[linux-2.6-block.git] / tools / perf / bench / futex-wake-parallel.c

// SPDX-License-Identifier: GPL-2.0
/*
 * Copyright (C) 2015 Davidlohr Bueso.
 *
 * Block a bunch of threads and let parallel waker threads wakeup an
 * equal amount of them. The program output reflects the avg latency
 * for each individual thread to service its share of work. Ultimately
 * it can be used to measure futex_wake() changes.
 */

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

#include <signal.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 "bench.h"
#include "futex.h"

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

struct thread_data {
	pthread_t worker;
	unsigned int nwoken;
	struct timeval runtime;
};

static unsigned int nwakes = 1;

/* all threads will block on the same futex -- hash bucket chaos ;) */
static u_int32_t futex = 0;

static pthread_t *blocked_worker;
static bool done = false, silent = false, fshared = false;
static unsigned int nblocked_threads = 0, nwaking_threads = 0;
static pthread_mutex_t thread_lock;
static pthread_cond_t thread_parent, thread_worker;
static struct stats waketime_stats, wakeup_stats;
static unsigned int ncpus, threads_starting;
static int futex_flag = 0;

static const struct option options[] = {
	OPT_UINTEGER('t', "threads", &nblocked_threads, "Specify amount of threads"),
	OPT_UINTEGER('w', "nwakers", &nwaking_threads, "Specify amount of waking threads"),
	OPT_BOOLEAN( 's', "silent",  &silent,   "Silent mode: do not display data/details"),
	OPT_BOOLEAN( 'S', "shared",  &fshared,  "Use shared futexes instead of private ones"),
	OPT_END()
};

static const char * const bench_futex_wake_parallel_usage[] = {
	"perf bench futex wake-parallel <options>",
	NULL
};

static void *waking_workerfn(void *arg)
{
	struct thread_data *waker = (struct thread_data *) arg;
	struct timeval start, end;

	gettimeofday(&start, NULL);

	waker->nwoken = futex_wake(&futex, nwakes, futex_flag);
	if (waker->nwoken != nwakes)
		warnx("couldn't wakeup all tasks (%d/%d)",
		      waker->nwoken, nwakes);

	gettimeofday(&end, NULL);
	timersub(&end, &start, &waker->runtime);

	pthread_exit(NULL);
	return NULL;
}

static void wakeup_threads(struct thread_data *td, pthread_attr_t thread_attr)
{
	unsigned int i;

	pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);

	/* create and block all threads */
	for (i = 0; i < nwaking_threads; i++) {
		/*
		 * Thread creation order will impact per-thread latency
		 * as it will affect the order to acquire the hb spinlock.
		 * For now let the scheduler decide.
		 */
		if (pthread_create(&td[i].worker, &thread_attr,
				   waking_workerfn, (void *)&td[i]))
			err(EXIT_FAILURE, "pthread_create");
	}

	for (i = 0; i < nwaking_threads; i++)
		if (pthread_join(td[i].worker, NULL))
			err(EXIT_FAILURE, "pthread_join");
}

static void *blocked_workerfn(void *arg __maybe_unused)
{
	pthread_mutex_lock(&thread_lock);
	threads_starting--;
	if (!threads_starting)
		pthread_cond_signal(&thread_parent);
	pthread_cond_wait(&thread_worker, &thread_lock);
	pthread_mutex_unlock(&thread_lock);

	while (1) { /* handle spurious wakeups */
		if (futex_wait(&futex, 0, NULL, futex_flag) != EINTR)
			break;
	}

	pthread_exit(NULL);
	return NULL;
}

static void block_threads(pthread_t *w, pthread_attr_t thread_attr)
{
	cpu_set_t cpu;
	unsigned int i;

	threads_starting = nblocked_threads;

	/* create and block all threads */
	for (i = 0; i < nblocked_threads; i++) {
		CPU_ZERO(&cpu);
		CPU_SET(i % ncpus, &cpu);

		if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu))
			err(EXIT_FAILURE, "pthread_attr_setaffinity_np");

		if (pthread_create(&w[i], &thread_attr, blocked_workerfn, NULL))
			err(EXIT_FAILURE, "pthread_create");
	}
}

static void print_run(struct thread_data *waking_worker, unsigned int run_num)
{
	unsigned int i, wakeup_avg;
	double waketime_avg, waketime_stddev;
	struct stats __waketime_stats, __wakeup_stats;

	init_stats(&__wakeup_stats);
	init_stats(&__waketime_stats);

	for (i = 0; i < nwaking_threads; i++) {
		update_stats(&__waketime_stats, waking_worker[i].runtime.tv_usec);
		update_stats(&__wakeup_stats, waking_worker[i].nwoken);
	}

	waketime_avg = avg_stats(&__waketime_stats);
	waketime_stddev = stddev_stats(&__waketime_stats);
	wakeup_avg = avg_stats(&__wakeup_stats);

	printf("[Run %d]: Avg per-thread latency (waking %d/%d threads) "
	       "in %.4f ms (+-%.2f%%)\n", run_num + 1, wakeup_avg,
	       nblocked_threads, waketime_avg / USEC_PER_MSEC,
	       rel_stddev_stats(waketime_stddev, waketime_avg));
}

static void print_summary(void)
{
	unsigned int wakeup_avg;
	double waketime_avg, waketime_stddev;

	waketime_avg = avg_stats(&waketime_stats);
	waketime_stddev = stddev_stats(&waketime_stats);
	wakeup_avg = avg_stats(&wakeup_stats);

	printf("Avg per-thread latency (waking %d/%d threads) in %.4f ms (+-%.2f%%)\n",
	       wakeup_avg,
	       nblocked_threads,
	       waketime_avg / USEC_PER_MSEC,
	       rel_stddev_stats(waketime_stddev, waketime_avg));
}


static void do_run_stats(struct thread_data *waking_worker)
{
	unsigned int i;

	for (i = 0; i < nwaking_threads; i++) {
		update_stats(&waketime_stats, waking_worker[i].runtime.tv_usec);
		update_stats(&wakeup_stats, waking_worker[i].nwoken);
	}

}

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

int bench_futex_wake_parallel(int argc, const char **argv)
{
	int ret = 0;
	unsigned int i, j;
	struct sigaction act;
	pthread_attr_t thread_attr;
	struct thread_data *waking_worker;

	argc = parse_options(argc, argv, options,
			     bench_futex_wake_parallel_usage, 0);
	if (argc) {
		usage_with_options(bench_futex_wake_parallel_usage, options);
		exit(EXIT_FAILURE);
	}

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

	ncpus = sysconf(_SC_NPROCESSORS_ONLN);
	if (!nblocked_threads)
		nblocked_threads = ncpus;

	/* some sanity checks */
	if (nwaking_threads > nblocked_threads || !nwaking_threads)
		nwaking_threads = nblocked_threads;

	if (nblocked_threads % nwaking_threads)
		errx(EXIT_FAILURE, "Must be perfectly divisible");
	/*
	 * Each thread will wakeup nwakes tasks in
	 * a single futex_wait call.
	 */
	nwakes = nblocked_threads/nwaking_threads;

	blocked_worker = calloc(nblocked_threads, sizeof(*blocked_worker));
	if (!blocked_worker)
		err(EXIT_FAILURE, "calloc");

	if (!fshared)
		futex_flag = FUTEX_PRIVATE_FLAG;

	printf("Run summary [PID %d]: blocking on %d threads (at [%s] "
	       "futex %p), %d threads waking up %d at a time.\n\n",
	       getpid(), nblocked_threads, fshared ? "shared":"private",
	       &futex, nwaking_threads, nwakes);

	init_stats(&wakeup_stats);
	init_stats(&waketime_stats);

	pthread_attr_init(&thread_attr);
	pthread_mutex_init(&thread_lock, NULL);
	pthread_cond_init(&thread_parent, NULL);
	pthread_cond_init(&thread_worker, NULL);

	for (j = 0; j < bench_repeat && !done; j++) {
		waking_worker = calloc(nwaking_threads, sizeof(*waking_worker));
		if (!waking_worker)
			err(EXIT_FAILURE, "calloc");

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

		/* make sure all threads are already blocked */
		pthread_mutex_lock(&thread_lock);
		while (threads_starting)
			pthread_cond_wait(&thread_parent, &thread_lock);
		pthread_cond_broadcast(&thread_worker);
		pthread_mutex_unlock(&thread_lock);

		usleep(100000);

		/* Ok, all threads are patiently blocked, start waking folks up */
		wakeup_threads(waking_worker, thread_attr);

		for (i = 0; i < nblocked_threads; i++) {
			ret = pthread_join(blocked_worker[i], NULL);
			if (ret)
				err(EXIT_FAILURE, "pthread_join");
		}

		do_run_stats(waking_worker);
		if (!silent)
			print_run(waking_worker, j);

		free(waking_worker);
	}

	/* cleanup & report results */
	pthread_cond_destroy(&thread_parent);
	pthread_cond_destroy(&thread_worker);
	pthread_mutex_destroy(&thread_lock);
	pthread_attr_destroy(&thread_attr);

	print_summary();

	free(blocked_worker);
	return ret;
}
Commit	Line	Data
b2441318	1	// SPDX-License-Identifier: GPL-2.0
d65817b4 DB	2	/*
	3	* Copyright (C) 2015 Davidlohr Bueso.
	4	*
	5	* Block a bunch of threads and let parallel waker threads wakeup an
	6	* equal amount of them. The program output reflects the avg latency
	7	* for each individual thread to service its share of work. Ultimately
	8	* it can be used to measure futex_wake() changes.
	9	*/
	10
8a158589	11	/* For the CLR_() macros */
a0f213e1	12	#include <string.h>
8a158589 ACM	13	#include <pthread.h>
8a158589 ACM	14
9c304f6c	15	#include <signal.h>
d65817b4	16	#include "../util/stat.h"
4b6ab94e	17	#include <subcmd/parse-options.h>
86695f59	18	#include <linux/compiler.h>
9c304f6c	19	#include <linux/kernel.h>
565e6911	20	#include <linux/time64.h>
9c304f6c	21	#include <errno.h>
d65817b4 DB	22	#include "bench.h"
	23	#include "futex.h"
	24
	25	#include <err.h>
	26	#include <stdlib.h>
	27	#include <sys/time.h>
d65817b4 DB	28
	29	struct thread_data {
	30	pthread_t worker;
	31	unsigned int nwoken;
	32	struct timeval runtime;
	33	};
	34
	35	static unsigned int nwakes = 1;
	36
	37	/* all threads will block on the same futex -- hash bucket chaos ;) */
	38	static u_int32_t futex = 0;
	39
	40	static pthread_t *blocked_worker;
	41	static bool done = false, silent = false, fshared = false;
	42	static unsigned int nblocked_threads = 0, nwaking_threads = 0;
	43	static pthread_mutex_t thread_lock;
	44	static pthread_cond_t thread_parent, thread_worker;
	45	static struct stats waketime_stats, wakeup_stats;
	46	static unsigned int ncpus, threads_starting;
	47	static int futex_flag = 0;
	48
	49	static const struct option options[] = {
	50	OPT_UINTEGER('t', "threads", &nblocked_threads, "Specify amount of threads"),
	51	OPT_UINTEGER('w', "nwakers", &nwaking_threads, "Specify amount of waking threads"),
	52	OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"),
	53	OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"),
	54	OPT_END()
	55	};
	56
	57	static const char * const bench_futex_wake_parallel_usage[] = {
	58	"perf bench futex wake-parallel <options>",
	59	NULL
	60	};
	61
	62	static void waking_workerfn(void arg)
	63	{
	64	struct thread_data waker = (struct thread_data ) arg;
	65	struct timeval start, end;
	66
	67	gettimeofday(&start, NULL);
	68
	69	waker->nwoken = futex_wake(&futex, nwakes, futex_flag);
	70	if (waker->nwoken != nwakes)
	71	warnx("couldn't wakeup all tasks (%d/%d)",
	72	waker->nwoken, nwakes);
	73
	74	gettimeofday(&end, NULL);
	75	timersub(&end, &start, &waker->runtime);
	76
	77	pthread_exit(NULL);
	78	return NULL;
	79	}
	80
	81	static void wakeup_threads(struct thread_data *td, pthread_attr_t thread_attr)
	82	{
	83	unsigned int i;
	84
	85	pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
	86
	87	/* create and block all threads */
	88	for (i = 0; i < nwaking_threads; i++) {
	89	/*
	90	* Thread creation order will impact per-thread latency
	91	* as it will affect the order to acquire the hb spinlock.
92	* For now let the scheduler decide.
93	*/
94	if (pthread_create(&td[i].worker, &thread_attr,
95	waking_workerfn, (void *)&td[i]))
96	err(EXIT_FAILURE, "pthread_create");
97	}
98
99	for (i = 0; i < nwaking_threads; i++)
100	if (pthread_join(td[i].worker, NULL))
101	err(EXIT_FAILURE, "pthread_join");
102	}
103
104	static void blocked_workerfn(void arg __maybe_unused)
105	{
106	pthread_mutex_lock(&thread_lock);
107	threads_starting--;
108	if (!threads_starting)
109	pthread_cond_signal(&thread_parent);
110	pthread_cond_wait(&thread_worker, &thread_lock);
111	pthread_mutex_unlock(&thread_lock);
112
113	while (1) { /* handle spurious wakeups */
114	if (futex_wait(&futex, 0, NULL, futex_flag) != EINTR)
115	break;
116	}
117
118	pthread_exit(NULL);
119	return NULL;
120	}
121
122	static void block_threads(pthread_t *w, pthread_attr_t thread_attr)
123	{
124	cpu_set_t cpu;
125	unsigned int i;
126
127	threads_starting = nblocked_threads;
128
129	/* create and block all threads */
130	for (i = 0; i < nblocked_threads; i++) {
131	CPU_ZERO(&cpu);
132	CPU_SET(i % ncpus, &cpu);
133
134	if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu))
135	err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
136
137	if (pthread_create(&w[i], &thread_attr, blocked_workerfn, NULL))
138	err(EXIT_FAILURE, "pthread_create");
139	}
140	}
141
142	static void print_run(struct thread_data *waking_worker, unsigned int run_num)
143	{
144	unsigned int i, wakeup_avg;
145	double waketime_avg, waketime_stddev;
146	struct stats __waketime_stats, __wakeup_stats;
147
148	init_stats(&__wakeup_stats);
149	init_stats(&__waketime_stats);
150
151	for (i = 0; i < nwaking_threads; i++) {
152	update_stats(&__waketime_stats, waking_worker[i].runtime.tv_usec);
153	update_stats(&__wakeup_stats, waking_worker[i].nwoken);
154	}
155
156	waketime_avg = avg_stats(&__waketime_stats);
157	waketime_stddev = stddev_stats(&__waketime_stats);
158	wakeup_avg = avg_stats(&__wakeup_stats);
159
160	printf("[Run %d]: Avg per-thread latency (waking %d/%d threads) "
161	"in %.4f ms (+-%.2f%%)\n", run_num + 1, wakeup_avg,
565e6911	162	nblocked_threads, waketime_avg / USEC_PER_MSEC,
d65817b4 DB	163	rel_stddev_stats(waketime_stddev, waketime_avg));
	164	}
	165
	166	static void print_summary(void)
	167	{
	168	unsigned int wakeup_avg;
	169	double waketime_avg, waketime_stddev;
	170
	171	waketime_avg = avg_stats(&waketime_stats);
	172	waketime_stddev = stddev_stats(&waketime_stats);
	173	wakeup_avg = avg_stats(&wakeup_stats);
	174
	175	printf("Avg per-thread latency (waking %d/%d threads) in %.4f ms (+-%.2f%%)\n",
	176	wakeup_avg,
	177	nblocked_threads,
565e6911	178	waketime_avg / USEC_PER_MSEC,
d65817b4 DB	179	rel_stddev_stats(waketime_stddev, waketime_avg));
	180	}
	181
	182
	183	static void do_run_stats(struct thread_data *waking_worker)
	184	{
	185	unsigned int i;
	186
	187	for (i = 0; i < nwaking_threads; i++) {
	188	update_stats(&waketime_stats, waking_worker[i].runtime.tv_usec);
	189	update_stats(&wakeup_stats, waking_worker[i].nwoken);
	190	}
	191
	192	}
	193
	194	static void toggle_done(int sig __maybe_unused,
	195	siginfo_t *info __maybe_unused,
	196	void *uc __maybe_unused)
	197	{
	198	done = true;
	199	}
	200
b0ad8ea6	201	int bench_futex_wake_parallel(int argc, const char **argv)
d65817b4 DB	202	{
	203	int ret = 0;
	204	unsigned int i, j;
	205	struct sigaction act;
	206	pthread_attr_t thread_attr;
	207	struct thread_data *waking_worker;
	208
	209	argc = parse_options(argc, argv, options,
	210	bench_futex_wake_parallel_usage, 0);
	211	if (argc) {
	212	usage_with_options(bench_futex_wake_parallel_usage, options);
	213	exit(EXIT_FAILURE);
	214	}
	215
	216	sigfillset(&act.sa_mask);
	217	act.sa_sigaction = toggle_done;
	218	sigaction(SIGINT, &act, NULL);
	219
	220	ncpus = sysconf(_SC_NPROCESSORS_ONLN);
	221	if (!nblocked_threads)
	222	nblocked_threads = ncpus;
	223
	224	/* some sanity checks */
	225	if (nwaking_threads > nblocked_threads \|\| !nwaking_threads)
	226	nwaking_threads = nblocked_threads;
	227
	228	if (nblocked_threads % nwaking_threads)
	229	errx(EXIT_FAILURE, "Must be perfectly divisible");
	230	/*
	231	* Each thread will wakeup nwakes tasks in
	232	* a single futex_wait call.
	233	*/
	234	nwakes = nblocked_threads/nwaking_threads;
	235
	236	blocked_worker = calloc(nblocked_threads, sizeof(*blocked_worker));
	237	if (!blocked_worker)
	238	err(EXIT_FAILURE, "calloc");
	239
	240	if (!fshared)
	241	futex_flag = FUTEX_PRIVATE_FLAG;
	242
	243	printf("Run summary [PID %d]: blocking on %d threads (at [%s] "
	244	"futex %p), %d threads waking up %d at a time.\n\n",
	245	getpid(), nblocked_threads, fshared ? "shared":"private",
	246	&futex, nwaking_threads, nwakes);
	247
	248	init_stats(&wakeup_stats);
	249	init_stats(&waketime_stats);
	250
	251	pthread_attr_init(&thread_attr);
	252	pthread_mutex_init(&thread_lock, NULL);
	253	pthread_cond_init(&thread_parent, NULL);
	254	pthread_cond_init(&thread_worker, NULL);
	255
	256	for (j = 0; j < bench_repeat && !done; j++) {
	257	waking_worker = calloc(nwaking_threads, sizeof(*waking_worker));
	258	if (!waking_worker)
	259	err(EXIT_FAILURE, "calloc");
	260
	261	/* create, launch & block all threads */
	262	block_threads(blocked_worker, thread_attr);
	263
	264	/* make sure all threads are already blocked */
	265	pthread_mutex_lock(&thread_lock);
266	while (threads_starting)
267	pthread_cond_wait(&thread_parent, &thread_lock);
268	pthread_cond_broadcast(&thread_worker);
269	pthread_mutex_unlock(&thread_lock);
270
271	usleep(100000);
272
273	/* Ok, all threads are patiently blocked, start waking folks up */
274	wakeup_threads(waking_worker, thread_attr);
275
276	for (i = 0; i < nblocked_threads; i++) {
277	ret = pthread_join(blocked_worker[i], NULL);
278	if (ret)
279	err(EXIT_FAILURE, "pthread_join");
280	}
281
282	do_run_stats(waking_worker);
283	if (!silent)
284	print_run(waking_worker, j);
285
286	free(waking_worker);
287	}
288
289	/* cleanup & report results */
290	pthread_cond_destroy(&thread_parent);
291	pthread_cond_destroy(&thread_worker);
292	pthread_mutex_destroy(&thread_lock);
293	pthread_attr_destroy(&thread_attr);
294
295	print_summary();
296
297	free(blocked_worker);
298	return ret;
299	}