[deliverable/linux.git] / net / rds / threads.c

/*
 * Copyright (c) 2006 Oracle.  All rights reserved.
 *
 * This software is available to you under a choice of one of two
 * licenses.  You may choose to be licensed under the terms of the GNU
 * General Public License (GPL) Version 2, available from the file
 * COPYING in the main directory of this source tree, or the
 * OpenIB.org BSD license below:
 *
 *     Redistribution and use in source and binary forms, with or
 *     without modification, are permitted provided that the following
 *     conditions are met:
 *
 *      - Redistributions of source code must retain the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer.
 *
 *      - Redistributions in binary form must reproduce the above
 *        copyright notice, this list of conditions and the following
 *        disclaimer in the documentation and/or other materials
 *        provided with the distribution.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 *
 */
#include <linux/kernel.h>
#include <linux/random.h>
#include <linux/export.h>

#include "rds.h"

/*
 * All of connection management is simplified by serializing it through
 * work queues that execute in a connection managing thread.
 *
 * TCP wants to send acks through sendpage() in response to data_ready(),
 * but it needs a process context to do so.
 *
 * The receive paths need to allocate but can't drop packets (!) so we have
 * a thread around to block allocating if the receive fast path sees an
 * allocation failure.
 */

/* Grand Unified Theory of connection life cycle:
 * At any point in time, the connection can be in one of these states:
 * DOWN, CONNECTING, UP, DISCONNECTING, ERROR
 *
 * The following transitions are possible:
 *  ANY		  -> ERROR
 *  UP		  -> DISCONNECTING
 *  ERROR	  -> DISCONNECTING
 *  DISCONNECTING -> DOWN
 *  DOWN	  -> CONNECTING
 *  CONNECTING	  -> UP
 *
 * Transition to state DISCONNECTING/DOWN:
 *  -	Inside the shutdown worker; synchronizes with xmit path
 *	through RDS_IN_XMIT, and with connection management callbacks
 *	via c_cm_lock.
 *
 *	For receive callbacks, we rely on the underlying transport
 *	(TCP, IB/RDMA) to provide the necessary synchronisation.
 */
struct workqueue_struct *rds_wq;
EXPORT_SYMBOL_GPL(rds_wq);

void rds_connect_path_complete(struct rds_connection *conn, int curr)
{
	if (!rds_conn_transition(conn, curr, RDS_CONN_UP)) {
		printk(KERN_WARNING "%s: Cannot transition to state UP, "
				"current state is %d\n",
				__func__,
				atomic_read(&conn->c_state));
		rds_conn_drop(conn);
		return;
	}

	rdsdebug("conn %p for %pI4 to %pI4 complete\n",
	  conn, &conn->c_laddr, &conn->c_faddr);

	conn->c_reconnect_jiffies = 0;
	set_bit(0, &conn->c_map_queued);
	queue_delayed_work(rds_wq, &conn->c_send_w, 0);
	queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
}
EXPORT_SYMBOL_GPL(rds_connect_path_complete);

void rds_connect_complete(struct rds_connection *conn)
{
	rds_connect_path_complete(conn, RDS_CONN_CONNECTING);
}
EXPORT_SYMBOL_GPL(rds_connect_complete);

/*
 * This random exponential backoff is relied on to eventually resolve racing
 * connects.
 *
 * If connect attempts race then both parties drop both connections and come
 * here to wait for a random amount of time before trying again.  Eventually
 * the backoff range will be so much greater than the time it takes to
 * establish a connection that one of the pair will establish the connection
 * before the other's random delay fires.
 *
 * Connection attempts that arrive while a connection is already established
 * are also considered to be racing connects.  This lets a connection from
 * a rebooted machine replace an existing stale connection before the transport
 * notices that the connection has failed.
 *
 * We should *always* start with a random backoff; otherwise a broken connection
 * will always take several iterations to be re-established.
 */
void rds_queue_reconnect(struct rds_connection *conn)
{
	unsigned long rand;

	rdsdebug("conn %p for %pI4 to %pI4 reconnect jiffies %lu\n",
	  conn, &conn->c_laddr, &conn->c_faddr,
	  conn->c_reconnect_jiffies);

	set_bit(RDS_RECONNECT_PENDING, &conn->c_flags);
	if (conn->c_reconnect_jiffies == 0) {
		conn->c_reconnect_jiffies = rds_sysctl_reconnect_min_jiffies;
		queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
		return;
	}

	get_random_bytes(&rand, sizeof(rand));
	rdsdebug("%lu delay %lu ceil conn %p for %pI4 -> %pI4\n",
		 rand % conn->c_reconnect_jiffies, conn->c_reconnect_jiffies,
		 conn, &conn->c_laddr, &conn->c_faddr);
	queue_delayed_work(rds_wq, &conn->c_conn_w,
			   rand % conn->c_reconnect_jiffies);

	conn->c_reconnect_jiffies = min(conn->c_reconnect_jiffies * 2,
					rds_sysctl_reconnect_max_jiffies);
}

void rds_connect_worker(struct work_struct *work)
{
	struct rds_connection *conn = container_of(work, struct rds_connection, c_conn_w.work);
	int ret;

	clear_bit(RDS_RECONNECT_PENDING, &conn->c_flags);
	if (rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
		ret = conn->c_trans->conn_connect(conn);
		rdsdebug("conn %p for %pI4 to %pI4 dispatched, ret %d\n",
			conn, &conn->c_laddr, &conn->c_faddr, ret);

		if (ret) {
			if (rds_conn_transition(conn, RDS_CONN_CONNECTING, RDS_CONN_DOWN))
				rds_queue_reconnect(conn);
			else
				rds_conn_error(conn, "RDS: connect failed\n");
		}
	}
}

void rds_send_worker(struct work_struct *work)
{
	struct rds_connection *conn = container_of(work, struct rds_connection, c_send_w.work);
	int ret;

	if (rds_conn_state(conn) == RDS_CONN_UP) {
		clear_bit(RDS_LL_SEND_FULL, &conn->c_flags);
		ret = rds_send_xmit(conn);
		cond_resched();
		rdsdebug("conn %p ret %d\n", conn, ret);
		switch (ret) {
		case -EAGAIN:
			rds_stats_inc(s_send_immediate_retry);
			queue_delayed_work(rds_wq, &conn->c_send_w, 0);
			break;
		case -ENOMEM:
			rds_stats_inc(s_send_delayed_retry);
			queue_delayed_work(rds_wq, &conn->c_send_w, 2);
		default:
			break;
		}
	}
}

void rds_recv_worker(struct work_struct *work)
{
	struct rds_connection *conn = container_of(work, struct rds_connection, c_recv_w.work);
	int ret;

	if (rds_conn_state(conn) == RDS_CONN_UP) {
		ret = conn->c_trans->recv(conn);
		rdsdebug("conn %p ret %d\n", conn, ret);
		switch (ret) {
		case -EAGAIN:
			rds_stats_inc(s_recv_immediate_retry);
			queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
			break;
		case -ENOMEM:
			rds_stats_inc(s_recv_delayed_retry);
			queue_delayed_work(rds_wq, &conn->c_recv_w, 2);
		default:
			break;
		}
	}
}

void rds_shutdown_worker(struct work_struct *work)
{
	struct rds_connection *conn = container_of(work, struct rds_connection, c_down_w);

	rds_conn_shutdown(conn);
}

void rds_threads_exit(void)
{
	destroy_workqueue(rds_wq);
}

int rds_threads_init(void)
{
	rds_wq = create_singlethread_workqueue("krdsd");
	if (!rds_wq)
		return -ENOMEM;

	return 0;
}
Commit	Line	Data
00e0f34c AG	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>
bc3b2d7f	35	#include <linux/export.h>
00e0f34c AG	36
	37	#include "rds.h"
	38
	39	/*
	40	* All of connection management is simplified by serializing it through
	41	* work queues that execute in a connection managing thread.
	42	*
	43	* TCP wants to send acks through sendpage() in response to data_ready(),
	44	* but it needs a process context to do so.
	45	*
	46	* The receive paths need to allocate but can't drop packets (!) so we have
	47	* a thread around to block allocating if the receive fast path sees an
	48	* allocation failure.
	49	*/
	50
	51	/* Grand Unified Theory of connection life cycle:
	52	* At any point in time, the connection can be in one of these states:
	53	* DOWN, CONNECTING, UP, DISCONNECTING, ERROR
	54	*
	55	* The following transitions are possible:
	56	* ANY -> ERROR
	57	* UP -> DISCONNECTING
	58	* ERROR -> DISCONNECTING
	59	* DISCONNECTING -> DOWN
	60	* DOWN -> CONNECTING
	61	* CONNECTING -> UP
	62	*
	63	* Transition to state DISCONNECTING/DOWN:
	64	* - Inside the shutdown worker; synchronizes with xmit path
0f4b1c7e	65	* through RDS_IN_XMIT, and with connection management callbacks
00e0f34c AG	66	* via c_cm_lock.
	67	*
	68	* For receive callbacks, we rely on the underlying transport
	69	* (TCP, IB/RDMA) to provide the necessary synchronisation.
	70	*/
	71	struct workqueue_struct *rds_wq;
616b757a	72	EXPORT_SYMBOL_GPL(rds_wq);
00e0f34c	73
9c79440e	74	void rds_connect_path_complete(struct rds_connection *conn, int curr)
00e0f34c	75	{
9c79440e	76	if (!rds_conn_transition(conn, curr, RDS_CONN_UP)) {
00e0f34c AG	77	printk(KERN_WARNING "%s: Cannot transition to state UP, "
	78	"current state is %d\n",
	79	__func__,
	80	atomic_read(&conn->c_state));
310886dd	81	rds_conn_drop(conn);
00e0f34c AG	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	}
9c79440e SV	93	EXPORT_SYMBOL_GPL(rds_connect_path_complete);
	94
	95	void rds_connect_complete(struct rds_connection *conn)
	96	{
	97	rds_connect_path_complete(conn, RDS_CONN_CONNECTING);
	98	}
616b757a	99	EXPORT_SYMBOL_GPL(rds_connect_complete);
00e0f34c AG	100
	101	/*
	102	* This random exponential backoff is relied on to eventually resolve racing
	103	* connects.
	104	*
	105	* If connect attempts race then both parties drop both connections and come
	106	* here to wait for a random amount of time before trying again. Eventually
	107	* the backoff range will be so much greater than the time it takes to
	108	* establish a connection that one of the pair will establish the connection
	109	* before the other's random delay fires.
	110	*
	111	* Connection attempts that arrive while a connection is already established
	112	* are also considered to be racing connects. This lets a connection from
	113	* a rebooted machine replace an existing stale connection before the transport
	114	* notices that the connection has failed.
	115	*
	116	* We should always start with a random backoff; otherwise a broken connection
	117	* will always take several iterations to be re-established.
	118	*/
2dc39357	119	void rds_queue_reconnect(struct rds_connection *conn)
00e0f34c AG	120	{
	121	unsigned long rand;
	122
	123	rdsdebug("conn %p for %pI4 to %pI4 reconnect jiffies %lu\n",
	124	conn, &conn->c_laddr, &conn->c_faddr,
	125	conn->c_reconnect_jiffies);
	126
	127	set_bit(RDS_RECONNECT_PENDING, &conn->c_flags);
	128	if (conn->c_reconnect_jiffies == 0) {
	129	conn->c_reconnect_jiffies = rds_sysctl_reconnect_min_jiffies;
	130	queue_delayed_work(rds_wq, &conn->c_conn_w, 0);
	131	return;
	132	}
	133
	134	get_random_bytes(&rand, sizeof(rand));
	135	rdsdebug("%lu delay %lu ceil conn %p for %pI4 -> %pI4\n",
	136	rand % conn->c_reconnect_jiffies, conn->c_reconnect_jiffies,
	137	conn, &conn->c_laddr, &conn->c_faddr);
	138	queue_delayed_work(rds_wq, &conn->c_conn_w,
	139	rand % conn->c_reconnect_jiffies);
	140
	141	conn->c_reconnect_jiffies = min(conn->c_reconnect_jiffies * 2,
	142	rds_sysctl_reconnect_max_jiffies);
	143	}
	144
	145	void rds_connect_worker(struct work_struct *work)
	146	{
	147	struct rds_connection *conn = container_of(work, struct rds_connection, c_conn_w.work);
	148	int ret;
	149
	150	clear_bit(RDS_RECONNECT_PENDING, &conn->c_flags);
	151	if (rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
	152	ret = conn->c_trans->conn_connect(conn);
	153	rdsdebug("conn %p for %pI4 to %pI4 dispatched, ret %d\n",
	154	conn, &conn->c_laddr, &conn->c_faddr, ret);
	155
	156	if (ret) {
	157	if (rds_conn_transition(conn, RDS_CONN_CONNECTING, RDS_CONN_DOWN))
	158	rds_queue_reconnect(conn);
	159	else
	160	rds_conn_error(conn, "RDS: connect failed\n");
	161	}
	162	}
	163	}
	164
00e0f34c AG	165	void rds_send_worker(struct work_struct *work)
	166	{
	167	struct rds_connection *conn = container_of(work, struct rds_connection, c_send_w.work);
	168	int ret;
	169
	170	if (rds_conn_state(conn) == RDS_CONN_UP) {
db6526dc	171	clear_bit(RDS_LL_SEND_FULL, &conn->c_flags);
00e0f34c	172	ret = rds_send_xmit(conn);
db6526dc	173	cond_resched();
00e0f34c AG	174	rdsdebug("conn %p ret %d\n", conn, ret);
	175	switch (ret) {
	176	case -EAGAIN:
	177	rds_stats_inc(s_send_immediate_retry);
	178	queue_delayed_work(rds_wq, &conn->c_send_w, 0);
	179	break;
	180	case -ENOMEM:
	181	rds_stats_inc(s_send_delayed_retry);
	182	queue_delayed_work(rds_wq, &conn->c_send_w, 2);
	183	default:
	184	break;
	185	}
	186	}
	187	}
	188
	189	void rds_recv_worker(struct work_struct *work)
	190	{
	191	struct rds_connection *conn = container_of(work, struct rds_connection, c_recv_w.work);
	192	int ret;
	193
	194	if (rds_conn_state(conn) == RDS_CONN_UP) {
	195	ret = conn->c_trans->recv(conn);
	196	rdsdebug("conn %p ret %d\n", conn, ret);
	197	switch (ret) {
	198	case -EAGAIN:
	199	rds_stats_inc(s_recv_immediate_retry);
	200	queue_delayed_work(rds_wq, &conn->c_recv_w, 0);
	201	break;
	202	case -ENOMEM:
	203	rds_stats_inc(s_recv_delayed_retry);
	204	queue_delayed_work(rds_wq, &conn->c_recv_w, 2);
	205	default:
	206	break;
	207	}
	208	}
	209	}
	210
2dc39357 AG	211	void rds_shutdown_worker(struct work_struct *work)
	212	{
	213	struct rds_connection *conn = container_of(work, struct rds_connection, c_down_w);
	214
	215	rds_conn_shutdown(conn);
	216	}
	217
00e0f34c AG	218	void rds_threads_exit(void)
	219	{
	220	destroy_workqueue(rds_wq);
	221	}
	222
ef87b7ea	223	int rds_threads_init(void)
00e0f34c	224	{
80c51be5	225	rds_wq = create_singlethread_workqueue("krdsd");
8690bfa1	226	if (!rds_wq)
00e0f34c AG	227	return -ENOMEM;
	228
	229	return 0;
	230	}