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

/*
 * Copyright (c) 2006, 2018 Oracle and/or its affiliates. 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_conn_path *cp, int curr)
{
	if (!rds_conn_path_transition(cp, curr, RDS_CONN_UP)) {
		printk(KERN_WARNING "%s: Cannot transition to state UP, "
				"current state is %d\n",
				__func__,
				atomic_read(&cp->cp_state));
		rds_conn_path_drop(cp, false);
		return;
	}

	rdsdebug("conn %p for %pI6c to %pI6c complete\n",
		 cp->cp_conn, &cp->cp_conn->c_laddr, &cp->cp_conn->c_faddr);

	cp->cp_reconnect_jiffies = 0;
	set_bit(0, &cp->cp_conn->c_map_queued);
	rcu_read_lock();
	if (!rds_destroy_pending(cp->cp_conn)) {
		queue_delayed_work(rds_wq, &cp->cp_send_w, 0);
		queue_delayed_work(rds_wq, &cp->cp_recv_w, 0);
	}
	rcu_read_unlock();
	cp->cp_conn->c_proposed_version = RDS_PROTOCOL_VERSION;
}
EXPORT_SYMBOL_GPL(rds_connect_path_complete);

void rds_connect_complete(struct rds_connection *conn)
{
	rds_connect_path_complete(&conn->c_path[0], 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_conn_path *cp)
{
	unsigned long rand;
	struct rds_connection *conn = cp->cp_conn;

	rdsdebug("conn %p for %pI6c to %pI6c reconnect jiffies %lu\n",
		 conn, &conn->c_laddr, &conn->c_faddr,
		 cp->cp_reconnect_jiffies);

	/* let peer with smaller addr initiate reconnect, to avoid duels */
	if (conn->c_trans->t_type == RDS_TRANS_TCP &&
	    rds_addr_cmp(&conn->c_laddr, &conn->c_faddr) >= 0)
		return;

	set_bit(RDS_RECONNECT_PENDING, &cp->cp_flags);
	if (cp->cp_reconnect_jiffies == 0) {
		cp->cp_reconnect_jiffies = rds_sysctl_reconnect_min_jiffies;
		rcu_read_lock();
		if (!rds_destroy_pending(cp->cp_conn))
			queue_delayed_work(rds_wq, &cp->cp_conn_w, 0);
		rcu_read_unlock();
		return;
	}

	get_random_bytes(&rand, sizeof(rand));
	rdsdebug("%lu delay %lu ceil conn %p for %pI6c -> %pI6c\n",
		 rand % cp->cp_reconnect_jiffies, cp->cp_reconnect_jiffies,
		 conn, &conn->c_laddr, &conn->c_faddr);
	rcu_read_lock();
	if (!rds_destroy_pending(cp->cp_conn))
		queue_delayed_work(rds_wq, &cp->cp_conn_w,
				   rand % cp->cp_reconnect_jiffies);
	rcu_read_unlock();

	cp->cp_reconnect_jiffies = min(cp->cp_reconnect_jiffies * 2,
					rds_sysctl_reconnect_max_jiffies);
}

void rds_connect_worker(struct work_struct *work)
{
	struct rds_conn_path *cp = container_of(work,
						struct rds_conn_path,
						cp_conn_w.work);
	struct rds_connection *conn = cp->cp_conn;
	int ret;

	if (cp->cp_index > 0 &&
	    rds_addr_cmp(&cp->cp_conn->c_laddr, &cp->cp_conn->c_faddr) >= 0)
		return;
	clear_bit(RDS_RECONNECT_PENDING, &cp->cp_flags);
	ret = rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_CONNECTING);
	if (ret) {
		ret = conn->c_trans->conn_path_connect(cp);
		rdsdebug("conn %p for %pI6c to %pI6c dispatched, ret %d\n",
			 conn, &conn->c_laddr, &conn->c_faddr, ret);

		if (ret) {
			if (rds_conn_path_transition(cp,
						     RDS_CONN_CONNECTING,
						     RDS_CONN_DOWN))
				rds_queue_reconnect(cp);
			else
				rds_conn_path_error(cp, "connect failed\n");
		}
	}
}

void rds_send_worker(struct work_struct *work)
{
	struct rds_conn_path *cp = container_of(work,
						struct rds_conn_path,
						cp_send_w.work);
	int ret;

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

void rds_recv_worker(struct work_struct *work)
{
	struct rds_conn_path *cp = container_of(work,
						struct rds_conn_path,
						cp_recv_w.work);
	int ret;

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

void rds_shutdown_worker(struct work_struct *work)
{
	struct rds_conn_path *cp = container_of(work,
						struct rds_conn_path,
						cp_down_w);

	rds_conn_shutdown(cp);
}

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;
}

/* Compare two IPv6 addresses.  Return 0 if the two addresses are equal.
 * Return 1 if the first is greater.  Return -1 if the second is greater.
 */
int rds_addr_cmp(const struct in6_addr *addr1,
		 const struct in6_addr *addr2)
{
#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && BITS_PER_LONG == 64
	const __be64 *a1, *a2;
	u64 x, y;

	a1 = (__be64 *)addr1;
	a2 = (__be64 *)addr2;

	if (*a1 != *a2) {
		if (be64_to_cpu(*a1) < be64_to_cpu(*a2))
			return -1;
		else
			return 1;
	} else {
		x = be64_to_cpu(*++a1);
		y = be64_to_cpu(*++a2);
		if (x < y)
			return -1;
		else if (x > y)
			return 1;
		else
			return 0;
	}
#else
	u32 a, b;
	int i;

	for (i = 0; i < 4; i++) {
		if (addr1->s6_addr32[i] != addr2->s6_addr32[i]) {
			a = ntohl(addr1->s6_addr32[i]);
			b = ntohl(addr2->s6_addr32[i]);
			if (a < b)
				return -1;
			else if (a > b)
				return 1;
		}
	}
	return 0;
#endif
}
EXPORT_SYMBOL_GPL(rds_addr_cmp);
Commit	Line	Data
00e0f34c	1	/*
eee2fa6a	2	* Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved.
00e0f34c AG	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
0cb43965	74	void rds_connect_path_complete(struct rds_conn_path *cp, int curr)
00e0f34c	75	{
0cb43965	76	if (!rds_conn_path_transition(cp, curr, RDS_CONN_UP)) {
00e0f34c AG	77	printk(KERN_WARNING "%s: Cannot transition to state UP, "
	78	"current state is %d\n",
	79	__func__,
0cb43965	80	atomic_read(&cp->cp_state));
aed20a53	81	rds_conn_path_drop(cp, false);
00e0f34c AG	82	return;
	83	}
	84
eee2fa6a KCP	85	rdsdebug("conn %p for %pI6c to %pI6c complete\n",
eee2fa6a KCP	86	cp->cp_conn, &cp->cp_conn->c_laddr, &cp->cp_conn->c_faddr);
00e0f34c	87
0cb43965 SV	88	cp->cp_reconnect_jiffies = 0;
0cb43965 SV	89	set_bit(0, &cp->cp_conn->c_map_queued);
3db6e0d1	90	rcu_read_lock();
ebeeb1ad	91	if (!rds_destroy_pending(cp->cp_conn)) {
3db6e0d1 SV	92	queue_delayed_work(rds_wq, &cp->cp_send_w, 0);
	93	queue_delayed_work(rds_wq, &cp->cp_recv_w, 0);
	94	}
	95	rcu_read_unlock();
cdc306a5	96	cp->cp_conn->c_proposed_version = RDS_PROTOCOL_VERSION;
00e0f34c	97	}
9c79440e SV	98	EXPORT_SYMBOL_GPL(rds_connect_path_complete);
	99
	100	void rds_connect_complete(struct rds_connection *conn)
	101	{
0cb43965	102	rds_connect_path_complete(&conn->c_path[0], RDS_CONN_CONNECTING);
9c79440e	103	}
616b757a	104	EXPORT_SYMBOL_GPL(rds_connect_complete);
00e0f34c AG	105
	106	/*
	107	* This random exponential backoff is relied on to eventually resolve racing
	108	* connects.
	109	*
	110	* If connect attempts race then both parties drop both connections and come
	111	* here to wait for a random amount of time before trying again. Eventually
	112	* the backoff range will be so much greater than the time it takes to
	113	* establish a connection that one of the pair will establish the connection
	114	* before the other's random delay fires.
	115	*
	116	* Connection attempts that arrive while a connection is already established
	117	* are also considered to be racing connects. This lets a connection from
	118	* a rebooted machine replace an existing stale connection before the transport
	119	* notices that the connection has failed.
	120	*
	121	* We should always start with a random backoff; otherwise a broken connection
	122	* will always take several iterations to be re-established.
	123	*/
0cb43965	124	void rds_queue_reconnect(struct rds_conn_path *cp)
00e0f34c AG	125	{
00e0f34c AG	126	unsigned long rand;
0cb43965	127	struct rds_connection *conn = cp->cp_conn;
00e0f34c	128
eee2fa6a KCP	129	rdsdebug("conn %p for %pI6c to %pI6c reconnect jiffies %lu\n",
	130	conn, &conn->c_laddr, &conn->c_faddr,
	131	cp->cp_reconnect_jiffies);
00e0f34c	132
8315011a SV	133	/* let peer with smaller addr initiate reconnect, to avoid duels */
8315011a SV	134	if (conn->c_trans->t_type == RDS_TRANS_TCP &&
eee2fa6a	135	rds_addr_cmp(&conn->c_laddr, &conn->c_faddr) >= 0)
8315011a SV	136	return;
8315011a SV	137
0cb43965 SV	138	set_bit(RDS_RECONNECT_PENDING, &cp->cp_flags);
	139	if (cp->cp_reconnect_jiffies == 0) {
	140	cp->cp_reconnect_jiffies = rds_sysctl_reconnect_min_jiffies;
3db6e0d1	141	rcu_read_lock();
ebeeb1ad	142	if (!rds_destroy_pending(cp->cp_conn))
3db6e0d1 SV	143	queue_delayed_work(rds_wq, &cp->cp_conn_w, 0);
3db6e0d1 SV	144	rcu_read_unlock();
00e0f34c AG	145	return;
	146	}
	147
	148	get_random_bytes(&rand, sizeof(rand));
eee2fa6a	149	rdsdebug("%lu delay %lu ceil conn %p for %pI6c -> %pI6c\n",
0cb43965	150	rand % cp->cp_reconnect_jiffies, cp->cp_reconnect_jiffies,
00e0f34c	151	conn, &conn->c_laddr, &conn->c_faddr);
3db6e0d1	152	rcu_read_lock();
ebeeb1ad	153	if (!rds_destroy_pending(cp->cp_conn))
3db6e0d1 SV	154	queue_delayed_work(rds_wq, &cp->cp_conn_w,
	155	rand % cp->cp_reconnect_jiffies);
	156	rcu_read_unlock();
00e0f34c	157
0cb43965	158	cp->cp_reconnect_jiffies = min(cp->cp_reconnect_jiffies * 2,
00e0f34c AG	159	rds_sysctl_reconnect_max_jiffies);
	160	}
	161
	162	void rds_connect_worker(struct work_struct *work)
	163	{
0cb43965 SV	164	struct rds_conn_path *cp = container_of(work,
	165	struct rds_conn_path,
	166	cp_conn_w.work);
	167	struct rds_connection *conn = cp->cp_conn;
00e0f34c AG	168	int ret;
00e0f34c AG	169
00354de5	170	if (cp->cp_index > 0 &&
eee2fa6a	171	rds_addr_cmp(&cp->cp_conn->c_laddr, &cp->cp_conn->c_faddr) >= 0)
5916e2c1	172	return;
0cb43965	173	clear_bit(RDS_RECONNECT_PENDING, &cp->cp_flags);
b04e8554 SV	174	ret = rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_CONNECTING);
	175	if (ret) {
	176	ret = conn->c_trans->conn_path_connect(cp);
eee2fa6a KCP	177	rdsdebug("conn %p for %pI6c to %pI6c dispatched, ret %d\n",
eee2fa6a KCP	178	conn, &conn->c_laddr, &conn->c_faddr, ret);
00e0f34c AG	179
00e0f34c AG	180	if (ret) {
0cb43965 SV	181	if (rds_conn_path_transition(cp,
	182	RDS_CONN_CONNECTING,
	183	RDS_CONN_DOWN))
	184	rds_queue_reconnect(cp);
00e0f34c	185	else
9c7cbcf5	186	rds_conn_path_error(cp, "connect failed\n");
00e0f34c AG	187	}
	188	}
	189	}
	190
00e0f34c AG	191	void rds_send_worker(struct work_struct *work)
00e0f34c AG	192	{
0cb43965 SV	193	struct rds_conn_path *cp = container_of(work,
	194	struct rds_conn_path,
	195	cp_send_w.work);
00e0f34c AG	196	int ret;
00e0f34c AG	197
0cb43965 SV	198	if (rds_conn_path_state(cp) == RDS_CONN_UP) {
0cb43965 SV	199	clear_bit(RDS_LL_SEND_FULL, &cp->cp_flags);
1f9ecd7e	200	ret = rds_send_xmit(cp);
db6526dc	201	cond_resched();
0cb43965	202	rdsdebug("conn %p ret %d\n", cp->cp_conn, ret);
00e0f34c AG	203	switch (ret) {
	204	case -EAGAIN:
	205	rds_stats_inc(s_send_immediate_retry);
0cb43965	206	queue_delayed_work(rds_wq, &cp->cp_send_w, 0);
00e0f34c AG	207	break;
	208	case -ENOMEM:
	209	rds_stats_inc(s_send_delayed_retry);
0cb43965	210	queue_delayed_work(rds_wq, &cp->cp_send_w, 2);
00e0f34c AG	211	default:
	212	break;
	213	}
	214	}
	215	}
	216
	217	void rds_recv_worker(struct work_struct *work)
	218	{
0cb43965 SV	219	struct rds_conn_path *cp = container_of(work,
	220	struct rds_conn_path,
	221	cp_recv_w.work);
00e0f34c AG	222	int ret;
00e0f34c AG	223
0cb43965	224	if (rds_conn_path_state(cp) == RDS_CONN_UP) {
2da43c4a	225	ret = cp->cp_conn->c_trans->recv_path(cp);
0cb43965	226	rdsdebug("conn %p ret %d\n", cp->cp_conn, ret);
00e0f34c AG	227	switch (ret) {
	228	case -EAGAIN:
	229	rds_stats_inc(s_recv_immediate_retry);
0cb43965	230	queue_delayed_work(rds_wq, &cp->cp_recv_w, 0);
00e0f34c AG	231	break;
	232	case -ENOMEM:
	233	rds_stats_inc(s_recv_delayed_retry);
0cb43965	234	queue_delayed_work(rds_wq, &cp->cp_recv_w, 2);
00e0f34c AG	235	default:
	236	break;
	237	}
	238	}
	239	}
	240
2dc39357 AG	241	void rds_shutdown_worker(struct work_struct *work)
2dc39357 AG	242	{
0cb43965 SV	243	struct rds_conn_path *cp = container_of(work,
	244	struct rds_conn_path,
	245	cp_down_w);
2dc39357	246
d769ef81	247	rds_conn_shutdown(cp);
2dc39357 AG	248	}
2dc39357 AG	249
00e0f34c AG	250	void rds_threads_exit(void)
	251	{
	252	destroy_workqueue(rds_wq);
	253	}
	254
ef87b7ea	255	int rds_threads_init(void)
00e0f34c	256	{
80c51be5	257	rds_wq = create_singlethread_workqueue("krdsd");
8690bfa1	258	if (!rds_wq)
00e0f34c AG	259	return -ENOMEM;
	260
	261	return 0;
	262	}
eee2fa6a KCP	263
	264	/* Compare two IPv6 addresses. Return 0 if the two addresses are equal.
	265	* Return 1 if the first is greater. Return -1 if the second is greater.
	266	*/
	267	int rds_addr_cmp(const struct in6_addr *addr1,
	268	const struct in6_addr *addr2)
	269	{
	270	#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && BITS_PER_LONG == 64
	271	const __be64 a1, a2;
	272	u64 x, y;
	273
	274	a1 = (__be64 *)addr1;
	275	a2 = (__be64 *)addr2;
	276
	277	if (a1 != a2) {
	278	if (be64_to_cpu(a1) < be64_to_cpu(a2))
	279	return -1;
	280	else
	281	return 1;
	282	} else {
	283	x = be64_to_cpu(*++a1);
	284	y = be64_to_cpu(*++a2);
	285	if (x < y)
	286	return -1;
	287	else if (x > y)
	288	return 1;
	289	else
	290	return 0;
	291	}
	292	#else
	293	u32 a, b;
	294	int i;
	295
	296	for (i = 0; i < 4; i++) {
	297	if (addr1->s6_addr32[i] != addr2->s6_addr32[i]) {
	298	a = ntohl(addr1->s6_addr32[i]);
	299	b = ntohl(addr2->s6_addr32[i]);
	300	if (a < b)
	301	return -1;
	302	else if (a > b)
	303	return 1;
	304	}
	305	}
	306	return 0;
	307	#endif
	308	}
	309	EXPORT_SYMBOL_GPL(rds_addr_cmp);