1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * Copyright (C) 2013  Davidlohr Bueso <davidlohr@hp.com>
  4 *
  5 * futex-wake: Block a bunch of threads on a futex and wake'em up, N at a time.
  6 *
  7 * This program is particularly useful to measure the latency of nthread wakeups
  8 * in non-error situations:  all waiters are queued and all wake calls wakeup
  9 * one or more tasks, and thus the waitqueue is never empty.
 10 */
 11
 12/* For the CLR_() macros */
 13#include <string.h>
 14#include <pthread.h>
 15
 16#include <signal.h>
 17#include "../util/stat.h"
 18#include <subcmd/parse-options.h>
 19#include <linux/compiler.h>
 20#include <linux/kernel.h>
 21#include <linux/time64.h>
 22#include <errno.h>
 23#include "bench.h"
 24#include "futex.h"
 25#include "cpumap.h"
 26
 27#include <err.h>
 28#include <stdlib.h>
 29#include <sys/time.h>
 30
 31/* all threads will block on the same futex */
 32static u_int32_t futex1 = 0;
 33
 34/*
 35 * How many wakeups to do at a time.
 36 * Default to 1 in order to make the kernel work more.
 37 */
 38static unsigned int nwakes = 1;
 39
 40pthread_t *worker;
 41static bool done = false, silent = false, fshared = false;
 42static pthread_mutex_t thread_lock;
 43static pthread_cond_t thread_parent, thread_worker;
 44static struct stats waketime_stats, wakeup_stats;
 45static unsigned int ncpus, threads_starting, nthreads = 0;
 46static int futex_flag = 0;
 47
 48static const struct option options[] = {
 49	OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
 50	OPT_UINTEGER('w', "nwakes",  &nwakes,   "Specify amount of threads to wake at once"),
 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
 56static const char * const bench_futex_wake_usage[] = {
 57	"perf bench futex wake <options>",
 58	NULL
 59};
 60
 61static void *workerfn(void *arg __maybe_unused)
 62{
 63	pthread_mutex_lock(&thread_lock);
 64	threads_starting--;
 65	if (!threads_starting)
 66		pthread_cond_signal(&thread_parent);
 67	pthread_cond_wait(&thread_worker, &thread_lock);
 68	pthread_mutex_unlock(&thread_lock);
 69
 70	while (1) {
 71		if (futex_wait(&futex1, 0, NULL, futex_flag) != EINTR)
 72			break;
 73	}
 74
 75	pthread_exit(NULL);
 76	return NULL;
 77}
 78
 79static void print_summary(void)
 80{
 81	double waketime_avg = avg_stats(&waketime_stats);
 82	double waketime_stddev = stddev_stats(&waketime_stats);
 83	unsigned int wakeup_avg = avg_stats(&wakeup_stats);
 84
 85	printf("Wokeup %d of %d threads in %.4f ms (+-%.2f%%)\n",
 86	       wakeup_avg,
 87	       nthreads,
 88	       waketime_avg / USEC_PER_MSEC,
 89	       rel_stddev_stats(waketime_stddev, waketime_avg));
 90}
 91
 92static void block_threads(pthread_t *w,
 93			  pthread_attr_t thread_attr, struct cpu_map *cpu)
 94{
 95	cpu_set_t cpuset;
 96	unsigned int i;
 97
 98	threads_starting = nthreads;
 99
100	/* create and block all threads */
101	for (i = 0; i < nthreads; i++) {
102		CPU_ZERO(&cpuset);
103		CPU_SET(cpu->map[i % cpu->nr], &cpuset);
104
105		if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset))
106			err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
107
108		if (pthread_create(&w[i], &thread_attr, workerfn, NULL))
109			err(EXIT_FAILURE, "pthread_create");
110	}
111}
112
113static void toggle_done(int sig __maybe_unused,
114			siginfo_t *info __maybe_unused,
115			void *uc __maybe_unused)
116{
117	done = true;
118}
119
120int bench_futex_wake(int argc, const char **argv)
121{
122	int ret = 0;
123	unsigned int i, j;
124	struct sigaction act;
125	pthread_attr_t thread_attr;
126	struct cpu_map *cpu;
127
128	argc = parse_options(argc, argv, options, bench_futex_wake_usage, 0);
129	if (argc) {
130		usage_with_options(bench_futex_wake_usage, options);
131		exit(EXIT_FAILURE);
132	}
133
134	cpu = cpu_map__new(NULL);
135	if (!cpu)
136		err(EXIT_FAILURE, "calloc");
137
138	sigfillset(&act.sa_mask);
139	act.sa_sigaction = toggle_done;
140	sigaction(SIGINT, &act, NULL);
141
142	if (!nthreads)
143		nthreads = ncpus;
144
145	worker = calloc(nthreads, sizeof(*worker));
146	if (!worker)
147		err(EXIT_FAILURE, "calloc");
148
149	if (!fshared)
150		futex_flag = FUTEX_PRIVATE_FLAG;
151
152	printf("Run summary [PID %d]: blocking on %d threads (at [%s] futex %p), "
153	       "waking up %d at a time.\n\n",
154	       getpid(), nthreads, fshared ? "shared":"private",  &futex1, nwakes);
155
156	init_stats(&wakeup_stats);
157	init_stats(&waketime_stats);
158	pthread_attr_init(&thread_attr);
159	pthread_mutex_init(&thread_lock, NULL);
160	pthread_cond_init(&thread_parent, NULL);
161	pthread_cond_init(&thread_worker, NULL);
162
163	for (j = 0; j < bench_repeat && !done; j++) {
164		unsigned int nwoken = 0;
165		struct timeval start, end, runtime;
166
167		/* create, launch & block all threads */
168		block_threads(worker, thread_attr, cpu);
169
170		/* make sure all threads are already blocked */
171		pthread_mutex_lock(&thread_lock);
172		while (threads_starting)
173			pthread_cond_wait(&thread_parent, &thread_lock);
174		pthread_cond_broadcast(&thread_worker);
175		pthread_mutex_unlock(&thread_lock);
176
177		usleep(100000);
178
179		/* Ok, all threads are patiently blocked, start waking folks up */
180		gettimeofday(&start, NULL);
181		while (nwoken != nthreads)
182			nwoken += futex_wake(&futex1, nwakes, futex_flag);
183		gettimeofday(&end, NULL);
184		timersub(&end, &start, &runtime);
185
186		update_stats(&wakeup_stats, nwoken);
187		update_stats(&waketime_stats, runtime.tv_usec);
188
189		if (!silent) {
190			printf("[Run %d]: Wokeup %d of %d threads in %.4f ms\n",
191			       j + 1, nwoken, nthreads, runtime.tv_usec / (double)USEC_PER_MSEC);
192		}
193
194		for (i = 0; i < nthreads; i++) {
195			ret = pthread_join(worker[i], NULL);
196			if (ret)
197				err(EXIT_FAILURE, "pthread_join");
198		}
199
200	}
201
202	/* cleanup & report results */
203	pthread_cond_destroy(&thread_parent);
204	pthread_cond_destroy(&thread_worker);
205	pthread_mutex_destroy(&thread_lock);
206	pthread_attr_destroy(&thread_attr);
207
208	print_summary();
209
210	free(worker);
211	return ret;
212}