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
39#include "rds.h"
40#include "tcp.h"
41
42/* only for info exporting */
43static DEFINE_SPINLOCK(rds_tcp_tc_list_lock);
44static LIST_HEAD(rds_tcp_tc_list);
45static unsigned int rds_tcp_tc_count;
46
47/* Track rds_tcp_connection structs so they can be cleaned up */
48static DEFINE_SPINLOCK(rds_tcp_conn_lock);
49static LIST_HEAD(rds_tcp_conn_list);
50
51static struct kmem_cache *rds_tcp_conn_slab;
52
53#define RDS_TCP_DEFAULT_BUFSIZE (128 * 1024)
54
55/* doing it this way avoids calling tcp_sk() */
56void rds_tcp_nonagle(struct socket *sock)
57{
58 mm_segment_t oldfs = get_fs();
59 int val = 1;
60
61 set_fs(KERNEL_DS);
62 sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY, (char __user *)&val,
63 sizeof(val));
64 set_fs(oldfs);
65}
66
67void rds_tcp_tune(struct socket *sock)
68{
69 struct sock *sk = sock->sk;
70
71 rds_tcp_nonagle(sock);
72
73 /*
74 * We're trying to saturate gigabit with the default,
75 * see svc_sock_setbufsize().
76 */
77 lock_sock(sk);
78 sk->sk_sndbuf = RDS_TCP_DEFAULT_BUFSIZE;
79 sk->sk_rcvbuf = RDS_TCP_DEFAULT_BUFSIZE;
80 sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
81 release_sock(sk);
82}
83
84u32 rds_tcp_snd_nxt(struct rds_tcp_connection *tc)
85{
86 return tcp_sk(tc->t_sock->sk)->snd_nxt;
87}
88
89u32 rds_tcp_snd_una(struct rds_tcp_connection *tc)
90{
91 return tcp_sk(tc->t_sock->sk)->snd_una;
92}
93
94void rds_tcp_restore_callbacks(struct socket *sock,
95 struct rds_tcp_connection *tc)
96{
97 rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc);
98 write_lock_bh(&sock->sk->sk_callback_lock);
99
100 /* done under the callback_lock to serialize with write_space */
101 spin_lock(&rds_tcp_tc_list_lock);
102 list_del_init(&tc->t_list_item);
103 rds_tcp_tc_count--;
104 spin_unlock(&rds_tcp_tc_list_lock);
105
106 tc->t_sock = NULL;
107
108 sock->sk->sk_write_space = tc->t_orig_write_space;
109 sock->sk->sk_data_ready = tc->t_orig_data_ready;
110 sock->sk->sk_state_change = tc->t_orig_state_change;
111 sock->sk->sk_user_data = NULL;
112
113 write_unlock_bh(&sock->sk->sk_callback_lock);
114}
115
116/*
117 * This is the only path that sets tc->t_sock. Send and receive trust that
118 * it is set. The RDS_CONN_CONNECTED bit protects those paths from being
119 * called while it isn't set.
120 */
121void rds_tcp_set_callbacks(struct socket *sock, struct rds_connection *conn)
122{
123 struct rds_tcp_connection *tc = conn->c_transport_data;
124
125 rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
126 write_lock_bh(&sock->sk->sk_callback_lock);
127
128 /* done under the callback_lock to serialize with write_space */
129 spin_lock(&rds_tcp_tc_list_lock);
130 list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
131 rds_tcp_tc_count++;
132 spin_unlock(&rds_tcp_tc_list_lock);
133
134 /* accepted sockets need our listen data ready undone */
135 if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
136 sock->sk->sk_data_ready = sock->sk->sk_user_data;
137
138 tc->t_sock = sock;
139 tc->conn = conn;
140 tc->t_orig_data_ready = sock->sk->sk_data_ready;
141 tc->t_orig_write_space = sock->sk->sk_write_space;
142 tc->t_orig_state_change = sock->sk->sk_state_change;
143
144 sock->sk->sk_user_data = conn;
145 sock->sk->sk_data_ready = rds_tcp_data_ready;
146 sock->sk->sk_write_space = rds_tcp_write_space;
147 sock->sk->sk_state_change = rds_tcp_state_change;
148
149 write_unlock_bh(&sock->sk->sk_callback_lock);
150}
151
152static void rds_tcp_tc_info(struct socket *sock, unsigned int len,
153 struct rds_info_iterator *iter,
154 struct rds_info_lengths *lens)
155{
156 struct rds_info_tcp_socket tsinfo;
157 struct rds_tcp_connection *tc;
158 unsigned long flags;
159 struct sockaddr_in sin;
160 int sinlen;
161
162 spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
163
164 if (len / sizeof(tsinfo) < rds_tcp_tc_count)
165 goto out;
166
167 list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
168
169 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 0);
170 tsinfo.local_addr = sin.sin_addr.s_addr;
171 tsinfo.local_port = sin.sin_port;
172 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 1);
173 tsinfo.peer_addr = sin.sin_addr.s_addr;
174 tsinfo.peer_port = sin.sin_port;
175
176 tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
177 tsinfo.data_rem = tc->t_tinc_data_rem;
178 tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
179 tsinfo.last_expected_una = tc->t_last_expected_una;
180 tsinfo.last_seen_una = tc->t_last_seen_una;
181
182 rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
183 }
184
185out:
186 lens->nr = rds_tcp_tc_count;
187 lens->each = sizeof(tsinfo);
188
189 spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
190}
191
192static int rds_tcp_laddr_check(__be32 addr)
193{
194 if (inet_addr_type(&init_net, addr) == RTN_LOCAL)
195 return 0;
196 return -EADDRNOTAVAIL;
197}
198
199static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
200{
201 struct rds_tcp_connection *tc;
202
203 tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
204 if (!tc)
205 return -ENOMEM;
206
207 tc->t_sock = NULL;
208 tc->t_tinc = NULL;
209 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
210 tc->t_tinc_data_rem = 0;
211
212 conn->c_transport_data = tc;
213
214 spin_lock_irq(&rds_tcp_conn_lock);
215 list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
216 spin_unlock_irq(&rds_tcp_conn_lock);
217
218 rdsdebug("alloced tc %p\n", conn->c_transport_data);
219 return 0;
220}
221
222static void rds_tcp_conn_free(void *arg)
223{
224 struct rds_tcp_connection *tc = arg;
225 unsigned long flags;
226 rdsdebug("freeing tc %p\n", tc);
227
228 spin_lock_irqsave(&rds_tcp_conn_lock, flags);
229 list_del(&tc->t_tcp_node);
230 spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
231
232 kmem_cache_free(rds_tcp_conn_slab, tc);
233}
234
235static void rds_tcp_destroy_conns(void)
236{
237 struct rds_tcp_connection *tc, *_tc;
238 LIST_HEAD(tmp_list);
239
240 /* avoid calling conn_destroy with irqs off */
241 spin_lock_irq(&rds_tcp_conn_lock);
242 list_splice(&rds_tcp_conn_list, &tmp_list);
243 INIT_LIST_HEAD(&rds_tcp_conn_list);
244 spin_unlock_irq(&rds_tcp_conn_lock);
245
246 list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
247 if (tc->conn->c_passive)
248 rds_conn_destroy(tc->conn->c_passive);
249 rds_conn_destroy(tc->conn);
250 }
251}
252
253static void rds_tcp_exit(void)
254{
255 rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
256 rds_tcp_listen_stop();
257 rds_tcp_destroy_conns();
258 rds_trans_unregister(&rds_tcp_transport);
259 rds_tcp_recv_exit();
260 kmem_cache_destroy(rds_tcp_conn_slab);
261}
262module_exit(rds_tcp_exit);
263
264struct rds_transport rds_tcp_transport = {
265 .laddr_check = rds_tcp_laddr_check,
266 .xmit_prepare = rds_tcp_xmit_prepare,
267 .xmit_complete = rds_tcp_xmit_complete,
268 .xmit = rds_tcp_xmit,
269 .recv = rds_tcp_recv,
270 .conn_alloc = rds_tcp_conn_alloc,
271 .conn_free = rds_tcp_conn_free,
272 .conn_connect = rds_tcp_conn_connect,
273 .conn_shutdown = rds_tcp_conn_shutdown,
274 .inc_copy_to_user = rds_tcp_inc_copy_to_user,
275 .inc_free = rds_tcp_inc_free,
276 .stats_info_copy = rds_tcp_stats_info_copy,
277 .exit = rds_tcp_exit,
278 .t_owner = THIS_MODULE,
279 .t_name = "tcp",
280 .t_type = RDS_TRANS_TCP,
281 .t_prefer_loopback = 1,
282};
283
284static int rds_tcp_init(void)
285{
286 int ret;
287
288 rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
289 sizeof(struct rds_tcp_connection),
290 0, 0, NULL);
291 if (!rds_tcp_conn_slab) {
292 ret = -ENOMEM;
293 goto out;
294 }
295
296 ret = rds_tcp_recv_init();
297 if (ret)
298 goto out_slab;
299
300 ret = rds_trans_register(&rds_tcp_transport);
301 if (ret)
302 goto out_recv;
303
304 ret = rds_tcp_listen_init();
305 if (ret)
306 goto out_register;
307
308 rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
309
310 goto out;
311
312out_register:
313 rds_trans_unregister(&rds_tcp_transport);
314out_recv:
315 rds_tcp_recv_exit();
316out_slab:
317 kmem_cache_destroy(rds_tcp_conn_slab);
318out:
319 return ret;
320}
321module_init(rds_tcp_init);
322
323MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
324MODULE_DESCRIPTION("RDS: TCP transport");
325MODULE_LICENSE("Dual BSD/GPL");
326
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");