[urcu.git] / urcu-defer.c

/*
 * urcu-defer.c
 *
 * Userspace RCU library - batch memory reclamation
 *
 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
 *
 * This library is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * This library is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with this library; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

#include <stdio.h>
#include <pthread.h>
#include <signal.h>
#include <assert.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>
#include <poll.h>
#include <sys/time.h>
#include <syscall.h>
#include <unistd.h>

#include "urcu/urcu-futex.h"
#include "urcu-defer-static.h"
/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
#include "urcu-defer.h"

void __attribute__((destructor)) urcu_defer_exit(void);

extern void synchronize_rcu(void);

/*
 * urcu_defer_mutex nests inside defer_thread_mutex.
 */
static pthread_mutex_t urcu_defer_mutex = PTHREAD_MUTEX_INITIALIZER;
static pthread_mutex_t defer_thread_mutex = PTHREAD_MUTEX_INITIALIZER;

static int defer_thread_futex;

/*
 * Written to only by each individual deferer. Read by both the deferer and
 * the reclamation tread.
 */
static struct defer_queue __thread defer_queue;
static LIST_HEAD(registry);
static pthread_t tid_defer;

static void internal_urcu_lock(pthread_mutex_t *mutex)
{
	int ret;

#ifndef DISTRUST_SIGNALS_EXTREME
	ret = pthread_mutex_lock(mutex);
	if (ret) {
		perror("Error in pthread mutex lock");
		exit(-1);
	}
#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
	while ((ret = pthread_mutex_trylock(mutex)) != 0) {
		if (ret != EBUSY && ret != EINTR) {
			printf("ret = %d, errno = %d\n", ret, errno);
			perror("Error in pthread mutex lock");
			exit(-1);
		}
		pthread_testcancel();
		poll(NULL,0,10);
	}
#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
}

static void internal_urcu_unlock(pthread_mutex_t *mutex)
{
	int ret;

	ret = pthread_mutex_unlock(mutex);
	if (ret) {
		perror("Error in pthread mutex unlock");
		exit(-1);
	}
}

/*
 * Wake-up any waiting defer thread. Called from many concurrent threads.
 */
static void wake_up_defer(void)
{
	if (unlikely(uatomic_read(&defer_thread_futex) == -1)) {
		uatomic_set(&defer_thread_futex, 0);
		futex_noasync(&defer_thread_futex, FUTEX_WAKE, 1,
		      NULL, NULL, 0);
	}
}

static unsigned long rcu_defer_num_callbacks(void)
{
	unsigned long num_items = 0, head;
	struct defer_queue *index;

	internal_urcu_lock(&urcu_defer_mutex);
	list_for_each_entry(index, &registry, list) {
		head = LOAD_SHARED(index->head);
		num_items += head - index->tail;
	}
	internal_urcu_unlock(&urcu_defer_mutex);
	return num_items;
}

/*
 * Defer thread waiting. Single thread.
 */
static void wait_defer(void)
{
	uatomic_dec(&defer_thread_futex);
	smp_mb();	/* Write futex before read queue */
	if (rcu_defer_num_callbacks()) {
		smp_mb();	/* Read queue before write futex */
		/* Callbacks are queued, don't wait. */
		uatomic_set(&defer_thread_futex, 0);
	} else {
		smp_rmb();	/* Read queue before read futex */
		if (uatomic_read(&defer_thread_futex) == -1)
			futex_noasync(&defer_thread_futex, FUTEX_WAIT, -1,
			      NULL, NULL, 0);
	}
}

/*
 * Must be called after Q.S. is reached.
 */
static void rcu_defer_barrier_queue(struct defer_queue *queue,
				    unsigned long head)
{
	unsigned long i;
	void (*fct)(void *p);
	void *p;

	/*
	 * Tail is only modified when lock is held.
	 * Head is only modified by owner thread.
	 */

	for (i = queue->tail; i != head;) {
		smp_rmb();       /* read head before q[]. */
		p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
		if (unlikely(DQ_IS_FCT_BIT(p))) {
			DQ_CLEAR_FCT_BIT(p);
			queue->last_fct_out = p;
			p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
		} else if (unlikely(p == DQ_FCT_MARK)) {
			p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
			queue->last_fct_out = p;
			p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
		}
		fct = queue->last_fct_out;
		fct(p);
	}
	smp_mb();	/* push tail after having used q[] */
	STORE_SHARED(queue->tail, i);
}

static void _rcu_defer_barrier_thread(void)
{
	unsigned long head, num_items;

	head = defer_queue.head;
	num_items = head - defer_queue.tail;
	if (unlikely(!num_items))
		return;
	synchronize_rcu();
	rcu_defer_barrier_queue(&defer_queue, head);
}

void rcu_defer_barrier_thread(void)
{
	internal_urcu_lock(&urcu_defer_mutex);
	_rcu_defer_barrier_thread();
	internal_urcu_unlock(&urcu_defer_mutex);
}

/*
 * rcu_defer_barrier - Execute all queued rcu callbacks.
 *
 * Execute all RCU callbacks queued before rcu_defer_barrier() execution.
 * All callbacks queued on the local thread prior to a rcu_defer_barrier() call
 * are guaranteed to be executed.
 * Callbacks queued by other threads concurrently with rcu_defer_barrier()
 * execution are not guaranteed to be executed in the current batch (could
 * be left for the next batch). These callbacks queued by other threads are only
 * guaranteed to be executed if there is explicit synchronization between
 * the thread adding to the queue and the thread issuing the defer_barrier call.
 */

void rcu_defer_barrier(void)
{
	struct defer_queue *index;
	unsigned long num_items = 0;

	if (list_empty(&registry))
		return;

	internal_urcu_lock(&urcu_defer_mutex);
	list_for_each_entry(index, &registry, list) {
		index->last_head = LOAD_SHARED(index->head);
		num_items += index->last_head - index->tail;
	}
	if (likely(!num_items)) {
		/*
		 * We skip the grace period because there are no queued
		 * callbacks to execute.
		 */
		goto end;
	}
	synchronize_rcu();
	list_for_each_entry(index, &registry, list)
		rcu_defer_barrier_queue(index, index->last_head);
end:
	internal_urcu_unlock(&urcu_defer_mutex);
}

/*
 * _defer_rcu - Queue a RCU callback.
 */
void _defer_rcu_ratelimit(void (*fct)(void *p), void *p, int (*rl)(void *p))
{
	unsigned long head, tail;
	int sync;

	/*
	 * Verify if we reached the rate limiter threshold.
	 */
	sync = rl ? rl(p) : 0;

	/*
	 * Head is only modified by ourself. Tail can be modified by reclamation
	 * thread.
	 */
	head = defer_queue.head;
	tail = LOAD_SHARED(defer_queue.tail);

	/*
	 * If queue is full, or reached threshold. Empty queue ourself.
	 * Worse-case: must allow 2 supplementary entries for fct pointer.
	 */
	if (unlikely(sync || (head - tail >= DEFER_QUEUE_SIZE - 2))) {
		assert(head - tail <= DEFER_QUEUE_SIZE);
		rcu_defer_barrier_thread();
		assert(head - LOAD_SHARED(defer_queue.tail) == 0);
	}

	if (unlikely(defer_queue.last_fct_in != fct)) {
		defer_queue.last_fct_in = fct;
		if (unlikely(DQ_IS_FCT_BIT(fct) || fct == DQ_FCT_MARK)) {
			/*
			 * If the function to encode is not aligned or the
			 * marker, write DQ_FCT_MARK followed by the function
			 * pointer.
			 */
			_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
				      DQ_FCT_MARK);
			_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
				      fct);
		} else {
			DQ_SET_FCT_BIT(fct);
			_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
				      fct);
		}
	} else {
		if (unlikely(DQ_IS_FCT_BIT(p) || p == DQ_FCT_MARK)) {
			/*
			 * If the data to encode is not aligned or the marker,
			 * write DQ_FCT_MARK followed by the function pointer.
			 */
			_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
				      DQ_FCT_MARK);
			_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
				      fct);
		}
	}
	_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], p);
	smp_wmb();	/* Publish new pointer before head */
			/* Write q[] before head. */
	STORE_SHARED(defer_queue.head, head);
	smp_mb();	/* Write queue head before read futex */
	/*
	 * Wake-up any waiting defer thread.
	 */
	wake_up_defer();
}

void *thr_defer(void *args)
{
	for (;;) {
		pthread_testcancel();
		/*
		 * "Be green". Don't wake up the CPU if there is no RCU work
		 * to perform whatsoever. Aims at saving laptop battery life by
		 * leaving the processor in sleep state when idle.
		 */
		wait_defer();
		/* Sleeping after wait_defer to let many callbacks enqueue */
		poll(NULL,0,100);	/* wait for 100ms */
		rcu_defer_barrier();
	}

	return NULL;
}

/*
 * library wrappers to be used by non-LGPL compatible source code.
 */

void defer_rcu_ratelimit(void (*fct)(void *p), void *p, int (*rl)(void *p))
{
	_defer_rcu_ratelimit(fct, p, rl);
}

static void start_defer_thread(void)
{
	int ret;

	ret = pthread_create(&tid_defer, NULL, thr_defer, NULL);
	assert(!ret);
}

static void stop_defer_thread(void)
{
	int ret;
	void *tret;

	pthread_cancel(tid_defer);
	wake_up_defer();
	ret = pthread_join(tid_defer, &tret);
	assert(!ret);
}

void rcu_defer_register_thread(void)
{
	int was_empty;

	assert(defer_queue.last_head == 0);
	assert(defer_queue.q == NULL);
	defer_queue.q = malloc(sizeof(void *) * DEFER_QUEUE_SIZE);

	internal_urcu_lock(&defer_thread_mutex);
	internal_urcu_lock(&urcu_defer_mutex);
	was_empty = list_empty(&registry);
	list_add(&defer_queue.list, &registry);
	internal_urcu_unlock(&urcu_defer_mutex);

	if (was_empty)
		start_defer_thread();
	internal_urcu_unlock(&defer_thread_mutex);
}

void rcu_defer_unregister_thread(void)
{
	int is_empty;

	internal_urcu_lock(&defer_thread_mutex);
	internal_urcu_lock(&urcu_defer_mutex);
	list_del(&defer_queue.list);
	_rcu_defer_barrier_thread();
	free(defer_queue.q);
	defer_queue.q = NULL;
	is_empty = list_empty(&registry);
	internal_urcu_unlock(&urcu_defer_mutex);

	if (is_empty)
		stop_defer_thread();
	internal_urcu_unlock(&defer_thread_mutex);
}

void urcu_defer_exit(void)
{
	assert(list_empty(&registry));
}
Commit	Line	Data
786ee85b MD	1	/*
	2	* urcu-defer.c
	3	*
	4	* Userspace RCU library - batch memory reclamation
	5	*
	6	* Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
	7	*
	8	* This library is free software; you can redistribute it and/or
	9	* modify it under the terms of the GNU Lesser General Public
	10	* License as published by the Free Software Foundation; either
	11	* version 2.1 of the License, or (at your option) any later version.
	12	*
	13	* This library is distributed in the hope that it will be useful,
	14	* but WITHOUT ANY WARRANTY; without even the implied warranty of
	15	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
	16	* Lesser General Public License for more details.
	17	*
	18	* You should have received a copy of the GNU Lesser General Public
	19	* License along with this library; if not, write to the Free Software
	20	* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
	21	*/
	22
	23	#include <stdio.h>
	24	#include <pthread.h>
	25	#include <signal.h>
	26	#include <assert.h>
	27	#include <stdlib.h>
	28	#include <string.h>
	29	#include <errno.h>
	30	#include <poll.h>
4ce9e4f2 MD	31	#include <sys/time.h>
	32	#include <syscall.h>
	33	#include <unistd.h>
786ee85b	34
0854ccff	35	#include "urcu/urcu-futex.h"
786ee85b MD	36	#include "urcu-defer-static.h"
	37	/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
	38	#include "urcu-defer.h"
	39
	40	void __attribute__((destructor)) urcu_defer_exit(void);
	41
	42	extern void synchronize_rcu(void);
	43
	44	/*
	45	* urcu_defer_mutex nests inside defer_thread_mutex.
	46	*/
	47	static pthread_mutex_t urcu_defer_mutex = PTHREAD_MUTEX_INITIALIZER;
	48	static pthread_mutex_t defer_thread_mutex = PTHREAD_MUTEX_INITIALIZER;
	49
3dce4bfa	50	static int defer_thread_futex;
4ce9e4f2	51
786ee85b MD	52	/*
	53	* Written to only by each individual deferer. Read by both the deferer and
	54	* the reclamation tread.
	55	*/
3dce4bfa	56	static struct defer_queue __thread defer_queue;
dbc6128f	57	static LIST_HEAD(registry);
786ee85b	58	static pthread_t tid_defer;
4ce9e4f2	59
786ee85b MD	60	static void internal_urcu_lock(pthread_mutex_t *mutex)
	61	{
	62	int ret;
	63
	64	#ifndef DISTRUST_SIGNALS_EXTREME
	65	ret = pthread_mutex_lock(mutex);
	66	if (ret) {
	67	perror("Error in pthread mutex lock");
	68	exit(-1);
	69	}
	70	#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
	71	while ((ret = pthread_mutex_trylock(mutex)) != 0) {
	72	if (ret != EBUSY && ret != EINTR) {
	73	printf("ret = %d, errno = %d\n", ret, errno);
	74	perror("Error in pthread mutex lock");
	75	exit(-1);
	76	}
4ce9e4f2	77	pthread_testcancel();
786ee85b MD	78	poll(NULL,0,10);
	79	}
	80	#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
	81	}
	82
	83	static void internal_urcu_unlock(pthread_mutex_t *mutex)
	84	{
	85	int ret;
	86
	87	ret = pthread_mutex_unlock(mutex);
	88	if (ret) {
	89	perror("Error in pthread mutex unlock");
	90	exit(-1);
	91	}
	92	}
	93
04eb9c4f MD	94	/*
	95	* Wake-up any waiting defer thread. Called from many concurrent threads.
	96	*/
	97	static void wake_up_defer(void)
	98	{
ec4e58a3 MD	99	if (unlikely(uatomic_read(&defer_thread_futex) == -1)) {
ec4e58a3 MD	100	uatomic_set(&defer_thread_futex, 0);
0854ccff	101	futex_noasync(&defer_thread_futex, FUTEX_WAKE, 1,
04eb9c4f MD	102	NULL, NULL, 0);
	103	}
	104	}
	105
	106	static unsigned long rcu_defer_num_callbacks(void)
	107	{
	108	unsigned long num_items = 0, head;
dbc6128f	109	struct defer_queue *index;
04eb9c4f MD	110
04eb9c4f MD	111	internal_urcu_lock(&urcu_defer_mutex);
dbc6128f MD	112	list_for_each_entry(index, &registry, list) {
	113	head = LOAD_SHARED(index->head);
	114	num_items += head - index->tail;
04eb9c4f MD	115	}
	116	internal_urcu_unlock(&urcu_defer_mutex);
	117	return num_items;
	118	}
	119
	120	/*
	121	* Defer thread waiting. Single thread.
	122	*/
	123	static void wait_defer(void)
	124	{
ec4e58a3	125	uatomic_dec(&defer_thread_futex);
04eb9c4f MD	126	smp_mb(); /* Write futex before read queue */
	127	if (rcu_defer_num_callbacks()) {
	128	smp_mb(); /* Read queue before write futex */
	129	/* Callbacks are queued, don't wait. */
ec4e58a3	130	uatomic_set(&defer_thread_futex, 0);
04eb9c4f MD	131	} else {
04eb9c4f MD	132	smp_rmb(); /* Read queue before read futex */
ec4e58a3	133	if (uatomic_read(&defer_thread_futex) == -1)
0854ccff	134	futex_noasync(&defer_thread_futex, FUTEX_WAIT, -1,
04eb9c4f MD	135	NULL, NULL, 0);
	136	}
	137	}
	138
786ee85b MD	139	/*
	140	* Must be called after Q.S. is reached.
	141	*/
	142	static void rcu_defer_barrier_queue(struct defer_queue *queue,
804b4375	143	unsigned long head)
786ee85b MD	144	{
786ee85b MD	145	unsigned long i;
804b4375 MD	146	void (fct)(void p);
804b4375 MD	147	void *p;
786ee85b MD	148
	149	/*
	150	* Tail is only modified when lock is held.
	151	* Head is only modified by owner thread.
	152	*/
	153
804b4375	154	for (i = queue->tail; i != head;) {
786ee85b	155	smp_rmb(); /* read head before q[]. */
804b4375 MD	156	p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
804b4375 MD	157	if (unlikely(DQ_IS_FCT_BIT(p))) {
804b4375 MD	158	DQ_CLEAR_FCT_BIT(p);
	159	queue->last_fct_out = p;
	160	p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
	161	} else if (unlikely(p == DQ_FCT_MARK)) {
804b4375 MD	162	p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
	163	queue->last_fct_out = p;
	164	p = LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
29cdb8d8	165	}
804b4375	166	fct = queue->last_fct_out;
804b4375	167	fct(p);
786ee85b MD	168	}
	169	smp_mb(); /* push tail after having used q[] */
	170	STORE_SHARED(queue->tail, i);
	171	}
	172
	173	static void _rcu_defer_barrier_thread(void)
	174	{
0d0e6c21	175	unsigned long head, num_items;
786ee85b MD	176
786ee85b MD	177	head = defer_queue.head;
0d0e6c21 MD	178	num_items = head - defer_queue.tail;
	179	if (unlikely(!num_items))
	180	return;
786ee85b MD	181	synchronize_rcu();
	182	rcu_defer_barrier_queue(&defer_queue, head);
	183	}
	184
786ee85b MD	185	void rcu_defer_barrier_thread(void)
	186	{
	187	internal_urcu_lock(&urcu_defer_mutex);
	188	_rcu_defer_barrier_thread();
	189	internal_urcu_unlock(&urcu_defer_mutex);
	190	}
	191
0d0e6c21 MD	192	/*
	193	* rcu_defer_barrier - Execute all queued rcu callbacks.
	194	*
	195	* Execute all RCU callbacks queued before rcu_defer_barrier() execution.
	196	* All callbacks queued on the local thread prior to a rcu_defer_barrier() call
	197	* are guaranteed to be executed.
	198	* Callbacks queued by other threads concurrently with rcu_defer_barrier()
	199	* execution are not guaranteed to be executed in the current batch (could
	200	* be left for the next batch). These callbacks queued by other threads are only
	201	* guaranteed to be executed if there is explicit synchronization between
	202	* the thread adding to the queue and the thread issuing the defer_barrier call.
	203	*/
	204
786ee85b MD	205	void rcu_defer_barrier(void)
786ee85b MD	206	{
dbc6128f	207	struct defer_queue *index;
0d0e6c21	208	unsigned long num_items = 0;
786ee85b	209
dbc6128f	210	if (list_empty(&registry))
786ee85b MD	211	return;
	212
	213	internal_urcu_lock(&urcu_defer_mutex);
dbc6128f MD	214	list_for_each_entry(index, &registry, list) {
	215	index->last_head = LOAD_SHARED(index->head);
	216	num_items += index->last_head - index->tail;
0d0e6c21 MD	217	}
	218	if (likely(!num_items)) {
	219	/*
	220	* We skip the grace period because there are no queued
	221	* callbacks to execute.
	222	*/
	223	goto end;
	224	}
786ee85b	225	synchronize_rcu();
dbc6128f MD	226	list_for_each_entry(index, &registry, list)
dbc6128f MD	227	rcu_defer_barrier_queue(index, index->last_head);
0d0e6c21	228	end:
786ee85b MD	229	internal_urcu_unlock(&urcu_defer_mutex);
	230	}
	231
2c22932b	232	/*
b4f313b7	233	* _defer_rcu - Queue a RCU callback.
2c22932b	234	*/
ec8e44cf	235	void _defer_rcu_ratelimit(void (fct)(void p), void p, int (rl)(void *p))
2c22932b MD	236	{
2c22932b MD	237	unsigned long head, tail;
ec8e44cf MD	238	int sync;
	239
	240	/*
	241	* Verify if we reached the rate limiter threshold.
	242	*/
	243	sync = rl ? rl(p) : 0;
2c22932b MD	244
	245	/*
	246	* Head is only modified by ourself. Tail can be modified by reclamation
	247	* thread.
	248	*/
	249	head = defer_queue.head;
	250	tail = LOAD_SHARED(defer_queue.tail);
	251
	252	/*
ec8e44cf	253	* If queue is full, or reached threshold. Empty queue ourself.
2c22932b MD	254	* Worse-case: must allow 2 supplementary entries for fct pointer.
2c22932b MD	255	*/
ec8e44cf	256	if (unlikely(sync \|\| (head - tail >= DEFER_QUEUE_SIZE - 2))) {
2c22932b MD	257	assert(head - tail <= DEFER_QUEUE_SIZE);
	258	rcu_defer_barrier_thread();
	259	assert(head - LOAD_SHARED(defer_queue.tail) == 0);
	260	}
	261
	262	if (unlikely(defer_queue.last_fct_in != fct)) {
	263	defer_queue.last_fct_in = fct;
	264	if (unlikely(DQ_IS_FCT_BIT(fct) \|\| fct == DQ_FCT_MARK)) {
	265	/*
	266	* If the function to encode is not aligned or the
	267	* marker, write DQ_FCT_MARK followed by the function
	268	* pointer.
	269	*/
	270	_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
	271	DQ_FCT_MARK);
	272	_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
	273	fct);
	274	} else {
	275	DQ_SET_FCT_BIT(fct);
	276	_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
	277	fct);
	278	}
	279	} else {
	280	if (unlikely(DQ_IS_FCT_BIT(p) \|\| p == DQ_FCT_MARK)) {
	281	/*
	282	* If the data to encode is not aligned or the marker,
	283	* write DQ_FCT_MARK followed by the function pointer.
	284	*/
	285	_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
	286	DQ_FCT_MARK);
	287	_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
	288	fct);
	289	}
	290	}
	291	_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], p);
	292	smp_wmb(); /* Publish new pointer before head */
	293	/* Write q[] before head. */
	294	STORE_SHARED(defer_queue.head, head);
04eb9c4f	295	smp_mb(); /* Write queue head before read futex */
2c22932b MD	296	/*
	297	* Wake-up any waiting defer thread.
	298	*/
	299	wake_up_defer();
	300	}
	301
786ee85b MD	302	void thr_defer(void args)
	303	{
	304	for (;;) {
4ce9e4f2	305	pthread_testcancel();
4ce9e4f2 MD	306	/*
	307	* "Be green". Don't wake up the CPU if there is no RCU work
	308	* to perform whatsoever. Aims at saving laptop battery life by
	309	* leaving the processor in sleep state when idle.
	310	*/
4ce9e4f2	311	wait_defer();
4ce9e4f2	312	/* Sleeping after wait_defer to let many callbacks enqueue */
71df5ef4	313	poll(NULL,0,100); /* wait for 100ms */
786ee85b MD	314	rcu_defer_barrier();
	315	}
	316
	317	return NULL;
	318	}
	319
	320	/*
	321	* library wrappers to be used by non-LGPL compatible source code.
	322	*/
	323
ec8e44cf	324	void defer_rcu_ratelimit(void (fct)(void p), void p, int (rl)(void *p))
786ee85b	325	{
ec8e44cf	326	_defer_rcu_ratelimit(fct, p, rl);
786ee85b MD	327	}
786ee85b MD	328
786ee85b MD	329	static void start_defer_thread(void)
	330	{
	331	int ret;
	332
dbc6128f	333	ret = pthread_create(&tid_defer, NULL, thr_defer, NULL);
786ee85b MD	334	assert(!ret);
	335	}
	336
	337	static void stop_defer_thread(void)
	338	{
	339	int ret;
	340	void *tret;
	341
4ce9e4f2 MD	342	pthread_cancel(tid_defer);
4ce9e4f2 MD	343	wake_up_defer();
786ee85b MD	344	ret = pthread_join(tid_defer, &tret);
	345	assert(!ret);
	346	}
	347
	348	void rcu_defer_register_thread(void)
	349	{
dbc6128f MD	350	int was_empty;
	351
	352	assert(defer_queue.last_head == 0);
	353	assert(defer_queue.q == NULL);
	354	defer_queue.q = malloc(sizeof(void ) DEFER_QUEUE_SIZE);
786ee85b MD	355
	356	internal_urcu_lock(&defer_thread_mutex);
	357	internal_urcu_lock(&urcu_defer_mutex);
dbc6128f MD	358	was_empty = list_empty(&registry);
dbc6128f MD	359	list_add(&defer_queue.list, &registry);
786ee85b MD	360	internal_urcu_unlock(&urcu_defer_mutex);
786ee85b MD	361
dbc6128f	362	if (was_empty)
786ee85b MD	363	start_defer_thread();
	364	internal_urcu_unlock(&defer_thread_mutex);
	365	}
	366
	367	void rcu_defer_unregister_thread(void)
	368	{
dbc6128f	369	int is_empty;
786ee85b MD	370
	371	internal_urcu_lock(&defer_thread_mutex);
	372	internal_urcu_lock(&urcu_defer_mutex);
dbc6128f	373	list_del(&defer_queue.list);
786ee85b MD	374	_rcu_defer_barrier_thread();
	375	free(defer_queue.q);
	376	defer_queue.q = NULL;
dbc6128f	377	is_empty = list_empty(&registry);
786ee85b MD	378	internal_urcu_unlock(&urcu_defer_mutex);
786ee85b MD	379
dbc6128f	380	if (is_empty)
786ee85b MD	381	stop_defer_thread();
	382	internal_urcu_unlock(&defer_thread_mutex);
	383	}
	384
	385	void urcu_defer_exit(void)
	386	{
dbc6128f	387	assert(list_empty(&registry));
786ee85b	388	}