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 */
71 struct workqueue_struct *rds_wq;
72 EXPORT_SYMBOL_GPL(rds_wq);
73
rds_connect_complete(struct rds_connection * conn)74 void 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 }
93 EXPORT_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 */
rds_queue_reconnect(struct rds_connection * conn)113 void 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
rds_connect_worker(struct work_struct * work)139 void 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
rds_send_worker(struct work_struct * work)159 void 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
rds_recv_worker(struct work_struct * work)183 void 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
rds_shutdown_worker(struct work_struct * work)205 void 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
rds_threads_exit(void)212 void rds_threads_exit(void)
213 {
214 destroy_workqueue(rds_wq);
215 }
216
rds_threads_init(void)217 int rds_threads_init(void)
218 {
219 rds_wq = create_singlethread_workqueue("krdsd");
220 if (!rds_wq)
221 return -ENOMEM;
222
223 return 0;
224 }
225