2 * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com>
4 * futex-hash: Stress the hell out of the Linux kernel futex uaddr hashing.
6 * This program is particularly useful for measuring the kernel's futex hash
7 * table/function implementation. In order for it to make sense, use with as
8 * many threads and futexes as possible.
11 /* For the CLR_() macros */
17 #include <linux/compiler.h>
18 #include <linux/kernel.h>
21 #include "../util/stat.h"
22 #include <subcmd/parse-options.h>
29 static unsigned int nthreads = 0;
30 static unsigned int nsecs = 10;
31 /* amount of futexes per thread */
32 static unsigned int nfutexes = 1024;
33 static bool fshared = false, done = false, silent = false;
34 static int futex_flag = 0;
36 struct timeval start, end, runtime;
37 static pthread_mutex_t thread_lock;
38 static unsigned int threads_starting;
39 static struct stats throughput_stats;
40 static pthread_cond_t thread_parent, thread_worker;
49 static const struct option options[] = {
50 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
51 OPT_UINTEGER('r', "runtime", &nsecs, "Specify runtime (in seconds)"),
52 OPT_UINTEGER('f', "futexes", &nfutexes, "Specify amount of futexes per threads"),
53 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"),
54 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"),
58 static const char * const bench_futex_hash_usage[] = {
59 "perf bench futex hash <options>",
63 static void *workerfn(void *arg)
66 struct worker *w = (struct worker *) arg;
68 unsigned long ops = w->ops; /* avoid cacheline bouncing */
70 pthread_mutex_lock(&thread_lock);
72 if (!threads_starting)
73 pthread_cond_signal(&thread_parent);
74 pthread_cond_wait(&thread_worker, &thread_lock);
75 pthread_mutex_unlock(&thread_lock);
78 for (i = 0; i < nfutexes; i++, ops++) {
80 * We want the futex calls to fail in order to stress
81 * the hashing of uaddr and not measure other steps,
82 * such as internal waitqueue handling, thus enlarging
83 * the critical region protected by hb->lock.
85 ret = futex_wait(&w->futex[i], 1234, NULL, futex_flag);
87 (!ret || errno != EAGAIN || errno != EWOULDBLOCK))
88 warn("Non-expected futex return call");
96 static void toggle_done(int sig __maybe_unused,
97 siginfo_t *info __maybe_unused,
98 void *uc __maybe_unused)
100 /* inform all threads that we're done for the day */
102 gettimeofday(&end, NULL);
103 timersub(&end, &start, &runtime);
106 static void print_summary(void)
108 unsigned long avg = avg_stats(&throughput_stats);
109 double stddev = stddev_stats(&throughput_stats);
111 printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
112 !silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg),
113 (int) runtime.tv_sec);
116 int bench_futex_hash(int argc, const char **argv,
117 const char *prefix __maybe_unused)
121 struct sigaction act;
122 unsigned int i, ncpus;
123 pthread_attr_t thread_attr;
124 struct worker *worker = NULL;
126 argc = parse_options(argc, argv, options, bench_futex_hash_usage, 0);
128 usage_with_options(bench_futex_hash_usage, options);
132 ncpus = sysconf(_SC_NPROCESSORS_ONLN);
133 nsecs = futexbench_sanitize_numeric(nsecs);
134 nfutexes = futexbench_sanitize_numeric(nfutexes);
136 sigfillset(&act.sa_mask);
137 act.sa_sigaction = toggle_done;
138 sigaction(SIGINT, &act, NULL);
140 if (!nthreads) /* default to the number of CPUs */
143 nthreads = futexbench_sanitize_numeric(nthreads);
145 worker = calloc(nthreads, sizeof(*worker));
150 futex_flag = FUTEX_PRIVATE_FLAG;
152 printf("Run summary [PID %d]: %d threads, each operating on %d [%s] futexes for %d secs.\n\n",
153 getpid(), nthreads, nfutexes, fshared ? "shared":"private", nsecs);
155 init_stats(&throughput_stats);
156 pthread_mutex_init(&thread_lock, NULL);
157 pthread_cond_init(&thread_parent, NULL);
158 pthread_cond_init(&thread_worker, NULL);
160 threads_starting = nthreads;
161 pthread_attr_init(&thread_attr);
162 gettimeofday(&start, NULL);
163 for (i = 0; i < nthreads; i++) {
165 worker[i].futex = calloc(nfutexes, sizeof(*worker[i].futex));
166 if (!worker[i].futex)
170 CPU_SET(i % ncpus, &cpu);
172 ret = pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu);
174 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
176 ret = pthread_create(&worker[i].thread, &thread_attr, workerfn,
177 (void *)(struct worker *) &worker[i]);
179 err(EXIT_FAILURE, "pthread_create");
182 pthread_attr_destroy(&thread_attr);
184 pthread_mutex_lock(&thread_lock);
185 while (threads_starting)
186 pthread_cond_wait(&thread_parent, &thread_lock);
187 pthread_cond_broadcast(&thread_worker);
188 pthread_mutex_unlock(&thread_lock);
191 toggle_done(0, NULL, NULL);
193 for (i = 0; i < nthreads; i++) {
194 ret = pthread_join(worker[i].thread, NULL);
196 err(EXIT_FAILURE, "pthread_join");
199 /* cleanup & report results */
200 pthread_cond_destroy(&thread_parent);
201 pthread_cond_destroy(&thread_worker);
202 pthread_mutex_destroy(&thread_lock);
204 for (i = 0; i < nthreads; i++) {
205 unsigned long t = worker[i].ops/runtime.tv_sec;
206 update_stats(&throughput_stats, t);
209 printf("[thread %2d] futex: %p [ %ld ops/sec ]\n",
210 worker[i].tid, &worker[i].futex[0], t);
212 printf("[thread %2d] futexes: %p ... %p [ %ld ops/sec ]\n",
213 worker[i].tid, &worker[i].futex[0],
214 &worker[i].futex[nfutexes-1], t);
217 free(worker[i].futex);
225 err(EXIT_FAILURE, "calloc");