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 #include <sys/containerof.h>
47
48 #include <sys/ib/clients/rdsv3/rdsv3.h>
49 #include <sys/ib/clients/rdsv3/rdsv3_debug.h>
50
51 /*
52 * All of connection management is simplified by serializing it through
53 * work queues that execute in a connection managing thread.
54 *
55 * TCP wants to send acks through sendpage() in response to data_ready(),
56 * but it needs a process context to do so.
57 *
58 * The receive paths need to allocate but can't drop packets (!) so we have
59 * a thread around to block allocating if the receive fast path sees an
60 * allocation failure.
61 */
62
63 /*
64 * Grand Unified Theory of connection life cycle:
65 * At any point in time, the connection can be in one of these states:
66 * DOWN, CONNECTING, UP, DISCONNECTING, ERROR
67 *
68 * The following transitions are possible:
69 * ANY -> ERROR
70 * UP -> DISCONNECTING
71 * ERROR -> DISCONNECTING
72 * DISCONNECTING -> DOWN
73 * DOWN -> CONNECTING
74 * CONNECTING -> UP
75 *
76 * Transition to state DISCONNECTING/DOWN:
77 * - Inside the shutdown worker; synchronizes with xmit path
78 * through c_send_lock, and with connection management callbacks
79 * via c_cm_lock.
80 *
81 * For receive callbacks, we rely on the underlying transport
82 * (TCP, IB/RDMA) to provide the necessary synchronisation.
83 */
84 struct rdsv3_workqueue_struct_s *rdsv3_wq;
85
86 void
rdsv3_connect_complete(struct rdsv3_connection * conn)87 rdsv3_connect_complete(struct rdsv3_connection *conn)
88 {
89 RDSV3_DPRINTF4("rdsv3_connect_complete", "Enter(conn: %p)", conn);
90
91 if (!rdsv3_conn_transition(conn, RDSV3_CONN_CONNECTING,
92 RDSV3_CONN_UP)) {
93 #ifndef __lock_lint
94 RDSV3_DPRINTF2("rdsv3_connect_complete",
95 "%s: Cannot transition to state UP, "
96 "current state is %d",
97 __func__,
98 atomic_get(&conn->c_state));
99 #endif
100 conn->c_state = RDSV3_CONN_ERROR;
101 rdsv3_queue_work(rdsv3_wq, &conn->c_down_w);
102 return;
103 }
104
105 RDSV3_DPRINTF2("rdsv3_connect_complete",
106 "conn %p for %u.%u.%u.%u to %u.%u.%u.%u complete",
107 conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr));
108
109 conn->c_reconnect_jiffies = 0;
110 conn->c_last_connect_jiffies = ddi_get_lbolt();
111
112 set_bit(0, &conn->c_map_queued);
113 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_send_w, 0);
114 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_recv_w, 0);
115
116 RDSV3_DPRINTF4("rdsv3_connect_complete", "Return(conn: %p)", conn);
117 }
118
119 /*
120 * This random exponential backoff is relied on to eventually resolve racing
121 * connects.
122 *
123 * If connect attempts race then both parties drop both connections and come
124 * here to wait for a random amount of time before trying again. Eventually
125 * the backoff range will be so much greater than the time it takes to
126 * establish a connection that one of the pair will establish the connection
127 * before the other's random delay fires.
128 *
129 * Connection attempts that arrive while a connection is already established
130 * are also considered to be racing connects. This lets a connection from
131 * a rebooted machine replace an existing stale connection before the transport
132 * notices that the connection has failed.
133 *
134 * We should *always* start with a random backoff; otherwise a broken connection
135 * will always take several iterations to be re-established.
136 */
137 void
rdsv3_queue_reconnect(struct rdsv3_connection * conn)138 rdsv3_queue_reconnect(struct rdsv3_connection *conn)
139 {
140 unsigned long rand;
141
142 RDSV3_DPRINTF2("rdsv3_queue_reconnect",
143 "conn %p for %u.%u.%u.%u to %u.%u.%u.%u reconnect jiffies %lu",
144 conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr),
145 conn->c_reconnect_jiffies);
146
147 set_bit(RDSV3_RECONNECT_PENDING, &conn->c_flags);
148 if (conn->c_reconnect_jiffies == 0) {
149 conn->c_reconnect_jiffies = rdsv3_sysctl_reconnect_min_jiffies;
150 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_conn_w, 0);
151 return;
152 }
153
154 (void) random_get_pseudo_bytes((uint8_t *)&rand, sizeof (rand));
155
156 RDSV3_DPRINTF5("rdsv3",
157 "%lu delay %lu ceil conn %p for %u.%u.%u.%u -> %u.%u.%u.%u",
158 rand % conn->c_reconnect_jiffies, conn->c_reconnect_jiffies,
159 conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr));
160
161 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_conn_w,
162 rand % conn->c_reconnect_jiffies);
163
164 conn->c_reconnect_jiffies = min(conn->c_reconnect_jiffies * 2,
165 rdsv3_sysctl_reconnect_max_jiffies);
166 }
167
168 void
rdsv3_connect_worker(struct rdsv3_work_s * work)169 rdsv3_connect_worker(struct rdsv3_work_s *work)
170 {
171 struct rdsv3_connection *conn = __containerof(work,
172 struct rdsv3_connection, c_conn_w.work);
173 int ret;
174
175 RDSV3_DPRINTF2("rdsv3_connect_worker", "Enter(work: %p)", work);
176
177 clear_bit(RDSV3_RECONNECT_PENDING, &conn->c_flags);
178 if (rdsv3_conn_transition(conn, RDSV3_CONN_DOWN,
179 RDSV3_CONN_CONNECTING)) {
180 ret = conn->c_trans->conn_connect(conn);
181
182 RDSV3_DPRINTF5("rdsv3",
183 "connect conn %p for %u.%u.%u.%u -> %u.%u.%u.%u "
184 "ret %d", conn, NIPQUAD(conn->c_laddr),
185 NIPQUAD(conn->c_faddr), ret);
186
187 RDSV3_DPRINTF2("rdsv3_connect_worker",
188 "conn %p for %u.%u.%u.%u to %u.%u.%u.%u dispatched, ret %d",
189 conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr), ret);
190
191 if (ret) {
192 if (rdsv3_conn_transition(conn, RDSV3_CONN_CONNECTING,
193 RDSV3_CONN_DOWN))
194 rdsv3_queue_reconnect(conn);
195 else {
196 RDSV3_DPRINTF2("rdsv3_connect_worker",
197 "RDS: connect failed: %p", conn);
198 rdsv3_conn_drop(conn);
199 }
200 }
201 }
202
203 RDSV3_DPRINTF2("rdsv3_connect_worker", "Return(work: %p)", work);
204 }
205
206 void
rdsv3_send_worker(struct rdsv3_work_s * work)207 rdsv3_send_worker(struct rdsv3_work_s *work)
208 {
209 struct rdsv3_connection *conn = __containerof(work,
210 struct rdsv3_connection, c_send_w.work);
211 int ret;
212
213 RDSV3_DPRINTF4("rdsv3_send_worker", "Enter(work: %p)", work);
214
215 if (rdsv3_conn_state(conn) == RDSV3_CONN_UP) {
216 ret = rdsv3_send_xmit(conn);
217 RDSV3_DPRINTF5("rdsv3", "conn %p ret %d", conn, ret);
218 switch (ret) {
219 case -EAGAIN:
220 rdsv3_stats_inc(s_send_immediate_retry);
221 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_send_w, 0);
222 break;
223 case -ENOMEM:
224 rdsv3_stats_inc(s_send_delayed_retry);
225 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_send_w, 2);
226 default:
227 break;
228 }
229 }
230
231 RDSV3_DPRINTF4("rdsv3_send_worker", "Return(work: %p)", work);
232 }
233
234 void
rdsv3_recv_worker(struct rdsv3_work_s * work)235 rdsv3_recv_worker(struct rdsv3_work_s *work)
236 {
237 struct rdsv3_connection *conn = __containerof(work,
238 struct rdsv3_connection, c_recv_w.work);
239 int ret;
240
241 RDSV3_DPRINTF4("rdsv3_recv_worker", "Enter(work: %p)", work);
242
243 if (rdsv3_conn_state(conn) == RDSV3_CONN_UP) {
244 ret = conn->c_trans->recv(conn);
245 RDSV3_DPRINTF5("rdsv3", "conn %p ret %d", conn, ret);
246 switch (ret) {
247 case -EAGAIN:
248 rdsv3_stats_inc(s_recv_immediate_retry);
249 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_recv_w, 0);
250 break;
251 case -ENOMEM:
252 rdsv3_stats_inc(s_recv_delayed_retry);
253 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_recv_w, 2);
254 default:
255 break;
256 }
257 }
258
259 RDSV3_DPRINTF4("rdsv3_recv_worker", "Return(work: %p)", work);
260 }
261
262 void
rdsv3_shutdown_worker(struct rdsv3_work_s * work)263 rdsv3_shutdown_worker(struct rdsv3_work_s *work)
264 {
265 struct rdsv3_connection *conn = __containerof(work,
266 struct rdsv3_connection, c_down_w);
267 rdsv3_conn_shutdown(conn);
268 }
269
270 #define time_after(a, b) ((long)(b) - (long)(a) < 0)
271
272 void
rdsv3_reaper_worker(struct rdsv3_work_s * work)273 rdsv3_reaper_worker(struct rdsv3_work_s *work)
274 {
275 struct rdsv3_connection *conn = __containerof(work,
276 struct rdsv3_connection, c_reap_w.work);
277
278 if (rdsv3_conn_state(conn) != RDSV3_CONN_UP &&
279 !time_after(conn->c_last_connect_jiffies,
280 ddi_get_lbolt() - RDSV3_REAPER_WAIT_JIFFIES)) {
281 rdsv3_conn_destroy(conn);
282 } else {
283 rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_reap_w,
284 RDSV3_REAPER_WAIT_JIFFIES);
285 }
286 }
287
288 void
rdsv3_threads_exit(void)289 rdsv3_threads_exit(void)
290 {
291 rdsv3_destroy_task_workqueue(rdsv3_wq);
292 }
293
294 int
rdsv3_threads_init(void)295 rdsv3_threads_init(void)
296 {
297 rdsv3_wq = rdsv3_create_task_workqueue("krdsd");
298 if (!rdsv3_wq)
299 return (-ENOMEM);
300
301 return (0);
302 }
303