[urcu.git] / urcu.c

/*
 * urcu.c
 *
 * Userspace RCU library
 *
 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
 * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
 *
 * 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
 *
 * IBM's contributions to this file may be relicensed under LGPLv2 or later.
 */

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

pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;

/*
 * Global grace period counter.
 * Contains the current RCU_GP_CTR_BIT.
 * Also has a RCU_GP_CTR_BIT of 1, to accelerate the reader fast path.
 * Written to only by writer with mutex taken. Read by both writer and readers.
 */
long urcu_gp_ctr = RCU_GP_COUNT;

/*
 * Written to only by each individual reader. Read by both the reader and the
 * writers.
 */
long __thread urcu_active_readers;

/* Thread IDs of registered readers */
#define INIT_NUM_THREADS 4

struct reader_registry {
	pthread_t tid;
	long *urcu_active_readers;
	char *need_mb;
};

#ifdef DEBUG_YIELD
unsigned int yield_active;
unsigned int __thread rand_yield;
#endif

static struct reader_registry *registry;
static char __thread need_mb;
static int num_readers, alloc_readers;

void internal_urcu_lock(void)
{
	int ret;

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

void internal_urcu_unlock(void)
{
	int ret;

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

/*
 * called with urcu_mutex held.
 */
static void switch_next_urcu_qparity(void)
{
	STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
}

#ifdef DEBUG_FULL_MB
#ifdef HAS_INCOHERENT_CACHES
static void force_mb_single_thread(struct reader_registry *index)
{
	smp_mb();
}
#endif /* #ifdef HAS_INCOHERENT_CACHES */

static void force_mb_all_threads(void)
{
	smp_mb();
}
#else /* #ifdef DEBUG_FULL_MB */
#ifdef HAS_INCOHERENT_CACHES
static void force_mb_single_thread(struct reader_registry *index)
{
	assert(registry);
	/*
	 * pthread_kill has a smp_mb(). But beware, we assume it performs
	 * a cache flush on architectures with non-coherent cache. Let's play
	 * safe and don't assume anything : we use smp_mc() to make sure the
	 * cache flush is enforced.
	 */
	*index->need_mb = 1;
	smp_mc();	/* write ->need_mb before sending the signals */
	pthread_kill(index->tid, SIGURCU);
	smp_mb();
	/*
	 * Wait for sighandler (and thus mb()) to execute on every thread.
	 * BUSY-LOOP.
	 */
	while (*index->need_mb) {
		poll(NULL, 0, 1);
	}
	smp_mb();	/* read ->need_mb before ending the barrier */
}
#endif /* #ifdef HAS_INCOHERENT_CACHES */

static void force_mb_all_threads(void)
{
	struct reader_registry *index;
	/*
	 * Ask for each threads to execute a smp_mb() so we can consider the
	 * compiler barriers around rcu read lock as real memory barriers.
	 */
	if (!registry)
		return;
	/*
	 * pthread_kill has a smp_mb(). But beware, we assume it performs
	 * a cache flush on architectures with non-coherent cache. Let's play
	 * safe and don't assume anything : we use smp_mc() to make sure the
	 * cache flush is enforced.
	 */
	for (index = registry; index < registry + num_readers; index++) {
		*index->need_mb = 1;
		smp_mc();	/* write need_mb before sending the signal */
		pthread_kill(index->tid, SIGURCU);
	}
	/*
	 * Wait for sighandler (and thus mb()) to execute on every thread.
	 *
	 * Note that the pthread_kill() will never be executed on systems
	 * that correctly deliver signals in a timely manner.  However, it
	 * is not uncommon for kernels to have bugs that can result in
	 * lost or unduly delayed signals.
	 *
	 * If you are seeing the below pthread_kill() executing much at
	 * all, we suggest testing the underlying kernel and filing the
	 * relevant bug report.  For Linux kernels, we recommend getting
	 * the Linux Test Project (LTP).
	 */
	for (index = registry; index < registry + num_readers; index++) {
		while (*index->need_mb) {
			pthread_kill(index->tid, SIGURCU);
			poll(NULL, 0, 1);
		}
	}
	smp_mb();	/* read ->need_mb before ending the barrier */
}
#endif /* #else #ifdef DEBUG_FULL_MB */

void wait_for_quiescent_state(void)
{
	struct reader_registry *index;

	if (!registry)
		return;
	/*
	 * Wait for each thread urcu_active_readers count to become 0.
	 */
	for (index = registry; index < registry + num_readers; index++) {
#ifndef HAS_INCOHERENT_CACHES
		while (rcu_old_gp_ongoing(index->urcu_active_readers))
			cpu_relax();
#else /* #ifndef HAS_INCOHERENT_CACHES */
		int wait_loops = 0;
		/*
		 * BUSY-LOOP. Force the reader thread to commit its
		 * urcu_active_readers update to memory if we wait for too long.
		 */
		while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
			if (wait_loops++ == KICK_READER_LOOPS) {
				force_mb_single_thread(index);
				wait_loops = 0;
			} else {
				cpu_relax();
			}
		}
#endif /* #else #ifndef HAS_INCOHERENT_CACHES */
	}
}

void synchronize_rcu(void)
{
	internal_urcu_lock();

	/* All threads should read qparity before accessing data structure
	 * where new ptr points to. Must be done within internal_urcu_lock
	 * because it iterates on reader threads.*/
	/* Write new ptr before changing the qparity */
	force_mb_all_threads();

	switch_next_urcu_qparity();	/* 0 -> 1 */

	/*
	 * Must commit qparity update to memory before waiting for parity
	 * 0 quiescent state. Failure to do so could result in the writer
	 * waiting forever while new readers are always accessing data (no
	 * progress).
	 * Ensured by STORE_SHARED and LOAD_SHARED.
	 */

	/*
	 * Wait for previous parity to be empty of readers.
	 */
	wait_for_quiescent_state();	/* Wait readers in parity 0 */

	/*
	 * Must finish waiting for quiescent state for parity 0 before
	 * committing qparity update to memory. Failure to do so could result in
	 * the writer waiting forever while new readers are always accessing
	 * data (no progress).
	 * Ensured by STORE_SHARED and LOAD_SHARED.
	 */

	switch_next_urcu_qparity();	/* 1 -> 0 */

	/*
	 * Must commit qparity update to memory before waiting for parity
	 * 1 quiescent state. Failure to do so could result in the writer
	 * waiting forever while new readers are always accessing data (no
	 * progress).
	 * Ensured by STORE_SHARED and LOAD_SHARED.
	 */

	/*
	 * Wait for previous parity to be empty of readers.
	 */
	wait_for_quiescent_state();	/* Wait readers in parity 1 */

	/* Finish waiting for reader threads before letting the old ptr being
	 * freed. Must be done within internal_urcu_lock because it iterates on
	 * reader threads. */
	force_mb_all_threads();

	internal_urcu_unlock();
}

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

void rcu_read_lock(void)
{
	_rcu_read_lock();
}

void rcu_read_unlock(void)
{
	_rcu_read_unlock();
}

void *rcu_dereference(void *p)
{
	return _rcu_dereference(p);
}

void *rcu_assign_pointer_sym(void **p, void *v)
{
	wmb();
	return STORE_SHARED(p, v);
}

void *rcu_xchg_pointer_sym(void **p, void *v)
{
	wmb();
	return xchg(p, v);
}

void *rcu_publish_content_sym(void **p, void *v)
{
	void *oldptr;

	oldptr = _rcu_xchg_pointer(p, v);
	synchronize_rcu();
	return oldptr;
}

static void rcu_add_reader(pthread_t id)
{
	struct reader_registry *oldarray;

	if (!registry) {
		alloc_readers = INIT_NUM_THREADS;
		num_readers = 0;
		registry =
			malloc(sizeof(struct reader_registry) * alloc_readers);
	}
	if (alloc_readers < num_readers + 1) {
		oldarray = registry;
		registry = malloc(sizeof(struct reader_registry)
				* (alloc_readers << 1));
		memcpy(registry, oldarray,
			sizeof(struct reader_registry) * alloc_readers);
		alloc_readers <<= 1;
		free(oldarray);
	}
	registry[num_readers].tid = id;
	/* reference to the TLS of _this_ reader thread. */
	registry[num_readers].urcu_active_readers = &urcu_active_readers;
	registry[num_readers].need_mb = &need_mb;
	num_readers++;
}

/*
 * Never shrink (implementation limitation).
 * This is O(nb threads). Eventually use a hash table.
 */
static void rcu_remove_reader(pthread_t id)
{
	struct reader_registry *index;

	assert(registry != NULL);
	for (index = registry; index < registry + num_readers; index++) {
		if (pthread_equal(index->tid, id)) {
			memcpy(index, &registry[num_readers - 1],
				sizeof(struct reader_registry));
			registry[num_readers - 1].tid = 0;
			registry[num_readers - 1].urcu_active_readers = NULL;
			num_readers--;
			return;
		}
	}
	/* Hrm not found, forgot to register ? */
	assert(0);
}

void rcu_register_thread(void)
{
	internal_urcu_lock();
	rcu_add_reader(pthread_self());
	internal_urcu_unlock();
}

void rcu_unregister_thread(void)
{
	internal_urcu_lock();
	rcu_remove_reader(pthread_self());
	internal_urcu_unlock();
}

#ifndef DEBUG_FULL_MB
static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
{
	/*
	 * Executing this smp_mb() is the only purpose of this signal handler.
	 * It punctually promotes barrier() into smp_mb() on every thread it is
	 * executed on.
	 */
	smp_mb();
	need_mb = 0;
	smp_mb();
}

void __attribute__((constructor)) urcu_init(void)
{
	struct sigaction act;
	int ret;

	act.sa_sigaction = sigurcu_handler;
	ret = sigaction(SIGURCU, &act, NULL);
	if (ret) {
		perror("Error in sigaction");
		exit(-1);
	}
}

void __attribute__((destructor)) urcu_exit(void)
{
	struct sigaction act;
	int ret;

	ret = sigaction(SIGURCU, NULL, &act);
	if (ret) {
		perror("Error in sigaction");
		exit(-1);
	}
	assert(act.sa_sigaction == sigurcu_handler);
	free(registry);
}
#endif /* #ifndef DEBUG_FULL_MB */
Commit	Line	Data
	1	/*
	2	* urcu.c
	3	*
	4	* Userspace RCU library
	5	*
	6	* Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
	7	* Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
	8	*
	9	* This library is free software; you can redistribute it and/or
	10	* modify it under the terms of the GNU Lesser General Public
	11	* License as published by the Free Software Foundation; either
	12	* version 2.1 of the License, or (at your option) any later version.
	13	*
	14	* This library is distributed in the hope that it will be useful,
	15	* but WITHOUT ANY WARRANTY; without even the implied warranty of
	16	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
	17	* Lesser General Public License for more details.
	18	*
	19	* You should have received a copy of the GNU Lesser General Public
	20	* License along with this library; if not, write to the Free Software
	21	* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
	22	*
	23	* IBM's contributions to this file may be relicensed under LGPLv2 or later.
	24	*/
	25
	26	#include <stdio.h>
	27	#include <pthread.h>
	28	#include <signal.h>
	29	#include <assert.h>
	30	#include <stdlib.h>
	31	#include <string.h>
	32	#include <errno.h>
	33	#include <poll.h>
	34
	35	#include "urcu-static.h"
	36	/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
	37	#include "urcu.h"
	38
	39	pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
	40
	41	/*
	42	* Global grace period counter.
	43	* Contains the current RCU_GP_CTR_BIT.
	44	* Also has a RCU_GP_CTR_BIT of 1, to accelerate the reader fast path.
	45	* Written to only by writer with mutex taken. Read by both writer and readers.
	46	*/
	47	long urcu_gp_ctr = RCU_GP_COUNT;
	48
	49	/*
	50	* Written to only by each individual reader. Read by both the reader and the
	51	* writers.
	52	*/
	53	long __thread urcu_active_readers;
	54
	55	/* Thread IDs of registered readers */
	56	#define INIT_NUM_THREADS 4
	57
	58	struct reader_registry {
	59	pthread_t tid;
	60	long *urcu_active_readers;
	61	char *need_mb;
	62	};
	63
	64	#ifdef DEBUG_YIELD
	65	unsigned int yield_active;
	66	unsigned int __thread rand_yield;
	67	#endif
	68
	69	static struct reader_registry *registry;
	70	static char __thread need_mb;
	71	static int num_readers, alloc_readers;
	72
	73	void internal_urcu_lock(void)
	74	{
	75	int ret;
	76
	77	#ifndef DISTRUST_SIGNALS_EXTREME
	78	ret = pthread_mutex_lock(&urcu_mutex);
	79	if (ret) {
	80	perror("Error in pthread mutex lock");
	81	exit(-1);
	82	}
	83	#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
	84	while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) {
	85	if (ret != EBUSY && ret != EINTR) {
	86	printf("ret = %d, errno = %d\n", ret, errno);
	87	perror("Error in pthread mutex lock");
	88	exit(-1);
	89	}
	90	if (need_mb) {
	91	smp_mb();
	92	need_mb = 0;
	93	smp_mb();
	94	}
	95	poll(NULL,0,10);
	96	}
	97	#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
	98	}
	99
	100	void internal_urcu_unlock(void)
	101	{
	102	int ret;
	103
	104	ret = pthread_mutex_unlock(&urcu_mutex);
	105	if (ret) {
	106	perror("Error in pthread mutex unlock");
	107	exit(-1);
	108	}
	109	}
	110
	111	/*
	112	* called with urcu_mutex held.
	113	*/
	114	static void switch_next_urcu_qparity(void)
	115	{
	116	STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
	117	}
	118
	119	#ifdef DEBUG_FULL_MB
	120	#ifdef HAS_INCOHERENT_CACHES
	121	static void force_mb_single_thread(struct reader_registry *index)
	122	{
	123	smp_mb();
	124	}
	125	#endif /* #ifdef HAS_INCOHERENT_CACHES */
	126
	127	static void force_mb_all_threads(void)
	128	{
	129	smp_mb();
	130	}
	131	#else /* #ifdef DEBUG_FULL_MB */
	132	#ifdef HAS_INCOHERENT_CACHES
	133	static void force_mb_single_thread(struct reader_registry *index)
	134	{
	135	assert(registry);
	136	/*
	137	* pthread_kill has a smp_mb(). But beware, we assume it performs
	138	* a cache flush on architectures with non-coherent cache. Let's play
	139	* safe and don't assume anything : we use smp_mc() to make sure the
	140	* cache flush is enforced.
	141	*/
	142	*index->need_mb = 1;
	143	smp_mc(); /* write ->need_mb before sending the signals */
	144	pthread_kill(index->tid, SIGURCU);
	145	smp_mb();
	146	/*
	147	* Wait for sighandler (and thus mb()) to execute on every thread.
	148	* BUSY-LOOP.
	149	*/
	150	while (*index->need_mb) {
	151	poll(NULL, 0, 1);
	152	}
	153	smp_mb(); /* read ->need_mb before ending the barrier */
	154	}
	155	#endif /* #ifdef HAS_INCOHERENT_CACHES */
	156
	157	static void force_mb_all_threads(void)
	158	{
	159	struct reader_registry *index;
	160	/*
	161	* Ask for each threads to execute a smp_mb() so we can consider the
	162	* compiler barriers around rcu read lock as real memory barriers.
	163	*/
	164	if (!registry)
	165	return;
	166	/*
	167	* pthread_kill has a smp_mb(). But beware, we assume it performs
	168	* a cache flush on architectures with non-coherent cache. Let's play
	169	* safe and don't assume anything : we use smp_mc() to make sure the
	170	* cache flush is enforced.
	171	*/
	172	for (index = registry; index < registry + num_readers; index++) {
	173	*index->need_mb = 1;
	174	smp_mc(); /* write need_mb before sending the signal */
	175	pthread_kill(index->tid, SIGURCU);
	176	}
	177	/*
	178	* Wait for sighandler (and thus mb()) to execute on every thread.
	179	*
	180	* Note that the pthread_kill() will never be executed on systems
	181	* that correctly deliver signals in a timely manner. However, it
	182	* is not uncommon for kernels to have bugs that can result in
	183	* lost or unduly delayed signals.
	184	*
	185	* If you are seeing the below pthread_kill() executing much at
	186	* all, we suggest testing the underlying kernel and filing the
	187	* relevant bug report. For Linux kernels, we recommend getting
	188	* the Linux Test Project (LTP).
	189	*/
	190	for (index = registry; index < registry + num_readers; index++) {
	191	while (*index->need_mb) {
	192	pthread_kill(index->tid, SIGURCU);
	193	poll(NULL, 0, 1);
	194	}
	195	}
	196	smp_mb(); /* read ->need_mb before ending the barrier */
	197	}
	198	#endif /* #else #ifdef DEBUG_FULL_MB */
	199
	200	void wait_for_quiescent_state(void)
	201	{
	202	struct reader_registry *index;
	203
	204	if (!registry)
	205	return;
	206	/*
	207	* Wait for each thread urcu_active_readers count to become 0.
	208	*/
	209	for (index = registry; index < registry + num_readers; index++) {
	210	#ifndef HAS_INCOHERENT_CACHES
	211	while (rcu_old_gp_ongoing(index->urcu_active_readers))
	212	cpu_relax();
	213	#else /* #ifndef HAS_INCOHERENT_CACHES */
	214	int wait_loops = 0;
	215	/*
	216	* BUSY-LOOP. Force the reader thread to commit its
	217	* urcu_active_readers update to memory if we wait for too long.
	218	*/
	219	while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
	220	if (wait_loops++ == KICK_READER_LOOPS) {
	221	force_mb_single_thread(index);
	222	wait_loops = 0;
	223	} else {
	224	cpu_relax();
	225	}
	226	}
	227	#endif /* #else #ifndef HAS_INCOHERENT_CACHES */
	228	}
	229	}
	230
	231	void synchronize_rcu(void)
	232	{
	233	internal_urcu_lock();
	234
	235	/* All threads should read qparity before accessing data structure
	236	* where new ptr points to. Must be done within internal_urcu_lock
	237	* because it iterates on reader threads.*/
	238	/* Write new ptr before changing the qparity */
	239	force_mb_all_threads();
	240
	241	switch_next_urcu_qparity(); /* 0 -> 1 */
	242
	243	/*
	244	* Must commit qparity update to memory before waiting for parity
	245	* 0 quiescent state. Failure to do so could result in the writer
	246	* waiting forever while new readers are always accessing data (no
	247	* progress).
	248	* Ensured by STORE_SHARED and LOAD_SHARED.
	249	*/
	250
	251	/*
	252	* Wait for previous parity to be empty of readers.
	253	*/
	254	wait_for_quiescent_state(); /* Wait readers in parity 0 */
	255
	256	/*
	257	* Must finish waiting for quiescent state for parity 0 before
	258	* committing qparity update to memory. Failure to do so could result in
	259	* the writer waiting forever while new readers are always accessing
	260	* data (no progress).
	261	* Ensured by STORE_SHARED and LOAD_SHARED.
	262	*/
	263
	264	switch_next_urcu_qparity(); /* 1 -> 0 */
	265
	266	/*
	267	* Must commit qparity update to memory before waiting for parity
	268	* 1 quiescent state. Failure to do so could result in the writer
	269	* waiting forever while new readers are always accessing data (no
	270	* progress).
	271	* Ensured by STORE_SHARED and LOAD_SHARED.
	272	*/
	273
	274	/*
	275	* Wait for previous parity to be empty of readers.
	276	*/
	277	wait_for_quiescent_state(); /* Wait readers in parity 1 */
	278
	279	/* Finish waiting for reader threads before letting the old ptr being
	280	* freed. Must be done within internal_urcu_lock because it iterates on
	281	* reader threads. */
	282	force_mb_all_threads();
	283
	284	internal_urcu_unlock();
	285	}
	286
	287	/*
	288	* library wrappers to be used by non-LGPL compatible source code.
	289	*/
	290
	291	void rcu_read_lock(void)
	292	{
	293	_rcu_read_lock();
	294	}
	295
	296	void rcu_read_unlock(void)
	297	{
	298	_rcu_read_unlock();
	299	}
	300
	301	void rcu_dereference(void p)
	302	{
	303	return _rcu_dereference(p);
	304	}
	305
	306	void rcu_assign_pointer_sym(void p, void v)
	307	{
	308	wmb();
	309	return STORE_SHARED(p, v);
	310	}
	311
	312	void rcu_xchg_pointer_sym(void p, void v)
	313	{
	314	wmb();
	315	return xchg(p, v);
	316	}
	317
	318	void rcu_publish_content_sym(void p, void v)
	319	{
	320	void *oldptr;
	321
	322	oldptr = _rcu_xchg_pointer(p, v);
	323	synchronize_rcu();
	324	return oldptr;
	325	}
	326
	327	static void rcu_add_reader(pthread_t id)
	328	{
	329	struct reader_registry *oldarray;
	330
	331	if (!registry) {
	332	alloc_readers = INIT_NUM_THREADS;
	333	num_readers = 0;
	334	registry =
	335	malloc(sizeof(struct reader_registry) * alloc_readers);
	336	}
	337	if (alloc_readers < num_readers + 1) {
	338	oldarray = registry;
	339	registry = malloc(sizeof(struct reader_registry)
	340	* (alloc_readers << 1));
	341	memcpy(registry, oldarray,
	342	sizeof(struct reader_registry) * alloc_readers);
	343	alloc_readers <<= 1;
	344	free(oldarray);
	345	}
	346	registry[num_readers].tid = id;
	347	/* reference to the TLS of _this_ reader thread. */
	348	registry[num_readers].urcu_active_readers = &urcu_active_readers;
	349	registry[num_readers].need_mb = &need_mb;
	350	num_readers++;
	351	}
	352
	353	/*
	354	* Never shrink (implementation limitation).
	355	* This is O(nb threads). Eventually use a hash table.
	356	*/
	357	static void rcu_remove_reader(pthread_t id)
	358	{
	359	struct reader_registry *index;
	360
	361	assert(registry != NULL);
	362	for (index = registry; index < registry + num_readers; index++) {
	363	if (pthread_equal(index->tid, id)) {
	364	memcpy(index, &registry[num_readers - 1],
	365	sizeof(struct reader_registry));
	366	registry[num_readers - 1].tid = 0;
	367	registry[num_readers - 1].urcu_active_readers = NULL;
	368	num_readers--;
	369	return;
	370	}
	371	}
	372	/* Hrm not found, forgot to register ? */
	373	assert(0);
	374	}
	375
	376	void rcu_register_thread(void)
	377	{
	378	internal_urcu_lock();
	379	rcu_add_reader(pthread_self());
	380	internal_urcu_unlock();
	381	}
	382
	383	void rcu_unregister_thread(void)
	384	{
	385	internal_urcu_lock();
	386	rcu_remove_reader(pthread_self());
	387	internal_urcu_unlock();
	388	}
	389
	390	#ifndef DEBUG_FULL_MB
	391	static void sigurcu_handler(int signo, siginfo_t siginfo, void context)
	392	{
	393	/*
	394	* Executing this smp_mb() is the only purpose of this signal handler.
	395	* It punctually promotes barrier() into smp_mb() on every thread it is
	396	* executed on.
	397	*/
	398	smp_mb();
	399	need_mb = 0;
	400	smp_mb();
	401	}
	402
	403	void __attribute__((constructor)) urcu_init(void)
	404	{
	405	struct sigaction act;
	406	int ret;
	407
	408	act.sa_sigaction = sigurcu_handler;
	409	ret = sigaction(SIGURCU, &act, NULL);
	410	if (ret) {
	411	perror("Error in sigaction");
	412	exit(-1);
	413	}
	414	}
	415
	416	void __attribute__((destructor)) urcu_exit(void)
	417	{
	418	struct sigaction act;
	419	int ret;
	420
	421	ret = sigaction(SIGURCU, NULL, &act);
	422	if (ret) {
	423	perror("Error in sigaction");
	424	exit(-1);
	425	}
	426	assert(act.sa_sigaction == sigurcu_handler);
	427	free(registry);
	428	}
	429	#endif /* #ifndef DEBUG_FULL_MB */