1/*
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>
35
36#include "rds.h"
37
38/*
39 * All of connection management is simplified by serializing it through
40 * work queues that execute in a connection managing thread.
41 *
42 * TCP wants to send acks through sendpage() in response to data_ready(),
43 * but it needs a process context to do so.
44 *
45 * The receive paths need to allocate but can't drop packets (!) so we have
46 * a thread around to block allocating if the receive fast path sees an
47 * allocation failure.
48 */
49
50/* Grand Unified Theory of connection life cycle:
51 * At any point in time, the connection can be in one of these states:
52 * DOWN, CONNECTING, UP, DISCONNECTING, ERROR
53 *
54 * The following transitions are possible:
55 *  ANY		  -> ERROR
56 *  UP		  -> DISCONNECTING
57 *  ERROR	  -> DISCONNECTING
58 *  DISCONNECTING -> DOWN
59 *  DOWN	  -> CONNECTING
60 *  CONNECTING	  -> UP
61 *
62 * Transition to state DISCONNECTING/DOWN:
63 *  -	Inside the shutdown worker; synchronizes with xmit path
64 *	through c_send_lock, and with connection management callbacks
65 *	via c_cm_lock.
66 *
67 *	For receive callbacks, we rely on the underlying transport
68 *	(TCP, IB/RDMA) to provide the necessary synchronisation.
69 */
70struct workqueue_struct *rds_wq;
71EXPORT_SYMBOL_GPL(rds_wq);
72
73void rds_connect_complete(struct rds_connection *conn)
74{
75	if (!rds_conn_transition(conn, RDS_CONN_CONNECTING, RDS_CONN_UP)) {
76		printk(KERN_WARNING "%s: Cannot transition to state UP, "
77				"current state is %d\n",
78				__func__,
79				atomic_read(&conn->c_state));
80		atomic_set(&conn->c_state, RDS_CONN_ERROR);
81		queue_work(rds_wq, &conn->c_down_w);
82		return;
83	}
84
85	rdsdebug("conn %p for %pI4 to %pI4 complete\n",
86	  conn, &conn->c_laddr, &conn->c_faddr);
87
88	conn->c_reconnect_jiffies = 0;
89	set_bit(0, &conn->c_map_queued);
90	queue_delayed_work(rds_wq, &conn->c_send_w, 0);
91	queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
92}
93EXPORT_SYMBOL_GPL(rds_connect_complete);
94
95/*
96 * This random exponential backoff is relied on to eventually resolve racing
97 * connects.
98 *
99 * If connect attempts race then both parties drop both connections and come
100 * here to wait for a random amount of time before trying again.  Eventually
101 * the backoff range will be so much greater than the time it takes to
102 * establish a connection that one of the pair will establish the connection
103 * before the other's random delay fires.
104 *
105 * Connection attempts that arrive while a connection is already established
106 * are also considered to be racing connects.  This lets a connection from
107 * a rebooted machine replace an existing stale connection before the transport
108 * notices that the connection has failed.
109 *
110 * We should *always* start with a random backoff; otherwise a broken connection
111 * will always take several iterations to be re-established.
112 */
113static void rds_queue_reconnect(struct rds_connection *conn)
114{
115	unsigned long rand;
116
117	rdsdebug("conn %p for %pI4 to %pI4 reconnect jiffies %lu\n",
118	  conn, &conn->c_laddr, &conn->c_faddr,
119	  conn->c_reconnect_jiffies);
120
121	set_bit(RDS_RECONNECT_PENDING, &conn->c_flags);
122	if (conn->c_reconnect_jiffies == 0) {
123		conn->c_reconnect_jiffies = rds_sysctl_reconnect_min_jiffies;
124		queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
125		return;
126	}
127
128	get_random_bytes(&rand, sizeof(rand));
129	rdsdebug("%lu delay %lu ceil conn %p for %pI4 -> %pI4\n",
130		 rand % conn->c_reconnect_jiffies, conn->c_reconnect_jiffies,
131		 conn, &conn->c_laddr, &conn->c_faddr);
132	queue_delayed_work(rds_wq, &conn->c_conn_w,
133			   rand % conn->c_reconnect_jiffies);
134
135	conn->c_reconnect_jiffies = min(conn->c_reconnect_jiffies * 2,
136					rds_sysctl_reconnect_max_jiffies);
137}
138
139void rds_connect_worker(struct work_struct *work)
140{
141	struct rds_connection *conn = container_of(work, struct rds_connection, c_conn_w.work);
142	int ret;
143
144	clear_bit(RDS_RECONNECT_PENDING, &conn->c_flags);
145	if (rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
146		ret = conn->c_trans->conn_connect(conn);
147		rdsdebug("conn %p for %pI4 to %pI4 dispatched, ret %d\n",
148			conn, &conn->c_laddr, &conn->c_faddr, ret);
149
150		if (ret) {
151			if (rds_conn_transition(conn, RDS_CONN_CONNECTING, RDS_CONN_DOWN))
152				rds_queue_reconnect(conn);
153			else
154				rds_conn_error(conn, "RDS: connect failed\n");
155		}
156	}
157}
158
159void rds_shutdown_worker(struct work_struct *work)
160{
161	struct rds_connection *conn = container_of(work, struct rds_connection, c_down_w);
162
163	/* shut it down unless it's down already */
164	if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) {
165		/*
166		 * Quiesce the connection mgmt handlers before we start tearing
167		 * things down. We don't hold the mutex for the entire
168		 * duration of the shutdown operation, else we may be
169		 * deadlocking with the CM handler. Instead, the CM event
170		 * handler is supposed to check for state DISCONNECTING
171		 */
172		mutex_lock(&conn->c_cm_lock);
173		if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING) &&
174		    !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) {
175			rds_conn_error(conn, "shutdown called in state %d\n",
176					atomic_read(&conn->c_state));
177			mutex_unlock(&conn->c_cm_lock);
178			return;
179		}
180		mutex_unlock(&conn->c_cm_lock);
181
182		mutex_lock(&conn->c_send_lock);
183		conn->c_trans->conn_shutdown(conn);
184		rds_conn_reset(conn);
185		mutex_unlock(&conn->c_send_lock);
186
187		if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) {
188			/* This can happen - eg when we're in the middle of tearing
189			 * down the connection, and someone unloads the rds module.
190			 * Quite reproduceable with loopback connections.
191			 * Mostly harmless.
192			 */
193			rds_conn_error(conn,
194				"%s: failed to transition to state DOWN, "
195				"current state is %d\n",
196				__func__,
197				atomic_read(&conn->c_state));
198			return;
199		}
200	}
201
202	/* Then reconnect if it's still live.
203	 * The passive side of an IB loopback connection is never added
204	 * to the conn hash, so we never trigger a reconnect on this
205	 * conn - the reconnect is always triggered by the active peer. */
206	cancel_delayed_work(&conn->c_conn_w);
207	if (!hlist_unhashed(&conn->c_hash_node))
208		rds_queue_reconnect(conn);
209}
210
211void rds_send_worker(struct work_struct *work)
212{
213	struct rds_connection *conn = container_of(work, struct rds_connection, c_send_w.work);
214	int ret;
215
216	if (rds_conn_state(conn) == RDS_CONN_UP) {
217		ret = rds_send_xmit(conn);
218		rdsdebug("conn %p ret %d\n", conn, ret);
219		switch (ret) {
220		case -EAGAIN:
221			rds_stats_inc(s_send_immediate_retry);
222			queue_delayed_work(rds_wq, &conn->c_send_w, 0);
223			break;
224		case -ENOMEM:
225			rds_stats_inc(s_send_delayed_retry);
226			queue_delayed_work(rds_wq, &conn->c_send_w, 2);
227		default:
228			break;
229		}
230	}
231}
232
233void rds_recv_worker(struct work_struct *work)
234{
235	struct rds_connection *conn = container_of(work, struct rds_connection, c_recv_w.work);
236	int ret;
237
238	if (rds_conn_state(conn) == RDS_CONN_UP) {
239		ret = conn->c_trans->recv(conn);
240		rdsdebug("conn %p ret %d\n", conn, ret);
241		switch (ret) {
242		case -EAGAIN:
243			rds_stats_inc(s_recv_immediate_retry);
244			queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
245			break;
246		case -ENOMEM:
247			rds_stats_inc(s_recv_delayed_retry);
248			queue_delayed_work(rds_wq, &conn->c_recv_w, 2);
249		default:
250			break;
251		}
252	}
253}
254
255void rds_threads_exit(void)
256{
257	destroy_workqueue(rds_wq);
258}
259
260int __init rds_threads_init(void)
261{
262	rds_wq = create_workqueue("krdsd");
263	if (rds_wq == NULL)
264		return -ENOMEM;
265
266	return 0;
267}
268