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/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_complete(struct rds_connection *conn)
75{
76 if (!rds_conn_transition(conn, RDS_CONN_CONNECTING, RDS_CONN_UP)) {
77 printk(KERN_WARNING "%s: Cannot transition to state UP, "
78 "current state is %d\n",
79 __func__,
80 atomic_read(&conn->c_state));
81 atomic_set(&conn->c_state, RDS_CONN_ERROR);
82 queue_work(rds_wq, &conn->c_down_w);
83 return;
84 }
85
86 rdsdebug("conn %p for %pI4 to %pI4 complete\n",
87 conn, &conn->c_laddr, &conn->c_faddr);
88
89 conn->c_reconnect_jiffies = 0;
90 set_bit(0, &conn->c_map_queued);
91 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
92 queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
93}
94EXPORT_SYMBOL_GPL(rds_connect_complete);
95
96/*
97 * This random exponential backoff is relied on to eventually resolve racing
98 * connects.
99 *
100 * If connect attempts race then both parties drop both connections and come
101 * here to wait for a random amount of time before trying again. Eventually
102 * the backoff range will be so much greater than the time it takes to
103 * establish a connection that one of the pair will establish the connection
104 * before the other's random delay fires.
105 *
106 * Connection attempts that arrive while a connection is already established
107 * are also considered to be racing connects. This lets a connection from
108 * a rebooted machine replace an existing stale connection before the transport
109 * notices that the connection has failed.
110 *
111 * We should *always* start with a random backoff; otherwise a broken connection
112 * will always take several iterations to be re-established.
113 */
114void rds_queue_reconnect(struct rds_connection *conn)
115{
116 unsigned long rand;
117
118 rdsdebug("conn %p for %pI4 to %pI4 reconnect jiffies %lu\n",
119 conn, &conn->c_laddr, &conn->c_faddr,
120 conn->c_reconnect_jiffies);
121
122 set_bit(RDS_RECONNECT_PENDING, &conn->c_flags);
123 if (conn->c_reconnect_jiffies == 0) {
124 conn->c_reconnect_jiffies = rds_sysctl_reconnect_min_jiffies;
125 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
126 return;
127 }
128
129 get_random_bytes(&rand, sizeof(rand));
130 rdsdebug("%lu delay %lu ceil conn %p for %pI4 -> %pI4\n",
131 rand % conn->c_reconnect_jiffies, conn->c_reconnect_jiffies,
132 conn, &conn->c_laddr, &conn->c_faddr);
133 queue_delayed_work(rds_wq, &conn->c_conn_w,
134 rand % conn->c_reconnect_jiffies);
135
136 conn->c_reconnect_jiffies = min(conn->c_reconnect_jiffies * 2,
137 rds_sysctl_reconnect_max_jiffies);
138}
139
140void rds_connect_worker(struct work_struct *work)
141{
142 struct rds_connection *conn = container_of(work, struct rds_connection, c_conn_w.work);
143 int ret;
144
145 clear_bit(RDS_RECONNECT_PENDING, &conn->c_flags);
146 if (rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
147 ret = conn->c_trans->conn_connect(conn);
148 rdsdebug("conn %p for %pI4 to %pI4 dispatched, ret %d\n",
149 conn, &conn->c_laddr, &conn->c_faddr, ret);
150
151 if (ret) {
152 if (rds_conn_transition(conn, RDS_CONN_CONNECTING, RDS_CONN_DOWN))
153 rds_queue_reconnect(conn);
154 else
155 rds_conn_error(conn, "RDS: connect failed\n");
156 }
157 }
158}
159
160void rds_send_worker(struct work_struct *work)
161{
162 struct rds_connection *conn = container_of(work, struct rds_connection, c_send_w.work);
163 int ret;
164
165 if (rds_conn_state(conn) == RDS_CONN_UP) {
166 ret = rds_send_xmit(conn);
167 rdsdebug("conn %p ret %d\n", conn, ret);
168 switch (ret) {
169 case -EAGAIN:
170 rds_stats_inc(s_send_immediate_retry);
171 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
172 break;
173 case -ENOMEM:
174 rds_stats_inc(s_send_delayed_retry);
175 queue_delayed_work(rds_wq, &conn->c_send_w, 2);
176 default:
177 break;
178 }
179 }
180}
181
182void rds_recv_worker(struct work_struct *work)
183{
184 struct rds_connection *conn = container_of(work, struct rds_connection, c_recv_w.work);
185 int ret;
186
187 if (rds_conn_state(conn) == RDS_CONN_UP) {
188 ret = conn->c_trans->recv(conn);
189 rdsdebug("conn %p ret %d\n", conn, ret);
190 switch (ret) {
191 case -EAGAIN:
192 rds_stats_inc(s_recv_immediate_retry);
193 queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
194 break;
195 case -ENOMEM:
196 rds_stats_inc(s_recv_delayed_retry);
197 queue_delayed_work(rds_wq, &conn->c_recv_w, 2);
198 default:
199 break;
200 }
201 }
202}
203
204void rds_shutdown_worker(struct work_struct *work)
205{
206 struct rds_connection *conn = container_of(work, struct rds_connection, c_down_w);
207
208 rds_conn_shutdown(conn);
209}
210
211void rds_threads_exit(void)
212{
213 destroy_workqueue(rds_wq);
214}
215
216int rds_threads_init(void)
217{
218 rds_wq = create_singlethread_workqueue("krdsd");
219 if (!rds_wq)
220 return -ENOMEM;
221
222 return 0;
223}
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#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 %pI4 to %pI4 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}
97EXPORT_SYMBOL_GPL(rds_connect_path_complete);
98
99void rds_connect_complete(struct rds_connection *conn)
100{
101 rds_connect_path_complete(&conn->c_path[0], RDS_CONN_CONNECTING);
102}
103EXPORT_SYMBOL_GPL(rds_connect_complete);
104
105/*
106 * This random exponential backoff is relied on to eventually resolve racing
107 * connects.
108 *
109 * If connect attempts race then both parties drop both connections and come
110 * here to wait for a random amount of time before trying again. Eventually
111 * the backoff range will be so much greater than the time it takes to
112 * establish a connection that one of the pair will establish the connection
113 * before the other's random delay fires.
114 *
115 * Connection attempts that arrive while a connection is already established
116 * are also considered to be racing connects. This lets a connection from
117 * a rebooted machine replace an existing stale connection before the transport
118 * notices that the connection has failed.
119 *
120 * We should *always* start with a random backoff; otherwise a broken connection
121 * will always take several iterations to be re-established.
122 */
123void rds_queue_reconnect(struct rds_conn_path *cp)
124{
125 unsigned long rand;
126 struct rds_connection *conn = cp->cp_conn;
127
128 rdsdebug("conn %p for %pI4 to %pI4 reconnect jiffies %lu\n",
129 conn, &conn->c_laddr, &conn->c_faddr,
130 cp->cp_reconnect_jiffies);
131
132 /* let peer with smaller addr initiate reconnect, to avoid duels */
133 if (conn->c_trans->t_type == RDS_TRANS_TCP &&
134 !IS_CANONICAL(conn->c_laddr, conn->c_faddr))
135 return;
136
137 set_bit(RDS_RECONNECT_PENDING, &cp->cp_flags);
138 if (cp->cp_reconnect_jiffies == 0) {
139 cp->cp_reconnect_jiffies = rds_sysctl_reconnect_min_jiffies;
140 rcu_read_lock();
141 if (!rds_destroy_pending(cp->cp_conn))
142 queue_delayed_work(rds_wq, &cp->cp_conn_w, 0);
143 rcu_read_unlock();
144 return;
145 }
146
147 get_random_bytes(&rand, sizeof(rand));
148 rdsdebug("%lu delay %lu ceil conn %p for %pI4 -> %pI4\n",
149 rand % cp->cp_reconnect_jiffies, cp->cp_reconnect_jiffies,
150 conn, &conn->c_laddr, &conn->c_faddr);
151 rcu_read_lock();
152 if (!rds_destroy_pending(cp->cp_conn))
153 queue_delayed_work(rds_wq, &cp->cp_conn_w,
154 rand % cp->cp_reconnect_jiffies);
155 rcu_read_unlock();
156
157 cp->cp_reconnect_jiffies = min(cp->cp_reconnect_jiffies * 2,
158 rds_sysctl_reconnect_max_jiffies);
159}
160
161void rds_connect_worker(struct work_struct *work)
162{
163 struct rds_conn_path *cp = container_of(work,
164 struct rds_conn_path,
165 cp_conn_w.work);
166 struct rds_connection *conn = cp->cp_conn;
167 int ret;
168
169 if (cp->cp_index > 0 &&
170 !IS_CANONICAL(cp->cp_conn->c_laddr, cp->cp_conn->c_faddr))
171 return;
172 clear_bit(RDS_RECONNECT_PENDING, &cp->cp_flags);
173 ret = rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_CONNECTING);
174 if (ret) {
175 ret = conn->c_trans->conn_path_connect(cp);
176 rdsdebug("conn %p for %pI4 to %pI4 dispatched, ret %d\n",
177 conn, &conn->c_laddr, &conn->c_faddr, ret);
178
179 if (ret) {
180 if (rds_conn_path_transition(cp,
181 RDS_CONN_CONNECTING,
182 RDS_CONN_DOWN))
183 rds_queue_reconnect(cp);
184 else
185 rds_conn_path_error(cp, "connect failed\n");
186 }
187 }
188}
189
190void rds_send_worker(struct work_struct *work)
191{
192 struct rds_conn_path *cp = container_of(work,
193 struct rds_conn_path,
194 cp_send_w.work);
195 int ret;
196
197 if (rds_conn_path_state(cp) == RDS_CONN_UP) {
198 clear_bit(RDS_LL_SEND_FULL, &cp->cp_flags);
199 ret = rds_send_xmit(cp);
200 cond_resched();
201 rdsdebug("conn %p ret %d\n", cp->cp_conn, ret);
202 switch (ret) {
203 case -EAGAIN:
204 rds_stats_inc(s_send_immediate_retry);
205 queue_delayed_work(rds_wq, &cp->cp_send_w, 0);
206 break;
207 case -ENOMEM:
208 rds_stats_inc(s_send_delayed_retry);
209 queue_delayed_work(rds_wq, &cp->cp_send_w, 2);
210 default:
211 break;
212 }
213 }
214}
215
216void rds_recv_worker(struct work_struct *work)
217{
218 struct rds_conn_path *cp = container_of(work,
219 struct rds_conn_path,
220 cp_recv_w.work);
221 int ret;
222
223 if (rds_conn_path_state(cp) == RDS_CONN_UP) {
224 ret = cp->cp_conn->c_trans->recv_path(cp);
225 rdsdebug("conn %p ret %d\n", cp->cp_conn, ret);
226 switch (ret) {
227 case -EAGAIN:
228 rds_stats_inc(s_recv_immediate_retry);
229 queue_delayed_work(rds_wq, &cp->cp_recv_w, 0);
230 break;
231 case -ENOMEM:
232 rds_stats_inc(s_recv_delayed_retry);
233 queue_delayed_work(rds_wq, &cp->cp_recv_w, 2);
234 default:
235 break;
236 }
237 }
238}
239
240void rds_shutdown_worker(struct work_struct *work)
241{
242 struct rds_conn_path *cp = container_of(work,
243 struct rds_conn_path,
244 cp_down_w);
245
246 rds_conn_shutdown(cp);
247}
248
249void rds_threads_exit(void)
250{
251 destroy_workqueue(rds_wq);
252}
253
254int rds_threads_init(void)
255{
256 rds_wq = create_singlethread_workqueue("krdsd");
257 if (!rds_wq)
258 return -ENOMEM;
259
260 return 0;
261}