[urcu.git] / tests / benchmark / test_urcu_lfs.c

// SPDX-FileCopyrightText: 2010-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
// SPDX-FileCopyrightText: 2010 Paolo Bonzini <pbonzini@redhat.com>
//
// SPDX-License-Identifier: GPL-2.0-or-later

/*
 * Userspace RCU library - example lock-free stack
 */

#include <stdio.h>
#include <pthread.h>
#include <stdlib.h>
#include <stdint.h>
#include <stdbool.h>
#include <string.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <unistd.h>
#include <stdio.h>
#include <errno.h>

#include <urcu/arch.h>
#include <urcu/assert.h>
#include <urcu/tls-compat.h>
#include "thread-id.h"

/* hardcoded number of CPUs */
#define NR_CPUS 16384

#ifndef DYNAMIC_LINK_TEST
#define _LGPL_SOURCE
#endif
#include <urcu.h>
#include <urcu/cds.h>

#define POISON_PTR	((void *) 0x42UL)

/*
 * External synchronization used.
 */
enum test_sync {
	TEST_SYNC_NONE = 0,
	TEST_SYNC_RCU,
};

static enum test_sync test_sync;

static unsigned long rduration;

static unsigned long duration;

/* read-side C.S. duration, in loops */
static unsigned long wdelay;

static inline void loop_sleep(unsigned long loops)
{
	while (loops-- != 0)
		caa_cpu_relax();
}

static int verbose_mode;

static int test_pop, test_pop_all;

#define printf_verbose(fmt, args...)		\
	do {					\
		if (verbose_mode)		\
			printf(fmt, ## args);	\
	} while (0)

static unsigned int cpu_affinities[NR_CPUS];
static unsigned int next_aff = 0;
static int use_affinity = 0;

pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;

static void set_affinity(void)
{
#ifdef HAVE_SCHED_SETAFFINITY
	cpu_set_t mask;
	int cpu, ret;
#endif /* HAVE_SCHED_SETAFFINITY */

	if (!use_affinity)
		return;

#ifdef HAVE_SCHED_SETAFFINITY
	ret = pthread_mutex_lock(&affinity_mutex);
	if (ret) {
		perror("Error in pthread mutex lock");
		exit(-1);
	}
	cpu = cpu_affinities[next_aff++];
	ret = pthread_mutex_unlock(&affinity_mutex);
	if (ret) {
		perror("Error in pthread mutex unlock");
		exit(-1);
	}

	CPU_ZERO(&mask);
	CPU_SET(cpu, &mask);
	sched_setaffinity(0, sizeof(mask), &mask);
#endif /* HAVE_SCHED_SETAFFINITY */
}

/*
 * returns 0 if test should end.
 */
static int test_duration_dequeue(void)
{
	return test_duration_read();
}

static int test_duration_enqueue(void)
{
	return test_duration_write();
}

static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);

static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);

static unsigned int nr_enqueuers;
static unsigned int nr_dequeuers;

struct test {
	struct cds_lfs_node list;
	struct rcu_head rcu;
};

static struct cds_lfs_stack s;

static void *thr_enqueuer(void *_count)
{
	unsigned long long *count = _count;

	printf_verbose("thread_begin %s, tid %lu\n",
			"enqueuer", urcu_get_thread_id());

	set_affinity();

	rcu_register_thread();

	wait_until_go();

	for (;;) {
		struct test *node = malloc(sizeof(*node));
		if (!node)
			goto fail;
		cds_lfs_node_init(&node->list);
		cds_lfs_push(&s, &node->list);
		URCU_TLS(nr_successful_enqueues)++;

		if (caa_unlikely(wdelay))
			loop_sleep(wdelay);
fail:
		URCU_TLS(nr_enqueues)++;
		if (caa_unlikely(!test_duration_enqueue()))
			break;
	}

	rcu_unregister_thread();

	count[0] = URCU_TLS(nr_enqueues);
	count[1] = URCU_TLS(nr_successful_enqueues);
	printf_verbose("enqueuer thread_end, tid %lu, "
			"enqueues %llu successful_enqueues %llu\n",
			urcu_get_thread_id(),
			URCU_TLS(nr_enqueues),
			URCU_TLS(nr_successful_enqueues));
	return ((void*)1);

}

static
void free_node_cb(struct rcu_head *head)
{
	struct test *node =
		caa_container_of(head, struct test, rcu);
	free(node);
}

static
void do_test_pop(enum test_sync sync)
{
	struct cds_lfs_node *snode;

	if (sync == TEST_SYNC_RCU)
		rcu_read_lock();
	snode = __cds_lfs_pop(&s);
	if (sync == TEST_SYNC_RCU)
		rcu_read_unlock();
	if (snode) {
		struct test *node;

		snode->next = POISON_PTR;
		node = caa_container_of(snode,
			struct test, list);
		if (sync == TEST_SYNC_RCU)
			call_rcu(&node->rcu, free_node_cb);
		else
			free(node);
		URCU_TLS(nr_successful_dequeues)++;
	}
	URCU_TLS(nr_dequeues)++;
}

static
void do_test_pop_all(enum test_sync sync)
{
	struct cds_lfs_node *snode;
	struct cds_lfs_head *head;
	struct cds_lfs_node *n;

	head = __cds_lfs_pop_all(&s);
	cds_lfs_for_each_safe(head, snode, n) {
		struct test *node;

		snode->next = POISON_PTR;
		node = caa_container_of(snode, struct test, list);
		if (sync == TEST_SYNC_RCU)
			call_rcu(&node->rcu, free_node_cb);
		else
			free(node);
		URCU_TLS(nr_successful_dequeues)++;
		URCU_TLS(nr_dequeues)++;
	}

}

static void *thr_dequeuer(void *_count)
{
	unsigned long long *count = _count;
	unsigned int counter = 0;

	printf_verbose("thread_begin %s, tid %lu\n",
			"dequeuer", urcu_get_thread_id());

	set_affinity();

	rcu_register_thread();

	wait_until_go();

	urcu_posix_assert(test_pop || test_pop_all);

	for (;;) {
		if (test_pop && test_pop_all) {
			/* both pop and pop all */
			if (counter & 1)
				do_test_pop(test_sync);
			else
				do_test_pop_all(test_sync);
			counter++;
		} else {
			if (test_pop)
				do_test_pop(test_sync);
			else
				do_test_pop_all(test_sync);
		}

		if (caa_unlikely(!test_duration_dequeue()))
			break;
		if (caa_unlikely(rduration))
			loop_sleep(rduration);
	}

	rcu_unregister_thread();

	printf_verbose("dequeuer thread_end, tid %lu, "
			"dequeues %llu, successful_dequeues %llu\n",
			urcu_get_thread_id(),
			URCU_TLS(nr_dequeues),
			URCU_TLS(nr_successful_dequeues));
	count[0] = URCU_TLS(nr_dequeues);
	count[1] = URCU_TLS(nr_successful_dequeues);
	return ((void*)2);
}

static void test_end(unsigned long long *nr_dequeues_l)
{
	struct cds_lfs_node *snode;

	do {
		snode = __cds_lfs_pop(&s);
		if (snode) {
			struct test *node;

			node = caa_container_of(snode, struct test, list);
			free(node);
			(*nr_dequeues_l)++;
		}
	} while (snode);
}

static void show_usage(char **argv)
{
	printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
		argv[0]);
	printf("OPTIONS:\n");
	printf("	[-d delay] (enqueuer period (in loops))\n");
	printf("	[-c duration] (dequeuer period (in loops))\n");
	printf("	[-v] (verbose output)\n");
	printf("	[-a cpu#] [-a cpu#]... (affinity)\n");
	printf("	[-p] (test pop)\n");
	printf("	[-P] (test pop_all, enabled by default)\n");
	printf("	[-R] (use RCU external synchronization)\n");
	printf("		Note: default: no external synchronization used.\n");
	printf("\n");
}

int main(int argc, char **argv)
{
	int err;
	pthread_t *tid_enqueuer, *tid_dequeuer;
	void *tret;
	unsigned long long *count_enqueuer, *count_dequeuer;
	unsigned long long tot_enqueues = 0, tot_dequeues = 0;
	unsigned long long tot_successful_enqueues = 0,
			   tot_successful_dequeues = 0;
	unsigned long long end_dequeues = 0;
	int i, a;
	unsigned int i_thr;

	if (argc < 4) {
		show_usage(argv);
		return -1;
	}

	err = sscanf(argv[1], "%u", &nr_dequeuers);
	if (err != 1) {
		show_usage(argv);
		return -1;
	}

	err = sscanf(argv[2], "%u", &nr_enqueuers);
	if (err != 1) {
		show_usage(argv);
		return -1;
	}

	err = sscanf(argv[3], "%lu", &duration);
	if (err != 1) {
		show_usage(argv);
		return -1;
	}

	for (i = 4; i < argc; i++) {
		if (argv[i][0] != '-')
			continue;
		switch (argv[i][1]) {
		case 'a':
			if (argc < i + 2) {
				show_usage(argv);
				return -1;
			}
			a = atoi(argv[++i]);
			cpu_affinities[next_aff++] = a;
			use_affinity = 1;
			printf_verbose("Adding CPU %d affinity\n", a);
			break;
		case 'c':
			if (argc < i + 2) {
				show_usage(argv);
				return -1;
			}
			rduration = atol(argv[++i]);
			break;
		case 'd':
			if (argc < i + 2) {
				show_usage(argv);
				return -1;
			}
			wdelay = atol(argv[++i]);
			break;
		case 'v':
			verbose_mode = 1;
			break;
		case 'p':
			test_pop = 1;
			break;
		case 'P':
			test_pop_all = 1;
			break;
		case 'R':
			test_sync = TEST_SYNC_RCU;
			break;
		}
	}

	/* activate pop_all test by default */
	if (!test_pop && !test_pop_all)
		test_pop_all = 1;

	printf_verbose("running test for %lu seconds, %u enqueuers, "
		       "%u dequeuers.\n",
		       duration, nr_enqueuers, nr_dequeuers);
	if (test_pop)
		printf_verbose("pop test activated.\n");
	if (test_pop_all)
		printf_verbose("pop_all test activated.\n");
	if (test_sync == TEST_SYNC_RCU)
		printf_verbose("External sync: RCU.\n");
	else
		printf_verbose("External sync: none.\n");
	printf_verbose("Writer delay : %lu loops.\n", rduration);
	printf_verbose("Reader duration : %lu loops.\n", wdelay);
	printf_verbose("thread %-6s, tid %lu\n",
			"main", urcu_get_thread_id());

	tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
	tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
	count_enqueuer = calloc(nr_enqueuers, 2 * sizeof(*count_enqueuer));
	count_dequeuer = calloc(nr_dequeuers, 2 * sizeof(*count_dequeuer));
	cds_lfs_init(&s);
	err = create_all_cpu_call_rcu_data(0);
	if (err) {
		printf("Per-CPU call_rcu() worker threads unavailable. Using default global worker thread.\n");
	}

	next_aff = 0;

	for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
		err = pthread_create(&tid_enqueuer[i_thr], NULL, thr_enqueuer,
				     &count_enqueuer[2 * i_thr]);
		if (err != 0)
			exit(1);
	}
	for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
		err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer,
				     &count_dequeuer[2 * i_thr]);
		if (err != 0)
			exit(1);
	}

	cmm_smp_mb();

	begin_test();

	for (i_thr = 0; i_thr < duration; i_thr++) {
		sleep(1);
		if (verbose_mode) {
			fwrite(".", sizeof(char), 1, stdout);
			fflush(stdout);
		}
	}

	end_test();

	for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
		err = pthread_join(tid_enqueuer[i_thr], &tret);
		if (err != 0)
			exit(1);
		tot_enqueues += count_enqueuer[2 * i_thr];
		tot_successful_enqueues += count_enqueuer[2 * i_thr + 1];
	}
	for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
		err = pthread_join(tid_dequeuer[i_thr], &tret);
		if (err != 0)
			exit(1);
		tot_dequeues += count_dequeuer[2 * i_thr];
		tot_successful_dequeues += count_dequeuer[2 * i_thr + 1];
	}

	test_end(&end_dequeues);

	printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
		       tot_enqueues, tot_dequeues);
	printf_verbose("total number of successful enqueues : %llu, "
		       "successful dequeues %llu\n",
		       tot_successful_enqueues, tot_successful_dequeues);
	printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
		"nr_dequeuers %3u "
		"rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
		"successful enqueues %12llu successful dequeues %12llu "
		"end_dequeues %llu nr_ops %12llu\n",
		argv[0], duration, nr_enqueuers, wdelay,
		nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
		tot_successful_enqueues,
		tot_successful_dequeues, end_dequeues,
		tot_enqueues + tot_dequeues);
	if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues)
		printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
		       "succ. dequeues + end dequeues %llu.\n",
		       tot_successful_enqueues,
		       tot_successful_dequeues + end_dequeues);

	free_all_cpu_call_rcu_data();
	cds_lfs_destroy(&s);
	free(count_enqueuer);
	free(count_dequeuer);
	free(tid_enqueuer);
	free(tid_dequeuer);

	return 0;
}
Commit	Line	Data
ce29b371 MJ	1	// SPDX-FileCopyrightText: 2010-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
	2	// SPDX-FileCopyrightText: 2010 Paolo Bonzini <pbonzini@redhat.com>
	3	//
	4	// SPDX-License-Identifier: GPL-2.0-or-later
	5
453629a9	6	/*
cb8818f0	7	* Userspace RCU library - example lock-free stack
453629a9 MD	8	*/
453629a9 MD	9
453629a9 MD	10	#include <stdio.h>
	11	#include <pthread.h>
	12	#include <stdlib.h>
	13	#include <stdint.h>
	14	#include <stdbool.h>
	15	#include <string.h>
	16	#include <sys/types.h>
	17	#include <sys/wait.h>
	18	#include <unistd.h>
	19	#include <stdio.h>
453629a9 MD	20	#include <errno.h>
	21
	22	#include <urcu/arch.h>
01477510	23	#include <urcu/assert.h>
bd252a04	24	#include <urcu/tls-compat.h>
94df6318	25	#include "thread-id.h"
65fcc7e9	26
453629a9 MD	27	/* hardcoded number of CPUs */
	28	#define NR_CPUS 16384
	29
453629a9 MD	30	#ifndef DYNAMIC_LINK_TEST
	31	#define _LGPL_SOURCE
	32	#endif
	33	#include <urcu.h>
e5a7d709	34	#include <urcu/cds.h>
453629a9	35
f9b5c2b6 MD	36	#define POISON_PTR ((void *) 0x42UL)
f9b5c2b6 MD	37
111ce0c3 MD	38	/*
	39	* External synchronization used.
	40	*/
	41	enum test_sync {
	42	TEST_SYNC_NONE = 0,
	43	TEST_SYNC_RCU,
	44	};
	45
	46	static enum test_sync test_sync;
	47
453629a9 MD	48	static unsigned long rduration;
	49
	50	static unsigned long duration;
	51
	52	/* read-side C.S. duration, in loops */
	53	static unsigned long wdelay;
	54
ab0aacbe	55	static inline void loop_sleep(unsigned long loops)
453629a9	56	{
ab0aacbe	57	while (loops-- != 0)
06f22bdb	58	caa_cpu_relax();
453629a9 MD	59	}
	60
	61	static int verbose_mode;
	62
111ce0c3 MD	63	static int test_pop, test_pop_all;
111ce0c3 MD	64
453629a9 MD	65	#define printf_verbose(fmt, args...) \
	66	do { \
	67	if (verbose_mode) \
111ce0c3	68	printf(fmt, ## args); \
453629a9 MD	69	} while (0)
	70
	71	static unsigned int cpu_affinities[NR_CPUS];
	72	static unsigned int next_aff = 0;
	73	static int use_affinity = 0;
	74
	75	pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
	76
453629a9 MD	77	static void set_affinity(void)
453629a9 MD	78	{
2388c075	79	#ifdef HAVE_SCHED_SETAFFINITY
453629a9	80	cpu_set_t mask;
95bc7fb9 MD	81	int cpu, ret;
95bc7fb9 MD	82	#endif /* HAVE_SCHED_SETAFFINITY */
453629a9 MD	83
	84	if (!use_affinity)
	85	return;
	86
2388c075	87	#ifdef HAVE_SCHED_SETAFFINITY
453629a9 MD	88	ret = pthread_mutex_lock(&affinity_mutex);
	89	if (ret) {
	90	perror("Error in pthread mutex lock");
	91	exit(-1);
	92	}
	93	cpu = cpu_affinities[next_aff++];
	94	ret = pthread_mutex_unlock(&affinity_mutex);
	95	if (ret) {
	96	perror("Error in pthread mutex unlock");
	97	exit(-1);
	98	}
	99
	100	CPU_ZERO(&mask);
	101	CPU_SET(cpu, &mask);
453629a9	102	sched_setaffinity(0, sizeof(mask), &mask);
453629a9 MD	103	#endif /* HAVE_SCHED_SETAFFINITY */
	104	}
	105
	106	/*
	107	* returns 0 if test should end.
	108	*/
	109	static int test_duration_dequeue(void)
	110	{
9e4e7ad1	111	return test_duration_read();
453629a9 MD	112	}
	113
	114	static int test_duration_enqueue(void)
	115	{
9e4e7ad1	116	return test_duration_write();
453629a9 MD	117	}
453629a9 MD	118
bd252a04 MD	119	static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
bd252a04 MD	120	static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
453629a9	121
bd252a04 MD	122	static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
bd252a04 MD	123	static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
453629a9 MD	124
	125	static unsigned int nr_enqueuers;
	126	static unsigned int nr_dequeuers;
	127
94aec122	128	struct test {
cb8818f0	129	struct cds_lfs_node list;
94aec122 MD	130	struct rcu_head rcu;
	131	};
	132
cb8818f0	133	static struct cds_lfs_stack s;
453629a9	134
eb314468	135	static void thr_enqueuer(void _count)
453629a9 MD	136	{
	137	unsigned long long *count = _count;
	138
94df6318 MD	139	printf_verbose("thread_begin %s, tid %lu\n",
94df6318 MD	140	"enqueuer", urcu_get_thread_id());
453629a9 MD	141
	142	set_affinity();
	143
	144	rcu_register_thread();
	145
9e4e7ad1	146	wait_until_go();
453629a9 MD	147
453629a9 MD	148	for (;;) {
94aec122	149	struct test node = malloc(sizeof(node));
453629a9 MD	150	if (!node)
453629a9 MD	151	goto fail;
cb8818f0 MD	152	cds_lfs_node_init(&node->list);
cb8818f0 MD	153	cds_lfs_push(&s, &node->list);
bd252a04	154	URCU_TLS(nr_successful_enqueues)++;
453629a9	155
a0b7f7ea	156	if (caa_unlikely(wdelay))
453629a9 MD	157	loop_sleep(wdelay);
453629a9 MD	158	fail:
bd252a04	159	URCU_TLS(nr_enqueues)++;
a0b7f7ea	160	if (caa_unlikely(!test_duration_enqueue()))
453629a9 MD	161	break;
	162	}
	163
	164	rcu_unregister_thread();
	165
bd252a04 MD	166	count[0] = URCU_TLS(nr_enqueues);
bd252a04 MD	167	count[1] = URCU_TLS(nr_successful_enqueues);
94df6318 MD	168	printf_verbose("enqueuer thread_end, tid %lu, "
	169	"enqueues %llu successful_enqueues %llu\n",
	170	urcu_get_thread_id(),
	171	URCU_TLS(nr_enqueues),
	172	URCU_TLS(nr_successful_enqueues));
453629a9 MD	173	return ((void*)1);
	174
	175	}
	176
94aec122 MD	177	static
	178	void free_node_cb(struct rcu_head *head)
	179	{
	180	struct test *node =
	181	caa_container_of(head, struct test, rcu);
	182	free(node);
	183	}
	184
111ce0c3 MD	185	static
	186	void do_test_pop(enum test_sync sync)
	187	{
	188	struct cds_lfs_node *snode;
	189
	190	if (sync == TEST_SYNC_RCU)
	191	rcu_read_lock();
	192	snode = __cds_lfs_pop(&s);
	193	if (sync == TEST_SYNC_RCU)
	194	rcu_read_unlock();
	195	if (snode) {
	196	struct test *node;
	197
f9b5c2b6	198	snode->next = POISON_PTR;
111ce0c3 MD	199	node = caa_container_of(snode,
	200	struct test, list);
	201	if (sync == TEST_SYNC_RCU)
	202	call_rcu(&node->rcu, free_node_cb);
	203	else
	204	free(node);
	205	URCU_TLS(nr_successful_dequeues)++;
	206	}
	207	URCU_TLS(nr_dequeues)++;
	208	}
	209
	210	static
	211	void do_test_pop_all(enum test_sync sync)
	212	{
	213	struct cds_lfs_node *snode;
	214	struct cds_lfs_head *head;
	215	struct cds_lfs_node *n;
	216
	217	head = __cds_lfs_pop_all(&s);
	218	cds_lfs_for_each_safe(head, snode, n) {
	219	struct test *node;
	220
f9b5c2b6	221	snode->next = POISON_PTR;
111ce0c3 MD	222	node = caa_container_of(snode, struct test, list);
	223	if (sync == TEST_SYNC_RCU)
	224	call_rcu(&node->rcu, free_node_cb);
	225	else
	226	free(node);
	227	URCU_TLS(nr_successful_dequeues)++;
	228	URCU_TLS(nr_dequeues)++;
	229	}
	230
	231	}
	232
eb314468	233	static void thr_dequeuer(void _count)
453629a9 MD	234	{
453629a9 MD	235	unsigned long long *count = _count;
f42bc7ef	236	unsigned int counter = 0;
453629a9	237
94df6318 MD	238	printf_verbose("thread_begin %s, tid %lu\n",
94df6318 MD	239	"dequeuer", urcu_get_thread_id());
453629a9 MD	240
	241	set_affinity();
	242
453629a9 MD	243	rcu_register_thread();
453629a9 MD	244
9e4e7ad1	245	wait_until_go();
453629a9	246
01477510	247	urcu_posix_assert(test_pop \|\| test_pop_all);
f3eb6ef1	248
111ce0c3	249	for (;;) {
111ce0c3 MD	250	if (test_pop && test_pop_all) {
	251	/* both pop and pop all */
	252	if (counter & 1)
	253	do_test_pop(test_sync);
	254	else
	255	do_test_pop_all(test_sync);
	256	counter++;
	257	} else {
	258	if (test_pop)
	259	do_test_pop(test_sync);
	260	else
	261	do_test_pop_all(test_sync);
453629a9	262	}
111ce0c3	263
a0b7f7ea	264	if (caa_unlikely(!test_duration_dequeue()))
453629a9	265	break;
a0b7f7ea	266	if (caa_unlikely(rduration))
453629a9 MD	267	loop_sleep(rduration);
	268	}
	269
	270	rcu_unregister_thread();
453629a9	271
94df6318 MD	272	printf_verbose("dequeuer thread_end, tid %lu, "
	273	"dequeues %llu, successful_dequeues %llu\n",
	274	urcu_get_thread_id(),
	275	URCU_TLS(nr_dequeues),
	276	URCU_TLS(nr_successful_dequeues));
bd252a04 MD	277	count[0] = URCU_TLS(nr_dequeues);
bd252a04 MD	278	count[1] = URCU_TLS(nr_successful_dequeues);
453629a9 MD	279	return ((void*)2);
	280	}
	281
bd23a6c6	282	static void test_end(unsigned long long *nr_dequeues_l)
453629a9	283	{
cb8818f0	284	struct cds_lfs_node *snode;
453629a9 MD	285
453629a9 MD	286	do {
bd23a6c6	287	snode = __cds_lfs_pop(&s);
94aec122 MD	288	if (snode) {
	289	struct test *node;
	290
	291	node = caa_container_of(snode, struct test, list);
453629a9	292	free(node);
bd23a6c6	293	(*nr_dequeues_l)++;
453629a9	294	}
94aec122	295	} while (snode);
453629a9 MD	296	}
453629a9 MD	297
70469b43	298	static void show_usage(char **argv)
453629a9	299	{
06637138 MD	300	printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
	301	argv[0]);
	302	printf("OPTIONS:\n");
	303	printf(" [-d delay] (enqueuer period (in loops))\n");
	304	printf(" [-c duration] (dequeuer period (in loops))\n");
	305	printf(" [-v] (verbose output)\n");
	306	printf(" [-a cpu#] [-a cpu#]... (affinity)\n");
	307	printf(" [-p] (test pop)\n");
	308	printf(" [-P] (test pop_all, enabled by default)\n");
	309	printf(" [-R] (use RCU external synchronization)\n");
	310	printf(" Note: default: no external synchronization used.\n");
453629a9 MD	311	printf("\n");
	312	}
	313
	314	int main(int argc, char **argv)
	315	{
	316	int err;
	317	pthread_t tid_enqueuer, tid_dequeuer;
	318	void *tret;
	319	unsigned long long count_enqueuer, count_dequeuer;
	320	unsigned long long tot_enqueues = 0, tot_dequeues = 0;
	321	unsigned long long tot_successful_enqueues = 0,
	322	tot_successful_dequeues = 0;
	323	unsigned long long end_dequeues = 0;
	324	int i, a;
83e334d0	325	unsigned int i_thr;
453629a9 MD	326
453629a9 MD	327	if (argc < 4) {
70469b43	328	show_usage(argv);
453629a9 MD	329	return -1;
	330	}
	331
	332	err = sscanf(argv[1], "%u", &nr_dequeuers);
	333	if (err != 1) {
70469b43	334	show_usage(argv);
453629a9 MD	335	return -1;
	336	}
	337
	338	err = sscanf(argv[2], "%u", &nr_enqueuers);
	339	if (err != 1) {
70469b43	340	show_usage(argv);
453629a9 MD	341	return -1;
453629a9 MD	342	}
83e334d0	343
453629a9 MD	344	err = sscanf(argv[3], "%lu", &duration);
453629a9 MD	345	if (err != 1) {
70469b43	346	show_usage(argv);
453629a9 MD	347	return -1;
	348	}
	349
	350	for (i = 4; i < argc; i++) {
	351	if (argv[i][0] != '-')
	352	continue;
	353	switch (argv[i][1]) {
	354	case 'a':
	355	if (argc < i + 2) {
70469b43	356	show_usage(argv);
453629a9 MD	357	return -1;
	358	}
	359	a = atoi(argv[++i]);
	360	cpu_affinities[next_aff++] = a;
	361	use_affinity = 1;
	362	printf_verbose("Adding CPU %d affinity\n", a);
	363	break;
	364	case 'c':
	365	if (argc < i + 2) {
70469b43	366	show_usage(argv);
453629a9 MD	367	return -1;
	368	}
	369	rduration = atol(argv[++i]);
	370	break;
	371	case 'd':
	372	if (argc < i + 2) {
70469b43	373	show_usage(argv);
453629a9 MD	374	return -1;
	375	}
	376	wdelay = atol(argv[++i]);
	377	break;
	378	case 'v':
	379	verbose_mode = 1;
	380	break;
111ce0c3 MD	381	case 'p':
	382	test_pop = 1;
	383	break;
	384	case 'P':
	385	test_pop_all = 1;
	386	break;
	387	case 'R':
	388	test_sync = TEST_SYNC_RCU;
	389	break;
453629a9 MD	390	}
	391	}
	392
111ce0c3 MD	393	/* activate pop_all test by default */
	394	if (!test_pop && !test_pop_all)
	395	test_pop_all = 1;
	396
453629a9 MD	397	printf_verbose("running test for %lu seconds, %u enqueuers, "
	398	"%u dequeuers.\n",
	399	duration, nr_enqueuers, nr_dequeuers);
111ce0c3 MD	400	if (test_pop)
	401	printf_verbose("pop test activated.\n");
	402	if (test_pop_all)
	403	printf_verbose("pop_all test activated.\n");
eb314468 MD	404	if (test_sync == TEST_SYNC_RCU)
	405	printf_verbose("External sync: RCU.\n");
	406	else
	407	printf_verbose("External sync: none.\n");
453629a9 MD	408	printf_verbose("Writer delay : %lu loops.\n", rduration);
453629a9 MD	409	printf_verbose("Reader duration : %lu loops.\n", wdelay);
94df6318 MD	410	printf_verbose("thread %-6s, tid %lu\n",
94df6318 MD	411	"main", urcu_get_thread_id());
453629a9	412
9aa14175 MD	413	tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
	414	tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
	415	count_enqueuer = calloc(nr_enqueuers, 2 * sizeof(*count_enqueuer));
	416	count_dequeuer = calloc(nr_dequeuers, 2 * sizeof(*count_dequeuer));
cb8818f0	417	cds_lfs_init(&s);
94aec122	418	err = create_all_cpu_call_rcu_data(0);
8bcbd94a MD	419	if (err) {
	420	printf("Per-CPU call_rcu() worker threads unavailable. Using default global worker thread.\n");
	421	}
453629a9 MD	422
	423	next_aff = 0;
	424
83e334d0 MJ	425	for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
	426	err = pthread_create(&tid_enqueuer[i_thr], NULL, thr_enqueuer,
	427	&count_enqueuer[2 * i_thr]);
453629a9 MD	428	if (err != 0)
	429	exit(1);
	430	}
83e334d0 MJ	431	for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
	432	err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer,
	433	&count_dequeuer[2 * i_thr]);
453629a9 MD	434	if (err != 0)
	435	exit(1);
	436	}
	437
5481ddb3	438	cmm_smp_mb();
453629a9	439
9e4e7ad1	440	begin_test();
453629a9	441
83e334d0	442	for (i_thr = 0; i_thr < duration; i_thr++) {
453629a9	443	sleep(1);
4b665350 MD	444	if (verbose_mode) {
	445	fwrite(".", sizeof(char), 1, stdout);
	446	fflush(stdout);
	447	}
453629a9 MD	448	}
453629a9 MD	449
9e4e7ad1	450	end_test();
453629a9	451
83e334d0 MJ	452	for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
83e334d0 MJ	453	err = pthread_join(tid_enqueuer[i_thr], &tret);
453629a9 MD	454	if (err != 0)
453629a9 MD	455	exit(1);
83e334d0 MJ	456	tot_enqueues += count_enqueuer[2 * i_thr];
83e334d0 MJ	457	tot_successful_enqueues += count_enqueuer[2 * i_thr + 1];
453629a9	458	}
83e334d0 MJ	459	for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
83e334d0 MJ	460	err = pthread_join(tid_dequeuer[i_thr], &tret);
453629a9 MD	461	if (err != 0)
453629a9 MD	462	exit(1);
83e334d0 MJ	463	tot_dequeues += count_dequeuer[2 * i_thr];
83e334d0 MJ	464	tot_successful_dequeues += count_dequeuer[2 * i_thr + 1];
453629a9	465	}
83e334d0	466
bd23a6c6	467	test_end(&end_dequeues);
453629a9 MD	468
	469	printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
	470	tot_enqueues, tot_dequeues);
	471	printf_verbose("total number of successful enqueues : %llu, "
	472	"successful dequeues %llu\n",
	473	tot_successful_enqueues, tot_successful_dequeues);
	474	printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
	475	"nr_dequeuers %3u "
	476	"rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
	477	"successful enqueues %12llu successful dequeues %12llu "
	478	"end_dequeues %llu nr_ops %12llu\n",
	479	argv[0], duration, nr_enqueuers, wdelay,
	480	nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
	481	tot_successful_enqueues,
	482	tot_successful_dequeues, end_dequeues,
	483	tot_enqueues + tot_dequeues);
	484	if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues)
	485	printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
	486	"succ. dequeues + end dequeues %llu.\n",
	487	tot_successful_enqueues,
	488	tot_successful_dequeues + end_dequeues);
	489
0938c541	490	free_all_cpu_call_rcu_data();
2af1c19e	491	cds_lfs_destroy(&s);
453629a9 MD	492	free(count_enqueuer);
	493	free(count_dequeuer);
	494	free(tid_enqueuer);
	495	free(tid_dequeuer);
83e334d0	496
453629a9 MD	497	return 0;
453629a9 MD	498	}