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 <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
1/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/slab.h>
35#include <linux/in.h>
36#include <linux/module.h>
37#include <net/tcp.h>
38#include <net/net_namespace.h>
39#include <net/netns/generic.h>
40
41#include "rds.h"
42#include "tcp.h"
43
44/* only for info exporting */
45static DEFINE_SPINLOCK(rds_tcp_tc_list_lock);
46static LIST_HEAD(rds_tcp_tc_list);
47static unsigned int rds_tcp_tc_count;
48
49/* Track rds_tcp_connection structs so they can be cleaned up */
50static DEFINE_SPINLOCK(rds_tcp_conn_lock);
51static LIST_HEAD(rds_tcp_conn_list);
52
53static struct kmem_cache *rds_tcp_conn_slab;
54
55static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
56 void __user *buffer, size_t *lenp,
57 loff_t *fpos);
58
59static int rds_tcp_min_sndbuf = SOCK_MIN_SNDBUF;
60static int rds_tcp_min_rcvbuf = SOCK_MIN_RCVBUF;
61
62static struct ctl_table rds_tcp_sysctl_table[] = {
63#define RDS_TCP_SNDBUF 0
64 {
65 .procname = "rds_tcp_sndbuf",
66 /* data is per-net pointer */
67 .maxlen = sizeof(int),
68 .mode = 0644,
69 .proc_handler = rds_tcp_skbuf_handler,
70 .extra1 = &rds_tcp_min_sndbuf,
71 },
72#define RDS_TCP_RCVBUF 1
73 {
74 .procname = "rds_tcp_rcvbuf",
75 /* data is per-net pointer */
76 .maxlen = sizeof(int),
77 .mode = 0644,
78 .proc_handler = rds_tcp_skbuf_handler,
79 .extra1 = &rds_tcp_min_rcvbuf,
80 },
81 { }
82};
83
84/* doing it this way avoids calling tcp_sk() */
85void rds_tcp_nonagle(struct socket *sock)
86{
87 mm_segment_t oldfs = get_fs();
88 int val = 1;
89
90 set_fs(KERNEL_DS);
91 sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY, (char __user *)&val,
92 sizeof(val));
93 set_fs(oldfs);
94}
95
96u32 rds_tcp_snd_nxt(struct rds_tcp_connection *tc)
97{
98 return tcp_sk(tc->t_sock->sk)->snd_nxt;
99}
100
101u32 rds_tcp_snd_una(struct rds_tcp_connection *tc)
102{
103 return tcp_sk(tc->t_sock->sk)->snd_una;
104}
105
106void rds_tcp_restore_callbacks(struct socket *sock,
107 struct rds_tcp_connection *tc)
108{
109 rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc);
110 write_lock_bh(&sock->sk->sk_callback_lock);
111
112 /* done under the callback_lock to serialize with write_space */
113 spin_lock(&rds_tcp_tc_list_lock);
114 list_del_init(&tc->t_list_item);
115 rds_tcp_tc_count--;
116 spin_unlock(&rds_tcp_tc_list_lock);
117
118 tc->t_sock = NULL;
119
120 sock->sk->sk_write_space = tc->t_orig_write_space;
121 sock->sk->sk_data_ready = tc->t_orig_data_ready;
122 sock->sk->sk_state_change = tc->t_orig_state_change;
123 sock->sk->sk_user_data = NULL;
124
125 write_unlock_bh(&sock->sk->sk_callback_lock);
126}
127
128/*
129 * rds_tcp_reset_callbacks() switches the to the new sock and
130 * returns the existing tc->t_sock.
131 *
132 * The only functions that set tc->t_sock are rds_tcp_set_callbacks
133 * and rds_tcp_reset_callbacks. Send and receive trust that
134 * it is set. The absence of RDS_CONN_UP bit protects those paths
135 * from being called while it isn't set.
136 */
137void rds_tcp_reset_callbacks(struct socket *sock,
138 struct rds_conn_path *cp)
139{
140 struct rds_tcp_connection *tc = cp->cp_transport_data;
141 struct socket *osock = tc->t_sock;
142
143 if (!osock)
144 goto newsock;
145
146 /* Need to resolve a duelling SYN between peers.
147 * We have an outstanding SYN to this peer, which may
148 * potentially have transitioned to the RDS_CONN_UP state,
149 * so we must quiesce any send threads before resetting
150 * cp_transport_data. We quiesce these threads by setting
151 * cp_state to something other than RDS_CONN_UP, and then
152 * waiting for any existing threads in rds_send_xmit to
153 * complete release_in_xmit(). (Subsequent threads entering
154 * rds_send_xmit() will bail on !rds_conn_up().
155 *
156 * However an incoming syn-ack at this point would end up
157 * marking the conn as RDS_CONN_UP, and would again permit
158 * rds_send_xmi() threads through, so ideally we would
159 * synchronize on RDS_CONN_UP after lock_sock(), but cannot
160 * do that: waiting on !RDS_IN_XMIT after lock_sock() may
161 * end up deadlocking with tcp_sendmsg(), and the RDS_IN_XMIT
162 * would not get set. As a result, we set c_state to
163 * RDS_CONN_RESETTTING, to ensure that rds_tcp_state_change
164 * cannot mark rds_conn_path_up() in the window before lock_sock()
165 */
166 atomic_set(&cp->cp_state, RDS_CONN_RESETTING);
167 wait_event(cp->cp_waitq, !test_bit(RDS_IN_XMIT, &cp->cp_flags));
168 lock_sock(osock->sk);
169 /* reset receive side state for rds_tcp_data_recv() for osock */
170 cancel_delayed_work_sync(&cp->cp_send_w);
171 cancel_delayed_work_sync(&cp->cp_recv_w);
172 if (tc->t_tinc) {
173 rds_inc_put(&tc->t_tinc->ti_inc);
174 tc->t_tinc = NULL;
175 }
176 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
177 tc->t_tinc_data_rem = 0;
178 rds_tcp_restore_callbacks(osock, tc);
179 release_sock(osock->sk);
180 sock_release(osock);
181newsock:
182 rds_send_path_reset(cp);
183 lock_sock(sock->sk);
184 rds_tcp_set_callbacks(sock, cp);
185 release_sock(sock->sk);
186}
187
188/* Add tc to rds_tcp_tc_list and set tc->t_sock. See comments
189 * above rds_tcp_reset_callbacks for notes about synchronization
190 * with data path
191 */
192void rds_tcp_set_callbacks(struct socket *sock, struct rds_conn_path *cp)
193{
194 struct rds_tcp_connection *tc = cp->cp_transport_data;
195
196 rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
197 write_lock_bh(&sock->sk->sk_callback_lock);
198
199 /* done under the callback_lock to serialize with write_space */
200 spin_lock(&rds_tcp_tc_list_lock);
201 list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
202 rds_tcp_tc_count++;
203 spin_unlock(&rds_tcp_tc_list_lock);
204
205 /* accepted sockets need our listen data ready undone */
206 if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
207 sock->sk->sk_data_ready = sock->sk->sk_user_data;
208
209 tc->t_sock = sock;
210 tc->t_cpath = cp;
211 tc->t_orig_data_ready = sock->sk->sk_data_ready;
212 tc->t_orig_write_space = sock->sk->sk_write_space;
213 tc->t_orig_state_change = sock->sk->sk_state_change;
214
215 sock->sk->sk_user_data = cp;
216 sock->sk->sk_data_ready = rds_tcp_data_ready;
217 sock->sk->sk_write_space = rds_tcp_write_space;
218 sock->sk->sk_state_change = rds_tcp_state_change;
219
220 write_unlock_bh(&sock->sk->sk_callback_lock);
221}
222
223static void rds_tcp_tc_info(struct socket *rds_sock, unsigned int len,
224 struct rds_info_iterator *iter,
225 struct rds_info_lengths *lens)
226{
227 struct rds_info_tcp_socket tsinfo;
228 struct rds_tcp_connection *tc;
229 unsigned long flags;
230 struct sockaddr_in sin;
231 int sinlen;
232 struct socket *sock;
233
234 spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
235
236 if (len / sizeof(tsinfo) < rds_tcp_tc_count)
237 goto out;
238
239 list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
240
241 sock = tc->t_sock;
242 if (sock) {
243 sock->ops->getname(sock, (struct sockaddr *)&sin,
244 &sinlen, 0);
245 tsinfo.local_addr = sin.sin_addr.s_addr;
246 tsinfo.local_port = sin.sin_port;
247 sock->ops->getname(sock, (struct sockaddr *)&sin,
248 &sinlen, 1);
249 tsinfo.peer_addr = sin.sin_addr.s_addr;
250 tsinfo.peer_port = sin.sin_port;
251 }
252
253 tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
254 tsinfo.data_rem = tc->t_tinc_data_rem;
255 tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
256 tsinfo.last_expected_una = tc->t_last_expected_una;
257 tsinfo.last_seen_una = tc->t_last_seen_una;
258
259 rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
260 }
261
262out:
263 lens->nr = rds_tcp_tc_count;
264 lens->each = sizeof(tsinfo);
265
266 spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
267}
268
269static int rds_tcp_laddr_check(struct net *net, __be32 addr)
270{
271 if (inet_addr_type(net, addr) == RTN_LOCAL)
272 return 0;
273 return -EADDRNOTAVAIL;
274}
275
276static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
277{
278 struct rds_tcp_connection *tc;
279 int i;
280
281 for (i = 0; i < RDS_MPATH_WORKERS; i++) {
282 tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
283 if (!tc)
284 return -ENOMEM;
285
286 mutex_init(&tc->t_conn_path_lock);
287 tc->t_sock = NULL;
288 tc->t_tinc = NULL;
289 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
290 tc->t_tinc_data_rem = 0;
291
292 conn->c_path[i].cp_transport_data = tc;
293 tc->t_cpath = &conn->c_path[i];
294
295 spin_lock_irq(&rds_tcp_conn_lock);
296 list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
297 spin_unlock_irq(&rds_tcp_conn_lock);
298 rdsdebug("rds_conn_path [%d] tc %p\n", i,
299 conn->c_path[i].cp_transport_data);
300 }
301
302 return 0;
303}
304
305static void rds_tcp_conn_free(void *arg)
306{
307 struct rds_tcp_connection *tc = arg;
308 unsigned long flags;
309 rdsdebug("freeing tc %p\n", tc);
310
311 spin_lock_irqsave(&rds_tcp_conn_lock, flags);
312 list_del(&tc->t_tcp_node);
313 spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
314
315 kmem_cache_free(rds_tcp_conn_slab, tc);
316}
317
318static bool list_has_conn(struct list_head *list, struct rds_connection *conn)
319{
320 struct rds_tcp_connection *tc, *_tc;
321
322 list_for_each_entry_safe(tc, _tc, list, t_tcp_node) {
323 if (tc->t_cpath->cp_conn == conn)
324 return true;
325 }
326 return false;
327}
328
329static void rds_tcp_destroy_conns(void)
330{
331 struct rds_tcp_connection *tc, *_tc;
332 LIST_HEAD(tmp_list);
333
334 /* avoid calling conn_destroy with irqs off */
335 spin_lock_irq(&rds_tcp_conn_lock);
336 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
337 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn))
338 list_move_tail(&tc->t_tcp_node, &tmp_list);
339 }
340 spin_unlock_irq(&rds_tcp_conn_lock);
341
342 list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node)
343 rds_conn_destroy(tc->t_cpath->cp_conn);
344}
345
346static void rds_tcp_exit(void);
347
348struct rds_transport rds_tcp_transport = {
349 .laddr_check = rds_tcp_laddr_check,
350 .xmit_path_prepare = rds_tcp_xmit_path_prepare,
351 .xmit_path_complete = rds_tcp_xmit_path_complete,
352 .xmit = rds_tcp_xmit,
353 .recv_path = rds_tcp_recv_path,
354 .conn_alloc = rds_tcp_conn_alloc,
355 .conn_free = rds_tcp_conn_free,
356 .conn_path_connect = rds_tcp_conn_path_connect,
357 .conn_path_shutdown = rds_tcp_conn_path_shutdown,
358 .inc_copy_to_user = rds_tcp_inc_copy_to_user,
359 .inc_free = rds_tcp_inc_free,
360 .stats_info_copy = rds_tcp_stats_info_copy,
361 .exit = rds_tcp_exit,
362 .t_owner = THIS_MODULE,
363 .t_name = "tcp",
364 .t_type = RDS_TRANS_TCP,
365 .t_prefer_loopback = 1,
366 .t_mp_capable = 1,
367};
368
369static unsigned int rds_tcp_netid;
370
371/* per-network namespace private data for this module */
372struct rds_tcp_net {
373 struct socket *rds_tcp_listen_sock;
374 struct work_struct rds_tcp_accept_w;
375 struct ctl_table_header *rds_tcp_sysctl;
376 struct ctl_table *ctl_table;
377 int sndbuf_size;
378 int rcvbuf_size;
379};
380
381/* All module specific customizations to the RDS-TCP socket should be done in
382 * rds_tcp_tune() and applied after socket creation.
383 */
384void rds_tcp_tune(struct socket *sock)
385{
386 struct sock *sk = sock->sk;
387 struct net *net = sock_net(sk);
388 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
389
390 rds_tcp_nonagle(sock);
391 lock_sock(sk);
392 if (rtn->sndbuf_size > 0) {
393 sk->sk_sndbuf = rtn->sndbuf_size;
394 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
395 }
396 if (rtn->rcvbuf_size > 0) {
397 sk->sk_sndbuf = rtn->rcvbuf_size;
398 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
399 }
400 release_sock(sk);
401}
402
403static void rds_tcp_accept_worker(struct work_struct *work)
404{
405 struct rds_tcp_net *rtn = container_of(work,
406 struct rds_tcp_net,
407 rds_tcp_accept_w);
408
409 while (rds_tcp_accept_one(rtn->rds_tcp_listen_sock) == 0)
410 cond_resched();
411}
412
413void rds_tcp_accept_work(struct sock *sk)
414{
415 struct net *net = sock_net(sk);
416 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
417
418 queue_work(rds_wq, &rtn->rds_tcp_accept_w);
419}
420
421static __net_init int rds_tcp_init_net(struct net *net)
422{
423 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
424 struct ctl_table *tbl;
425 int err = 0;
426
427 memset(rtn, 0, sizeof(*rtn));
428
429 /* {snd, rcv}buf_size default to 0, which implies we let the
430 * stack pick the value, and permit auto-tuning of buffer size.
431 */
432 if (net == &init_net) {
433 tbl = rds_tcp_sysctl_table;
434 } else {
435 tbl = kmemdup(rds_tcp_sysctl_table,
436 sizeof(rds_tcp_sysctl_table), GFP_KERNEL);
437 if (!tbl) {
438 pr_warn("could not set allocate syctl table\n");
439 return -ENOMEM;
440 }
441 rtn->ctl_table = tbl;
442 }
443 tbl[RDS_TCP_SNDBUF].data = &rtn->sndbuf_size;
444 tbl[RDS_TCP_RCVBUF].data = &rtn->rcvbuf_size;
445 rtn->rds_tcp_sysctl = register_net_sysctl(net, "net/rds/tcp", tbl);
446 if (!rtn->rds_tcp_sysctl) {
447 pr_warn("could not register sysctl\n");
448 err = -ENOMEM;
449 goto fail;
450 }
451 rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net);
452 if (!rtn->rds_tcp_listen_sock) {
453 pr_warn("could not set up listen sock\n");
454 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
455 rtn->rds_tcp_sysctl = NULL;
456 err = -EAFNOSUPPORT;
457 goto fail;
458 }
459 INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
460 return 0;
461
462fail:
463 if (net != &init_net)
464 kfree(tbl);
465 return err;
466}
467
468static void __net_exit rds_tcp_exit_net(struct net *net)
469{
470 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
471
472 if (rtn->rds_tcp_sysctl)
473 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
474
475 if (net != &init_net && rtn->ctl_table)
476 kfree(rtn->ctl_table);
477
478 /* If rds_tcp_exit_net() is called as a result of netns deletion,
479 * the rds_tcp_kill_sock() device notifier would already have cleaned
480 * up the listen socket, thus there is no work to do in this function.
481 *
482 * If rds_tcp_exit_net() is called as a result of module unload,
483 * i.e., due to rds_tcp_exit() -> unregister_pernet_subsys(), then
484 * we do need to clean up the listen socket here.
485 */
486 if (rtn->rds_tcp_listen_sock) {
487 rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
488 rtn->rds_tcp_listen_sock = NULL;
489 flush_work(&rtn->rds_tcp_accept_w);
490 }
491}
492
493static struct pernet_operations rds_tcp_net_ops = {
494 .init = rds_tcp_init_net,
495 .exit = rds_tcp_exit_net,
496 .id = &rds_tcp_netid,
497 .size = sizeof(struct rds_tcp_net),
498};
499
500/* explicitly send a RST on each socket, thereby releasing any socket refcnts
501 * that may otherwise hold up netns deletion.
502 */
503static void rds_tcp_conn_paths_destroy(struct rds_connection *conn)
504{
505 struct rds_conn_path *cp;
506 struct rds_tcp_connection *tc;
507 int i;
508 struct sock *sk;
509
510 for (i = 0; i < RDS_MPATH_WORKERS; i++) {
511 cp = &conn->c_path[i];
512 tc = cp->cp_transport_data;
513 if (!tc->t_sock)
514 continue;
515 sk = tc->t_sock->sk;
516 sk->sk_prot->disconnect(sk, 0);
517 tcp_done(sk);
518 }
519}
520
521static void rds_tcp_kill_sock(struct net *net)
522{
523 struct rds_tcp_connection *tc, *_tc;
524 LIST_HEAD(tmp_list);
525 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
526
527 rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
528 rtn->rds_tcp_listen_sock = NULL;
529 flush_work(&rtn->rds_tcp_accept_w);
530 spin_lock_irq(&rds_tcp_conn_lock);
531 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
532 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
533
534 if (net != c_net || !tc->t_sock)
535 continue;
536 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn))
537 list_move_tail(&tc->t_tcp_node, &tmp_list);
538 }
539 spin_unlock_irq(&rds_tcp_conn_lock);
540 list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
541 rds_tcp_conn_paths_destroy(tc->t_cpath->cp_conn);
542 rds_conn_destroy(tc->t_cpath->cp_conn);
543 }
544}
545
546void *rds_tcp_listen_sock_def_readable(struct net *net)
547{
548 struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
549
550 return rtn->rds_tcp_listen_sock->sk->sk_user_data;
551}
552
553static int rds_tcp_dev_event(struct notifier_block *this,
554 unsigned long event, void *ptr)
555{
556 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
557
558 /* rds-tcp registers as a pernet subys, so the ->exit will only
559 * get invoked after network acitivity has quiesced. We need to
560 * clean up all sockets to quiesce network activity, and use
561 * the unregistration of the per-net loopback device as a trigger
562 * to start that cleanup.
563 */
564 if (event == NETDEV_UNREGISTER_FINAL &&
565 dev->ifindex == LOOPBACK_IFINDEX)
566 rds_tcp_kill_sock(dev_net(dev));
567
568 return NOTIFY_DONE;
569}
570
571static struct notifier_block rds_tcp_dev_notifier = {
572 .notifier_call = rds_tcp_dev_event,
573 .priority = -10, /* must be called after other network notifiers */
574};
575
576/* when sysctl is used to modify some kernel socket parameters,this
577 * function resets the RDS connections in that netns so that we can
578 * restart with new parameters. The assumption is that such reset
579 * events are few and far-between.
580 */
581static void rds_tcp_sysctl_reset(struct net *net)
582{
583 struct rds_tcp_connection *tc, *_tc;
584
585 spin_lock_irq(&rds_tcp_conn_lock);
586 list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
587 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
588
589 if (net != c_net || !tc->t_sock)
590 continue;
591
592 /* reconnect with new parameters */
593 rds_conn_path_drop(tc->t_cpath);
594 }
595 spin_unlock_irq(&rds_tcp_conn_lock);
596}
597
598static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
599 void __user *buffer, size_t *lenp,
600 loff_t *fpos)
601{
602 struct net *net = current->nsproxy->net_ns;
603 int err;
604
605 err = proc_dointvec_minmax(ctl, write, buffer, lenp, fpos);
606 if (err < 0) {
607 pr_warn("Invalid input. Must be >= %d\n",
608 *(int *)(ctl->extra1));
609 return err;
610 }
611 if (write)
612 rds_tcp_sysctl_reset(net);
613 return 0;
614}
615
616static void rds_tcp_exit(void)
617{
618 rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
619 unregister_pernet_subsys(&rds_tcp_net_ops);
620 if (unregister_netdevice_notifier(&rds_tcp_dev_notifier))
621 pr_warn("could not unregister rds_tcp_dev_notifier\n");
622 rds_tcp_destroy_conns();
623 rds_trans_unregister(&rds_tcp_transport);
624 rds_tcp_recv_exit();
625 kmem_cache_destroy(rds_tcp_conn_slab);
626}
627module_exit(rds_tcp_exit);
628
629static int rds_tcp_init(void)
630{
631 int ret;
632
633 rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
634 sizeof(struct rds_tcp_connection),
635 0, 0, NULL);
636 if (!rds_tcp_conn_slab) {
637 ret = -ENOMEM;
638 goto out;
639 }
640
641 ret = register_netdevice_notifier(&rds_tcp_dev_notifier);
642 if (ret) {
643 pr_warn("could not register rds_tcp_dev_notifier\n");
644 goto out;
645 }
646
647 ret = register_pernet_subsys(&rds_tcp_net_ops);
648 if (ret)
649 goto out_slab;
650
651 ret = rds_tcp_recv_init();
652 if (ret)
653 goto out_pernet;
654
655 ret = rds_trans_register(&rds_tcp_transport);
656 if (ret)
657 goto out_recv;
658
659 rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
660
661 goto out;
662
663out_recv:
664 rds_tcp_recv_exit();
665out_pernet:
666 unregister_pernet_subsys(&rds_tcp_net_ops);
667out_slab:
668 if (unregister_netdevice_notifier(&rds_tcp_dev_notifier))
669 pr_warn("could not unregister rds_tcp_dev_notifier\n");
670 kmem_cache_destroy(rds_tcp_conn_slab);
671out:
672 return ret;
673}
674module_init(rds_tcp_init);
675
676MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
677MODULE_DESCRIPTION("RDS: TCP transport");
678MODULE_LICENSE("Dual BSD/GPL");
679