[userspace-rcu.git] / tests / benchmark / test_urcu_wfs.c

// SPDX-FileCopyrightText: 2010 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 RCU-based 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 <urcu/uatomic.h>
#include "thread-id.h"

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

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

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

static enum test_sync test_sync;

static int test_force_sync;

static volatile int test_stop_enqueue, test_stop_dequeue;

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, test_wait_empty;
static unsigned int test_enqueue_stopped;

#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 !uatomic_load(&test_stop_dequeue, CMM_RELAXED);
}

static int test_duration_enqueue(void)
{
	return !uatomic_load(&test_stop_enqueue, CMM_RELAXED);
}

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 DEFINE_URCU_TLS(unsigned long long, nr_empty_dest_enqueues);
static DEFINE_URCU_TLS(unsigned long long, nr_pop_all);
static DEFINE_URCU_TLS(unsigned long long, nr_pop_last);

static unsigned int nr_enqueuers;
static unsigned int nr_dequeuers;

static struct cds_wfs_stack s;

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

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

	set_affinity();

	wait_until_go();

	for (;;) {
		struct cds_wfs_node *node = malloc(sizeof(*node));
		if (!node)
			goto fail;
		cds_wfs_node_init(node);
		was_nonempty = cds_wfs_push(&s, node);
		URCU_TLS(nr_successful_enqueues)++;
		if (!was_nonempty)
			URCU_TLS(nr_empty_dest_enqueues)++;

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

	uatomic_inc(&test_enqueue_stopped);
	count[0] = URCU_TLS(nr_enqueues);
	count[1] = URCU_TLS(nr_successful_enqueues);
	count[2] = URCU_TLS(nr_empty_dest_enqueues);
	printf_verbose("enqueuer thread_end, tid %lu, "
			"enqueues %llu successful_enqueues %llu, "
			"empty_dest_enqueues %llu\n",
			urcu_get_thread_id(),
			URCU_TLS(nr_enqueues),
			URCU_TLS(nr_successful_enqueues),
			URCU_TLS(nr_empty_dest_enqueues));
	return ((void*)1);

}

static void do_test_pop(enum test_sync sync)
{
	struct cds_wfs_node *node;
	int state;

	if (sync == TEST_SYNC_MUTEX)
		cds_wfs_pop_lock(&s);
	node = __cds_wfs_pop_with_state_blocking(&s, &state);
	if (sync == TEST_SYNC_MUTEX)
		cds_wfs_pop_unlock(&s);

	if (node) {
		if (state & CDS_WFS_STATE_LAST)
			URCU_TLS(nr_pop_last)++;
		free(node);
		URCU_TLS(nr_successful_dequeues)++;
	}
	URCU_TLS(nr_dequeues)++;
}

static void do_test_pop_all(enum test_sync sync)
{
	struct cds_wfs_head *head;
	struct cds_wfs_node *node, *n;

	if (sync == TEST_SYNC_MUTEX)
		cds_wfs_pop_lock(&s);
	head = __cds_wfs_pop_all(&s);
	if (sync == TEST_SYNC_MUTEX)
		cds_wfs_pop_unlock(&s);

	/* Check if empty */
	if (cds_wfs_first(head) == NULL)
		return;

	URCU_TLS(nr_pop_all)++;
	URCU_TLS(nr_pop_last)++;

	cds_wfs_for_each_blocking_safe(head, node, n) {
		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();

	wait_until_go();

	urcu_posix_assert(test_pop || test_pop_all);

	for (;;) {
		if (test_pop && test_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);
	}

	printf_verbose("dequeuer thread_end, tid %lu, "
			"dequeues %llu, successful_dequeues %llu "
			"pop_all %llu pop_last %llu\n",
			urcu_get_thread_id(),
			URCU_TLS(nr_dequeues), URCU_TLS(nr_successful_dequeues),
			URCU_TLS(nr_pop_all),
			URCU_TLS(nr_pop_last));
	count[0] = URCU_TLS(nr_dequeues);
	count[1] = URCU_TLS(nr_successful_dequeues);
	count[2] = URCU_TLS(nr_pop_all);
	count[3] = URCU_TLS(nr_pop_last);
	return ((void*)2);
}

static void test_end(unsigned long long *nr_dequeues_l,
		unsigned long long *nr_pop_last_l)
{
	struct cds_wfs_node *node;
	int state;

	do {
		node = cds_wfs_pop_with_state_blocking(&s, &state);
		if (node) {
			if (state & CDS_WFS_STATE_LAST)
				(*nr_pop_last_l)++;
			free(node);
			(*nr_dequeues_l)++;
		}
	} while (node);
}

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("	[-M] (use mutex external synchronization)\n");
	printf("		Note: default: no external synchronization used.\n");
	printf("	[-f] (force user-provided synchronization)\n");
	printf("	[-w] Wait for dequeuer to empty stack\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,
			   tot_empty_dest_enqueues = 0,
			   tot_pop_all = 0, tot_pop_last = 0;
	unsigned long long end_dequeues = 0;
	int i, a, retval = 0;
	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 'M':
			test_sync = TEST_SYNC_MUTEX;
			break;
		case 'w':
			test_wait_empty = 1;
			break;
		case 'f':
			test_force_sync = 1;
			break;
		}
	}

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

	if (test_sync == TEST_SYNC_NONE && nr_dequeuers > 1 && test_pop) {
		if (test_force_sync) {
			fprintf(stderr, "[WARNING] Using pop concurrently "
				"with other pop or pop_all without external "
				"synchronization. Expect run-time failure.\n");
		} else {
			printf("Enforcing mutex synchronization\n");
			test_sync = TEST_SYNC_MUTEX;
		}
	}

	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_MUTEX)
		printf_verbose("External sync: mutex.\n");
	else
		printf_verbose("External sync: none.\n");
	if (test_wait_empty)
		printf_verbose("Wait for dequeuers to empty stack.\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, 3 * sizeof(*count_enqueuer));
	count_dequeuer = calloc(nr_dequeuers, 4 * sizeof(*count_dequeuer));
	cds_wfs_init(&s);

	next_aff = 0;

	for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
		err = pthread_create(&tid_enqueuer[i], NULL, thr_enqueuer,
				     &count_enqueuer[3 * 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[4 * 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);
		}
	}

	uatomic_store(&test_stop_enqueue, 1, CMM_RELEASE);

	if (test_wait_empty) {
		while (nr_enqueuers != uatomic_read(&test_enqueue_stopped)) {
			sleep(1);
		}
		while (!cds_wfs_empty(&s)) {
			sleep(1);
		}
	}

	uatomic_store(&test_stop_dequeue, 1, CMM_RELAXED);

	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[3 * i_thr];
		tot_successful_enqueues += count_enqueuer[3 * i_thr + 1];
		tot_empty_dest_enqueues += count_enqueuer[3 * i_thr + 2];
	}
	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[4 * i_thr];
		tot_successful_dequeues += count_dequeuer[4 * i_thr + 1];
		tot_pop_all += count_dequeuer[4 * i_thr + 2];
		tot_pop_last += count_dequeuer[4 * i_thr + 3];
	}

	test_end(&end_dequeues, &tot_pop_last);

	printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
		       tot_enqueues, tot_dequeues);
	printf_verbose("total number of successful enqueues : %llu, "
		       "enqueues to empty dest : %llu, "
		       "successful dequeues %llu, "
		       "pop_all : %llu, pop_last : %llu\n",
		       tot_successful_enqueues,
		       tot_empty_dest_enqueues,
		       tot_successful_dequeues,
		       tot_pop_all, tot_pop_last);
	printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
		"nr_dequeuers %3u "
		"rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
		"successful enqueues %12llu enqueues to empty dest %12llu "
		"successful dequeues %12llu pop_all %12llu "
		"pop_last %llu end_dequeues %llu nr_ops %12llu\n",
		argv[0], duration, nr_enqueuers, wdelay,
		nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
		tot_successful_enqueues,
		tot_empty_dest_enqueues,
		tot_successful_dequeues, tot_pop_all, tot_pop_last,
		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);
		retval = 1;
	}
	/*
	 * The enqueuer should see exactly as many empty queues than the
	 * number of non-empty stacks dequeued.
	 */
	if (tot_empty_dest_enqueues != tot_pop_last) {
		printf("WARNING! Discrepancy between empty enqueue (%llu) and "
			"number of pop last (%llu)\n",
			tot_empty_dest_enqueues,
			tot_pop_last);
		retval = 1;
	}

	cds_wfs_destroy(&s);
	free(count_enqueuer);
	free(count_dequeuer);
	free(tid_enqueuer);
	free(tid_dequeuer);

	return retval;
}
Commit	Line	Data
ce29b371 MJ	1	// SPDX-FileCopyrightText: 2010 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
	6	/*
a5eb5ffc	7	* Userspace RCU library - example RCU-based lock-free stack
a5eb5ffc MD	8	*/
a5eb5ffc MD	9
a5eb5ffc 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>
a5eb5ffc MD	20	#include <errno.h>
	21
	22	#include <urcu/arch.h>
01477510	23	#include <urcu/assert.h>
bd252a04	24	#include <urcu/tls-compat.h>
96a0c0cd	25	#include <urcu/uatomic.h>
94df6318	26	#include "thread-id.h"
65fcc7e9	27
a5eb5ffc MD	28	/* hardcoded number of CPUs */
	29	#define NR_CPUS 16384
	30
a5eb5ffc MD	31	#ifndef DYNAMIC_LINK_TEST
	32	#define _LGPL_SOURCE
	33	#endif
cb3f3d6b	34	#include <urcu/wfstack.h>
a5eb5ffc	35
69328034 MD	36	/*
	37	* External synchronization used.
	38	*/
	39	enum test_sync {
	40	TEST_SYNC_NONE = 0,
	41	TEST_SYNC_MUTEX,
	42	};
	43
	44	static enum test_sync test_sync;
	45
0ee400bd MD	46	static int test_force_sync;
0ee400bd MD	47
9e4e7ad1	48	static volatile int test_stop_enqueue, test_stop_dequeue;
a5eb5ffc MD	49
	50	static unsigned long rduration;
	51
	52	static unsigned long duration;
	53
	54	/* read-side C.S. duration, in loops */
	55	static unsigned long wdelay;
	56
ab0aacbe	57	static inline void loop_sleep(unsigned long loops)
a5eb5ffc	58	{
ab0aacbe	59	while (loops-- != 0)
06f22bdb	60	caa_cpu_relax();
a5eb5ffc MD	61	}
	62
	63	static int verbose_mode;
	64
d05e9936	65	static int test_pop, test_pop_all, test_wait_empty;
83e334d0	66	static unsigned int test_enqueue_stopped;
69328034	67
a5eb5ffc MD	68	#define printf_verbose(fmt, args...) \
	69	do { \
	70	if (verbose_mode) \
69328034	71	printf(fmt, ## args); \
a5eb5ffc MD	72	} while (0)
	73
	74	static unsigned int cpu_affinities[NR_CPUS];
	75	static unsigned int next_aff = 0;
	76	static int use_affinity = 0;
	77
	78	pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
	79
a5eb5ffc MD	80	static void set_affinity(void)
a5eb5ffc MD	81	{
2388c075	82	#ifdef HAVE_SCHED_SETAFFINITY
a5eb5ffc	83	cpu_set_t mask;
95bc7fb9 MD	84	int cpu, ret;
95bc7fb9 MD	85	#endif /* HAVE_SCHED_SETAFFINITY */
a5eb5ffc MD	86
	87	if (!use_affinity)
	88	return;
	89
2388c075	90	#ifdef HAVE_SCHED_SETAFFINITY
a5eb5ffc MD	91	ret = pthread_mutex_lock(&affinity_mutex);
	92	if (ret) {
	93	perror("Error in pthread mutex lock");
	94	exit(-1);
	95	}
	96	cpu = cpu_affinities[next_aff++];
	97	ret = pthread_mutex_unlock(&affinity_mutex);
	98	if (ret) {
	99	perror("Error in pthread mutex unlock");
	100	exit(-1);
	101	}
	102
	103	CPU_ZERO(&mask);
	104	CPU_SET(cpu, &mask);
a5eb5ffc	105	sched_setaffinity(0, sizeof(mask), &mask);
a5eb5ffc MD	106	#endif /* HAVE_SCHED_SETAFFINITY */
	107	}
	108
	109	/*
	110	* returns 0 if test should end.
	111	*/
	112	static int test_duration_dequeue(void)
	113	{
9e4e7ad1	114	return !uatomic_load(&test_stop_dequeue, CMM_RELAXED);
a5eb5ffc MD	115	}
	116
	117	static int test_duration_enqueue(void)
	118	{
9e4e7ad1	119	return !uatomic_load(&test_stop_enqueue, CMM_RELAXED);
a5eb5ffc MD	120	}
a5eb5ffc MD	121
bd252a04 MD	122	static DEFINE_URCU_TLS(unsigned long long, nr_dequeues);
bd252a04 MD	123	static DEFINE_URCU_TLS(unsigned long long, nr_enqueues);
a5eb5ffc	124
bd252a04 MD	125	static DEFINE_URCU_TLS(unsigned long long, nr_successful_dequeues);
bd252a04 MD	126	static DEFINE_URCU_TLS(unsigned long long, nr_successful_enqueues);
d05e9936 MD	127	static DEFINE_URCU_TLS(unsigned long long, nr_empty_dest_enqueues);
d05e9936 MD	128	static DEFINE_URCU_TLS(unsigned long long, nr_pop_all);
d89f0c51	129	static DEFINE_URCU_TLS(unsigned long long, nr_pop_last);
a5eb5ffc MD	130
	131	static unsigned int nr_enqueuers;
	132	static unsigned int nr_dequeuers;
	133
16aa9ee8	134	static struct cds_wfs_stack s;
a5eb5ffc	135
98fbc848	136	static void thr_enqueuer(void _count)
a5eb5ffc MD	137	{
a5eb5ffc MD	138	unsigned long long *count = _count;
d05e9936	139	bool was_nonempty;
a5eb5ffc	140
94df6318 MD	141	printf_verbose("thread_begin %s, tid %lu\n",
94df6318 MD	142	"enqueuer", urcu_get_thread_id());
a5eb5ffc MD	143
	144	set_affinity();
	145
9e4e7ad1	146	wait_until_go();
a5eb5ffc MD	147
a5eb5ffc MD	148	for (;;) {
16aa9ee8	149	struct cds_wfs_node node = malloc(sizeof(node));
a5eb5ffc MD	150	if (!node)
a5eb5ffc MD	151	goto fail;
16aa9ee8	152	cds_wfs_node_init(node);
d05e9936	153	was_nonempty = cds_wfs_push(&s, node);
bd252a04	154	URCU_TLS(nr_successful_enqueues)++;
d05e9936 MD	155	if (!was_nonempty)
d05e9936 MD	156	URCU_TLS(nr_empty_dest_enqueues)++;
a5eb5ffc	157
a0b7f7ea	158	if (caa_unlikely(wdelay))
a5eb5ffc MD	159	loop_sleep(wdelay);
a5eb5ffc MD	160	fail:
bd252a04	161	URCU_TLS(nr_enqueues)++;
a0b7f7ea	162	if (caa_unlikely(!test_duration_enqueue()))
a5eb5ffc MD	163	break;
	164	}
	165
d05e9936	166	uatomic_inc(&test_enqueue_stopped);
bd252a04 MD	167	count[0] = URCU_TLS(nr_enqueues);
bd252a04 MD	168	count[1] = URCU_TLS(nr_successful_enqueues);
d05e9936	169	count[2] = URCU_TLS(nr_empty_dest_enqueues);
94df6318 MD	170	printf_verbose("enqueuer thread_end, tid %lu, "
	171	"enqueues %llu successful_enqueues %llu, "
	172	"empty_dest_enqueues %llu\n",
	173	urcu_get_thread_id(),
	174	URCU_TLS(nr_enqueues),
	175	URCU_TLS(nr_successful_enqueues),
	176	URCU_TLS(nr_empty_dest_enqueues));
a5eb5ffc MD	177	return ((void*)1);
	178
	179	}
	180
69328034 MD	181	static void do_test_pop(enum test_sync sync)
	182	{
	183	struct cds_wfs_node *node;
d89f0c51	184	int state;
69328034 MD	185
	186	if (sync == TEST_SYNC_MUTEX)
	187	cds_wfs_pop_lock(&s);
d89f0c51	188	node = __cds_wfs_pop_with_state_blocking(&s, &state);
69328034 MD	189	if (sync == TEST_SYNC_MUTEX)
	190	cds_wfs_pop_unlock(&s);
	191
	192	if (node) {
d89f0c51 MD	193	if (state & CDS_WFS_STATE_LAST)
d89f0c51 MD	194	URCU_TLS(nr_pop_last)++;
69328034 MD	195	free(node);
	196	URCU_TLS(nr_successful_dequeues)++;
	197	}
	198	URCU_TLS(nr_dequeues)++;
	199	}
	200
	201	static void do_test_pop_all(enum test_sync sync)
	202	{
	203	struct cds_wfs_head *head;
	204	struct cds_wfs_node node, n;
	205
	206	if (sync == TEST_SYNC_MUTEX)
	207	cds_wfs_pop_lock(&s);
	208	head = __cds_wfs_pop_all(&s);
	209	if (sync == TEST_SYNC_MUTEX)
	210	cds_wfs_pop_unlock(&s);
	211
d05e9936	212	/* Check if empty */
c7ba06ba	213	if (cds_wfs_first(head) == NULL)
d05e9936 MD	214	return;
	215
	216	URCU_TLS(nr_pop_all)++;
d89f0c51	217	URCU_TLS(nr_pop_last)++;
d05e9936	218
69328034 MD	219	cds_wfs_for_each_blocking_safe(head, node, n) {
	220	free(node);
	221	URCU_TLS(nr_successful_dequeues)++;
	222	URCU_TLS(nr_dequeues)++;
	223	}
	224	}
	225
98fbc848	226	static void thr_dequeuer(void _count)
a5eb5ffc MD	227	{
a5eb5ffc MD	228	unsigned long long *count = _count;
f42bc7ef	229	unsigned int counter = 0;
a5eb5ffc	230
94df6318 MD	231	printf_verbose("thread_begin %s, tid %lu\n",
94df6318 MD	232	"dequeuer", urcu_get_thread_id());
a5eb5ffc MD	233
	234	set_affinity();
	235
9e4e7ad1	236	wait_until_go();
a5eb5ffc	237
01477510	238	urcu_posix_assert(test_pop \|\| test_pop_all);
a5eb5ffc	239
69328034 MD	240	for (;;) {
	241	if (test_pop && test_pop_all) {
	242	if (counter & 1)
	243	do_test_pop(test_sync);
	244	else
	245	do_test_pop_all(test_sync);
	246	counter++;
	247	} else {
	248	if (test_pop)
	249	do_test_pop(test_sync);
	250	else
	251	do_test_pop_all(test_sync);
a5eb5ffc MD	252	}
a5eb5ffc MD	253
a0b7f7ea	254	if (caa_unlikely(!test_duration_dequeue()))
a5eb5ffc	255	break;
a0b7f7ea	256	if (caa_unlikely(rduration))
a5eb5ffc MD	257	loop_sleep(rduration);
	258	}
	259
94df6318 MD	260	printf_verbose("dequeuer thread_end, tid %lu, "
	261	"dequeues %llu, successful_dequeues %llu "
	262	"pop_all %llu pop_last %llu\n",
	263	urcu_get_thread_id(),
	264	URCU_TLS(nr_dequeues), URCU_TLS(nr_successful_dequeues),
	265	URCU_TLS(nr_pop_all),
	266	URCU_TLS(nr_pop_last));
bd252a04 MD	267	count[0] = URCU_TLS(nr_dequeues);
bd252a04 MD	268	count[1] = URCU_TLS(nr_successful_dequeues);
d05e9936	269	count[2] = URCU_TLS(nr_pop_all);
d89f0c51	270	count[3] = URCU_TLS(nr_pop_last);
a5eb5ffc MD	271	return ((void*)2);
	272	}
	273
bd23a6c6 MJ	274	static void test_end(unsigned long long *nr_dequeues_l,
bd23a6c6 MJ	275	unsigned long long *nr_pop_last_l)
a5eb5ffc	276	{
16aa9ee8	277	struct cds_wfs_node *node;
d89f0c51	278	int state;
a5eb5ffc MD	279
a5eb5ffc MD	280	do {
bd23a6c6	281	node = cds_wfs_pop_with_state_blocking(&s, &state);
a5eb5ffc	282	if (node) {
d89f0c51	283	if (state & CDS_WFS_STATE_LAST)
bd23a6c6	284	(*nr_pop_last_l)++;
a5eb5ffc	285	free(node);
bd23a6c6	286	(*nr_dequeues_l)++;
a5eb5ffc MD	287	}
	288	} while (node);
	289	}
	290
70469b43	291	static void show_usage(char **argv)
a5eb5ffc	292	{
06637138 MD	293	printf("Usage : %s nr_dequeuers nr_enqueuers duration (s) <OPTIONS>\n",
	294	argv[0]);
	295	printf("OPTIONS:\n");
	296	printf(" [-d delay] (enqueuer period (in loops))\n");
	297	printf(" [-c duration] (dequeuer period (in loops))\n");
	298	printf(" [-v] (verbose output)\n");
	299	printf(" [-a cpu#] [-a cpu#]... (affinity)\n");
	300	printf(" [-p] (test pop)\n");
	301	printf(" [-P] (test pop_all, enabled by default)\n");
	302	printf(" [-M] (use mutex external synchronization)\n");
	303	printf(" Note: default: no external synchronization used.\n");
	304	printf(" [-f] (force user-provided synchronization)\n");
	305	printf(" [-w] Wait for dequeuer to empty stack\n");
a5eb5ffc MD	306	printf("\n");
	307	}
	308
	309	int main(int argc, char **argv)
	310	{
	311	int err;
	312	pthread_t tid_enqueuer, tid_dequeuer;
	313	void *tret;
	314	unsigned long long count_enqueuer, count_dequeuer;
	315	unsigned long long tot_enqueues = 0, tot_dequeues = 0;
	316	unsigned long long tot_successful_enqueues = 0,
d05e9936 MD	317	tot_successful_dequeues = 0,
d05e9936 MD	318	tot_empty_dest_enqueues = 0,
d89f0c51	319	tot_pop_all = 0, tot_pop_last = 0;
a5eb5ffc	320	unsigned long long end_dequeues = 0;
d05e9936	321	int i, a, retval = 0;
83e334d0	322	unsigned int i_thr;
a5eb5ffc MD	323
a5eb5ffc MD	324	if (argc < 4) {
70469b43	325	show_usage(argv);
a5eb5ffc MD	326	return -1;
	327	}
	328
	329	err = sscanf(argv[1], "%u", &nr_dequeuers);
	330	if (err != 1) {
70469b43	331	show_usage(argv);
a5eb5ffc MD	332	return -1;
	333	}
	334
	335	err = sscanf(argv[2], "%u", &nr_enqueuers);
	336	if (err != 1) {
70469b43	337	show_usage(argv);
a5eb5ffc MD	338	return -1;
a5eb5ffc MD	339	}
83e334d0	340
a5eb5ffc MD	341	err = sscanf(argv[3], "%lu", &duration);
a5eb5ffc MD	342	if (err != 1) {
70469b43	343	show_usage(argv);
a5eb5ffc MD	344	return -1;
	345	}
	346
	347	for (i = 4; i < argc; i++) {
	348	if (argv[i][0] != '-')
	349	continue;
	350	switch (argv[i][1]) {
	351	case 'a':
	352	if (argc < i + 2) {
70469b43	353	show_usage(argv);
a5eb5ffc MD	354	return -1;
	355	}
	356	a = atoi(argv[++i]);
	357	cpu_affinities[next_aff++] = a;
	358	use_affinity = 1;
	359	printf_verbose("Adding CPU %d affinity\n", a);
	360	break;
	361	case 'c':
	362	if (argc < i + 2) {
70469b43	363	show_usage(argv);
a5eb5ffc MD	364	return -1;
	365	}
	366	rduration = atol(argv[++i]);
	367	break;
	368	case 'd':
	369	if (argc < i + 2) {
70469b43	370	show_usage(argv);
a5eb5ffc MD	371	return -1;
	372	}
	373	wdelay = atol(argv[++i]);
	374	break;
	375	case 'v':
	376	verbose_mode = 1;
	377	break;
69328034 MD	378	case 'p':
	379	test_pop = 1;
	380	break;
	381	case 'P':
	382	test_pop_all = 1;
	383	break;
	384	case 'M':
	385	test_sync = TEST_SYNC_MUTEX;
	386	break;
d05e9936 MD	387	case 'w':
	388	test_wait_empty = 1;
	389	break;
0ee400bd MD	390	case 'f':
	391	test_force_sync = 1;
	392	break;
a5eb5ffc MD	393	}
	394	}
	395
69328034 MD	396	/* activate pop_all test by default */
	397	if (!test_pop && !test_pop_all)
	398	test_pop_all = 1;
	399
0ee400bd MD	400	if (test_sync == TEST_SYNC_NONE && nr_dequeuers > 1 && test_pop) {
	401	if (test_force_sync) {
	402	fprintf(stderr, "[WARNING] Using pop concurrently "
	403	"with other pop or pop_all without external "
	404	"synchronization. Expect run-time failure.\n");
	405	} else {
	406	printf("Enforcing mutex synchronization\n");
	407	test_sync = TEST_SYNC_MUTEX;
	408	}
	409	}
	410
a5eb5ffc MD	411	printf_verbose("running test for %lu seconds, %u enqueuers, "
	412	"%u dequeuers.\n",
	413	duration, nr_enqueuers, nr_dequeuers);
69328034 MD	414	if (test_pop)
	415	printf_verbose("pop test activated.\n");
	416	if (test_pop_all)
	417	printf_verbose("pop_all test activated.\n");
98fbc848 MD	418	if (test_sync == TEST_SYNC_MUTEX)
	419	printf_verbose("External sync: mutex.\n");
	420	else
	421	printf_verbose("External sync: none.\n");
d05e9936 MD	422	if (test_wait_empty)
d05e9936 MD	423	printf_verbose("Wait for dequeuers to empty stack.\n");
a5eb5ffc MD	424	printf_verbose("Writer delay : %lu loops.\n", rduration);
a5eb5ffc MD	425	printf_verbose("Reader duration : %lu loops.\n", wdelay);
94df6318 MD	426	printf_verbose("thread %-6s, tid %lu\n",
94df6318 MD	427	"main", urcu_get_thread_id());
a5eb5ffc	428
9aa14175 MD	429	tid_enqueuer = calloc(nr_enqueuers, sizeof(*tid_enqueuer));
	430	tid_dequeuer = calloc(nr_dequeuers, sizeof(*tid_dequeuer));
	431	count_enqueuer = calloc(nr_enqueuers, 3 * sizeof(*count_enqueuer));
	432	count_dequeuer = calloc(nr_dequeuers, 4 * sizeof(*count_dequeuer));
16aa9ee8	433	cds_wfs_init(&s);
a5eb5ffc MD	434
	435	next_aff = 0;
	436
83e334d0	437	for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
a5eb5ffc	438	err = pthread_create(&tid_enqueuer[i], NULL, thr_enqueuer,
83e334d0	439	&count_enqueuer[3 * i_thr]);
a5eb5ffc MD	440	if (err != 0)
	441	exit(1);
	442	}
83e334d0 MJ	443	for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
	444	err = pthread_create(&tid_dequeuer[i_thr], NULL, thr_dequeuer,
	445	&count_dequeuer[4 * i_thr]);
a5eb5ffc MD	446	if (err != 0)
	447	exit(1);
	448	}
	449
5481ddb3	450	cmm_smp_mb();
a5eb5ffc	451
9e4e7ad1	452	begin_test();
a5eb5ffc	453
83e334d0	454	for (i_thr = 0; i_thr < duration; i_thr++) {
a5eb5ffc	455	sleep(1);
4b665350 MD	456	if (verbose_mode) {
	457	fwrite(".", sizeof(char), 1, stdout);
	458	fflush(stdout);
	459	}
a5eb5ffc MD	460	}
a5eb5ffc MD	461
9e4e7ad1	462	uatomic_store(&test_stop_enqueue, 1, CMM_RELEASE);
d05e9936 MD	463
	464	if (test_wait_empty) {
	465	while (nr_enqueuers != uatomic_read(&test_enqueue_stopped)) {
	466	sleep(1);
	467	}
	468	while (!cds_wfs_empty(&s)) {
	469	sleep(1);
	470	}
	471	}
	472
9e4e7ad1	473	uatomic_store(&test_stop_dequeue, 1, CMM_RELAXED);
a5eb5ffc	474
83e334d0 MJ	475	for (i_thr = 0; i_thr < nr_enqueuers; i_thr++) {
83e334d0 MJ	476	err = pthread_join(tid_enqueuer[i_thr], &tret);
a5eb5ffc MD	477	if (err != 0)
a5eb5ffc MD	478	exit(1);
83e334d0 MJ	479	tot_enqueues += count_enqueuer[3 * i_thr];
	480	tot_successful_enqueues += count_enqueuer[3 * i_thr + 1];
	481	tot_empty_dest_enqueues += count_enqueuer[3 * i_thr + 2];
a5eb5ffc	482	}
83e334d0 MJ	483	for (i_thr = 0; i_thr < nr_dequeuers; i_thr++) {
83e334d0 MJ	484	err = pthread_join(tid_dequeuer[i_thr], &tret);
a5eb5ffc MD	485	if (err != 0)
a5eb5ffc MD	486	exit(1);
83e334d0 MJ	487	tot_dequeues += count_dequeuer[4 * i_thr];
	488	tot_successful_dequeues += count_dequeuer[4 * i_thr + 1];
	489	tot_pop_all += count_dequeuer[4 * i_thr + 2];
	490	tot_pop_last += count_dequeuer[4 * i_thr + 3];
a5eb5ffc	491	}
83e334d0	492
bd23a6c6	493	test_end(&end_dequeues, &tot_pop_last);
a5eb5ffc MD	494
	495	printf_verbose("total number of enqueues : %llu, dequeues %llu\n",
	496	tot_enqueues, tot_dequeues);
	497	printf_verbose("total number of successful enqueues : %llu, "
d05e9936 MD	498	"enqueues to empty dest : %llu, "
d05e9936 MD	499	"successful dequeues %llu, "
d89f0c51	500	"pop_all : %llu, pop_last : %llu\n",
d05e9936 MD	501	tot_successful_enqueues,
	502	tot_empty_dest_enqueues,
	503	tot_successful_dequeues,
d89f0c51	504	tot_pop_all, tot_pop_last);
a5eb5ffc MD	505	printf("SUMMARY %-25s testdur %4lu nr_enqueuers %3u wdelay %6lu "
	506	"nr_dequeuers %3u "
	507	"rdur %6lu nr_enqueues %12llu nr_dequeues %12llu "
d05e9936 MD	508	"successful enqueues %12llu enqueues to empty dest %12llu "
d05e9936 MD	509	"successful dequeues %12llu pop_all %12llu "
d89f0c51	510	"pop_last %llu end_dequeues %llu nr_ops %12llu\n",
a5eb5ffc MD	511	argv[0], duration, nr_enqueuers, wdelay,
	512	nr_dequeuers, rduration, tot_enqueues, tot_dequeues,
	513	tot_successful_enqueues,
d05e9936	514	tot_empty_dest_enqueues,
d89f0c51 MD	515	tot_successful_dequeues, tot_pop_all, tot_pop_last,
d89f0c51 MD	516	end_dequeues,
a5eb5ffc	517	tot_enqueues + tot_dequeues);
d05e9936	518	if (tot_successful_enqueues != tot_successful_dequeues + end_dequeues) {
a5eb5ffc MD	519	printf("WARNING! Discrepancy between nr succ. enqueues %llu vs "
	520	"succ. dequeues + end dequeues %llu.\n",
	521	tot_successful_enqueues,
	522	tot_successful_dequeues + end_dequeues);
d05e9936 MD	523	retval = 1;
	524	}
	525	/*
d89f0c51 MD	526	* The enqueuer should see exactly as many empty queues than the
d89f0c51 MD	527	* number of non-empty stacks dequeued.
d05e9936	528	*/
d89f0c51	529	if (tot_empty_dest_enqueues != tot_pop_last) {
d05e9936	530	printf("WARNING! Discrepancy between empty enqueue (%llu) and "
d89f0c51	531	"number of pop last (%llu)\n",
d05e9936	532	tot_empty_dest_enqueues,
d89f0c51	533	tot_pop_last);
d05e9936 MD	534	retval = 1;
d05e9936 MD	535	}
83e334d0	536
2af1c19e	537	cds_wfs_destroy(&s);
a5eb5ffc MD	538	free(count_enqueuer);
	539	free(count_dequeuer);
	540	free(tid_enqueuer);
	541	free(tid_dequeuer);
83e334d0	542
d05e9936	543	return retval;
a5eb5ffc	544	}