[userspace-rcu.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
#define _LGPL_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/wfqueue.h"
#include "urcu/map/urcu-qsbr.h"
#define BUILD_QSBR_LIB
#include "urcu/static/urcu-qsbr.h"
#include "urcu-pointer.h"

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

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_set(&gp_futex, -1);
			/*
			 * Write futex before write waiting (the other side
			 * reads them in the opposite order).
			 */
			cmm_smp_wmb();
			cds_list_for_each_entry(index, &registry, node) {
				_CMM_STORE_SHARED(index->waiting, 1);
			}
			/* 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.  In the "then" case, rcu_thread_offline
	 * includes a memory barrier.
	 *
	 * 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)
		rcu_thread_offline();
	else
		cmm_smp_mb();

	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)
		rcu_thread_online();
	else
		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.
	 */
	if (was_online)
		rcu_thread_offline();
	else
		cmm_smp_mb();

	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)
		rcu_thread_online();
	else
		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)
{
	/*
	 * Assertion disabled because call_rcu threads are now rcu
	 * readers, and left running at exit.
	 * assert(cds_list_empty(&registry));
	 */
}

DEFINE_RCU_FLAVOR()

#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	#define _LGPL_SOURCE
	28	#include <stdio.h>
	29	#include <pthread.h>
	30	#include <signal.h>
	31	#include <assert.h>
	32	#include <stdlib.h>
	33	#include <stdint.h>
	34	#include <string.h>
	35	#include <errno.h>
	36	#include <poll.h>
	37
	38	#include "urcu/wfqueue.h"
	39	#include "urcu/map/urcu-qsbr.h"
	40	#define BUILD_QSBR_LIB
	41	#include "urcu/static/urcu-qsbr.h"
	42	#include "urcu-pointer.h"
	43
	44	/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
	45	#undef _LGPL_SOURCE
	46	#include "urcu-qsbr.h"
	47	#define _LGPL_SOURCE
	48
	49	void __attribute__((destructor)) rcu_exit(void);
	50
	51	static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
	52
	53	int32_t gp_futex;
	54
	55	/*
	56	* Global grace period counter.
	57	*/
	58	unsigned long rcu_gp_ctr = RCU_GP_ONLINE;
	59
	60	/*
	61	* Active attempts to check for reader Q.S. before calling futex().
	62	*/
	63	#define RCU_QS_ACTIVE_ATTEMPTS 100
	64
	65	/*
	66	* Written to only by each individual reader. Read by both the reader and the
	67	* writers.
	68	*/
	69	struct rcu_reader __thread rcu_reader;
	70
	71	#ifdef DEBUG_YIELD
	72	unsigned int yield_active;
	73	unsigned int __thread rand_yield;
	74	#endif
	75
	76	static CDS_LIST_HEAD(registry);
	77
	78	static void mutex_lock(pthread_mutex_t *mutex)
	79	{
	80	int ret;
	81
	82	#ifndef DISTRUST_SIGNALS_EXTREME
	83	ret = pthread_mutex_lock(mutex);
	84	if (ret) {
	85	perror("Error in pthread mutex lock");
	86	exit(-1);
	87	}
	88	#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
	89	while ((ret = pthread_mutex_trylock(mutex)) != 0) {
	90	if (ret != EBUSY && ret != EINTR) {
	91	printf("ret = %d, errno = %d\n", ret, errno);
	92	perror("Error in pthread mutex lock");
	93	exit(-1);
	94	}
	95	poll(NULL,0,10);
	96	}
	97	#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
	98	}
	99
	100	static void mutex_unlock(pthread_mutex_t *mutex)
	101	{
	102	int ret;
	103
	104	ret = pthread_mutex_unlock(mutex);
	105	if (ret) {
	106	perror("Error in pthread mutex unlock");
	107	exit(-1);
	108	}
	109	}
	110
	111	/*
	112	* synchronize_rcu() waiting. Single thread.
	113	*/
	114	static void wait_gp(void)
	115	{
	116	/* Read reader_gp before read futex */
	117	cmm_smp_rmb();
	118	if (uatomic_read(&gp_futex) == -1)
	119	futex_noasync(&gp_futex, FUTEX_WAIT, -1,
	120	NULL, NULL, 0);
	121	}
	122
	123	static void update_counter_and_wait(void)
	124	{
	125	CDS_LIST_HEAD(qsreaders);
	126	int wait_loops = 0;
	127	struct rcu_reader index, tmp;
	128
	129	#if (CAA_BITS_PER_LONG < 64)
	130	/* Switch parity: 0 -> 1, 1 -> 0 */
	131	CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
	132	#else /* !(CAA_BITS_PER_LONG < 64) */
	133	/* Increment current G.P. */
	134	CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
	135	#endif /* !(CAA_BITS_PER_LONG < 64) */
	136
	137	/*
	138	* Must commit rcu_gp_ctr update to memory before waiting for
	139	* quiescent state. Failure to do so could result in the writer
	140	* waiting forever while new readers are always accessing data
	141	* (no progress). Enforce compiler-order of store to rcu_gp_ctr
	142	* before load rcu_reader ctr.
	143	*/
	144	cmm_barrier();
	145
	146	/*
	147	* Adding a cmm_smp_mb() which is _not_ formally required, but makes the
	148	* model easier to understand. It does not have a big performance impact
	149	* anyway, given this is the write-side.
	150	*/
	151	cmm_smp_mb();
	152
	153	/*
	154	* Wait for each thread rcu_reader_qs_gp count to become 0.
	155	*/
	156	for (;;) {
	157	wait_loops++;
	158	if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
	159	uatomic_set(&gp_futex, -1);
	160	/*
	161	* Write futex before write waiting (the other side
	162	* reads them in the opposite order).
	163	*/
	164	cmm_smp_wmb();
	165	cds_list_for_each_entry(index, &registry, node) {
	166	_CMM_STORE_SHARED(index->waiting, 1);
	167	}
	168	/* Write futex before read reader_gp */
	169	cmm_smp_mb();
	170	}
	171	cds_list_for_each_entry_safe(index, tmp, &registry, node) {
	172	if (!rcu_gp_ongoing(&index->ctr))
	173	cds_list_move(&index->node, &qsreaders);
	174	}
	175
	176	if (cds_list_empty(&registry)) {
	177	if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
	178	/* Read reader_gp before write futex */
	179	cmm_smp_mb();
	180	uatomic_set(&gp_futex, 0);
	181	}
	182	break;
	183	} else {
	184	if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
	185	wait_gp();
	186	} else {
	187	#ifndef HAS_INCOHERENT_CACHES
	188	caa_cpu_relax();
	189	#else /* #ifndef HAS_INCOHERENT_CACHES */
	190	cmm_smp_mb();
	191	#endif /* #else #ifndef HAS_INCOHERENT_CACHES */
	192	}
	193	}
	194	}
	195	/* put back the reader list in the registry */
	196	cds_list_splice(&qsreaders, &registry);
	197	}
	198
	199	/*
	200	* Using a two-subphases algorithm for architectures with smaller than 64-bit
	201	* long-size to ensure we do not encounter an overflow bug.
	202	*/
	203
	204	#if (CAA_BITS_PER_LONG < 64)
	205	void synchronize_rcu(void)
	206	{
	207	unsigned long was_online;
	208
	209	was_online = rcu_reader.ctr;
	210
	211	/* All threads should read qparity before accessing data structure
	212	* where new ptr points to. In the "then" case, rcu_thread_offline
	213	* includes a memory barrier.
	214	*
	215	* Mark the writer thread offline to make sure we don't wait for
	216	* our own quiescent state. This allows using synchronize_rcu()
	217	* in threads registered as readers.
	218	*/
	219	if (was_online)
	220	rcu_thread_offline();
	221	else
	222	cmm_smp_mb();
	223
	224	mutex_lock(&rcu_gp_lock);
	225
	226	if (cds_list_empty(&registry))
	227	goto out;
	228
	229	/*
	230	* Wait for previous parity to be empty of readers.
	231	*/
	232	update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */
	233
	234	/*
	235	* Must finish waiting for quiescent state for parity 0 before
	236	* committing next rcu_gp_ctr update to memory. Failure to
	237	* do so could result in the writer waiting forever while new
	238	* readers are always accessing data (no progress). Enforce
	239	* compiler-order of load rcu_reader ctr before store to
	240	* rcu_gp_ctr.
	241	*/
	242	cmm_barrier();
	243
	244	/*
	245	* Adding a cmm_smp_mb() which is _not_ formally required, but makes the
	246	* model easier to understand. It does not have a big performance impact
	247	* anyway, given this is the write-side.
	248	*/
	249	cmm_smp_mb();
	250
	251	/*
	252	* Wait for previous parity to be empty of readers.
	253	*/
	254	update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */
	255	out:
	256	mutex_unlock(&rcu_gp_lock);
	257
	258	/*
	259	* Finish waiting for reader threads before letting the old ptr being
	260	* freed.
	261	*/
	262	if (was_online)
	263	rcu_thread_online();
	264	else
	265	cmm_smp_mb();
	266	}
	267	#else /* !(CAA_BITS_PER_LONG < 64) */
	268	void synchronize_rcu(void)
	269	{
	270	unsigned long was_online;
	271
	272	was_online = rcu_reader.ctr;
	273
	274	/*
	275	* Mark the writer thread offline to make sure we don't wait for
	276	* our own quiescent state. This allows using synchronize_rcu()
	277	* in threads registered as readers.
	278	*/
	279	if (was_online)
	280	rcu_thread_offline();
	281	else
	282	cmm_smp_mb();
	283
	284	mutex_lock(&rcu_gp_lock);
	285	if (cds_list_empty(&registry))
	286	goto out;
	287	update_counter_and_wait();
	288	out:
	289	mutex_unlock(&rcu_gp_lock);
	290
	291	if (was_online)
	292	rcu_thread_online();
	293	else
	294	cmm_smp_mb();
	295	}
	296	#endif /* !(CAA_BITS_PER_LONG < 64) */
	297
	298	/*
	299	* library wrappers to be used by non-LGPL compatible source code.
	300	*/
	301
	302	void rcu_read_lock(void)
	303	{
	304	_rcu_read_lock();
	305	}
	306
	307	void rcu_read_unlock(void)
	308	{
	309	_rcu_read_unlock();
	310	}
	311
	312	void rcu_quiescent_state(void)
	313	{
	314	_rcu_quiescent_state();
	315	}
	316
	317	void rcu_thread_offline(void)
	318	{
	319	_rcu_thread_offline();
	320	}
	321
	322	void rcu_thread_online(void)
	323	{
	324	_rcu_thread_online();
	325	}
	326
	327	void rcu_register_thread(void)
	328	{
	329	rcu_reader.tid = pthread_self();
	330	assert(rcu_reader.ctr == 0);
	331
	332	mutex_lock(&rcu_gp_lock);
	333	cds_list_add(&rcu_reader.node, &registry);
	334	mutex_unlock(&rcu_gp_lock);
	335	_rcu_thread_online();
	336	}
	337
	338	void rcu_unregister_thread(void)
	339	{
	340	/*
	341	* We have to make the thread offline otherwise we end up dealocking
	342	* with a waiting writer.
	343	*/
	344	_rcu_thread_offline();
	345	mutex_lock(&rcu_gp_lock);
	346	cds_list_del(&rcu_reader.node);
	347	mutex_unlock(&rcu_gp_lock);
	348	}
	349
	350	void rcu_exit(void)
	351	{
	352	/*
	353	* Assertion disabled because call_rcu threads are now rcu
	354	* readers, and left running at exit.
	355	* assert(cds_list_empty(&registry));
	356	*/
	357	}
	358
	359	DEFINE_RCU_FLAVOR()
	360
	361	#include "urcu-call-rcu-impl.h"
	362	#include "urcu-defer-impl.h"