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/mutex.h"
18#include "../util/stat.h"
19#include <subcmd/parse-options.h>
20#include <linux/compiler.h>
21#include <linux/kernel.h>
22#include <linux/time64.h>
23#include <errno.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#include <sys/mman.h>
32
33/* all threads will block on the same futex */
34static u_int32_t futex1 = 0;
35
36static pthread_t *worker;
37static bool done = false;
38static struct mutex thread_lock;
39static struct cond thread_parent, thread_worker;
40static struct stats waketime_stats, wakeup_stats;
41static unsigned int threads_starting;
42static int futex_flag = 0;
43
44static struct bench_futex_parameters params = {
45 /*
46 * How many wakeups to do at a time.
47 * Default to 1 in order to make the kernel work more.
48 */
49 .nwakes = 1,
50};
51
52static const struct option options[] = {
53 OPT_UINTEGER('t', "threads", ¶ms.nthreads, "Specify amount of threads"),
54 OPT_UINTEGER('w', "nwakes", ¶ms.nwakes, "Specify amount of threads to wake at once"),
55 OPT_BOOLEAN( 's', "silent", ¶ms.silent, "Silent mode: do not display data/details"),
56 OPT_BOOLEAN( 'S', "shared", ¶ms.fshared, "Use shared futexes instead of private ones"),
57 OPT_BOOLEAN( 'm', "mlockall", ¶ms.mlockall, "Lock all current and future memory"),
58
59 OPT_END()
60};
61
62static const char * const bench_futex_wake_usage[] = {
63 "perf bench futex wake <options>",
64 NULL
65};
66
67static void *workerfn(void *arg __maybe_unused)
68{
69 mutex_lock(&thread_lock);
70 threads_starting--;
71 if (!threads_starting)
72 cond_signal(&thread_parent);
73 cond_wait(&thread_worker, &thread_lock);
74 mutex_unlock(&thread_lock);
75
76 while (1) {
77 if (futex_wait(&futex1, 0, NULL, futex_flag) != EINTR)
78 break;
79 }
80
81 pthread_exit(NULL);
82 return NULL;
83}
84
85static void print_summary(void)
86{
87 double waketime_avg = avg_stats(&waketime_stats);
88 double waketime_stddev = stddev_stats(&waketime_stats);
89 unsigned int wakeup_avg = avg_stats(&wakeup_stats);
90
91 printf("Wokeup %d of %d threads in %.4f ms (+-%.2f%%)\n",
92 wakeup_avg,
93 params.nthreads,
94 waketime_avg / USEC_PER_MSEC,
95 rel_stddev_stats(waketime_stddev, waketime_avg));
96}
97
98static void block_threads(pthread_t *w, struct perf_cpu_map *cpu)
99{
100 cpu_set_t *cpuset;
101 unsigned int i;
102 size_t size;
103 int nrcpus = perf_cpu_map__nr(cpu);
104 threads_starting = params.nthreads;
105
106 cpuset = CPU_ALLOC(nrcpus);
107 BUG_ON(!cpuset);
108 size = CPU_ALLOC_SIZE(nrcpus);
109
110 /* create and block all threads */
111 for (i = 0; i < params.nthreads; i++) {
112 pthread_attr_t thread_attr;
113
114 pthread_attr_init(&thread_attr);
115 CPU_ZERO_S(size, cpuset);
116 CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset);
117
118 if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) {
119 CPU_FREE(cpuset);
120 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
121 }
122
123 if (pthread_create(&w[i], &thread_attr, workerfn, NULL)) {
124 CPU_FREE(cpuset);
125 err(EXIT_FAILURE, "pthread_create");
126 }
127 pthread_attr_destroy(&thread_attr);
128 }
129 CPU_FREE(cpuset);
130}
131
132static void toggle_done(int sig __maybe_unused,
133 siginfo_t *info __maybe_unused,
134 void *uc __maybe_unused)
135{
136 done = true;
137}
138
139int bench_futex_wake(int argc, const char **argv)
140{
141 int ret = 0;
142 unsigned int i, j;
143 struct sigaction act;
144 struct perf_cpu_map *cpu;
145
146 argc = parse_options(argc, argv, options, bench_futex_wake_usage, 0);
147 if (argc) {
148 usage_with_options(bench_futex_wake_usage, options);
149 exit(EXIT_FAILURE);
150 }
151
152 cpu = perf_cpu_map__new_online_cpus();
153 if (!cpu)
154 err(EXIT_FAILURE, "calloc");
155
156 memset(&act, 0, sizeof(act));
157 sigfillset(&act.sa_mask);
158 act.sa_sigaction = toggle_done;
159 sigaction(SIGINT, &act, NULL);
160
161 if (params.mlockall) {
162 if (mlockall(MCL_CURRENT | MCL_FUTURE))
163 err(EXIT_FAILURE, "mlockall");
164 }
165
166 if (!params.nthreads)
167 params.nthreads = perf_cpu_map__nr(cpu);
168
169 worker = calloc(params.nthreads, sizeof(*worker));
170 if (!worker)
171 err(EXIT_FAILURE, "calloc");
172
173 if (!params.fshared)
174 futex_flag = FUTEX_PRIVATE_FLAG;
175
176 printf("Run summary [PID %d]: blocking on %d threads (at [%s] futex %p), "
177 "waking up %d at a time.\n\n",
178 getpid(), params.nthreads, params.fshared ? "shared":"private",
179 &futex1, params.nwakes);
180
181 init_stats(&wakeup_stats);
182 init_stats(&waketime_stats);
183 mutex_init(&thread_lock);
184 cond_init(&thread_parent);
185 cond_init(&thread_worker);
186
187 for (j = 0; j < bench_repeat && !done; j++) {
188 unsigned int nwoken = 0;
189 struct timeval start, end, runtime;
190
191 /* create, launch & block all threads */
192 block_threads(worker, cpu);
193
194 /* make sure all threads are already blocked */
195 mutex_lock(&thread_lock);
196 while (threads_starting)
197 cond_wait(&thread_parent, &thread_lock);
198 cond_broadcast(&thread_worker);
199 mutex_unlock(&thread_lock);
200
201 usleep(100000);
202
203 /* Ok, all threads are patiently blocked, start waking folks up */
204 gettimeofday(&start, NULL);
205 while (nwoken != params.nthreads)
206 nwoken += futex_wake(&futex1,
207 params.nwakes, futex_flag);
208 gettimeofday(&end, NULL);
209 timersub(&end, &start, &runtime);
210
211 update_stats(&wakeup_stats, nwoken);
212 update_stats(&waketime_stats, runtime.tv_usec);
213
214 if (!params.silent) {
215 printf("[Run %d]: Wokeup %d of %d threads in %.4f ms\n",
216 j + 1, nwoken, params.nthreads,
217 runtime.tv_usec / (double)USEC_PER_MSEC);
218 }
219
220 for (i = 0; i < params.nthreads; i++) {
221 ret = pthread_join(worker[i], NULL);
222 if (ret)
223 err(EXIT_FAILURE, "pthread_join");
224 }
225
226 }
227
228 /* cleanup & report results */
229 cond_destroy(&thread_parent);
230 cond_destroy(&thread_worker);
231 mutex_destroy(&thread_lock);
232
233 print_summary();
234
235 free(worker);
236 perf_cpu_map__put(cpu);
237 return ret;
238}