[urcu.git] / urcu-qsbr.c

/*
 * urcu-qsbr.c
 *
 * Userspace RCU QSBR library
 *
 * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
 * 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.
 */

#define _GNU_SOURCE
#include <stdio.h>
#include <pthread.h>
#include <signal.h>
#include <assert.h>
#include <stdlib.h>
#include <stdint.h>
#include <string.h>
#include <errno.h>
#include <poll.h>

#include "urcu/map/urcu-qsbr.h"

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

void __attribute__((destructor)) rcu_exit(void);

static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;

int32_t gp_futex;

/*
 * Global grace period counter.
 */
unsigned long rcu_gp_ctr = RCU_GP_ONLINE;

/*
 * Active attempts to check for reader Q.S. before calling futex().
 */
#define RCU_QS_ACTIVE_ATTEMPTS 100

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

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

static CDS_LIST_HEAD(registry);

static void mutex_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);
		}
		poll(NULL,0,10);
	}
#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
}

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

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

/*
 * synchronize_rcu() waiting. Single thread.
 */
static void wait_gp(void)
{
	/* Read reader_gp before read futex */
	cmm_smp_rmb();
	if (uatomic_read(&gp_futex) == -1)
		futex_noasync(&gp_futex, FUTEX_WAIT, -1,
		      NULL, NULL, 0);
}

static void update_counter_and_wait(void)
{
	CDS_LIST_HEAD(qsreaders);
	int wait_loops = 0;
	struct rcu_reader *index, *tmp;

#if (CAA_BITS_PER_LONG < 64)
	/* Switch parity: 0 -> 1, 1 -> 0 */
	CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
#else	/* !(CAA_BITS_PER_LONG < 64) */
	/* Increment current G.P. */
	CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
#endif	/* !(CAA_BITS_PER_LONG < 64) */

	/*
	 * Must commit rcu_gp_ctr update to memory before waiting for
	 * quiescent state. Failure to do so could result in the writer
	 * waiting forever while new readers are always accessing data
	 * (no progress). Enforce compiler-order of store to rcu_gp_ctr
	 * before load rcu_reader ctr.
	 */
	cmm_barrier();

	/*
	 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
	 * model easier to understand. It does not have a big performance impact
	 * anyway, given this is the write-side.
	 */
	cmm_smp_mb();

	/*
	 * Wait for each thread rcu_reader_qs_gp count to become 0.
	 */
	for (;;) {
		wait_loops++;
		if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
			uatomic_dec(&gp_futex);
			/* Write futex before read reader_gp */
			cmm_smp_mb();
		}

		cds_list_for_each_entry_safe(index, tmp, &registry, node) {
			if (!rcu_gp_ongoing(&index->ctr))
				cds_list_move(&index->node, &qsreaders);
		}

		if (cds_list_empty(&registry)) {
			if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
				/* Read reader_gp before write futex */
				cmm_smp_mb();
				uatomic_set(&gp_futex, 0);
			}
			break;
		} else {
			if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
				wait_gp();
			} else {
#ifndef HAS_INCOHERENT_CACHES
				caa_cpu_relax();
#else /* #ifndef HAS_INCOHERENT_CACHES */
				cmm_smp_mb();
#endif /* #else #ifndef HAS_INCOHERENT_CACHES */
			}
		}
	}
	/* put back the reader list in the registry */
	cds_list_splice(&qsreaders, &registry);
}

/*
 * Using a two-subphases algorithm for architectures with smaller than 64-bit
 * long-size to ensure we do not encounter an overflow bug.
 */

#if (CAA_BITS_PER_LONG < 64)
void synchronize_rcu(void)
{
	unsigned long was_online;

	was_online = rcu_reader.ctr;

	/* All threads should read qparity before accessing data structure
	 * where new ptr points to.
	 */
	/* Write new ptr before changing the qparity */
	cmm_smp_mb();

	/*
	 * Mark the writer thread offline to make sure we don't wait for
	 * our own quiescent state. This allows using synchronize_rcu()
	 * in threads registered as readers.
	 */
	if (was_online)
		CMM_STORE_SHARED(rcu_reader.ctr, 0);

	mutex_lock(&rcu_gp_lock);

	if (cds_list_empty(&registry))
		goto out;

	/*
	 * Wait for previous parity to be empty of readers.
	 */
	update_counter_and_wait();	/* 0 -> 1, wait readers in parity 0 */

	/*
	 * Must finish waiting for quiescent state for parity 0 before
	 * committing next rcu_gp_ctr update to memory. Failure to
	 * do so could result in the writer waiting forever while new
	 * readers are always accessing data (no progress).  Enforce
	 * compiler-order of load rcu_reader ctr before store to
	 * rcu_gp_ctr.
	 */
	cmm_barrier();

	/*
	 * Adding a cmm_smp_mb() which is _not_ formally required, but makes the
	 * model easier to understand. It does not have a big performance impact
	 * anyway, given this is the write-side.
	 */
	cmm_smp_mb();

	/*
	 * Wait for previous parity to be empty of readers.
	 */
	update_counter_and_wait();	/* 1 -> 0, wait readers in parity 1 */
out:
	mutex_unlock(&rcu_gp_lock);

	/*
	 * Finish waiting for reader threads before letting the old ptr being
	 * freed.
	 */
	if (was_online)
		_CMM_STORE_SHARED(rcu_reader.ctr,
				  CMM_LOAD_SHARED(rcu_gp_ctr));
	cmm_smp_mb();
}
#else /* !(CAA_BITS_PER_LONG < 64) */
void synchronize_rcu(void)
{
	unsigned long was_online;

	was_online = rcu_reader.ctr;

	/*
	 * Mark the writer thread offline to make sure we don't wait for
	 * our own quiescent state. This allows using synchronize_rcu()
	 * in threads registered as readers.
	 */
	cmm_smp_mb();
	if (was_online)
		CMM_STORE_SHARED(rcu_reader.ctr, 0);

	mutex_lock(&rcu_gp_lock);
	if (cds_list_empty(&registry))
		goto out;
	update_counter_and_wait();
out:
	mutex_unlock(&rcu_gp_lock);

	if (was_online)
		_CMM_STORE_SHARED(rcu_reader.ctr,
				  CMM_LOAD_SHARED(rcu_gp_ctr));
	cmm_smp_mb();
}
#endif  /* !(CAA_BITS_PER_LONG < 64) */

/*
 * 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_quiescent_state(void)
{
	_rcu_quiescent_state();
}

void rcu_thread_offline(void)
{
	_rcu_thread_offline();
}

void rcu_thread_online(void)
{
	_rcu_thread_online();
}

void rcu_register_thread(void)
{
	rcu_reader.tid = pthread_self();
	assert(rcu_reader.ctr == 0);

	mutex_lock(&rcu_gp_lock);
	cds_list_add(&rcu_reader.node, &registry);
	mutex_unlock(&rcu_gp_lock);
	_rcu_thread_online();
}

void rcu_unregister_thread(void)
{
	/*
	 * We have to make the thread offline otherwise we end up dealocking
	 * with a waiting writer.
	 */
	_rcu_thread_offline();
	mutex_lock(&rcu_gp_lock);
	cds_list_del(&rcu_reader.node);
	mutex_unlock(&rcu_gp_lock);
}

void rcu_exit(void)
{
	assert(cds_list_empty(&registry));
}

#include "urcu-call-rcu-impl.h"
#include "urcu-defer-impl.h"
Commit	Line	Data
	1	/*
	2	* urcu-qsbr.c
	3	*
	4	* Userspace RCU QSBR library
	5	*
	6	* Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
	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	#define _GNU_SOURCE
	27	#include <stdio.h>
	28	#include <pthread.h>
	29	#include <signal.h>
	30	#include <assert.h>
	31	#include <stdlib.h>
	32	#include <stdint.h>
	33	#include <string.h>
	34	#include <errno.h>
	35	#include <poll.h>
	36
	37	#include "urcu/map/urcu-qsbr.h"
	38
	39	#define BUILD_QSBR_LIB
	40	#include "urcu/static/urcu-qsbr.h"
	41	/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
	42	#include "urcu-qsbr.h"
	43
	44	void __attribute__((destructor)) rcu_exit(void);
	45
	46	static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
	47
	48	int32_t gp_futex;
	49
	50	/*
	51	* Global grace period counter.
	52	*/
	53	unsigned long rcu_gp_ctr = RCU_GP_ONLINE;
	54
	55	/*
	56	* Active attempts to check for reader Q.S. before calling futex().
	57	*/
	58	#define RCU_QS_ACTIVE_ATTEMPTS 100
	59
	60	/*
	61	* Written to only by each individual reader. Read by both the reader and the
	62	* writers.
	63	*/
	64	struct rcu_reader __thread rcu_reader;
	65
	66	#ifdef DEBUG_YIELD
	67	unsigned int yield_active;
	68	unsigned int __thread rand_yield;
	69	#endif
	70
	71	static CDS_LIST_HEAD(registry);
	72
	73	static void mutex_lock(pthread_mutex_t *mutex)
	74	{
	75	int ret;
	76
	77	#ifndef DISTRUST_SIGNALS_EXTREME
	78	ret = pthread_mutex_lock(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(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	poll(NULL,0,10);
	91	}
	92	#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
	93	}
	94
	95	static void mutex_unlock(pthread_mutex_t *mutex)
	96	{
	97	int ret;
	98
	99	ret = pthread_mutex_unlock(mutex);
	100	if (ret) {
	101	perror("Error in pthread mutex unlock");
	102	exit(-1);
	103	}
	104	}
	105
	106	/*
	107	* synchronize_rcu() waiting. Single thread.
	108	*/
	109	static void wait_gp(void)
	110	{
	111	/* Read reader_gp before read futex */
	112	cmm_smp_rmb();
	113	if (uatomic_read(&gp_futex) == -1)
	114	futex_noasync(&gp_futex, FUTEX_WAIT, -1,
	115	NULL, NULL, 0);
	116	}
	117
	118	static void update_counter_and_wait(void)
	119	{
	120	CDS_LIST_HEAD(qsreaders);
	121	int wait_loops = 0;
	122	struct rcu_reader index, tmp;
	123
	124	#if (CAA_BITS_PER_LONG < 64)
	125	/* Switch parity: 0 -> 1, 1 -> 0 */
	126	CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
	127	#else /* !(CAA_BITS_PER_LONG < 64) */
	128	/* Increment current G.P. */
	129	CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
	130	#endif /* !(CAA_BITS_PER_LONG < 64) */
	131
	132	/*
	133	* Must commit rcu_gp_ctr update to memory before waiting for
	134	* quiescent state. Failure to do so could result in the writer
	135	* waiting forever while new readers are always accessing data
	136	* (no progress). Enforce compiler-order of store to rcu_gp_ctr
	137	* before load rcu_reader ctr.
	138	*/
	139	cmm_barrier();
	140
	141	/*
	142	* Adding a cmm_smp_mb() which is _not_ formally required, but makes the
	143	* model easier to understand. It does not have a big performance impact
	144	* anyway, given this is the write-side.
	145	*/
	146	cmm_smp_mb();
	147
	148	/*
	149	* Wait for each thread rcu_reader_qs_gp count to become 0.
	150	*/
	151	for (;;) {
	152	wait_loops++;
	153	if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
	154	uatomic_dec(&gp_futex);
	155	/* Write futex before read reader_gp */
	156	cmm_smp_mb();
	157	}
	158
	159	cds_list_for_each_entry_safe(index, tmp, &registry, node) {
	160	if (!rcu_gp_ongoing(&index->ctr))
	161	cds_list_move(&index->node, &qsreaders);
	162	}
	163
	164	if (cds_list_empty(&registry)) {
	165	if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
	166	/* Read reader_gp before write futex */
	167	cmm_smp_mb();
	168	uatomic_set(&gp_futex, 0);
	169	}
	170	break;
	171	} else {
	172	if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
	173	wait_gp();
	174	} else {
	175	#ifndef HAS_INCOHERENT_CACHES
	176	caa_cpu_relax();
	177	#else /* #ifndef HAS_INCOHERENT_CACHES */
	178	cmm_smp_mb();
	179	#endif /* #else #ifndef HAS_INCOHERENT_CACHES */
	180	}
	181	}
	182	}
	183	/* put back the reader list in the registry */
	184	cds_list_splice(&qsreaders, &registry);
	185	}
	186
	187	/*
	188	* Using a two-subphases algorithm for architectures with smaller than 64-bit
	189	* long-size to ensure we do not encounter an overflow bug.
	190	*/
	191
	192	#if (CAA_BITS_PER_LONG < 64)
	193	void synchronize_rcu(void)
	194	{
	195	unsigned long was_online;
	196
	197	was_online = rcu_reader.ctr;
	198
	199	/* All threads should read qparity before accessing data structure
	200	* where new ptr points to.
	201	*/
	202	/* Write new ptr before changing the qparity */
	203	cmm_smp_mb();
	204
	205	/*
	206	* Mark the writer thread offline to make sure we don't wait for
	207	* our own quiescent state. This allows using synchronize_rcu()
	208	* in threads registered as readers.
	209	*/
	210	if (was_online)
	211	CMM_STORE_SHARED(rcu_reader.ctr, 0);
	212
	213	mutex_lock(&rcu_gp_lock);
	214
	215	if (cds_list_empty(&registry))
	216	goto out;
	217
	218	/*
	219	* Wait for previous parity to be empty of readers.
	220	*/
	221	update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */
	222
	223	/*
	224	* Must finish waiting for quiescent state for parity 0 before
	225	* committing next rcu_gp_ctr update to memory. Failure to
	226	* do so could result in the writer waiting forever while new
	227	* readers are always accessing data (no progress). Enforce
	228	* compiler-order of load rcu_reader ctr before store to
	229	* rcu_gp_ctr.
	230	*/
	231	cmm_barrier();
	232
	233	/*
	234	* Adding a cmm_smp_mb() which is _not_ formally required, but makes the
	235	* model easier to understand. It does not have a big performance impact
	236	* anyway, given this is the write-side.
	237	*/
	238	cmm_smp_mb();
	239
	240	/*
	241	* Wait for previous parity to be empty of readers.
	242	*/
	243	update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */
	244	out:
	245	mutex_unlock(&rcu_gp_lock);
	246
	247	/*
	248	* Finish waiting for reader threads before letting the old ptr being
	249	* freed.
	250	*/
	251	if (was_online)
	252	_CMM_STORE_SHARED(rcu_reader.ctr,
	253	CMM_LOAD_SHARED(rcu_gp_ctr));
	254	cmm_smp_mb();
	255	}
	256	#else /* !(CAA_BITS_PER_LONG < 64) */
	257	void synchronize_rcu(void)
	258	{
	259	unsigned long was_online;
	260
	261	was_online = rcu_reader.ctr;
	262
	263	/*
	264	* Mark the writer thread offline to make sure we don't wait for
	265	* our own quiescent state. This allows using synchronize_rcu()
	266	* in threads registered as readers.
	267	*/
	268	cmm_smp_mb();
	269	if (was_online)
	270	CMM_STORE_SHARED(rcu_reader.ctr, 0);
	271
	272	mutex_lock(&rcu_gp_lock);
	273	if (cds_list_empty(&registry))
	274	goto out;
	275	update_counter_and_wait();
	276	out:
	277	mutex_unlock(&rcu_gp_lock);
	278
	279	if (was_online)
	280	_CMM_STORE_SHARED(rcu_reader.ctr,
	281	CMM_LOAD_SHARED(rcu_gp_ctr));
	282	cmm_smp_mb();
	283	}
	284	#endif /* !(CAA_BITS_PER_LONG < 64) */
	285
	286	/*
	287	* library wrappers to be used by non-LGPL compatible source code.
	288	*/
	289
	290	void rcu_read_lock(void)
	291	{
	292	_rcu_read_lock();
	293	}
	294
	295	void rcu_read_unlock(void)
	296	{
	297	_rcu_read_unlock();
	298	}
	299
	300	void rcu_quiescent_state(void)
	301	{
	302	_rcu_quiescent_state();
	303	}
	304
	305	void rcu_thread_offline(void)
	306	{
	307	_rcu_thread_offline();
	308	}
	309
	310	void rcu_thread_online(void)
	311	{
	312	_rcu_thread_online();
	313	}
	314
	315	void rcu_register_thread(void)
	316	{
	317	rcu_reader.tid = pthread_self();
	318	assert(rcu_reader.ctr == 0);
	319
	320	mutex_lock(&rcu_gp_lock);
	321	cds_list_add(&rcu_reader.node, &registry);
	322	mutex_unlock(&rcu_gp_lock);
	323	_rcu_thread_online();
	324	}
	325
	326	void rcu_unregister_thread(void)
	327	{
	328	/*
	329	* We have to make the thread offline otherwise we end up dealocking
	330	* with a waiting writer.
	331	*/
	332	_rcu_thread_offline();
	333	mutex_lock(&rcu_gp_lock);
	334	cds_list_del(&rcu_reader.node);
	335	mutex_unlock(&rcu_gp_lock);
	336	}
	337
	338	void rcu_exit(void)
	339	{
	340	assert(cds_list_empty(&registry));
	341	}
	342
	343	#include "urcu-call-rcu-impl.h"
	344	#include "urcu-defer-impl.h"