[urcu.git] / urcu.c

/*
 * urcu.c
 *
 * Userspace RCU library
 *
 * Copyright February 2009 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
 *
 * Distributed under GPLv2
 */

#include <stdio.h>
#include <pthread.h>
#include <signal.h>
#include <assert.h>
#include <stdlib.h>
#include <string.h>
#include <errno.h>

#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
static void force_mb_single_thread(struct reader_registry *index)
{
	smp_mb();
}

static void force_mb_all_threads(void)
{
	smp_mb();
}
#else

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 */
}

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

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++) {
		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();
			}
		}
	}
}

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();
}

void urcu_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.
 */
void urcu_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 urcu_register_thread(void)
{
	internal_urcu_lock();
	urcu_add_reader(pthread_self());
	internal_urcu_unlock();
}

void urcu_unregister_thread(void)
{
	internal_urcu_lock();
	urcu_remove_reader(pthread_self());
	internal_urcu_unlock();
}

#ifndef DEBUG_FULL_MB
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
Commit	Line	Data
	1	/*
	2	* urcu.c
	3	*
	4	* Userspace RCU library
	5	*
	6	* Copyright February 2009 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
	7	*
	8	* Distributed under GPLv2
	9	*/
	10
	11	#include <stdio.h>
	12	#include <pthread.h>
	13	#include <signal.h>
	14	#include <assert.h>
	15	#include <stdlib.h>
	16	#include <string.h>
	17	#include <errno.h>
	18
	19	#include "urcu.h"
	20
	21	pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
	22
	23	/*
	24	* Global grace period counter.
	25	* Contains the current RCU_GP_CTR_BIT.
	26	* Also has a RCU_GP_CTR_BIT of 1, to accelerate the reader fast path.
	27	* Written to only by writer with mutex taken. Read by both writer and readers.
	28	*/
	29	long urcu_gp_ctr = RCU_GP_COUNT;
	30
	31	/*
	32	* Written to only by each individual reader. Read by both the reader and the
	33	* writers.
	34	*/
	35	long __thread urcu_active_readers;
	36
	37	/* Thread IDs of registered readers */
	38	#define INIT_NUM_THREADS 4
	39
	40	struct reader_registry {
	41	pthread_t tid;
	42	long *urcu_active_readers;
	43	char *need_mb;
	44	};
	45
	46	#ifdef DEBUG_YIELD
	47	unsigned int yield_active;
	48	unsigned int __thread rand_yield;
	49	#endif
	50
	51	static struct reader_registry *registry;
	52	static char __thread need_mb;
	53	static int num_readers, alloc_readers;
	54
	55	void internal_urcu_lock(void)
	56	{
	57	int ret;
	58
	59	#ifndef DISTRUST_SIGNALS_EXTREME
	60	ret = pthread_mutex_lock(&urcu_mutex);
	61	if (ret) {
	62	perror("Error in pthread mutex lock");
	63	exit(-1);
	64	}
	65	#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
	66	while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) {
	67	if (ret != EBUSY && ret != EINTR) {
	68	printf("ret = %d, errno = %d\n", ret, errno);
	69	perror("Error in pthread mutex lock");
	70	exit(-1);
	71	}
	72	if (need_mb) {
	73	smp_mb();
	74	need_mb = 0;
	75	smp_mb();
	76	}
	77	poll(NULL,0,10);
	78	}
	79	#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
	80	}
	81
	82	void internal_urcu_unlock(void)
	83	{
	84	int ret;
	85
	86	ret = pthread_mutex_unlock(&urcu_mutex);
	87	if (ret) {
	88	perror("Error in pthread mutex unlock");
	89	exit(-1);
	90	}
	91	}
	92
	93	/*
	94	* called with urcu_mutex held.
	95	*/
	96	static void switch_next_urcu_qparity(void)
	97	{
	98	STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
	99	}
	100
	101	#ifdef DEBUG_FULL_MB
	102	static void force_mb_single_thread(struct reader_registry *index)
	103	{
	104	smp_mb();
	105	}
	106
	107	static void force_mb_all_threads(void)
	108	{
	109	smp_mb();
	110	}
	111	#else
	112
	113	static void force_mb_single_thread(struct reader_registry *index)
	114	{
	115	assert(registry);
	116	/*
	117	* pthread_kill has a smp_mb(). But beware, we assume it performs
	118	* a cache flush on architectures with non-coherent cache. Let's play
	119	* safe and don't assume anything : we use smp_mc() to make sure the
	120	* cache flush is enforced.
	121	*/
	122	*index->need_mb = 1;
	123	smp_mc(); /* write ->need_mb before sending the signals */
	124	pthread_kill(index->tid, SIGURCU);
	125	smp_mb();
	126	/*
	127	* Wait for sighandler (and thus mb()) to execute on every thread.
	128	* BUSY-LOOP.
	129	*/
	130	while (*index->need_mb) {
	131	poll(NULL, 0, 1);
	132	}
	133	smp_mb(); /* read ->need_mb before ending the barrier */
	134	}
	135
	136	static void force_mb_all_threads(void)
	137	{
	138	struct reader_registry *index;
	139	/*
	140	* Ask for each threads to execute a smp_mb() so we can consider the
	141	* compiler barriers around rcu read lock as real memory barriers.
	142	*/
	143	if (!registry)
	144	return;
	145	/*
	146	* pthread_kill has a smp_mb(). But beware, we assume it performs
	147	* a cache flush on architectures with non-coherent cache. Let's play
	148	* safe and don't assume anything : we use smp_mc() to make sure the
	149	* cache flush is enforced.
	150	*/
	151	for (index = registry; index < registry + num_readers; index++) {
	152	*index->need_mb = 1;
	153	smp_mc(); /* write need_mb before sending the signal */
	154	pthread_kill(index->tid, SIGURCU);
	155	}
	156	/*
	157	* Wait for sighandler (and thus mb()) to execute on every thread.
	158	*
	159	* Note that the pthread_kill() will never be executed on systems
	160	* that correctly deliver signals in a timely manner. However, it
	161	* is not uncommon for kernels to have bugs that can result in
	162	* lost or unduly delayed signals.
	163	*
	164	* If you are seeing the below pthread_kill() executing much at
	165	* all, we suggest testing the underlying kernel and filing the
	166	* relevant bug report. For Linux kernels, we recommend getting
	167	* the Linux Test Project (LTP).
	168	*/
	169	for (index = registry; index < registry + num_readers; index++) {
	170	while (*index->need_mb) {
	171	pthread_kill(index->tid, SIGURCU);
	172	poll(NULL, 0, 1);
	173	}
	174	}
	175	smp_mb(); /* read ->need_mb before ending the barrier */
	176	}
	177	#endif
	178
	179	void wait_for_quiescent_state(void)
	180	{
	181	struct reader_registry *index;
	182
	183	if (!registry)
	184	return;
	185	/*
	186	* Wait for each thread urcu_active_readers count to become 0.
	187	*/
	188	for (index = registry; index < registry + num_readers; index++) {
	189	int wait_loops = 0;
	190	/*
	191	* BUSY-LOOP. Force the reader thread to commit its
	192	* urcu_active_readers update to memory if we wait for too long.
	193	*/
	194	while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
	195	if (wait_loops++ == KICK_READER_LOOPS) {
	196	force_mb_single_thread(index);
	197	wait_loops = 0;
	198	} else {
	199	cpu_relax();
	200	}
	201	}
	202	}
	203	}
	204
	205	void synchronize_rcu(void)
	206	{
	207	internal_urcu_lock();
	208
	209	/* All threads should read qparity before accessing data structure
	210	* where new ptr points to. Must be done within internal_urcu_lock
	211	* because it iterates on reader threads.*/
	212	/* Write new ptr before changing the qparity */
	213	force_mb_all_threads();
	214
	215	switch_next_urcu_qparity(); /* 0 -> 1 */
	216
	217	/*
	218	* Must commit qparity update to memory before waiting for parity
	219	* 0 quiescent state. Failure to do so could result in the writer
	220	* waiting forever while new readers are always accessing data (no
	221	* progress).
	222	* Ensured by STORE_SHARED and LOAD_SHARED.
	223	*/
	224
	225	/*
	226	* Wait for previous parity to be empty of readers.
	227	*/
	228	wait_for_quiescent_state(); /* Wait readers in parity 0 */
	229
	230	/*
	231	* Must finish waiting for quiescent state for parity 0 before
	232	* committing qparity update to memory. Failure to do so could result in
	233	* the writer waiting forever while new readers are always accessing
	234	* data (no progress).
	235	* Ensured by STORE_SHARED and LOAD_SHARED.
	236	*/
	237
	238	switch_next_urcu_qparity(); /* 1 -> 0 */
	239
	240	/*
	241	* Must commit qparity update to memory before waiting for parity
	242	* 1 quiescent state. Failure to do so could result in the writer
	243	* waiting forever while new readers are always accessing data (no
	244	* progress).
	245	* Ensured by STORE_SHARED and LOAD_SHARED.
	246	*/
	247
	248	/*
	249	* Wait for previous parity to be empty of readers.
	250	*/
	251	wait_for_quiescent_state(); /* Wait readers in parity 1 */
	252
	253	/* Finish waiting for reader threads before letting the old ptr being
	254	* freed. Must be done within internal_urcu_lock because it iterates on
	255	* reader threads. */
	256	force_mb_all_threads();
	257
	258	internal_urcu_unlock();
	259	}
	260
	261	void urcu_add_reader(pthread_t id)
	262	{
	263	struct reader_registry *oldarray;
	264
	265	if (!registry) {
	266	alloc_readers = INIT_NUM_THREADS;
	267	num_readers = 0;
	268	registry =
	269	malloc(sizeof(struct reader_registry) * alloc_readers);
	270	}
	271	if (alloc_readers < num_readers + 1) {
	272	oldarray = registry;
	273	registry = malloc(sizeof(struct reader_registry)
	274	* (alloc_readers << 1));
	275	memcpy(registry, oldarray,
	276	sizeof(struct reader_registry) * alloc_readers);
	277	alloc_readers <<= 1;
	278	free(oldarray);
	279	}
	280	registry[num_readers].tid = id;
	281	/* reference to the TLS of _this_ reader thread. */
	282	registry[num_readers].urcu_active_readers = &urcu_active_readers;
	283	registry[num_readers].need_mb = &need_mb;
	284	num_readers++;
	285	}
	286
	287	/*
	288	* Never shrink (implementation limitation).
	289	* This is O(nb threads). Eventually use a hash table.
	290	*/
	291	void urcu_remove_reader(pthread_t id)
	292	{
	293	struct reader_registry *index;
	294
	295	assert(registry != NULL);
	296	for (index = registry; index < registry + num_readers; index++) {
	297	if (pthread_equal(index->tid, id)) {
	298	memcpy(index, &registry[num_readers - 1],
	299	sizeof(struct reader_registry));
	300	registry[num_readers - 1].tid = 0;
	301	registry[num_readers - 1].urcu_active_readers = NULL;
	302	num_readers--;
	303	return;
	304	}
	305	}
	306	/* Hrm not found, forgot to register ? */
	307	assert(0);
	308	}
	309
	310	void urcu_register_thread(void)
	311	{
	312	internal_urcu_lock();
	313	urcu_add_reader(pthread_self());
	314	internal_urcu_unlock();
	315	}
	316
	317	void urcu_unregister_thread(void)
	318	{
	319	internal_urcu_lock();
	320	urcu_remove_reader(pthread_self());
	321	internal_urcu_unlock();
	322	}
	323
	324	#ifndef DEBUG_FULL_MB
	325	void sigurcu_handler(int signo, siginfo_t siginfo, void context)
	326	{
	327	/*
	328	* Executing this smp_mb() is the only purpose of this signal handler.
	329	* It punctually promotes barrier() into smp_mb() on every thread it is
	330	* executed on.
	331	*/
	332	smp_mb();
	333	need_mb = 0;
	334	smp_mb();
	335	}
	336
	337	void __attribute__((constructor)) urcu_init(void)
	338	{
	339	struct sigaction act;
	340	int ret;
	341
	342	act.sa_sigaction = sigurcu_handler;
	343	ret = sigaction(SIGURCU, &act, NULL);
	344	if (ret) {
	345	perror("Error in sigaction");
	346	exit(-1);
	347	}
	348	}
	349
	350	void __attribute__((destructor)) urcu_exit(void)
	351	{
	352	struct sigaction act;
	353	int ret;
	354
	355	ret = sigaction(SIGURCU, NULL, &act);
	356	if (ret) {
	357	perror("Error in sigaction");
	358	exit(-1);
	359	}
	360	assert(act.sa_sigaction == sigurcu_handler);
	361	free(registry);
	362	}
	363	#endif