[linux-block.git] / tools / perf / bench / futex-requeue.c

// SPDX-License-Identifier: GPL-2.0
/*
 * Copyright (C) 2013  Davidlohr Bueso <davidlohr@hp.com>
 *
 * futex-requeue: Block a bunch of threads on futex1 and requeue them
 *                on futex2, N at a time.
 *
 * This program is particularly useful to measure the latency of nthread
 * requeues without waking up any tasks (in the non-pi case) -- thus
 * mimicking a regular futex_wait.
 */

/* For the CLR_() macros */
#include <string.h>
#include <pthread.h>

#include <signal.h>
#include "../util/mutex.h"
#include "../util/stat.h"
#include <subcmd/parse-options.h>
#include <linux/compiler.h>
#include <linux/kernel.h>
#include <linux/time64.h>
#include <errno.h>
#include <perf/cpumap.h>
#include "bench.h"
#include "futex.h"

#include <err.h>
#include <stdlib.h>
#include <sys/time.h>
#include <sys/mman.h>

static u_int32_t futex1 = 0, futex2 = 0;

static pthread_t *worker;
static bool done = false;
static struct mutex thread_lock;
static struct cond thread_parent, thread_worker;
static struct stats requeuetime_stats, requeued_stats;
static unsigned int threads_starting;
static int futex_flag = 0;

static struct bench_futex_parameters params = {
	/*
	 * How many tasks to requeue at a time.
	 * Default to 1 in order to make the kernel work more.
	 */
	.nrequeue = 1,
};

static const struct option options[] = {
	OPT_UINTEGER('t', "threads",  &params.nthreads, "Specify amount of threads"),
	OPT_UINTEGER('q', "nrequeue", &params.nrequeue, "Specify amount of threads to requeue at once"),
	OPT_BOOLEAN( 's', "silent",   &params.silent, "Silent mode: do not display data/details"),
	OPT_BOOLEAN( 'S', "shared",   &params.fshared, "Use shared futexes instead of private ones"),
	OPT_BOOLEAN( 'm', "mlockall", &params.mlockall, "Lock all current and future memory"),
	OPT_BOOLEAN( 'B', "broadcast", &params.broadcast, "Requeue all threads at once"),
	OPT_BOOLEAN( 'p', "pi", &params.pi, "Use PI-aware variants of FUTEX_CMP_REQUEUE"),

	OPT_END()
};

static const char * const bench_futex_requeue_usage[] = {
	"perf bench futex requeue <options>",
	NULL
};

static void print_summary(void)
{
	double requeuetime_avg = avg_stats(&requeuetime_stats);
	double requeuetime_stddev = stddev_stats(&requeuetime_stats);
	unsigned int requeued_avg = avg_stats(&requeued_stats);

	printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n",
	       requeued_avg,
	       params.nthreads,
	       requeuetime_avg / USEC_PER_MSEC,
	       rel_stddev_stats(requeuetime_stddev, requeuetime_avg));
}

static void *workerfn(void *arg __maybe_unused)
{
	int ret;

	mutex_lock(&thread_lock);
	threads_starting--;
	if (!threads_starting)
		cond_signal(&thread_parent);
	cond_wait(&thread_worker, &thread_lock);
	mutex_unlock(&thread_lock);

	while (1) {
		if (!params.pi) {
			ret = futex_wait(&futex1, 0, NULL, futex_flag);
			if (!ret)
				break;

			if (ret && errno != EAGAIN) {
				if (!params.silent)
					warnx("futex_wait");
				break;
			}
		} else {
			ret = futex_wait_requeue_pi(&futex1, 0, &futex2,
						    NULL, futex_flag);
			if (!ret) {
				/* got the lock at futex2 */
				futex_unlock_pi(&futex2, futex_flag);
				break;
			}

			if (ret && errno != EAGAIN) {
				if (!params.silent)
					warnx("futex_wait_requeue_pi");
				break;
			}
		}
	}

	return NULL;
}

static void block_threads(pthread_t *w,
			  pthread_attr_t thread_attr, struct perf_cpu_map *cpu)
{
	cpu_set_t *cpuset;
	unsigned int i;
	int nrcpus = perf_cpu_map__nr(cpu);
	size_t size;

	threads_starting = params.nthreads;

	cpuset = CPU_ALLOC(nrcpus);
	BUG_ON(!cpuset);
	size = CPU_ALLOC_SIZE(nrcpus);

	/* create and block all threads */
	for (i = 0; i < params.nthreads; i++) {
		CPU_ZERO_S(size, cpuset);
		CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset);

		if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) {
			CPU_FREE(cpuset);
			err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
		}

		if (pthread_create(&w[i], &thread_attr, workerfn, NULL)) {
			CPU_FREE(cpuset);
			err(EXIT_FAILURE, "pthread_create");
		}
	}
	CPU_FREE(cpuset);
}

static void toggle_done(int sig __maybe_unused,
			siginfo_t *info __maybe_unused,
			void *uc __maybe_unused)
{
	done = true;
}

int bench_futex_requeue(int argc, const char **argv)
{
	int ret = 0;
	unsigned int i, j;
	struct sigaction act;
	pthread_attr_t thread_attr;
	struct perf_cpu_map *cpu;

	argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0);
	if (argc)
		goto err;

	cpu = perf_cpu_map__new(NULL);
	if (!cpu)
		err(EXIT_FAILURE, "cpu_map__new");

	memset(&act, 0, sizeof(act));
	sigfillset(&act.sa_mask);
	act.sa_sigaction = toggle_done;
	sigaction(SIGINT, &act, NULL);

	if (params.mlockall) {
		if (mlockall(MCL_CURRENT | MCL_FUTURE))
			err(EXIT_FAILURE, "mlockall");
	}

	if (!params.nthreads)
		params.nthreads = perf_cpu_map__nr(cpu);

	worker = calloc(params.nthreads, sizeof(*worker));
	if (!worker)
		err(EXIT_FAILURE, "calloc");

	if (!params.fshared)
		futex_flag = FUTEX_PRIVATE_FLAG;

	if (params.nrequeue > params.nthreads)
		params.nrequeue = params.nthreads;

	if (params.broadcast)
		params.nrequeue = params.nthreads;

	printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %s%p), "
	       "%d at a time.\n\n",  getpid(), params.nthreads,
	       params.fshared ? "shared":"private", &futex1,
	       params.pi ? "PI ": "", &futex2, params.nrequeue);

	init_stats(&requeued_stats);
	init_stats(&requeuetime_stats);
	pthread_attr_init(&thread_attr);
	mutex_init(&thread_lock);
	cond_init(&thread_parent);
	cond_init(&thread_worker);

	for (j = 0; j < bench_repeat && !done; j++) {
		unsigned int nrequeued = 0, wakeups = 0;
		struct timeval start, end, runtime;

		/* create, launch & block all threads */
		block_threads(worker, thread_attr, cpu);

		/* make sure all threads are already blocked */
		mutex_lock(&thread_lock);
		while (threads_starting)
			cond_wait(&thread_parent, &thread_lock);
		cond_broadcast(&thread_worker);
		mutex_unlock(&thread_lock);

		usleep(100000);

		/* Ok, all threads are patiently blocked, start requeueing */
		gettimeofday(&start, NULL);
		while (nrequeued < params.nthreads) {
			int r;

			/*
			 * For the regular non-pi case, do not wakeup any tasks
			 * blocked on futex1, allowing us to really measure
			 * futex_wait functionality. For the PI case the first
			 * waiter is always awoken.
			 */
			if (!params.pi) {
				r = futex_cmp_requeue(&futex1, 0, &futex2, 0,
						      params.nrequeue,
						      futex_flag);
			} else {
				r = futex_cmp_requeue_pi(&futex1, 0, &futex2,
							 params.nrequeue,
							 futex_flag);
				wakeups++; /* assume no error */
			}

			if (r < 0)
				err(EXIT_FAILURE, "couldn't requeue from %p to %p",
				    &futex1, &futex2);

			nrequeued += r;
		}

		gettimeofday(&end, NULL);
		timersub(&end, &start, &runtime);

		update_stats(&requeued_stats, nrequeued);
		update_stats(&requeuetime_stats, runtime.tv_usec);

		if (!params.silent) {
			if (!params.pi)
				printf("[Run %d]: Requeued %d of %d threads in "
				       "%.4f ms\n", j + 1, nrequeued,
				       params.nthreads,
				       runtime.tv_usec / (double)USEC_PER_MSEC);
			else {
				nrequeued -= wakeups;
				printf("[Run %d]: Awoke and Requeued (%d+%d) of "
				       "%d threads in %.4f ms\n",
				       j + 1, wakeups, nrequeued,
				       params.nthreads,
				       runtime.tv_usec / (double)USEC_PER_MSEC);
			}

		}

		if (!params.pi) {
			/* everybody should be blocked on futex2, wake'em up */
			nrequeued = futex_wake(&futex2, nrequeued, futex_flag);
			if (params.nthreads != nrequeued)
				warnx("couldn't wakeup all tasks (%d/%d)",
				      nrequeued, params.nthreads);
		}

		for (i = 0; i < params.nthreads; i++) {
			ret = pthread_join(worker[i], NULL);
			if (ret)
				err(EXIT_FAILURE, "pthread_join");
		}
	}

	/* cleanup & report results */
	cond_destroy(&thread_parent);
	cond_destroy(&thread_worker);
	mutex_destroy(&thread_lock);
	pthread_attr_destroy(&thread_attr);

	print_summary();

	free(worker);
	perf_cpu_map__put(cpu);
	return ret;
err:
	usage_with_options(bench_futex_requeue_usage, options);
	exit(EXIT_FAILURE);
}
Commit	Line	Data
b2441318	1	// SPDX-License-Identifier: GPL-2.0
0fb298cf DB	2	/*
	3	* Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com>
	4	*
	5	* futex-requeue: Block a bunch of threads on futex1 and requeue them
	6	* on futex2, N at a time.
	7	*
	8	* This program is particularly useful to measure the latency of nthread
46f81532 DB	9	* requeues without waking up any tasks (in the non-pi case) -- thus
46f81532 DB	10	* mimicking a regular futex_wait.
0fb298cf DB	11	*/
0fb298cf DB	12
8a158589	13	/* For the CLR_() macros */
a0f213e1	14	#include <string.h>
8a158589 ACM	15	#include <pthread.h>
8a158589 ACM	16
9c304f6c	17	#include <signal.h>
a64d3af5	18	#include "../util/mutex.h"
0fb298cf	19	#include "../util/stat.h"
4b6ab94e	20	#include <subcmd/parse-options.h>
86695f59	21	#include <linux/compiler.h>
9c304f6c	22	#include <linux/kernel.h>
565e6911	23	#include <linux/time64.h>
9c304f6c	24	#include <errno.h>
9c3516d1	25	#include <perf/cpumap.h>
0fb298cf DB	26	#include "bench.h"
	27	#include "futex.h"
	28
	29	#include <err.h>
	30	#include <stdlib.h>
	31	#include <sys/time.h>
9f9a3ffe	32	#include <sys/mman.h>
0fb298cf DB	33
	34	static u_int32_t futex1 = 0, futex2 = 0;
	35
0fb298cf	36	static pthread_t *worker;
09590463	37	static bool done = false;
a64d3af5 IR	38	static struct mutex thread_lock;
a64d3af5 IR	39	static struct cond thread_parent, thread_worker;
0fb298cf	40	static struct stats requeuetime_stats, requeued_stats;
09590463	41	static unsigned int threads_starting;
86c87e13	42	static int futex_flag = 0;
0fb298cf	43
09590463 DB	44	static struct bench_futex_parameters params = {
	45	/*
	46	* How many tasks to requeue at a time.
	47	* Default to 1 in order to make the kernel work more.
	48	*/
	49	.nrequeue = 1,
	50	};
	51
0fb298cf	52	static const struct option options[] = {
09590463 DB	53	OPT_UINTEGER('t', "threads", &params.nthreads, "Specify amount of threads"),
	54	OPT_UINTEGER('q', "nrequeue", &params.nrequeue, "Specify amount of threads to requeue at once"),
	55	OPT_BOOLEAN( 's', "silent", &params.silent, "Silent mode: do not display data/details"),
	56	OPT_BOOLEAN( 'S', "shared", &params.fshared, "Use shared futexes instead of private ones"),
9f9a3ffe	57	OPT_BOOLEAN( 'm', "mlockall", &params.mlockall, "Lock all current and future memory"),
d262e6a9	58	OPT_BOOLEAN( 'B', "broadcast", &params.broadcast, "Requeue all threads at once"),
46f81532 DB	59	OPT_BOOLEAN( 'p', "pi", &params.pi, "Use PI-aware variants of FUTEX_CMP_REQUEUE"),
46f81532 DB	60
0fb298cf DB	61	OPT_END()
	62	};
	63
	64	static const char * const bench_futex_requeue_usage[] = {
	65	"perf bench futex requeue <options>",
	66	NULL
	67	};
	68
	69	static void print_summary(void)
	70	{
	71	double requeuetime_avg = avg_stats(&requeuetime_stats);
	72	double requeuetime_stddev = stddev_stats(&requeuetime_stats);
	73	unsigned int requeued_avg = avg_stats(&requeued_stats);
	74
	75	printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n",
	76	requeued_avg,
09590463	77	params.nthreads,
565e6911	78	requeuetime_avg / USEC_PER_MSEC,
0fb298cf DB	79	rel_stddev_stats(requeuetime_stddev, requeuetime_avg));
	80	}
	81
	82	static void workerfn(void arg __maybe_unused)
	83	{
6f9661b2 DB	84	int ret;
6f9661b2 DB	85
a64d3af5	86	mutex_lock(&thread_lock);
0fb298cf DB	87	threads_starting--;
0fb298cf DB	88	if (!threads_starting)
a64d3af5 IR	89	cond_signal(&thread_parent);
	90	cond_wait(&thread_worker, &thread_lock);
	91	mutex_unlock(&thread_lock);
0fb298cf	92
6f9661b2	93	while (1) {
46f81532 DB	94	if (!params.pi) {
	95	ret = futex_wait(&futex1, 0, NULL, futex_flag);
	96	if (!ret)
	97	break;
	98
	99	if (ret && errno != EAGAIN) {
	100	if (!params.silent)
	101	warnx("futex_wait");
	102	break;
	103	}
	104	} else {
	105	ret = futex_wait_requeue_pi(&futex1, 0, &futex2,
	106	NULL, futex_flag);
	107	if (!ret) {
	108	/* got the lock at futex2 */
	109	futex_unlock_pi(&futex2, futex_flag);
	110	break;
	111	}
	112
	113	if (ret && errno != EAGAIN) {
	114	if (!params.silent)
	115	warnx("futex_wait_requeue_pi");
	116	break;
	117	}
6f9661b2 DB	118	}
	119	}
	120
0fb298cf DB	121	return NULL;
	122	}
	123
	124	static void block_threads(pthread_t *w,
f854839b	125	pthread_attr_t thread_attr, struct perf_cpu_map *cpu)
0fb298cf	126	{
c9c2a427	127	cpu_set_t *cpuset;
0fb298cf	128	unsigned int i;
c9c2a427 AR	129	int nrcpus = perf_cpu_map__nr(cpu);
c9c2a427 AR	130	size_t size;
0fb298cf	131
09590463	132	threads_starting = params.nthreads;
0fb298cf	133
c9c2a427 AR	134	cpuset = CPU_ALLOC(nrcpus);
	135	BUG_ON(!cpuset);
	136	size = CPU_ALLOC_SIZE(nrcpus);
	137
0fb298cf	138	/* create and block all threads */
09590463	139	for (i = 0; i < params.nthreads; i++) {
c9c2a427 AR	140	CPU_ZERO_S(size, cpuset);
c9c2a427 AR	141	CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset);
0fb298cf	142
c9c2a427 AR	143	if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) {
c9c2a427 AR	144	CPU_FREE(cpuset);
0fb298cf	145	err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
c9c2a427	146	}
0fb298cf	147
c9c2a427 AR	148	if (pthread_create(&w[i], &thread_attr, workerfn, NULL)) {
c9c2a427 AR	149	CPU_FREE(cpuset);
0fb298cf	150	err(EXIT_FAILURE, "pthread_create");
c9c2a427	151	}
0fb298cf	152	}
c9c2a427	153	CPU_FREE(cpuset);
0fb298cf DB	154	}
	155
	156	static void toggle_done(int sig __maybe_unused,
	157	siginfo_t *info __maybe_unused,
	158	void *uc __maybe_unused)
	159	{
	160	done = true;
	161	}
	162
b0ad8ea6	163	int bench_futex_requeue(int argc, const char **argv)
0fb298cf DB	164	{
	165	int ret = 0;
	166	unsigned int i, j;
	167	struct sigaction act;
	168	pthread_attr_t thread_attr;
f854839b	169	struct perf_cpu_map *cpu;
0fb298cf DB	170
	171	argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0);
	172	if (argc)
	173	goto err;
	174
9c3516d1	175	cpu = perf_cpu_map__new(NULL);
3b2323c2 DB	176	if (!cpu)
3b2323c2 DB	177	err(EXIT_FAILURE, "cpu_map__new");
0fb298cf	178
7b919a53	179	memset(&act, 0, sizeof(act));
0fb298cf DB	180	sigfillset(&act.sa_mask);
	181	act.sa_sigaction = toggle_done;
	182	sigaction(SIGINT, &act, NULL);
	183
9f9a3ffe DB	184	if (params.mlockall) {
	185	if (mlockall(MCL_CURRENT \| MCL_FUTURE))
	186	err(EXIT_FAILURE, "mlockall");
	187	}
	188
09590463	189	if (!params.nthreads)
44028699	190	params.nthreads = perf_cpu_map__nr(cpu);
0fb298cf	191
09590463	192	worker = calloc(params.nthreads, sizeof(*worker));
0fb298cf DB	193	if (!worker)
	194	err(EXIT_FAILURE, "calloc");
	195
09590463	196	if (!params.fshared)
86c87e13 DB	197	futex_flag = FUTEX_PRIVATE_FLAG;
86c87e13 DB	198
09590463 DB	199	if (params.nrequeue > params.nthreads)
09590463 DB	200	params.nrequeue = params.nthreads;
052b0f6e	201
d262e6a9 DB	202	if (params.broadcast)
	203	params.nrequeue = params.nthreads;
	204
46f81532	205	printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %s%p), "
09590463	206	"%d at a time.\n\n", getpid(), params.nthreads,
46f81532 DB	207	params.fshared ? "shared":"private", &futex1,
46f81532 DB	208	params.pi ? "PI ": "", &futex2, params.nrequeue);
0fb298cf DB	209
	210	init_stats(&requeued_stats);
	211	init_stats(&requeuetime_stats);
	212	pthread_attr_init(&thread_attr);
a64d3af5 IR	213	mutex_init(&thread_lock);
	214	cond_init(&thread_parent);
	215	cond_init(&thread_worker);
0fb298cf	216
d9de84af	217	for (j = 0; j < bench_repeat && !done; j++) {
46f81532	218	unsigned int nrequeued = 0, wakeups = 0;
0fb298cf DB	219	struct timeval start, end, runtime;
	220
	221	/* create, launch & block all threads */
3b2323c2	222	block_threads(worker, thread_attr, cpu);
0fb298cf DB	223
0fb298cf DB	224	/* make sure all threads are already blocked */
a64d3af5	225	mutex_lock(&thread_lock);
0fb298cf	226	while (threads_starting)
a64d3af5 IR	227	cond_wait(&thread_parent, &thread_lock);
	228	cond_broadcast(&thread_worker);
	229	mutex_unlock(&thread_lock);
0fb298cf DB	230
	231	usleep(100000);
	232
	233	/* Ok, all threads are patiently blocked, start requeueing */
	234	gettimeofday(&start, NULL);
09590463	235	while (nrequeued < params.nthreads) {
46f81532 DB	236	int r;
46f81532 DB	237
0fb298cf	238	/*
46f81532 DB	239	* For the regular non-pi case, do not wakeup any tasks
	240	* blocked on futex1, allowing us to really measure
	241	* futex_wait functionality. For the PI case the first
	242	* waiter is always awoken.
0fb298cf	243	*/
46f81532 DB	244	if (!params.pi) {
	245	r = futex_cmp_requeue(&futex1, 0, &futex2, 0,
	246	params.nrequeue,
	247	futex_flag);
	248	} else {
	249	r = futex_cmp_requeue_pi(&futex1, 0, &futex2,
	250	params.nrequeue,
	251	futex_flag);
	252	wakeups++; /* assume no error */
	253	}
	254
	255	if (r < 0)
	256	err(EXIT_FAILURE, "couldn't requeue from %p to %p",
	257	&futex1, &futex2);
	258
	259	nrequeued += r;
86c87e13	260	}
052b0f6e	261
0fb298cf DB	262	gettimeofday(&end, NULL);
	263	timersub(&end, &start, &runtime);
	264
	265	update_stats(&requeued_stats, nrequeued);
	266	update_stats(&requeuetime_stats, runtime.tv_usec);
	267
09590463	268	if (!params.silent) {
46f81532 DB	269	if (!params.pi)
	270	printf("[Run %d]: Requeued %d of %d threads in "
	271	"%.4f ms\n", j + 1, nrequeued,
	272	params.nthreads,
	273	runtime.tv_usec / (double)USEC_PER_MSEC);
	274	else {
	275	nrequeued -= wakeups;
	276	printf("[Run %d]: Awoke and Requeued (%d+%d) of "
	277	"%d threads in %.4f ms\n",
	278	j + 1, wakeups, nrequeued,
	279	params.nthreads,
	280	runtime.tv_usec / (double)USEC_PER_MSEC);
	281	}
	282
0fb298cf DB	283	}
0fb298cf DB	284
46f81532 DB	285	if (!params.pi) {
	286	/* everybody should be blocked on futex2, wake'em up */
	287	nrequeued = futex_wake(&futex2, nrequeued, futex_flag);
	288	if (params.nthreads != nrequeued)
	289	warnx("couldn't wakeup all tasks (%d/%d)",
	290	nrequeued, params.nthreads);
	291	}
0fb298cf	292
09590463	293	for (i = 0; i < params.nthreads; i++) {
0fb298cf DB	294	ret = pthread_join(worker[i], NULL);
	295	if (ret)
	296	err(EXIT_FAILURE, "pthread_join");
	297	}
0fb298cf DB	298	}
	299
	300	/* cleanup & report results */
a64d3af5 IR	301	cond_destroy(&thread_parent);
	302	cond_destroy(&thread_worker);
	303	mutex_destroy(&thread_lock);
0fb298cf DB	304	pthread_attr_destroy(&thread_attr);
	305
	306	print_summary();
	307
	308	free(worker);
88e48238	309	perf_cpu_map__put(cpu);
0fb298cf DB	310	return ret;
	311	err:
	312	usage_with_options(bench_futex_requeue_usage, options);
	313	exit(EXIT_FAILURE);
	314	}