Loading...
1/*
2 * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com>
3 *
4 * futex-wake: Block a bunch of threads on a futex and wake'em up, N at a time.
5 *
6 * This program is particularly useful to measure the latency of nthread wakeups
7 * in non-error situations: all waiters are queued and all wake calls wakeup
8 * one or more tasks, and thus the waitqueue is never empty.
9 */
10
11#include "../perf.h"
12#include "../util/util.h"
13#include "../util/stat.h"
14#include "../util/parse-options.h"
15#include "../util/header.h"
16#include "bench.h"
17#include "futex.h"
18
19#include <err.h>
20#include <stdlib.h>
21#include <sys/time.h>
22#include <pthread.h>
23
24/* all threads will block on the same futex */
25static u_int32_t futex1 = 0;
26
27/*
28 * How many wakeups to do at a time.
29 * Default to 1 in order to make the kernel work more.
30 */
31static unsigned int nwakes = 1;
32
33/*
34 * There can be significant variance from run to run,
35 * the more repeats, the more exact the overall avg and
36 * the better idea of the futex latency.
37 */
38static unsigned int repeat = 10;
39
40pthread_t *worker;
41static bool done = 0, silent = 0;
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;
46
47static const struct option options[] = {
48 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
49 OPT_UINTEGER('w', "nwakes", &nwakes, "Specify amount of threads to wake at once"),
50 OPT_UINTEGER('r', "repeat", &repeat, "Specify amount of times to repeat the run"),
51 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"),
52 OPT_END()
53};
54
55static const char * const bench_futex_wake_usage[] = {
56 "perf bench futex wake <options>",
57 NULL
58};
59
60static void *workerfn(void *arg __maybe_unused)
61{
62 pthread_mutex_lock(&thread_lock);
63 threads_starting--;
64 if (!threads_starting)
65 pthread_cond_signal(&thread_parent);
66 pthread_cond_wait(&thread_worker, &thread_lock);
67 pthread_mutex_unlock(&thread_lock);
68
69 futex_wait(&futex1, 0, NULL, FUTEX_PRIVATE_FLAG);
70 return NULL;
71}
72
73static void print_summary(void)
74{
75 double waketime_avg = avg_stats(&waketime_stats);
76 double waketime_stddev = stddev_stats(&waketime_stats);
77 unsigned int wakeup_avg = avg_stats(&wakeup_stats);
78
79 printf("Wokeup %d of %d threads in %.4f ms (+-%.2f%%)\n",
80 wakeup_avg,
81 nthreads,
82 waketime_avg/1e3,
83 rel_stddev_stats(waketime_stddev, waketime_avg));
84}
85
86static void block_threads(pthread_t *w,
87 pthread_attr_t thread_attr)
88{
89 cpu_set_t cpu;
90 unsigned int i;
91
92 threads_starting = nthreads;
93
94 /* create and block all threads */
95 for (i = 0; i < nthreads; i++) {
96 CPU_ZERO(&cpu);
97 CPU_SET(i % ncpus, &cpu);
98
99 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu))
100 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
101
102 if (pthread_create(&w[i], &thread_attr, workerfn, NULL))
103 err(EXIT_FAILURE, "pthread_create");
104 }
105}
106
107static void toggle_done(int sig __maybe_unused,
108 siginfo_t *info __maybe_unused,
109 void *uc __maybe_unused)
110{
111 done = true;
112}
113
114int bench_futex_wake(int argc, const char **argv,
115 const char *prefix __maybe_unused)
116{
117 int ret = 0;
118 unsigned int i, j;
119 struct sigaction act;
120 pthread_attr_t thread_attr;
121
122 argc = parse_options(argc, argv, options, bench_futex_wake_usage, 0);
123 if (argc) {
124 usage_with_options(bench_futex_wake_usage, options);
125 exit(EXIT_FAILURE);
126 }
127
128 ncpus = sysconf(_SC_NPROCESSORS_ONLN);
129
130 sigfillset(&act.sa_mask);
131 act.sa_sigaction = toggle_done;
132 sigaction(SIGINT, &act, NULL);
133
134 if (!nthreads)
135 nthreads = ncpus;
136
137 worker = calloc(nthreads, sizeof(*worker));
138 if (!worker)
139 err(EXIT_FAILURE, "calloc");
140
141 printf("Run summary [PID %d]: blocking on %d threads (at futex %p), "
142 "waking up %d at a time.\n\n",
143 getpid(), nthreads, &futex1, nwakes);
144
145 init_stats(&wakeup_stats);
146 init_stats(&waketime_stats);
147 pthread_attr_init(&thread_attr);
148 pthread_mutex_init(&thread_lock, NULL);
149 pthread_cond_init(&thread_parent, NULL);
150 pthread_cond_init(&thread_worker, NULL);
151
152 for (j = 0; j < repeat && !done; j++) {
153 unsigned int nwoken = 0;
154 struct timeval start, end, runtime;
155
156 /* create, launch & block all threads */
157 block_threads(worker, thread_attr);
158
159 /* make sure all threads are already blocked */
160 pthread_mutex_lock(&thread_lock);
161 while (threads_starting)
162 pthread_cond_wait(&thread_parent, &thread_lock);
163 pthread_cond_broadcast(&thread_worker);
164 pthread_mutex_unlock(&thread_lock);
165
166 usleep(100000);
167
168 /* Ok, all threads are patiently blocked, start waking folks up */
169 gettimeofday(&start, NULL);
170 while (nwoken != nthreads)
171 nwoken += futex_wake(&futex1, nwakes, FUTEX_PRIVATE_FLAG);
172 gettimeofday(&end, NULL);
173 timersub(&end, &start, &runtime);
174
175 update_stats(&wakeup_stats, nwoken);
176 update_stats(&waketime_stats, runtime.tv_usec);
177
178 if (!silent) {
179 printf("[Run %d]: Wokeup %d of %d threads in %.4f ms\n",
180 j + 1, nwoken, nthreads, runtime.tv_usec/1e3);
181 }
182
183 for (i = 0; i < nthreads; i++) {
184 ret = pthread_join(worker[i], NULL);
185 if (ret)
186 err(EXIT_FAILURE, "pthread_join");
187 }
188
189 }
190
191 /* cleanup & report results */
192 pthread_cond_destroy(&thread_parent);
193 pthread_cond_destroy(&thread_worker);
194 pthread_mutex_destroy(&thread_lock);
195 pthread_attr_destroy(&thread_attr);
196
197 print_summary();
198
199 free(worker);
200 return ret;
201}
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 <internal/cpumap.h>
24#include <perf/cpumap.h>
25#include "bench.h"
26#include "futex.h"
27
28#include <err.h>
29#include <stdlib.h>
30#include <sys/time.h>
31
32/* all threads will block on the same futex */
33static u_int32_t futex1 = 0;
34
35/*
36 * How many wakeups to do at a time.
37 * Default to 1 in order to make the kernel work more.
38 */
39static unsigned int nwakes = 1;
40
41pthread_t *worker;
42static bool done = false, silent = false, fshared = false;
43static pthread_mutex_t thread_lock;
44static pthread_cond_t thread_parent, thread_worker;
45static struct stats waketime_stats, wakeup_stats;
46static unsigned int threads_starting, nthreads = 0;
47static int futex_flag = 0;
48
49static const struct option options[] = {
50 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
51 OPT_UINTEGER('w', "nwakes", &nwakes, "Specify amount of threads to wake at once"),
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
57static const char * const bench_futex_wake_usage[] = {
58 "perf bench futex wake <options>",
59 NULL
60};
61
62static void *workerfn(void *arg __maybe_unused)
63{
64 pthread_mutex_lock(&thread_lock);
65 threads_starting--;
66 if (!threads_starting)
67 pthread_cond_signal(&thread_parent);
68 pthread_cond_wait(&thread_worker, &thread_lock);
69 pthread_mutex_unlock(&thread_lock);
70
71 while (1) {
72 if (futex_wait(&futex1, 0, NULL, futex_flag) != EINTR)
73 break;
74 }
75
76 pthread_exit(NULL);
77 return NULL;
78}
79
80static void print_summary(void)
81{
82 double waketime_avg = avg_stats(&waketime_stats);
83 double waketime_stddev = stddev_stats(&waketime_stats);
84 unsigned int wakeup_avg = avg_stats(&wakeup_stats);
85
86 printf("Wokeup %d of %d threads in %.4f ms (+-%.2f%%)\n",
87 wakeup_avg,
88 nthreads,
89 waketime_avg / USEC_PER_MSEC,
90 rel_stddev_stats(waketime_stddev, waketime_avg));
91}
92
93static void block_threads(pthread_t *w,
94 pthread_attr_t thread_attr, struct perf_cpu_map *cpu)
95{
96 cpu_set_t cpuset;
97 unsigned int i;
98
99 threads_starting = nthreads;
100
101 /* create and block all threads */
102 for (i = 0; i < nthreads; i++) {
103 CPU_ZERO(&cpuset);
104 CPU_SET(cpu->map[i % cpu->nr], &cpuset);
105
106 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset))
107 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
108
109 if (pthread_create(&w[i], &thread_attr, workerfn, NULL))
110 err(EXIT_FAILURE, "pthread_create");
111 }
112}
113
114static void toggle_done(int sig __maybe_unused,
115 siginfo_t *info __maybe_unused,
116 void *uc __maybe_unused)
117{
118 done = true;
119}
120
121int bench_futex_wake(int argc, const char **argv)
122{
123 int ret = 0;
124 unsigned int i, j;
125 struct sigaction act;
126 pthread_attr_t thread_attr;
127 struct perf_cpu_map *cpu;
128
129 argc = parse_options(argc, argv, options, bench_futex_wake_usage, 0);
130 if (argc) {
131 usage_with_options(bench_futex_wake_usage, options);
132 exit(EXIT_FAILURE);
133 }
134
135 cpu = perf_cpu_map__new(NULL);
136 if (!cpu)
137 err(EXIT_FAILURE, "calloc");
138
139 memset(&act, 0, sizeof(act));
140 sigfillset(&act.sa_mask);
141 act.sa_sigaction = toggle_done;
142 sigaction(SIGINT, &act, NULL);
143
144 if (!nthreads)
145 nthreads = cpu->nr;
146
147 worker = calloc(nthreads, sizeof(*worker));
148 if (!worker)
149 err(EXIT_FAILURE, "calloc");
150
151 if (!fshared)
152 futex_flag = FUTEX_PRIVATE_FLAG;
153
154 printf("Run summary [PID %d]: blocking on %d threads (at [%s] futex %p), "
155 "waking up %d at a time.\n\n",
156 getpid(), nthreads, fshared ? "shared":"private", &futex1, nwakes);
157
158 init_stats(&wakeup_stats);
159 init_stats(&waketime_stats);
160 pthread_attr_init(&thread_attr);
161 pthread_mutex_init(&thread_lock, NULL);
162 pthread_cond_init(&thread_parent, NULL);
163 pthread_cond_init(&thread_worker, NULL);
164
165 for (j = 0; j < bench_repeat && !done; j++) {
166 unsigned int nwoken = 0;
167 struct timeval start, end, runtime;
168
169 /* create, launch & block all threads */
170 block_threads(worker, thread_attr, cpu);
171
172 /* make sure all threads are already blocked */
173 pthread_mutex_lock(&thread_lock);
174 while (threads_starting)
175 pthread_cond_wait(&thread_parent, &thread_lock);
176 pthread_cond_broadcast(&thread_worker);
177 pthread_mutex_unlock(&thread_lock);
178
179 usleep(100000);
180
181 /* Ok, all threads are patiently blocked, start waking folks up */
182 gettimeofday(&start, NULL);
183 while (nwoken != nthreads)
184 nwoken += futex_wake(&futex1, nwakes, futex_flag);
185 gettimeofday(&end, NULL);
186 timersub(&end, &start, &runtime);
187
188 update_stats(&wakeup_stats, nwoken);
189 update_stats(&waketime_stats, runtime.tv_usec);
190
191 if (!silent) {
192 printf("[Run %d]: Wokeup %d of %d threads in %.4f ms\n",
193 j + 1, nwoken, nthreads, runtime.tv_usec / (double)USEC_PER_MSEC);
194 }
195
196 for (i = 0; i < nthreads; i++) {
197 ret = pthread_join(worker[i], NULL);
198 if (ret)
199 err(EXIT_FAILURE, "pthread_join");
200 }
201
202 }
203
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);
209
210 print_summary();
211
212 free(worker);
213 return ret;
214}