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/random.h>
35#include <linux/export.h>
36
37#include "rds.h"
38
39/*
40 * All of connection management is simplified by serializing it through
41 * work queues that execute in a connection managing thread.
42 *
43 * TCP wants to send acks through sendpage() in response to data_ready(),
44 * but it needs a process context to do so.
45 *
46 * The receive paths need to allocate but can't drop packets (!) so we have
47 * a thread around to block allocating if the receive fast path sees an
48 * allocation failure.
49 */
50
51/* Grand Unified Theory of connection life cycle:
52 * At any point in time, the connection can be in one of these states:
53 * DOWN, CONNECTING, UP, DISCONNECTING, ERROR
54 *
55 * The following transitions are possible:
56 * ANY -> ERROR
57 * UP -> DISCONNECTING
58 * ERROR -> DISCONNECTING
59 * DISCONNECTING -> DOWN
60 * DOWN -> CONNECTING
61 * CONNECTING -> UP
62 *
63 * Transition to state DISCONNECTING/DOWN:
64 * - Inside the shutdown worker; synchronizes with xmit path
65 * through RDS_IN_XMIT, and with connection management callbacks
66 * via c_cm_lock.
67 *
68 * For receive callbacks, we rely on the underlying transport
69 * (TCP, IB/RDMA) to provide the necessary synchronisation.
70 */
71struct workqueue_struct *rds_wq;
72EXPORT_SYMBOL_GPL(rds_wq);
73
74void rds_connect_path_complete(struct rds_conn_path *cp, int curr)
75{
76 if (!rds_conn_path_transition(cp, curr, RDS_CONN_UP)) {
77 printk(KERN_WARNING "%s: Cannot transition to state UP, "
78 "current state is %d\n",
79 __func__,
80 atomic_read(&cp->cp_state));
81 rds_conn_path_drop(cp, false);
82 return;
83 }
84
85 rdsdebug("conn %p for %pI6c to %pI6c complete\n",
86 cp->cp_conn, &cp->cp_conn->c_laddr, &cp->cp_conn->c_faddr);
87
88 cp->cp_reconnect_jiffies = 0;
89 set_bit(0, &cp->cp_conn->c_map_queued);
90 rcu_read_lock();
91 if (!rds_destroy_pending(cp->cp_conn)) {
92 queue_delayed_work(rds_wq, &cp->cp_send_w, 0);
93 queue_delayed_work(rds_wq, &cp->cp_recv_w, 0);
94 }
95 rcu_read_unlock();
96 cp->cp_conn->c_proposed_version = RDS_PROTOCOL_VERSION;
97}
98EXPORT_SYMBOL_GPL(rds_connect_path_complete);
99
100void rds_connect_complete(struct rds_connection *conn)
101{
102 rds_connect_path_complete(&conn->c_path[0], RDS_CONN_CONNECTING);
103}
104EXPORT_SYMBOL_GPL(rds_connect_complete);
105
106/*
107 * This random exponential backoff is relied on to eventually resolve racing
108 * connects.
109 *
110 * If connect attempts race then both parties drop both connections and come
111 * here to wait for a random amount of time before trying again. Eventually
112 * the backoff range will be so much greater than the time it takes to
113 * establish a connection that one of the pair will establish the connection
114 * before the other's random delay fires.
115 *
116 * Connection attempts that arrive while a connection is already established
117 * are also considered to be racing connects. This lets a connection from
118 * a rebooted machine replace an existing stale connection before the transport
119 * notices that the connection has failed.
120 *
121 * We should *always* start with a random backoff; otherwise a broken connection
122 * will always take several iterations to be re-established.
123 */
124void rds_queue_reconnect(struct rds_conn_path *cp)
125{
126 unsigned long rand;
127 struct rds_connection *conn = cp->cp_conn;
128
129 rdsdebug("conn %p for %pI6c to %pI6c reconnect jiffies %lu\n",
130 conn, &conn->c_laddr, &conn->c_faddr,
131 cp->cp_reconnect_jiffies);
132
133 /* let peer with smaller addr initiate reconnect, to avoid duels */
134 if (conn->c_trans->t_type == RDS_TRANS_TCP &&
135 rds_addr_cmp(&conn->c_laddr, &conn->c_faddr) >= 0)
136 return;
137
138 set_bit(RDS_RECONNECT_PENDING, &cp->cp_flags);
139 if (cp->cp_reconnect_jiffies == 0) {
140 cp->cp_reconnect_jiffies = rds_sysctl_reconnect_min_jiffies;
141 rcu_read_lock();
142 if (!rds_destroy_pending(cp->cp_conn))
143 queue_delayed_work(rds_wq, &cp->cp_conn_w, 0);
144 rcu_read_unlock();
145 return;
146 }
147
148 get_random_bytes(&rand, sizeof(rand));
149 rdsdebug("%lu delay %lu ceil conn %p for %pI6c -> %pI6c\n",
150 rand % cp->cp_reconnect_jiffies, cp->cp_reconnect_jiffies,
151 conn, &conn->c_laddr, &conn->c_faddr);
152 rcu_read_lock();
153 if (!rds_destroy_pending(cp->cp_conn))
154 queue_delayed_work(rds_wq, &cp->cp_conn_w,
155 rand % cp->cp_reconnect_jiffies);
156 rcu_read_unlock();
157
158 cp->cp_reconnect_jiffies = min(cp->cp_reconnect_jiffies * 2,
159 rds_sysctl_reconnect_max_jiffies);
160}
161
162void rds_connect_worker(struct work_struct *work)
163{
164 struct rds_conn_path *cp = container_of(work,
165 struct rds_conn_path,
166 cp_conn_w.work);
167 struct rds_connection *conn = cp->cp_conn;
168 int ret;
169
170 if (cp->cp_index > 0 &&
171 rds_addr_cmp(&cp->cp_conn->c_laddr, &cp->cp_conn->c_faddr) >= 0)
172 return;
173 clear_bit(RDS_RECONNECT_PENDING, &cp->cp_flags);
174 ret = rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_CONNECTING);
175 if (ret) {
176 ret = conn->c_trans->conn_path_connect(cp);
177 rdsdebug("conn %p for %pI6c to %pI6c dispatched, ret %d\n",
178 conn, &conn->c_laddr, &conn->c_faddr, ret);
179
180 if (ret) {
181 if (rds_conn_path_transition(cp,
182 RDS_CONN_CONNECTING,
183 RDS_CONN_DOWN))
184 rds_queue_reconnect(cp);
185 else
186 rds_conn_path_error(cp, "connect failed\n");
187 }
188 }
189}
190
191void rds_send_worker(struct work_struct *work)
192{
193 struct rds_conn_path *cp = container_of(work,
194 struct rds_conn_path,
195 cp_send_w.work);
196 int ret;
197
198 if (rds_conn_path_state(cp) == RDS_CONN_UP) {
199 clear_bit(RDS_LL_SEND_FULL, &cp->cp_flags);
200 ret = rds_send_xmit(cp);
201 cond_resched();
202 rdsdebug("conn %p ret %d\n", cp->cp_conn, ret);
203 switch (ret) {
204 case -EAGAIN:
205 rds_stats_inc(s_send_immediate_retry);
206 queue_delayed_work(rds_wq, &cp->cp_send_w, 0);
207 break;
208 case -ENOMEM:
209 rds_stats_inc(s_send_delayed_retry);
210 queue_delayed_work(rds_wq, &cp->cp_send_w, 2);
211 break;
212 default:
213 break;
214 }
215 }
216}
217
218void rds_recv_worker(struct work_struct *work)
219{
220 struct rds_conn_path *cp = container_of(work,
221 struct rds_conn_path,
222 cp_recv_w.work);
223 int ret;
224
225 if (rds_conn_path_state(cp) == RDS_CONN_UP) {
226 ret = cp->cp_conn->c_trans->recv_path(cp);
227 rdsdebug("conn %p ret %d\n", cp->cp_conn, ret);
228 switch (ret) {
229 case -EAGAIN:
230 rds_stats_inc(s_recv_immediate_retry);
231 queue_delayed_work(rds_wq, &cp->cp_recv_w, 0);
232 break;
233 case -ENOMEM:
234 rds_stats_inc(s_recv_delayed_retry);
235 queue_delayed_work(rds_wq, &cp->cp_recv_w, 2);
236 break;
237 default:
238 break;
239 }
240 }
241}
242
243void rds_shutdown_worker(struct work_struct *work)
244{
245 struct rds_conn_path *cp = container_of(work,
246 struct rds_conn_path,
247 cp_down_w);
248
249 rds_conn_shutdown(cp);
250}
251
252void rds_threads_exit(void)
253{
254 destroy_workqueue(rds_wq);
255}
256
257int rds_threads_init(void)
258{
259 rds_wq = create_singlethread_workqueue("krdsd");
260 if (!rds_wq)
261 return -ENOMEM;
262
263 return 0;
264}
265
266/* Compare two IPv6 addresses. Return 0 if the two addresses are equal.
267 * Return 1 if the first is greater. Return -1 if the second is greater.
268 */
269int rds_addr_cmp(const struct in6_addr *addr1,
270 const struct in6_addr *addr2)
271{
272#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && BITS_PER_LONG == 64
273 const __be64 *a1, *a2;
274 u64 x, y;
275
276 a1 = (__be64 *)addr1;
277 a2 = (__be64 *)addr2;
278
279 if (*a1 != *a2) {
280 if (be64_to_cpu(*a1) < be64_to_cpu(*a2))
281 return -1;
282 else
283 return 1;
284 } else {
285 x = be64_to_cpu(*++a1);
286 y = be64_to_cpu(*++a2);
287 if (x < y)
288 return -1;
289 else if (x > y)
290 return 1;
291 else
292 return 0;
293 }
294#else
295 u32 a, b;
296 int i;
297
298 for (i = 0; i < 4; i++) {
299 if (addr1->s6_addr32[i] != addr2->s6_addr32[i]) {
300 a = ntohl(addr1->s6_addr32[i]);
301 b = ntohl(addr2->s6_addr32[i]);
302 if (a < b)
303 return -1;
304 else if (a > b)
305 return 1;
306 }
307 }
308 return 0;
309#endif
310}
311EXPORT_SYMBOL_GPL(rds_addr_cmp);
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/random.h>
35
36#include "rds.h"
37
38/*
39 * All of connection management is simplified by serializing it through
40 * work queues that execute in a connection managing thread.
41 *
42 * TCP wants to send acks through sendpage() in response to data_ready(),
43 * but it needs a process context to do so.
44 *
45 * The receive paths need to allocate but can't drop packets (!) so we have
46 * a thread around to block allocating if the receive fast path sees an
47 * allocation failure.
48 */
49
50/* Grand Unified Theory of connection life cycle:
51 * At any point in time, the connection can be in one of these states:
52 * DOWN, CONNECTING, UP, DISCONNECTING, ERROR
53 *
54 * The following transitions are possible:
55 * ANY -> ERROR
56 * UP -> DISCONNECTING
57 * ERROR -> DISCONNECTING
58 * DISCONNECTING -> DOWN
59 * DOWN -> CONNECTING
60 * CONNECTING -> UP
61 *
62 * Transition to state DISCONNECTING/DOWN:
63 * - Inside the shutdown worker; synchronizes with xmit path
64 * through RDS_IN_XMIT, and with connection management callbacks
65 * via c_cm_lock.
66 *
67 * For receive callbacks, we rely on the underlying transport
68 * (TCP, IB/RDMA) to provide the necessary synchronisation.
69 */
70struct workqueue_struct *rds_wq;
71EXPORT_SYMBOL_GPL(rds_wq);
72
73void rds_connect_complete(struct rds_connection *conn)
74{
75 if (!rds_conn_transition(conn, RDS_CONN_CONNECTING, RDS_CONN_UP)) {
76 printk(KERN_WARNING "%s: Cannot transition to state UP, "
77 "current state is %d\n",
78 __func__,
79 atomic_read(&conn->c_state));
80 atomic_set(&conn->c_state, RDS_CONN_ERROR);
81 queue_work(rds_wq, &conn->c_down_w);
82 return;
83 }
84
85 rdsdebug("conn %p for %pI4 to %pI4 complete\n",
86 conn, &conn->c_laddr, &conn->c_faddr);
87
88 conn->c_reconnect_jiffies = 0;
89 set_bit(0, &conn->c_map_queued);
90 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
91 queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
92}
93EXPORT_SYMBOL_GPL(rds_connect_complete);
94
95/*
96 * This random exponential backoff is relied on to eventually resolve racing
97 * connects.
98 *
99 * If connect attempts race then both parties drop both connections and come
100 * here to wait for a random amount of time before trying again. Eventually
101 * the backoff range will be so much greater than the time it takes to
102 * establish a connection that one of the pair will establish the connection
103 * before the other's random delay fires.
104 *
105 * Connection attempts that arrive while a connection is already established
106 * are also considered to be racing connects. This lets a connection from
107 * a rebooted machine replace an existing stale connection before the transport
108 * notices that the connection has failed.
109 *
110 * We should *always* start with a random backoff; otherwise a broken connection
111 * will always take several iterations to be re-established.
112 */
113void rds_queue_reconnect(struct rds_connection *conn)
114{
115 unsigned long rand;
116
117 rdsdebug("conn %p for %pI4 to %pI4 reconnect jiffies %lu\n",
118 conn, &conn->c_laddr, &conn->c_faddr,
119 conn->c_reconnect_jiffies);
120
121 set_bit(RDS_RECONNECT_PENDING, &conn->c_flags);
122 if (conn->c_reconnect_jiffies == 0) {
123 conn->c_reconnect_jiffies = rds_sysctl_reconnect_min_jiffies;
124 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
125 return;
126 }
127
128 get_random_bytes(&rand, sizeof(rand));
129 rdsdebug("%lu delay %lu ceil conn %p for %pI4 -> %pI4\n",
130 rand % conn->c_reconnect_jiffies, conn->c_reconnect_jiffies,
131 conn, &conn->c_laddr, &conn->c_faddr);
132 queue_delayed_work(rds_wq, &conn->c_conn_w,
133 rand % conn->c_reconnect_jiffies);
134
135 conn->c_reconnect_jiffies = min(conn->c_reconnect_jiffies * 2,
136 rds_sysctl_reconnect_max_jiffies);
137}
138
139void rds_connect_worker(struct work_struct *work)
140{
141 struct rds_connection *conn = container_of(work, struct rds_connection, c_conn_w.work);
142 int ret;
143
144 clear_bit(RDS_RECONNECT_PENDING, &conn->c_flags);
145 if (rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
146 ret = conn->c_trans->conn_connect(conn);
147 rdsdebug("conn %p for %pI4 to %pI4 dispatched, ret %d\n",
148 conn, &conn->c_laddr, &conn->c_faddr, ret);
149
150 if (ret) {
151 if (rds_conn_transition(conn, RDS_CONN_CONNECTING, RDS_CONN_DOWN))
152 rds_queue_reconnect(conn);
153 else
154 rds_conn_error(conn, "RDS: connect failed\n");
155 }
156 }
157}
158
159void rds_send_worker(struct work_struct *work)
160{
161 struct rds_connection *conn = container_of(work, struct rds_connection, c_send_w.work);
162 int ret;
163
164 if (rds_conn_state(conn) == RDS_CONN_UP) {
165 ret = rds_send_xmit(conn);
166 rdsdebug("conn %p ret %d\n", conn, ret);
167 switch (ret) {
168 case -EAGAIN:
169 rds_stats_inc(s_send_immediate_retry);
170 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
171 break;
172 case -ENOMEM:
173 rds_stats_inc(s_send_delayed_retry);
174 queue_delayed_work(rds_wq, &conn->c_send_w, 2);
175 default:
176 break;
177 }
178 }
179}
180
181void rds_recv_worker(struct work_struct *work)
182{
183 struct rds_connection *conn = container_of(work, struct rds_connection, c_recv_w.work);
184 int ret;
185
186 if (rds_conn_state(conn) == RDS_CONN_UP) {
187 ret = conn->c_trans->recv(conn);
188 rdsdebug("conn %p ret %d\n", conn, ret);
189 switch (ret) {
190 case -EAGAIN:
191 rds_stats_inc(s_recv_immediate_retry);
192 queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
193 break;
194 case -ENOMEM:
195 rds_stats_inc(s_recv_delayed_retry);
196 queue_delayed_work(rds_wq, &conn->c_recv_w, 2);
197 default:
198 break;
199 }
200 }
201}
202
203void rds_shutdown_worker(struct work_struct *work)
204{
205 struct rds_connection *conn = container_of(work, struct rds_connection, c_down_w);
206
207 rds_conn_shutdown(conn);
208}
209
210void rds_threads_exit(void)
211{
212 destroy_workqueue(rds_wq);
213}
214
215int rds_threads_init(void)
216{
217 rds_wq = create_singlethread_workqueue("krdsd");
218 if (!rds_wq)
219 return -ENOMEM;
220
221 return 0;
222}