1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com>
5 * futex-requeue: Block a bunch of threads on futex1 and requeue them
6 * on futex2, N at a time.
8 * This program is particularly useful to measure the latency of nthread
9 * requeues without waking up any tasks -- thus mimicking a regular futex_wait.
12 /* For the CLR_() macros */
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>
30 static u_int32_t futex1 = 0, futex2 = 0;
33 * How many tasks to requeue at a time.
34 * Default to 1 in order to make the kernel work more.
36 static unsigned int nrequeue = 1;
38 static pthread_t *worker;
39 static bool done = false, silent = false, fshared = false;
40 static pthread_mutex_t thread_lock;
41 static pthread_cond_t thread_parent, thread_worker;
42 static struct stats requeuetime_stats, requeued_stats;
43 static unsigned int ncpus, threads_starting, nthreads = 0;
44 static int futex_flag = 0;
46 static const struct option options[] = {
47 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
48 OPT_UINTEGER('q', "nrequeue", &nrequeue, "Specify amount of threads to requeue at once"),
49 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"),
50 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"),
54 static const char * const bench_futex_requeue_usage[] = {
55 "perf bench futex requeue <options>",
59 static void print_summary(void)
61 double requeuetime_avg = avg_stats(&requeuetime_stats);
62 double requeuetime_stddev = stddev_stats(&requeuetime_stats);
63 unsigned int requeued_avg = avg_stats(&requeued_stats);
65 printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n",
68 requeuetime_avg / USEC_PER_MSEC,
69 rel_stddev_stats(requeuetime_stddev, requeuetime_avg));
72 static void *workerfn(void *arg __maybe_unused)
74 pthread_mutex_lock(&thread_lock);
76 if (!threads_starting)
77 pthread_cond_signal(&thread_parent);
78 pthread_cond_wait(&thread_worker, &thread_lock);
79 pthread_mutex_unlock(&thread_lock);
81 futex_wait(&futex1, 0, NULL, futex_flag);
85 static void block_threads(pthread_t *w,
86 pthread_attr_t thread_attr)
91 threads_starting = nthreads;
93 /* create and block all threads */
94 for (i = 0; i < nthreads; i++) {
96 CPU_SET(i % ncpus, &cpu);
98 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu))
99 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
101 if (pthread_create(&w[i], &thread_attr, workerfn, NULL))
102 err(EXIT_FAILURE, "pthread_create");
106 static void toggle_done(int sig __maybe_unused,
107 siginfo_t *info __maybe_unused,
108 void *uc __maybe_unused)
113 int bench_futex_requeue(int argc, const char **argv)
117 struct sigaction act;
118 pthread_attr_t thread_attr;
120 argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0);
124 ncpus = sysconf(_SC_NPROCESSORS_ONLN);
126 sigfillset(&act.sa_mask);
127 act.sa_sigaction = toggle_done;
128 sigaction(SIGINT, &act, NULL);
133 worker = calloc(nthreads, sizeof(*worker));
135 err(EXIT_FAILURE, "calloc");
138 futex_flag = FUTEX_PRIVATE_FLAG;
140 if (nrequeue > nthreads)
143 printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %p), "
144 "%d at a time.\n\n", getpid(), nthreads,
145 fshared ? "shared":"private", &futex1, &futex2, nrequeue);
147 init_stats(&requeued_stats);
148 init_stats(&requeuetime_stats);
149 pthread_attr_init(&thread_attr);
150 pthread_mutex_init(&thread_lock, NULL);
151 pthread_cond_init(&thread_parent, NULL);
152 pthread_cond_init(&thread_worker, NULL);
154 for (j = 0; j < bench_repeat && !done; j++) {
155 unsigned int nrequeued = 0;
156 struct timeval start, end, runtime;
158 /* create, launch & block all threads */
159 block_threads(worker, thread_attr);
161 /* make sure all threads are already blocked */
162 pthread_mutex_lock(&thread_lock);
163 while (threads_starting)
164 pthread_cond_wait(&thread_parent, &thread_lock);
165 pthread_cond_broadcast(&thread_worker);
166 pthread_mutex_unlock(&thread_lock);
170 /* Ok, all threads are patiently blocked, start requeueing */
171 gettimeofday(&start, NULL);
172 while (nrequeued < nthreads) {
174 * Do not wakeup any tasks blocked on futex1, allowing
175 * us to really measure futex_wait functionality.
177 nrequeued += futex_cmp_requeue(&futex1, 0, &futex2, 0,
178 nrequeue, futex_flag);
181 gettimeofday(&end, NULL);
182 timersub(&end, &start, &runtime);
184 update_stats(&requeued_stats, nrequeued);
185 update_stats(&requeuetime_stats, runtime.tv_usec);
188 printf("[Run %d]: Requeued %d of %d threads in %.4f ms\n",
189 j + 1, nrequeued, nthreads, runtime.tv_usec / (double)USEC_PER_MSEC);
192 /* everybody should be blocked on futex2, wake'em up */
193 nrequeued = futex_wake(&futex2, nrequeued, futex_flag);
194 if (nthreads != nrequeued)
195 warnx("couldn't wakeup all tasks (%d/%d)", nrequeued, nthreads);
197 for (i = 0; i < nthreads; i++) {
198 ret = pthread_join(worker[i], NULL);
200 err(EXIT_FAILURE, "pthread_join");
204 /* cleanup & report results */
205 pthread_cond_destroy(&thread_parent);
206 pthread_cond_destroy(&thread_worker);
207 pthread_mutex_destroy(&thread_lock);
208 pthread_attr_destroy(&thread_attr);
215 usage_with_options(bench_futex_requeue_usage, options);