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 rds_conn_drop(conn);
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 clear_bit(RDS_LL_SEND_FULL, &conn->c_flags);
166 ret = rds_send_xmit(conn);
167 cond_resched();
168 rdsdebug("conn %p ret %d\n", conn, ret);
169 switch (ret) {
170 case -EAGAIN:
171 rds_stats_inc(s_send_immediate_retry);
172 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
173 break;
174 case -ENOMEM:
175 rds_stats_inc(s_send_delayed_retry);
176 queue_delayed_work(rds_wq, &conn->c_send_w, 2);
177 default:
178 break;
179 }
180 }
181}
182
183void rds_recv_worker(struct work_struct *work)
184{
185 struct rds_connection *conn = container_of(work, struct rds_connection, c_recv_w.work);
186 int ret;
187
188 if (rds_conn_state(conn) == RDS_CONN_UP) {
189 ret = conn->c_trans->recv(conn);
190 rdsdebug("conn %p ret %d\n", conn, ret);
191 switch (ret) {
192 case -EAGAIN:
193 rds_stats_inc(s_recv_immediate_retry);
194 queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
195 break;
196 case -ENOMEM:
197 rds_stats_inc(s_recv_delayed_retry);
198 queue_delayed_work(rds_wq, &conn->c_recv_w, 2);
199 default:
200 break;
201 }
202 }
203}
204
205void rds_shutdown_worker(struct work_struct *work)
206{
207 struct rds_connection *conn = container_of(work, struct rds_connection, c_down_w);
208
209 rds_conn_shutdown(conn);
210}
211
212void rds_threads_exit(void)
213{
214 destroy_workqueue(rds_wq);
215}
216
217int rds_threads_init(void)
218{
219 rds_wq = create_singlethread_workqueue("krdsd");
220 if (!rds_wq)
221 return -ENOMEM;
222
223 return 0;
224}
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);
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 queue_delayed_work(rds_wq, &cp->cp_send_w, 0);
91 queue_delayed_work(rds_wq, &cp->cp_recv_w, 0);
92}
93EXPORT_SYMBOL_GPL(rds_connect_path_complete);
94
95void rds_connect_complete(struct rds_connection *conn)
96{
97 rds_connect_path_complete(&conn->c_path[0], RDS_CONN_CONNECTING);
98}
99EXPORT_SYMBOL_GPL(rds_connect_complete);
100
101/*
102 * This random exponential backoff is relied on to eventually resolve racing
103 * connects.
104 *
105 * If connect attempts race then both parties drop both connections and come
106 * here to wait for a random amount of time before trying again. Eventually
107 * the backoff range will be so much greater than the time it takes to
108 * establish a connection that one of the pair will establish the connection
109 * before the other's random delay fires.
110 *
111 * Connection attempts that arrive while a connection is already established
112 * are also considered to be racing connects. This lets a connection from
113 * a rebooted machine replace an existing stale connection before the transport
114 * notices that the connection has failed.
115 *
116 * We should *always* start with a random backoff; otherwise a broken connection
117 * will always take several iterations to be re-established.
118 */
119void rds_queue_reconnect(struct rds_conn_path *cp)
120{
121 unsigned long rand;
122 struct rds_connection *conn = cp->cp_conn;
123
124 rdsdebug("conn %p for %pI4 to %pI4 reconnect jiffies %lu\n",
125 conn, &conn->c_laddr, &conn->c_faddr,
126 cp->cp_reconnect_jiffies);
127
128 /* let peer with smaller addr initiate reconnect, to avoid duels */
129 if (conn->c_trans->t_type == RDS_TRANS_TCP &&
130 conn->c_laddr > conn->c_faddr)
131 return;
132
133 set_bit(RDS_RECONNECT_PENDING, &cp->cp_flags);
134 if (cp->cp_reconnect_jiffies == 0) {
135 cp->cp_reconnect_jiffies = rds_sysctl_reconnect_min_jiffies;
136 queue_delayed_work(rds_wq, &cp->cp_conn_w, 0);
137 return;
138 }
139
140 get_random_bytes(&rand, sizeof(rand));
141 rdsdebug("%lu delay %lu ceil conn %p for %pI4 -> %pI4\n",
142 rand % cp->cp_reconnect_jiffies, cp->cp_reconnect_jiffies,
143 conn, &conn->c_laddr, &conn->c_faddr);
144 queue_delayed_work(rds_wq, &cp->cp_conn_w,
145 rand % cp->cp_reconnect_jiffies);
146
147 cp->cp_reconnect_jiffies = min(cp->cp_reconnect_jiffies * 2,
148 rds_sysctl_reconnect_max_jiffies);
149}
150
151void rds_connect_worker(struct work_struct *work)
152{
153 struct rds_conn_path *cp = container_of(work,
154 struct rds_conn_path,
155 cp_conn_w.work);
156 struct rds_connection *conn = cp->cp_conn;
157 int ret;
158
159 if (cp->cp_index > 1 && cp->cp_conn->c_laddr > cp->cp_conn->c_faddr)
160 return;
161 clear_bit(RDS_RECONNECT_PENDING, &cp->cp_flags);
162 ret = rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_CONNECTING);
163 if (ret) {
164 ret = conn->c_trans->conn_path_connect(cp);
165 rdsdebug("conn %p for %pI4 to %pI4 dispatched, ret %d\n",
166 conn, &conn->c_laddr, &conn->c_faddr, ret);
167
168 if (ret) {
169 if (rds_conn_path_transition(cp,
170 RDS_CONN_CONNECTING,
171 RDS_CONN_DOWN))
172 rds_queue_reconnect(cp);
173 else
174 rds_conn_path_error(cp, "connect failed\n");
175 }
176 }
177}
178
179void rds_send_worker(struct work_struct *work)
180{
181 struct rds_conn_path *cp = container_of(work,
182 struct rds_conn_path,
183 cp_send_w.work);
184 int ret;
185
186 if (rds_conn_path_state(cp) == RDS_CONN_UP) {
187 clear_bit(RDS_LL_SEND_FULL, &cp->cp_flags);
188 ret = rds_send_xmit(cp);
189 cond_resched();
190 rdsdebug("conn %p ret %d\n", cp->cp_conn, ret);
191 switch (ret) {
192 case -EAGAIN:
193 rds_stats_inc(s_send_immediate_retry);
194 queue_delayed_work(rds_wq, &cp->cp_send_w, 0);
195 break;
196 case -ENOMEM:
197 rds_stats_inc(s_send_delayed_retry);
198 queue_delayed_work(rds_wq, &cp->cp_send_w, 2);
199 default:
200 break;
201 }
202 }
203}
204
205void rds_recv_worker(struct work_struct *work)
206{
207 struct rds_conn_path *cp = container_of(work,
208 struct rds_conn_path,
209 cp_recv_w.work);
210 int ret;
211
212 if (rds_conn_path_state(cp) == RDS_CONN_UP) {
213 ret = cp->cp_conn->c_trans->recv_path(cp);
214 rdsdebug("conn %p ret %d\n", cp->cp_conn, ret);
215 switch (ret) {
216 case -EAGAIN:
217 rds_stats_inc(s_recv_immediate_retry);
218 queue_delayed_work(rds_wq, &cp->cp_recv_w, 0);
219 break;
220 case -ENOMEM:
221 rds_stats_inc(s_recv_delayed_retry);
222 queue_delayed_work(rds_wq, &cp->cp_recv_w, 2);
223 default:
224 break;
225 }
226 }
227}
228
229void rds_shutdown_worker(struct work_struct *work)
230{
231 struct rds_conn_path *cp = container_of(work,
232 struct rds_conn_path,
233 cp_down_w);
234
235 rds_conn_shutdown(cp);
236}
237
238void rds_threads_exit(void)
239{
240 destroy_workqueue(rds_wq);
241}
242
243int rds_threads_init(void)
244{
245 rds_wq = create_singlethread_workqueue("krdsd");
246 if (!rds_wq)
247 return -ENOMEM;
248
249 return 0;
250}