blob: e42df11bf30a63870f509c69656176091ec8aa66 [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>
Paul Gortmakerbc3b2d72011-07-15 11:47:34 -040035#include <linux/export.h>
Andy Grover00e0f342009-02-24 15:30:23 +000036
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
Zach Brown0f4b1c72010-06-04 14:41:41 -070065 * through RDS_IN_XMIT, and with connection management callbacks
Andy Grover00e0f342009-02-24 15:30:23 +000066 * 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;
Andy Grover616b7572009-08-21 12:28:32 +000072EXPORT_SYMBOL_GPL(rds_wq);
Andy Grover00e0f342009-02-24 15:30:23 +000073
Sowmini Varadhan0cb43962016-06-13 09:44:26 -070074void rds_connect_path_complete(struct rds_conn_path *cp, int curr)
Andy Grover00e0f342009-02-24 15:30:23 +000075{
Sowmini Varadhan0cb43962016-06-13 09:44:26 -070076 if (!rds_conn_path_transition(cp, curr, RDS_CONN_UP)) {
Andy Grover00e0f342009-02-24 15:30:23 +000077 printk(KERN_WARNING "%s: Cannot transition to state UP, "
78 "current state is %d\n",
79 __func__,
Sowmini Varadhan0cb43962016-06-13 09:44:26 -070080 atomic_read(&cp->cp_state));
81 rds_conn_path_drop(cp);
Andy Grover00e0f342009-02-24 15:30:23 +000082 return;
83 }
84
85 rdsdebug("conn %p for %pI4 to %pI4 complete\n",
Sowmini Varadhan0cb43962016-06-13 09:44:26 -070086 cp->cp_conn, &cp->cp_conn->c_laddr, &cp->cp_conn->c_faddr);
Andy Grover00e0f342009-02-24 15:30:23 +000087
Sowmini Varadhan0cb43962016-06-13 09:44:26 -070088 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);
Andy Grover00e0f342009-02-24 15:30:23 +000092}
Sowmini Varadhan9c794402016-06-04 14:00:00 -070093EXPORT_SYMBOL_GPL(rds_connect_path_complete);
94
95void rds_connect_complete(struct rds_connection *conn)
96{
Sowmini Varadhan0cb43962016-06-13 09:44:26 -070097 rds_connect_path_complete(&conn->c_path[0], RDS_CONN_CONNECTING);
Sowmini Varadhan9c794402016-06-04 14:00:00 -070098}
Andy Grover616b7572009-08-21 12:28:32 +000099EXPORT_SYMBOL_GPL(rds_connect_complete);
Andy Grover00e0f342009-02-24 15:30:23 +0000100
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 */
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700119void rds_queue_reconnect(struct rds_conn_path *cp)
Andy Grover00e0f342009-02-24 15:30:23 +0000120{
121 unsigned long rand;
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700122 struct rds_connection *conn = cp->cp_conn;
Andy Grover00e0f342009-02-24 15:30:23 +0000123
124 rdsdebug("conn %p for %pI4 to %pI4 reconnect jiffies %lu\n",
125 conn, &conn->c_laddr, &conn->c_faddr,
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700126 cp->cp_reconnect_jiffies);
Andy Grover00e0f342009-02-24 15:30:23 +0000127
Sowmini Varadhan83150112016-06-30 16:11:17 -0700128 /* 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
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700133 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);
Andy Grover00e0f342009-02-24 15:30:23 +0000137 return;
138 }
139
140 get_random_bytes(&rand, sizeof(rand));
141 rdsdebug("%lu delay %lu ceil conn %p for %pI4 -> %pI4\n",
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700142 rand % cp->cp_reconnect_jiffies, cp->cp_reconnect_jiffies,
Andy Grover00e0f342009-02-24 15:30:23 +0000143 conn, &conn->c_laddr, &conn->c_faddr);
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700144 queue_delayed_work(rds_wq, &cp->cp_conn_w,
145 rand % cp->cp_reconnect_jiffies);
Andy Grover00e0f342009-02-24 15:30:23 +0000146
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700147 cp->cp_reconnect_jiffies = min(cp->cp_reconnect_jiffies * 2,
Andy Grover00e0f342009-02-24 15:30:23 +0000148 rds_sysctl_reconnect_max_jiffies);
149}
150
151void rds_connect_worker(struct work_struct *work)
152{
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700153 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;
Andy Grover00e0f342009-02-24 15:30:23 +0000157 int ret;
158
Sowmini Varadhan5916e2c2016-07-14 03:51:03 -0700159 if (cp->cp_index > 1 && cp->cp_conn->c_laddr > cp->cp_conn->c_faddr)
160 return;
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700161 clear_bit(RDS_RECONNECT_PENDING, &cp->cp_flags);
Sowmini Varadhanb04e8552016-06-30 16:11:16 -0700162 ret = rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_CONNECTING);
163 if (ret) {
164 ret = conn->c_trans->conn_path_connect(cp);
Andy Grover00e0f342009-02-24 15:30:23 +0000165 rdsdebug("conn %p for %pI4 to %pI4 dispatched, ret %d\n",
166 conn, &conn->c_laddr, &conn->c_faddr, ret);
167
168 if (ret) {
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700169 if (rds_conn_path_transition(cp,
170 RDS_CONN_CONNECTING,
171 RDS_CONN_DOWN))
172 rds_queue_reconnect(cp);
Andy Grover00e0f342009-02-24 15:30:23 +0000173 else
Sowmini Varadhanfb1b3dc2016-06-13 09:44:39 -0700174 rds_conn_path_error(cp,
175 "RDS: connect failed\n");
Andy Grover00e0f342009-02-24 15:30:23 +0000176 }
177 }
178}
179
Andy Grover00e0f342009-02-24 15:30:23 +0000180void rds_send_worker(struct work_struct *work)
181{
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700182 struct rds_conn_path *cp = container_of(work,
183 struct rds_conn_path,
184 cp_send_w.work);
Andy Grover00e0f342009-02-24 15:30:23 +0000185 int ret;
186
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700187 if (rds_conn_path_state(cp) == RDS_CONN_UP) {
188 clear_bit(RDS_LL_SEND_FULL, &cp->cp_flags);
Sowmini Varadhan1f9ecd72016-06-13 09:44:34 -0700189 ret = rds_send_xmit(cp);
Santosh Shilimkardb6526d2015-09-11 15:44:29 -0700190 cond_resched();
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700191 rdsdebug("conn %p ret %d\n", cp->cp_conn, ret);
Andy Grover00e0f342009-02-24 15:30:23 +0000192 switch (ret) {
193 case -EAGAIN:
194 rds_stats_inc(s_send_immediate_retry);
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700195 queue_delayed_work(rds_wq, &cp->cp_send_w, 0);
Andy Grover00e0f342009-02-24 15:30:23 +0000196 break;
197 case -ENOMEM:
198 rds_stats_inc(s_send_delayed_retry);
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700199 queue_delayed_work(rds_wq, &cp->cp_send_w, 2);
Andy Grover00e0f342009-02-24 15:30:23 +0000200 default:
201 break;
202 }
203 }
204}
205
206void rds_recv_worker(struct work_struct *work)
207{
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700208 struct rds_conn_path *cp = container_of(work,
209 struct rds_conn_path,
210 cp_recv_w.work);
Andy Grover00e0f342009-02-24 15:30:23 +0000211 int ret;
212
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700213 if (rds_conn_path_state(cp) == RDS_CONN_UP) {
Sowmini Varadhan2da43c42016-06-30 16:11:15 -0700214 ret = cp->cp_conn->c_trans->recv_path(cp);
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700215 rdsdebug("conn %p ret %d\n", cp->cp_conn, ret);
Andy Grover00e0f342009-02-24 15:30:23 +0000216 switch (ret) {
217 case -EAGAIN:
218 rds_stats_inc(s_recv_immediate_retry);
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700219 queue_delayed_work(rds_wq, &cp->cp_recv_w, 0);
Andy Grover00e0f342009-02-24 15:30:23 +0000220 break;
221 case -ENOMEM:
222 rds_stats_inc(s_recv_delayed_retry);
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700223 queue_delayed_work(rds_wq, &cp->cp_recv_w, 2);
Andy Grover00e0f342009-02-24 15:30:23 +0000224 default:
225 break;
226 }
227 }
228}
229
Andy Grover2dc39352010-06-11 13:49:13 -0700230void rds_shutdown_worker(struct work_struct *work)
231{
Sowmini Varadhan0cb43962016-06-13 09:44:26 -0700232 struct rds_conn_path *cp = container_of(work,
233 struct rds_conn_path,
234 cp_down_w);
Andy Grover2dc39352010-06-11 13:49:13 -0700235
Sowmini Varadhand769ef82016-06-13 09:44:41 -0700236 rds_conn_shutdown(cp);
Andy Grover2dc39352010-06-11 13:49:13 -0700237}
238
Andy Grover00e0f342009-02-24 15:30:23 +0000239void rds_threads_exit(void)
240{
241 destroy_workqueue(rds_wq);
242}
243
Zach Brownef87b7e2010-07-09 12:26:20 -0700244int rds_threads_init(void)
Andy Grover00e0f342009-02-24 15:30:23 +0000245{
Zach Brown80c51be2010-07-06 15:08:48 -0700246 rds_wq = create_singlethread_workqueue("krdsd");
Andy Grover8690bfa2010-01-12 11:56:44 -0800247 if (!rds_wq)
Andy Grover00e0f342009-02-24 15:30:23 +0000248 return -ENOMEM;
249
250 return 0;
251}