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

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