return 0;
694}
695
696static void rds_tcp_exit(void)
697{
698 rds_tcp_set_unloading();
699 synchronize_rcu();
700 rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
701#if IS_ENABLED(CONFIG_IPV6)
702 rds_info_deregister_func(RDS6_INFO_TCP_SOCKETS, rds6_tcp_tc_info);
703#endif
704 unregister_pernet_device(&rds_tcp_net_ops);
705 rds_tcp_destroy_conns();
706 rds_trans_unregister(&rds_tcp_transport);
707 rds_tcp_recv_exit();
708 kmem_cache_destroy(rds_tcp_conn_slab);
709}
710module_exit(rds_tcp_exit);
711
712static int rds_tcp_init(void)
713{
714 int ret;
715
716 rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
717 sizeof(struct rds_tcp_connection),
718 0, 0, NULL);
719 if (!rds_tcp_conn_slab) {
720 ret = -ENOMEM;
721 goto out;
722 }
723
724 ret = rds_tcp_recv_init();
725 if (ret)
726 goto out_slab;
727
728 ret = register_pernet_device(&rds_tcp_net_ops);
729 if (ret)
730 goto out_recv;
731
732 rds_trans_register(&rds_tcp_transport);
733
734 rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
735#if IS_ENABLED(CONFIG_IPV6)
736 rds_info_register_func(RDS6_INFO_TCP_SOCKETS, rds6_tcp_tc_info);
737#endif
738
739 goto out;
740out_recv:
741 rds_tcp_recv_exit();
742out_slab:
743 kmem_cache_destroy(rds_tcp_conn_slab);
744out:
745 return ret;
746}
747module_init(rds_tcp_init);
748
749MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
750MODULE_DESCRIPTION("RDS: TCP transport");
751MODULE_LICENSE("Dual BSD/GPL");