xref: /illumos-gate/usr/src/uts/common/io/ib/clients/rdsv3/threads.c (revision cadbfdc3bdb156e92d7a88978bc98ea87f6e037f)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 
25 /*
26  * Copyright (c) 2006 Oracle.  All rights reserved.
27  *
28  * This software is available to you under a choice of one of two
29  * licenses.  You may choose to be licensed under the terms of the GNU
30  * General Public License (GPL) Version 2, available from the file
31  * COPYING in the main directory of this source tree, or the
32  * OpenIB.org BSD license below:
33  *
34  *     Redistribution and use in source and binary forms, with or
35  *     without modification, are permitted provided that the following
36  *     conditions are met:
37  *
38  *      - Redistributions of source code must retain the above
39  *        copyright notice, this list of conditions and the following
40  *        disclaimer.
41  *
42  *      - Redistributions in binary form must reproduce the above
43  *        copyright notice, this list of conditions and the following
44  *        disclaimer in the documentation and/or other materials
45  *        provided with the distribution.
46  *
47  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
48  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
49  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
50  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
51  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
52  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
53  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
54  * SOFTWARE.
55  *
56  */
57 #include <sys/rds.h>
58 #include <sys/sunddi.h>
59 
60 #include <sys/ib/clients/rdsv3/rdsv3.h>
61 #include <sys/ib/clients/rdsv3/rdsv3_debug.h>
62 
63 /*
64  * All of connection management is simplified by serializing it through
65  * work queues that execute in a connection managing thread.
66  *
67  * TCP wants to send acks through sendpage() in response to data_ready(),
68  * but it needs a process context to do so.
69  *
70  * The receive paths need to allocate but can't drop packets (!) so we have
71  * a thread around to block allocating if the receive fast path sees an
72  * allocation failure.
73  */
74 
75 /*
76  * Grand Unified Theory of connection life cycle:
77  * At any point in time, the connection can be in one of these states:
78  * DOWN, CONNECTING, UP, DISCONNECTING, ERROR
79  *
80  * The following transitions are possible:
81  *  ANY		  -> ERROR
82  *  UP		  -> DISCONNECTING
83  *  ERROR	  -> DISCONNECTING
84  *  DISCONNECTING -> DOWN
85  *  DOWN	  -> CONNECTING
86  *  CONNECTING	  -> UP
87  *
88  * Transition to state DISCONNECTING/DOWN:
89  *  -	Inside the shutdown worker; synchronizes with xmit path
90  *	through c_send_lock, and with connection management callbacks
91  *	via c_cm_lock.
92  *
93  *	For receive callbacks, we rely on the underlying transport
94  *	(TCP, IB/RDMA) to provide the necessary synchronisation.
95  */
96 struct rdsv3_workqueue_struct_s *rdsv3_wq;
97 
98 void
99 rdsv3_connect_complete(struct rdsv3_connection *conn)
100 {
101 	RDSV3_DPRINTF4("rdsv3_connect_complete", "Enter(conn: %p)", conn);
102 
103 	if (!rdsv3_conn_transition(conn, RDSV3_CONN_CONNECTING,
104 	    RDSV3_CONN_UP)) {
105 #ifndef __lock_lint
106 		RDSV3_DPRINTF2("rdsv3_connect_complete",
107 		    "%s: Cannot transition to state UP, "
108 		    "current state is %d",
109 		    __func__,
110 		    atomic_get(&conn->c_state));
111 #endif
112 		conn->c_state = RDSV3_CONN_ERROR;
113 		rdsv3_queue_work(rdsv3_wq, &conn->c_down_w);
114 		return;
115 	}
116 
117 	RDSV3_DPRINTF2("rdsv3_connect_complete",
118 	    "conn %p for %u.%u.%u.%u to %u.%u.%u.%u complete",
119 	    conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr));
120 
121 	conn->c_reconnect_jiffies = 0;
122 	set_bit(0, &conn->c_map_queued);
123 	rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_send_w, 0);
124 	rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_recv_w, 0);
125 
126 	RDSV3_DPRINTF4("rdsv3_connect_complete", "Return(conn: %p)", conn);
127 }
128 
129 /*
130  * This random exponential backoff is relied on to eventually resolve racing
131  * connects.
132  *
133  * If connect attempts race then both parties drop both connections and come
134  * here to wait for a random amount of time before trying again.  Eventually
135  * the backoff range will be so much greater than the time it takes to
136  * establish a connection that one of the pair will establish the connection
137  * before the other's random delay fires.
138  *
139  * Connection attempts that arrive while a connection is already established
140  * are also considered to be racing connects.  This lets a connection from
141  * a rebooted machine replace an existing stale connection before the transport
142  * notices that the connection has failed.
143  *
144  * We should *always* start with a random backoff; otherwise a broken connection
145  * will always take several iterations to be re-established.
146  */
147 static void
148 rdsv3_queue_reconnect(struct rdsv3_connection *conn)
149 {
150 	unsigned long rand;
151 
152 	RDSV3_DPRINTF2("rdsv3_queue_reconnect",
153 	    "conn %p for %u.%u.%u.%u to %u.%u.%u.%u reconnect jiffies %lu",
154 	    conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr),
155 	    conn->c_reconnect_jiffies);
156 
157 	set_bit(RDSV3_RECONNECT_PENDING, &conn->c_flags);
158 	if (conn->c_reconnect_jiffies == 0) {
159 		conn->c_reconnect_jiffies = rdsv3_sysctl_reconnect_min_jiffies;
160 		rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_conn_w, 0);
161 		return;
162 	}
163 
164 	(void) random_get_pseudo_bytes((uint8_t *)&rand, sizeof (rand));
165 
166 	RDSV3_DPRINTF5("rdsv3",
167 	    "%lu delay %lu ceil conn %p for %u.%u.%u.%u -> %u.%u.%u.%u",
168 	    rand % conn->c_reconnect_jiffies, conn->c_reconnect_jiffies,
169 	    conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr));
170 
171 	rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_conn_w,
172 	    rand % conn->c_reconnect_jiffies);
173 
174 	conn->c_reconnect_jiffies = min(conn->c_reconnect_jiffies * 2,
175 	    rdsv3_sysctl_reconnect_max_jiffies);
176 }
177 
178 void
179 rdsv3_connect_worker(struct rdsv3_work_s *work)
180 {
181 	struct rdsv3_connection *conn = container_of(work,
182 	    struct rdsv3_connection, c_conn_w.work);
183 	int ret;
184 
185 	RDSV3_DPRINTF2("rdsv3_connect_worker", "Enter(work: %p)", work);
186 
187 	clear_bit(RDSV3_RECONNECT_PENDING, &conn->c_flags);
188 	if (rdsv3_conn_transition(conn, RDSV3_CONN_DOWN,
189 	    RDSV3_CONN_CONNECTING)) {
190 		ret = conn->c_trans->conn_connect(conn);
191 
192 		RDSV3_DPRINTF5("rdsv3",
193 		    "connect conn %p for %u.%u.%u.%u -> %u.%u.%u.%u "
194 		    "ret %d", conn, NIPQUAD(conn->c_laddr),
195 		    NIPQUAD(conn->c_faddr), ret);
196 
197 		RDSV3_DPRINTF2("rdsv3_connect_worker",
198 		    "conn %p for %u.%u.%u.%u to %u.%u.%u.%u dispatched, ret %d",
199 		    conn, NIPQUAD(conn->c_laddr), NIPQUAD(conn->c_faddr), ret);
200 
201 		if (ret) {
202 			if (rdsv3_conn_transition(conn, RDSV3_CONN_CONNECTING,
203 			    RDSV3_CONN_DOWN))
204 				rdsv3_queue_reconnect(conn);
205 			else {
206 				RDSV3_DPRINTF2("rdsv3_connect_worker",
207 				    "RDS: connect failed: %p", conn);
208 				rdsv3_conn_drop(conn);
209 			}
210 		}
211 	}
212 
213 	RDSV3_DPRINTF2("rdsv3_connect_worker", "Return(work: %p)", work);
214 }
215 
216 extern struct avl_tree	rdsv3_conn_hash;
217 
218 void
219 rdsv3_shutdown_worker(struct rdsv3_work_s *work)
220 {
221 	struct rdsv3_connection *conn = container_of(work,
222 	    struct rdsv3_connection, c_down_w);
223 	struct rdsv3_conn_info_s conn_info;
224 
225 	RDSV3_DPRINTF2("rdsv3_shutdown_worker", "Enter(work: %p)", work);
226 
227 	/* shut it down unless it's down already */
228 	if (!rdsv3_conn_transition(conn, RDSV3_CONN_DOWN, RDSV3_CONN_DOWN)) {
229 		/*
230 		 * Quiesce the connection mgmt handlers before we start tearing
231 		 * things down. We don't hold the mutex for the entire
232 		 * duration of the shutdown operation, else we may be
233 		 * deadlocking with the CM handler. Instead, the CM event
234 		 * handler is supposed to check for state DISCONNECTING
235 		 */
236 		mutex_enter(&conn->c_cm_lock);
237 		if (!rdsv3_conn_transition(conn, RDSV3_CONN_UP,
238 		    RDSV3_CONN_DISCONNECTING) &&
239 		    !rdsv3_conn_transition(conn, RDSV3_CONN_ERROR,
240 		    RDSV3_CONN_DISCONNECTING)) {
241 			RDSV3_DPRINTF2("rdsv3_shutdown_worker",
242 			    "RDS: connect failed: conn: %p, state: %d",
243 			    conn, atomic_get(&conn->c_state));
244 			rdsv3_conn_drop(conn);
245 			mutex_exit(&conn->c_cm_lock);
246 			return;
247 		}
248 		mutex_exit(&conn->c_cm_lock);
249 
250 		mutex_enter(&conn->c_send_lock);
251 		conn->c_trans->conn_shutdown(conn);
252 		rdsv3_conn_reset(conn);
253 		mutex_exit(&conn->c_send_lock);
254 
255 		if (!rdsv3_conn_transition(conn, RDSV3_CONN_DISCONNECTING,
256 		    RDSV3_CONN_DOWN)) {
257 			/*
258 			 * This can happen - eg when we're in the middle of
259 			 * tearing down the connection, and someone unloads
260 			 * the rds module. Quite reproduceable with loopback
261 			 * connections. Mostly harmless.
262 			 */
263 #ifndef __lock_lint
264 			RDSV3_DPRINTF2("rdsv3_shutdown_worker",
265 			    "failed to transition to state DOWN, "
266 			    "current statis is: %d conn: %p",
267 			    atomic_get(&conn->c_state), conn);
268 			rdsv3_conn_drop(conn);
269 #endif
270 			return;
271 		}
272 	}
273 
274 	/*
275 	 * Then reconnect if it's still live.
276 	 * The passive side of an IB loopback connection is never added
277 	 * to the conn hash, so we never trigger a reconnect on this
278 	 * conn - the reconnect is always triggered by the active peer.
279 	 */
280 	rdsv3_cancel_delayed_work(&conn->c_conn_w);
281 
282 	conn_info.c_laddr = conn->c_laddr;
283 	conn_info.c_faddr = conn->c_faddr;
284 	if (avl_find(&rdsv3_conn_hash, &conn_info, NULL) == conn)
285 		rdsv3_queue_reconnect(conn);
286 
287 	RDSV3_DPRINTF2("rdsv3_shutdown_worker", "Return(work: %p)", work);
288 }
289 
290 void
291 rdsv3_send_worker(struct rdsv3_work_s *work)
292 {
293 	struct rdsv3_connection *conn = container_of(work,
294 	    struct rdsv3_connection, c_send_w.work);
295 	int ret;
296 
297 	RDSV3_DPRINTF4("rdsv3_send_worker", "Enter(work: %p)", work);
298 
299 	if (rdsv3_conn_state(conn) == RDSV3_CONN_UP) {
300 		ret = rdsv3_send_xmit(conn);
301 		RDSV3_DPRINTF5("rdsv3", "conn %p ret %d", conn, ret);
302 		switch (ret) {
303 		case -EAGAIN:
304 			rdsv3_stats_inc(s_send_immediate_retry);
305 			rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_send_w, 0);
306 			break;
307 		case -ENOMEM:
308 			rdsv3_stats_inc(s_send_delayed_retry);
309 			rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_send_w, 2);
310 		default:
311 			break;
312 		}
313 	}
314 
315 	RDSV3_DPRINTF4("rdsv3_send_worker", "Return(work: %p)", work);
316 }
317 
318 void
319 rdsv3_recv_worker(struct rdsv3_work_s *work)
320 {
321 	struct rdsv3_connection *conn = container_of(work,
322 	    struct rdsv3_connection, c_recv_w.work);
323 	int ret;
324 
325 	RDSV3_DPRINTF4("rdsv3_recv_worker", "Enter(work: %p)", work);
326 
327 	if (rdsv3_conn_state(conn) == RDSV3_CONN_UP) {
328 		ret = conn->c_trans->recv(conn);
329 		RDSV3_DPRINTF5("rdsv3", "conn %p ret %d", conn, ret);
330 		switch (ret) {
331 		case -EAGAIN:
332 			rdsv3_stats_inc(s_recv_immediate_retry);
333 			rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_recv_w, 0);
334 			break;
335 		case -ENOMEM:
336 			rdsv3_stats_inc(s_recv_delayed_retry);
337 			rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_recv_w, 2);
338 		default:
339 			break;
340 		}
341 	}
342 
343 	RDSV3_DPRINTF4("rdsv3_recv_worker", "Return(work: %p)", work);
344 }
345 
346 void
347 rdsv3_threads_exit(void)
348 {
349 	rdsv3_destroy_task_workqueue(rdsv3_wq);
350 }
351 
352 int
353 rdsv3_threads_init(void)
354 {
355 	rdsv3_wq = rdsv3_create_task_workqueue("krdsd");
356 	if (rdsv3_wq == NULL)
357 		return (-ENOMEM);
358 
359 	return (0);
360 }
361