blob: 828a1bf9ea9236055956b9f4abdc4da80abc97d5 [file] [log] [blame]
Andy Grover00e0f342009-02-24 15:30:23 +00001/*
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 c_send_lock, 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;
71
72void rds_connect_complete(struct rds_connection *conn)
73{
74 if (!rds_conn_transition(conn, RDS_CONN_CONNECTING, RDS_CONN_UP)) {
75 printk(KERN_WARNING "%s: Cannot transition to state UP, "
76 "current state is %d\n",
77 __func__,
78 atomic_read(&conn->c_state));
79 atomic_set(&conn->c_state, RDS_CONN_ERROR);
80 queue_work(rds_wq, &conn->c_down_w);
81 return;
82 }
83
84 rdsdebug("conn %p for %pI4 to %pI4 complete\n",
85 conn, &conn->c_laddr, &conn->c_faddr);
86
87 conn->c_reconnect_jiffies = 0;
88 set_bit(0, &conn->c_map_queued);
89 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
90 queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
91}
92
93/*
94 * This random exponential backoff is relied on to eventually resolve racing
95 * connects.
96 *
97 * If connect attempts race then both parties drop both connections and come
98 * here to wait for a random amount of time before trying again. Eventually
99 * the backoff range will be so much greater than the time it takes to
100 * establish a connection that one of the pair will establish the connection
101 * before the other's random delay fires.
102 *
103 * Connection attempts that arrive while a connection is already established
104 * are also considered to be racing connects. This lets a connection from
105 * a rebooted machine replace an existing stale connection before the transport
106 * notices that the connection has failed.
107 *
108 * We should *always* start with a random backoff; otherwise a broken connection
109 * will always take several iterations to be re-established.
110 */
111static void rds_queue_reconnect(struct rds_connection *conn)
112{
113 unsigned long rand;
114
115 rdsdebug("conn %p for %pI4 to %pI4 reconnect jiffies %lu\n",
116 conn, &conn->c_laddr, &conn->c_faddr,
117 conn->c_reconnect_jiffies);
118
119 set_bit(RDS_RECONNECT_PENDING, &conn->c_flags);
120 if (conn->c_reconnect_jiffies == 0) {
121 conn->c_reconnect_jiffies = rds_sysctl_reconnect_min_jiffies;
122 queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
123 return;
124 }
125
126 get_random_bytes(&rand, sizeof(rand));
127 rdsdebug("%lu delay %lu ceil conn %p for %pI4 -> %pI4\n",
128 rand % conn->c_reconnect_jiffies, conn->c_reconnect_jiffies,
129 conn, &conn->c_laddr, &conn->c_faddr);
130 queue_delayed_work(rds_wq, &conn->c_conn_w,
131 rand % conn->c_reconnect_jiffies);
132
133 conn->c_reconnect_jiffies = min(conn->c_reconnect_jiffies * 2,
134 rds_sysctl_reconnect_max_jiffies);
135}
136
137void rds_connect_worker(struct work_struct *work)
138{
139 struct rds_connection *conn = container_of(work, struct rds_connection, c_conn_w.work);
140 int ret;
141
142 clear_bit(RDS_RECONNECT_PENDING, &conn->c_flags);
143 if (rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
144 ret = conn->c_trans->conn_connect(conn);
145 rdsdebug("conn %p for %pI4 to %pI4 dispatched, ret %d\n",
146 conn, &conn->c_laddr, &conn->c_faddr, ret);
147
148 if (ret) {
149 if (rds_conn_transition(conn, RDS_CONN_CONNECTING, RDS_CONN_DOWN))
150 rds_queue_reconnect(conn);
151 else
152 rds_conn_error(conn, "RDS: connect failed\n");
153 }
154 }
155}
156
157void rds_shutdown_worker(struct work_struct *work)
158{
159 struct rds_connection *conn = container_of(work, struct rds_connection, c_down_w);
160
161 /* shut it down unless it's down already */
162 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
163 /*
164 * Quiesce the connection mgmt handlers before we start tearing
165 * things down. We don't hold the mutex for the entire
166 * duration of the shutdown operation, else we may be
167 * deadlocking with the CM handler. Instead, the CM event
168 * handler is supposed to check for state DISCONNECTING
169 */
170 mutex_lock(&conn->c_cm_lock);
171 if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING)
172 && !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
173 rds_conn_error(conn, "shutdown called in state %d\n",
174 atomic_read(&conn->c_state));
175 mutex_unlock(&conn->c_cm_lock);
176 return;
177 }
178 mutex_unlock(&conn->c_cm_lock);
179
180 mutex_lock(&conn->c_send_lock);
181 conn->c_trans->conn_shutdown(conn);
182 rds_conn_reset(conn);
183 mutex_unlock(&conn->c_send_lock);
184
185 if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
186 /* This can happen - eg when we're in the middle of tearing
187 * down the connection, and someone unloads the rds module.
188 * Quite reproduceable with loopback connections.
189 * Mostly harmless.
190 */
191 rds_conn_error(conn,
192 "%s: failed to transition to state DOWN, "
193 "current state is %d\n",
194 __func__,
195 atomic_read(&conn->c_state));
196 return;
197 }
198 }
199
200 /* Then reconnect if it's still live.
201 * The passive side of an IB loopback connection is never added
202 * to the conn hash, so we never trigger a reconnect on this
203 * conn - the reconnect is always triggered by the active peer. */
204 cancel_delayed_work(&conn->c_conn_w);
205 if (!hlist_unhashed(&conn->c_hash_node))
206 rds_queue_reconnect(conn);
207}
208
209void rds_send_worker(struct work_struct *work)
210{
211 struct rds_connection *conn = container_of(work, struct rds_connection, c_send_w.work);
212 int ret;
213
214 if (rds_conn_state(conn) == RDS_CONN_UP) {
215 ret = rds_send_xmit(conn);
216 rdsdebug("conn %p ret %d\n", conn, ret);
217 switch (ret) {
218 case -EAGAIN:
219 rds_stats_inc(s_send_immediate_retry);
220 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
221 break;
222 case -ENOMEM:
223 rds_stats_inc(s_send_delayed_retry);
224 queue_delayed_work(rds_wq, &conn->c_send_w, 2);
225 default:
226 break;
227 }
228 }
229}
230
231void rds_recv_worker(struct work_struct *work)
232{
233 struct rds_connection *conn = container_of(work, struct rds_connection, c_recv_w.work);
234 int ret;
235
236 if (rds_conn_state(conn) == RDS_CONN_UP) {
237 ret = conn->c_trans->recv(conn);
238 rdsdebug("conn %p ret %d\n", conn, ret);
239 switch (ret) {
240 case -EAGAIN:
241 rds_stats_inc(s_recv_immediate_retry);
242 queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
243 break;
244 case -ENOMEM:
245 rds_stats_inc(s_recv_delayed_retry);
246 queue_delayed_work(rds_wq, &conn->c_recv_w, 2);
247 default:
248 break;
249 }
250 }
251}
252
253void rds_threads_exit(void)
254{
255 destroy_workqueue(rds_wq);
256}
257
258int __init rds_threads_init(void)
259{
260 rds_wq = create_singlethread_workqueue("krdsd");
261 if (rds_wq == NULL)
262 return -ENOMEM;
263
264 return 0;
265}