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
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 RDSV3_DPRINTF2("rdsv3_connect_complete",
93 "%s: Cannot transition to state UP, "
94 "current state is %d",
95 __func__,
96 atomic_get(&conn->c_state));
97 conn->c_state = RDSV3_CONN_ERROR;
98 rdsv3_queue_work(rdsv3_wq, &conn->c_down_w);
99 return;
100 }
101
102 RDSV3_DPRINTF2("rdsv3_connect_complete",
103 "conn %p for %u.%u.%u.%u to %u.%u.%u.%u complete",
104 conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr));
105
106 conn->c_reconnect_jiffies = 0;
107 conn->c_last_connect_jiffies = ddi_get_lbolt();
108
109 set_bit(0, &conn->c_map_queued);
110 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_send_w, 0);
111 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_recv_w, 0);
112
113 RDSV3_DPRINTF4("rdsv3_connect_complete", "Return(conn: %p)", conn);
114 }
115
116 /*
117 * This random exponential backoff is relied on to eventually resolve racing
118 * connects.
119 *
120 * If connect attempts race then both parties drop both connections and come
121 * here to wait for a random amount of time before trying again. Eventually
122 * the backoff range will be so much greater than the time it takes to
123 * establish a connection that one of the pair will establish the connection
124 * before the other's random delay fires.
125 *
126 * Connection attempts that arrive while a connection is already established
127 * are also considered to be racing connects. This lets a connection from
128 * a rebooted machine replace an existing stale connection before the transport
129 * notices that the connection has failed.
130 *
131 * We should *always* start with a random backoff; otherwise a broken connection
132 * will always take several iterations to be re-established.
133 */
134 void
135 rdsv3_queue_reconnect(struct rdsv3_connection *conn)
136 {
137 unsigned long rand;
138
139 RDSV3_DPRINTF2("rdsv3_queue_reconnect",
140 "conn %p for %u.%u.%u.%u to %u.%u.%u.%u reconnect jiffies %lu",
141 conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr),
142 conn->c_reconnect_jiffies);
143
144 set_bit(RDSV3_RECONNECT_PENDING, &conn->c_flags);
145 if (conn->c_reconnect_jiffies == 0) {
146 conn->c_reconnect_jiffies = rdsv3_sysctl_reconnect_min_jiffies;
147 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_conn_w, 0);
148 return;
149 }
150
151 (void) random_get_pseudo_bytes((uint8_t *)&rand, sizeof (rand));
152
153 RDSV3_DPRINTF5("rdsv3",
154 "%lu delay %lu ceil conn %p for %u.%u.%u.%u -> %u.%u.%u.%u",
155 rand % conn->c_reconnect_jiffies, conn->c_reconnect_jiffies,
156 conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr));
157
158 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_conn_w,
159 rand % conn->c_reconnect_jiffies);
160
161 conn->c_reconnect_jiffies = min(conn->c_reconnect_jiffies * 2,
162 rdsv3_sysctl_reconnect_max_jiffies);
163 }
164
165 void
166 rdsv3_connect_worker(struct rdsv3_work_s *work)
167 {
168 struct rdsv3_connection *conn = container_of(work,
169 struct rdsv3_connection, c_conn_w.work);
170 int ret;
171
172 RDSV3_DPRINTF2("rdsv3_connect_worker", "Enter(work: %p)", work);
173
174 clear_bit(RDSV3_RECONNECT_PENDING, &conn->c_flags);
175 if (rdsv3_conn_transition(conn, RDSV3_CONN_DOWN,
176 RDSV3_CONN_CONNECTING)) {
177 ret = conn->c_trans->conn_connect(conn);
178
179 RDSV3_DPRINTF5("rdsv3",
180 "connect conn %p for %u.%u.%u.%u -> %u.%u.%u.%u "
181 "ret %d", conn, NIPQUAD(conn->c_laddr),
182 NIPQUAD(conn->c_faddr), ret);
183
184 RDSV3_DPRINTF2("rdsv3_connect_worker",
185 "conn %p for %u.%u.%u.%u to %u.%u.%u.%u dispatched, ret %d",
186 conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr), ret);
187
188 if (ret) {
189 if (rdsv3_conn_transition(conn, RDSV3_CONN_CONNECTING,
190 RDSV3_CONN_DOWN))
191 rdsv3_queue_reconnect(conn);
192 else {
193 RDSV3_DPRINTF2("rdsv3_connect_worker",
194 "RDS: connect failed: %p", conn);
195 rdsv3_conn_drop(conn);
196 }
197 }
198 }
199
200 RDSV3_DPRINTF2("rdsv3_connect_worker", "Return(work: %p)", work);
201 }
202
203 void
204 rdsv3_send_worker(struct rdsv3_work_s *work)
205 {
206 struct rdsv3_connection *conn = container_of(work,
207 struct rdsv3_connection, c_send_w.work);
208 int ret;
209
210 RDSV3_DPRINTF4("rdsv3_send_worker", "Enter(work: %p)", work);
211
212 if (rdsv3_conn_state(conn) == RDSV3_CONN_UP) {
213 ret = rdsv3_send_xmit(conn);
214 RDSV3_DPRINTF5("rdsv3", "conn %p ret %d", conn, ret);
215 switch (ret) {
216 case -EAGAIN:
217 rdsv3_stats_inc(s_send_immediate_retry);
218 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_send_w, 0);
219 break;
220 case -ENOMEM:
221 rdsv3_stats_inc(s_send_delayed_retry);
222 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_send_w, 2);
223 default:
224 break;
225 }
226 }
227
228 RDSV3_DPRINTF4("rdsv3_send_worker", "Return(work: %p)", work);
229 }
230
231 void
232 rdsv3_recv_worker(struct rdsv3_work_s *work)
233 {
234 struct rdsv3_connection *conn = container_of(work,
235 struct rdsv3_connection, c_recv_w.work);
236 int ret;
237
238 RDSV3_DPRINTF4("rdsv3_recv_worker", "Enter(work: %p)", work);
239
240 if (rdsv3_conn_state(conn) == RDSV3_CONN_UP) {
241 ret = conn->c_trans->recv(conn);
242 RDSV3_DPRINTF5("rdsv3", "conn %p ret %d", conn, ret);
243 switch (ret) {
244 case -EAGAIN:
245 rdsv3_stats_inc(s_recv_immediate_retry);
246 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_recv_w, 0);
247 break;
248 case -ENOMEM:
249 rdsv3_stats_inc(s_recv_delayed_retry);
250 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_recv_w, 2);
251 default:
252 break;
253 }
254 }
255
256 RDSV3_DPRINTF4("rdsv3_recv_worker", "Return(work: %p)", work);
257 }
258
259 void
260 rdsv3_shutdown_worker(struct rdsv3_work_s *work)
261 {
262 struct rdsv3_connection *conn = container_of(work,
263 struct rdsv3_connection, c_down_w);
264 rdsv3_conn_shutdown(conn);
265 }
266
267 #define time_after(a, b) ((long)(b) - (long)(a) < 0)
268
269 void
270 rdsv3_reaper_worker(struct rdsv3_work_s *work)
271 {
272 struct rdsv3_connection *conn = container_of(work,
273 struct rdsv3_connection, c_reap_w.work);
274
275 if (rdsv3_conn_state(conn) != RDSV3_CONN_UP &&
276 !time_after(conn->c_last_connect_jiffies,
277 ddi_get_lbolt() - RDSV3_REAPER_WAIT_JIFFIES)) {
278 rdsv3_conn_destroy(conn);
279 } else {
280 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_reap_w,
281 RDSV3_REAPER_WAIT_JIFFIES);
282 }
283 }
284
285 void
286 rdsv3_threads_exit(void)
287 {
288 rdsv3_destroy_task_workqueue(rdsv3_wq);
289 }
290
291 int
292 rdsv3_threads_init(void)
293 {
294 rdsv3_wq = rdsv3_create_task_workqueue("krdsd");
295 if (!rdsv3_wq)
296 return (-ENOMEM);
297
298 return (0);
299 }