Loading...
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 ncpus, 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 sigfillset(&act.sa_mask);
140 act.sa_sigaction = toggle_done;
141 sigaction(SIGINT, &act, NULL);
142
143 if (!nthreads)
144 nthreads = ncpus;
145
146 worker = calloc(nthreads, sizeof(*worker));
147 if (!worker)
148 err(EXIT_FAILURE, "calloc");
149
150 if (!fshared)
151 futex_flag = FUTEX_PRIVATE_FLAG;
152
153 printf("Run summary [PID %d]: blocking on %d threads (at [%s] futex %p), "
154 "waking up %d at a time.\n\n",
155 getpid(), nthreads, fshared ? "shared":"private", &futex1, nwakes);
156
157 init_stats(&wakeup_stats);
158 init_stats(&waketime_stats);
159 pthread_attr_init(&thread_attr);
160 pthread_mutex_init(&thread_lock, NULL);
161 pthread_cond_init(&thread_parent, NULL);
162 pthread_cond_init(&thread_worker, NULL);
163
164 for (j = 0; j < bench_repeat && !done; j++) {
165 unsigned int nwoken = 0;
166 struct timeval start, end, runtime;
167
168 /* create, launch & block all threads */
169 block_threads(worker, thread_attr, cpu);
170
171 /* make sure all threads are already blocked */
172 pthread_mutex_lock(&thread_lock);
173 while (threads_starting)
174 pthread_cond_wait(&thread_parent, &thread_lock);
175 pthread_cond_broadcast(&thread_worker);
176 pthread_mutex_unlock(&thread_lock);
177
178 usleep(100000);
179
180 /* Ok, all threads are patiently blocked, start waking folks up */
181 gettimeofday(&start, NULL);
182 while (nwoken != nthreads)
183 nwoken += futex_wake(&futex1, nwakes, futex_flag);
184 gettimeofday(&end, NULL);
185 timersub(&end, &start, &runtime);
186
187 update_stats(&wakeup_stats, nwoken);
188 update_stats(&waketime_stats, runtime.tv_usec);
189
190 if (!silent) {
191 printf("[Run %d]: Wokeup %d of %d threads in %.4f ms\n",
192 j + 1, nwoken, nthreads, runtime.tv_usec / (double)USEC_PER_MSEC);
193 }
194
195 for (i = 0; i < nthreads; i++) {
196 ret = pthread_join(worker[i], NULL);
197 if (ret)
198 err(EXIT_FAILURE, "pthread_join");
199 }
200
201 }
202
203 /* cleanup & report results */
204 pthread_cond_destroy(&thread_parent);
205 pthread_cond_destroy(&thread_worker);
206 pthread_mutex_destroy(&thread_lock);
207 pthread_attr_destroy(&thread_attr);
208
209 print_summary();
210
211 free(worker);
212 return ret;
213}
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,
99 pthread_attr_t thread_attr, struct perf_cpu_map *cpu)
100{
101 cpu_set_t *cpuset;
102 unsigned int i;
103 size_t size;
104 int nrcpus = perf_cpu_map__nr(cpu);
105 threads_starting = params.nthreads;
106
107 cpuset = CPU_ALLOC(nrcpus);
108 BUG_ON(!cpuset);
109 size = CPU_ALLOC_SIZE(nrcpus);
110
111 /* create and block all threads */
112 for (i = 0; i < params.nthreads; i++) {
113 CPU_ZERO_S(size, cpuset);
114 CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset);
115
116 if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) {
117 CPU_FREE(cpuset);
118 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
119 }
120
121 if (pthread_create(&w[i], &thread_attr, workerfn, NULL)) {
122 CPU_FREE(cpuset);
123 err(EXIT_FAILURE, "pthread_create");
124 }
125 }
126 CPU_FREE(cpuset);
127}
128
129static void toggle_done(int sig __maybe_unused,
130 siginfo_t *info __maybe_unused,
131 void *uc __maybe_unused)
132{
133 done = true;
134}
135
136int bench_futex_wake(int argc, const char **argv)
137{
138 int ret = 0;
139 unsigned int i, j;
140 struct sigaction act;
141 pthread_attr_t thread_attr;
142 struct perf_cpu_map *cpu;
143
144 argc = parse_options(argc, argv, options, bench_futex_wake_usage, 0);
145 if (argc) {
146 usage_with_options(bench_futex_wake_usage, options);
147 exit(EXIT_FAILURE);
148 }
149
150 cpu = perf_cpu_map__new(NULL);
151 if (!cpu)
152 err(EXIT_FAILURE, "calloc");
153
154 memset(&act, 0, sizeof(act));
155 sigfillset(&act.sa_mask);
156 act.sa_sigaction = toggle_done;
157 sigaction(SIGINT, &act, NULL);
158
159 if (params.mlockall) {
160 if (mlockall(MCL_CURRENT | MCL_FUTURE))
161 err(EXIT_FAILURE, "mlockall");
162 }
163
164 if (!params.nthreads)
165 params.nthreads = perf_cpu_map__nr(cpu);
166
167 worker = calloc(params.nthreads, sizeof(*worker));
168 if (!worker)
169 err(EXIT_FAILURE, "calloc");
170
171 if (!params.fshared)
172 futex_flag = FUTEX_PRIVATE_FLAG;
173
174 printf("Run summary [PID %d]: blocking on %d threads (at [%s] futex %p), "
175 "waking up %d at a time.\n\n",
176 getpid(), params.nthreads, params.fshared ? "shared":"private",
177 &futex1, params.nwakes);
178
179 init_stats(&wakeup_stats);
180 init_stats(&waketime_stats);
181 pthread_attr_init(&thread_attr);
182 mutex_init(&thread_lock);
183 cond_init(&thread_parent);
184 cond_init(&thread_worker);
185
186 for (j = 0; j < bench_repeat && !done; j++) {
187 unsigned int nwoken = 0;
188 struct timeval start, end, runtime;
189
190 /* create, launch & block all threads */
191 block_threads(worker, thread_attr, cpu);
192
193 /* make sure all threads are already blocked */
194 mutex_lock(&thread_lock);
195 while (threads_starting)
196 cond_wait(&thread_parent, &thread_lock);
197 cond_broadcast(&thread_worker);
198 mutex_unlock(&thread_lock);
199
200 usleep(100000);
201
202 /* Ok, all threads are patiently blocked, start waking folks up */
203 gettimeofday(&start, NULL);
204 while (nwoken != params.nthreads)
205 nwoken += futex_wake(&futex1,
206 params.nwakes, futex_flag);
207 gettimeofday(&end, NULL);
208 timersub(&end, &start, &runtime);
209
210 update_stats(&wakeup_stats, nwoken);
211 update_stats(&waketime_stats, runtime.tv_usec);
212
213 if (!params.silent) {
214 printf("[Run %d]: Wokeup %d of %d threads in %.4f ms\n",
215 j + 1, nwoken, params.nthreads,
216 runtime.tv_usec / (double)USEC_PER_MSEC);
217 }
218
219 for (i = 0; i < params.nthreads; i++) {
220 ret = pthread_join(worker[i], NULL);
221 if (ret)
222 err(EXIT_FAILURE, "pthread_join");
223 }
224
225 }
226
227 /* cleanup & report results */
228 cond_destroy(&thread_parent);
229 cond_destroy(&thread_worker);
230 mutex_destroy(&thread_lock);
231 pthread_attr_destroy(&thread_attr);
232
233 print_summary();
234
235 free(worker);
236 perf_cpu_map__put(cpu);
237 return ret;
238}