1 /*
2 * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
3 */
4
5 /*
6 * This file contains code imported from the OFED rds source file threads.c
7 * Oracle elects to have and use the contents of threads.c under and governed
8 * by the OpenIB.org BSD license (see below for full license text). However,
9 * the following notice accompanied the original version of this file:
10 */
11
12 /*
13 * Copyright (c) 2006 Oracle. All rights reserved.
14 *
15 * This software is available to you under a choice of one of two
16 * licenses. You may choose to be licensed under the terms of the GNU
17 * General Public License (GPL) Version 2, available from the file
18 * COPYING in the main directory of this source tree, or the
19 * OpenIB.org BSD license below:
20 *
21 * Redistribution and use in source and binary forms, with or
22 * without modification, are permitted provided that the following
23 * conditions are met:
24 *
25 * - Redistributions of source code must retain the above
26 * copyright notice, this list of conditions and the following
27 * disclaimer.
28 *
29 * - Redistributions in binary form must reproduce the above
30 * copyright notice, this list of conditions and the following
31 * disclaimer in the documentation and/or other materials
32 * provided with the distribution.
33 *
34 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
35 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
36 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
37 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
38 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
39 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
40 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
41 * SOFTWARE.
42 *
43 */
44 #include <sys/rds.h>
45 #include <sys/sunddi.h>
46
47 #include <sys/ib/clients/rdsv3/rdsv3.h>
48 #include <sys/ib/clients/rdsv3/rdsv3_debug.h>
49
50 /*
51 * All of connection management is simplified by serializing it through
52 * work queues that execute in a connection managing thread.
53 *
54 * TCP wants to send acks through sendpage() in response to data_ready(),
55 * but it needs a process context to do so.
56 *
57 * The receive paths need to allocate but can't drop packets (!) so we have
58 * a thread around to block allocating if the receive fast path sees an
59 * allocation failure.
60 */
61
62 /*
63 * Grand Unified Theory of connection life cycle:
64 * At any point in time, the connection can be in one of these states:
65 * DOWN, CONNECTING, UP, DISCONNECTING, ERROR
66 *
67 * The following transitions are possible:
68 * ANY -> ERROR
69 * UP -> DISCONNECTING
70 * ERROR -> DISCONNECTING
71 * DISCONNECTING -> DOWN
72 * DOWN -> CONNECTING
73 * CONNECTING -> UP
74 *
75 * Transition to state DISCONNECTING/DOWN:
76 * - Inside the shutdown worker; synchronizes with xmit path
77 * through c_send_lock, and with connection management callbacks
78 * via c_cm_lock.
79 *
80 * For receive callbacks, we rely on the underlying transport
81 * (TCP, IB/RDMA) to provide the necessary synchronisation.
82 */
83 struct rdsv3_workqueue_struct_s *rdsv3_wq;
84
85 void
rdsv3_connect_complete(struct rdsv3_connection * conn)86 rdsv3_connect_complete(struct rdsv3_connection *conn)
87 {
88 RDSV3_DPRINTF4("rdsv3_connect_complete", "Enter(conn: %p)", conn);
89
90 if (!rdsv3_conn_transition(conn, RDSV3_CONN_CONNECTING,
91 RDSV3_CONN_UP)) {
92 #ifndef __lock_lint
93 RDSV3_DPRINTF2("rdsv3_connect_complete",
94 "%s: Cannot transition to state UP, "
95 "current state is %d",
96 __func__,
97 atomic_get(&conn->c_state));
98 #endif
99 conn->c_state = RDSV3_CONN_ERROR;
100 rdsv3_queue_work(rdsv3_wq, &conn->c_down_w);
101 return;
102 }
103
104 RDSV3_DPRINTF2("rdsv3_connect_complete",
105 "conn %p for %u.%u.%u.%u to %u.%u.%u.%u complete",
106 conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr));
107
108 conn->c_reconnect_jiffies = 0;
109 conn->c_last_connect_jiffies = ddi_get_lbolt();
110
111 set_bit(0, &conn->c_map_queued);
112 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_send_w, 0);
113 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_recv_w, 0);
114
115 RDSV3_DPRINTF4("rdsv3_connect_complete", "Return(conn: %p)", conn);
116 }
117
118 /*
119 * This random exponential backoff is relied on to eventually resolve racing
120 * connects.
121 *
122 * If connect attempts race then both parties drop both connections and come
123 * here to wait for a random amount of time before trying again. Eventually
124 * the backoff range will be so much greater than the time it takes to
125 * establish a connection that one of the pair will establish the connection
126 * before the other's random delay fires.
127 *
128 * Connection attempts that arrive while a connection is already established
129 * are also considered to be racing connects. This lets a connection from
130 * a rebooted machine replace an existing stale connection before the transport
131 * notices that the connection has failed.
132 *
133 * We should *always* start with a random backoff; otherwise a broken connection
134 * will always take several iterations to be re-established.
135 */
136 void
rdsv3_queue_reconnect(struct rdsv3_connection * conn)137 rdsv3_queue_reconnect(struct rdsv3_connection *conn)
138 {
139 unsigned long rand;
140
141 RDSV3_DPRINTF2("rdsv3_queue_reconnect",
142 "conn %p for %u.%u.%u.%u to %u.%u.%u.%u reconnect jiffies %lu",
143 conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr),
144 conn->c_reconnect_jiffies);
145
146 set_bit(RDSV3_RECONNECT_PENDING, &conn->c_flags);
147 if (conn->c_reconnect_jiffies == 0) {
148 conn->c_reconnect_jiffies = rdsv3_sysctl_reconnect_min_jiffies;
149 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_conn_w, 0);
150 return;
151 }
152
153 (void) random_get_pseudo_bytes((uint8_t *)&rand, sizeof (rand));
154
155 RDSV3_DPRINTF5("rdsv3",
156 "%lu delay %lu ceil conn %p for %u.%u.%u.%u -> %u.%u.%u.%u",
157 rand % conn->c_reconnect_jiffies, conn->c_reconnect_jiffies,
158 conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr));
159
160 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_conn_w,
161 rand % conn->c_reconnect_jiffies);
162
163 conn->c_reconnect_jiffies = min(conn->c_reconnect_jiffies * 2,
164 rdsv3_sysctl_reconnect_max_jiffies);
165 }
166
167 void
rdsv3_connect_worker(struct rdsv3_work_s * work)168 rdsv3_connect_worker(struct rdsv3_work_s *work)
169 {
170 struct rdsv3_connection *conn = container_of(work,
171 struct rdsv3_connection, c_conn_w.work);
172 int ret;
173
174 RDSV3_DPRINTF2("rdsv3_connect_worker", "Enter(work: %p)", work);
175
176 clear_bit(RDSV3_RECONNECT_PENDING, &conn->c_flags);
177 if (rdsv3_conn_transition(conn, RDSV3_CONN_DOWN,
178 RDSV3_CONN_CONNECTING)) {
179 ret = conn->c_trans->conn_connect(conn);
180
181 RDSV3_DPRINTF5("rdsv3",
182 "connect conn %p for %u.%u.%u.%u -> %u.%u.%u.%u "
183 "ret %d", conn, NIPQUAD(conn->c_laddr),
184 NIPQUAD(conn->c_faddr), ret);
185
186 RDSV3_DPRINTF2("rdsv3_connect_worker",
187 "conn %p for %u.%u.%u.%u to %u.%u.%u.%u dispatched, ret %d",
188 conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr), ret);
189
190 if (ret) {
191 if (rdsv3_conn_transition(conn, RDSV3_CONN_CONNECTING,
192 RDSV3_CONN_DOWN))
193 rdsv3_queue_reconnect(conn);
194 else {
195 RDSV3_DPRINTF2("rdsv3_connect_worker",
196 "RDS: connect failed: %p", conn);
197 rdsv3_conn_drop(conn);
198 }
199 }
200 }
201
202 RDSV3_DPRINTF2("rdsv3_connect_worker", "Return(work: %p)", work);
203 }
204
205 void
rdsv3_send_worker(struct rdsv3_work_s * work)206 rdsv3_send_worker(struct rdsv3_work_s *work)
207 {
208 struct rdsv3_connection *conn = container_of(work,
209 struct rdsv3_connection, c_send_w.work);
210 int ret;
211
212 RDSV3_DPRINTF4("rdsv3_send_worker", "Enter(work: %p)", work);
213
214 if (rdsv3_conn_state(conn) == RDSV3_CONN_UP) {
215 ret = rdsv3_send_xmit(conn);
216 RDSV3_DPRINTF5("rdsv3", "conn %p ret %d", conn, ret);
217 switch (ret) {
218 case -EAGAIN:
219 rdsv3_stats_inc(s_send_immediate_retry);
220 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_send_w, 0);
221 break;
222 case -ENOMEM:
223 rdsv3_stats_inc(s_send_delayed_retry);
224 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_send_w, 2);
225 default:
226 break;
227 }
228 }
229
230 RDSV3_DPRINTF4("rdsv3_send_worker", "Return(work: %p)", work);
231 }
232
233 void
rdsv3_recv_worker(struct rdsv3_work_s * work)234 rdsv3_recv_worker(struct rdsv3_work_s *work)
235 {
236 struct rdsv3_connection *conn = container_of(work,
237 struct rdsv3_connection, c_recv_w.work);
238 int ret;
239
240 RDSV3_DPRINTF4("rdsv3_recv_worker", "Enter(work: %p)", work);
241
242 if (rdsv3_conn_state(conn) == RDSV3_CONN_UP) {
243 ret = conn->c_trans->recv(conn);
244 RDSV3_DPRINTF5("rdsv3", "conn %p ret %d", conn, ret);
245 switch (ret) {
246 case -EAGAIN:
247 rdsv3_stats_inc(s_recv_immediate_retry);
248 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_recv_w, 0);
249 break;
250 case -ENOMEM:
251 rdsv3_stats_inc(s_recv_delayed_retry);
252 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_recv_w, 2);
253 default:
254 break;
255 }
256 }
257
258 RDSV3_DPRINTF4("rdsv3_recv_worker", "Return(work: %p)", work);
259 }
260
261 void
rdsv3_shutdown_worker(struct rdsv3_work_s * work)262 rdsv3_shutdown_worker(struct rdsv3_work_s *work)
263 {
264 struct rdsv3_connection *conn = container_of(work,
265 struct rdsv3_connection, c_down_w);
266 rdsv3_conn_shutdown(conn);
267 }
268
269 #define time_after(a, b) ((long)(b) - (long)(a) < 0)
270
271 void
rdsv3_reaper_worker(struct rdsv3_work_s * work)272 rdsv3_reaper_worker(struct rdsv3_work_s *work)
273 {
274 struct rdsv3_connection *conn = container_of(work,
275 struct rdsv3_connection, c_reap_w.work);
276
277 if (rdsv3_conn_state(conn) != RDSV3_CONN_UP &&
278 !time_after(conn->c_last_connect_jiffies,
279 ddi_get_lbolt() - RDSV3_REAPER_WAIT_JIFFIES)) {
280 rdsv3_conn_destroy(conn);
281 } else {
282 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_reap_w,
283 RDSV3_REAPER_WAIT_JIFFIES);
284 }
285 }
286
287 void
rdsv3_threads_exit(void)288 rdsv3_threads_exit(void)
289 {
290 rdsv3_destroy_task_workqueue(rdsv3_wq);
291 }
292
293 int
rdsv3_threads_init(void)294 rdsv3_threads_init(void)
295 {
296 rdsv3_wq = rdsv3_create_task_workqueue("krdsd");
297 if (!rdsv3_wq)
298 return (-ENOMEM);
299
300 return (0);
301 }
302