Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com>
4 *
5 * futex-requeue: Block a bunch of threads on futex1 and requeue them
6 * on futex2, N at a time.
7 *
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.
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
32static u_int32_t futex1 = 0, futex2 = 0;
33
34/*
35 * How many tasks to requeue at a time.
36 * Default to 1 in order to make the kernel work more.
37 */
38static unsigned int nrequeue = 1;
39
40static pthread_t *worker;
41static bool done = false, silent = false, fshared = false;
42static pthread_mutex_t thread_lock;
43static pthread_cond_t thread_parent, thread_worker;
44static struct stats requeuetime_stats, requeued_stats;
45static unsigned int threads_starting, nthreads = 0;
46static int futex_flag = 0;
47
48static const struct option options[] = {
49 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
50 OPT_UINTEGER('q', "nrequeue", &nrequeue, "Specify amount of threads to requeue at once"),
51 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"),
52 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"),
53 OPT_END()
54};
55
56static const char * const bench_futex_requeue_usage[] = {
57 "perf bench futex requeue <options>",
58 NULL
59};
60
61static void print_summary(void)
62{
63 double requeuetime_avg = avg_stats(&requeuetime_stats);
64 double requeuetime_stddev = stddev_stats(&requeuetime_stats);
65 unsigned int requeued_avg = avg_stats(&requeued_stats);
66
67 printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n",
68 requeued_avg,
69 nthreads,
70 requeuetime_avg / USEC_PER_MSEC,
71 rel_stddev_stats(requeuetime_stddev, requeuetime_avg));
72}
73
74static void *workerfn(void *arg __maybe_unused)
75{
76 pthread_mutex_lock(&thread_lock);
77 threads_starting--;
78 if (!threads_starting)
79 pthread_cond_signal(&thread_parent);
80 pthread_cond_wait(&thread_worker, &thread_lock);
81 pthread_mutex_unlock(&thread_lock);
82
83 futex_wait(&futex1, 0, NULL, futex_flag);
84 return NULL;
85}
86
87static void block_threads(pthread_t *w,
88 pthread_attr_t thread_attr, struct perf_cpu_map *cpu)
89{
90 cpu_set_t cpuset;
91 unsigned int i;
92
93 threads_starting = nthreads;
94
95 /* create and block all threads */
96 for (i = 0; i < nthreads; i++) {
97 CPU_ZERO(&cpuset);
98 CPU_SET(cpu->map[i % cpu->nr], &cpuset);
99
100 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset))
101 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
102
103 if (pthread_create(&w[i], &thread_attr, workerfn, NULL))
104 err(EXIT_FAILURE, "pthread_create");
105 }
106}
107
108static void toggle_done(int sig __maybe_unused,
109 siginfo_t *info __maybe_unused,
110 void *uc __maybe_unused)
111{
112 done = true;
113}
114
115int bench_futex_requeue(int argc, const char **argv)
116{
117 int ret = 0;
118 unsigned int i, j;
119 struct sigaction act;
120 pthread_attr_t thread_attr;
121 struct perf_cpu_map *cpu;
122
123 argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0);
124 if (argc)
125 goto err;
126
127 cpu = perf_cpu_map__new(NULL);
128 if (!cpu)
129 err(EXIT_FAILURE, "cpu_map__new");
130
131 sigfillset(&act.sa_mask);
132 act.sa_sigaction = toggle_done;
133 sigaction(SIGINT, &act, NULL);
134
135 if (!nthreads)
136 nthreads = cpu->nr;
137
138 worker = calloc(nthreads, sizeof(*worker));
139 if (!worker)
140 err(EXIT_FAILURE, "calloc");
141
142 if (!fshared)
143 futex_flag = FUTEX_PRIVATE_FLAG;
144
145 if (nrequeue > nthreads)
146 nrequeue = nthreads;
147
148 printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %p), "
149 "%d at a time.\n\n", getpid(), nthreads,
150 fshared ? "shared":"private", &futex1, &futex2, nrequeue);
151
152 init_stats(&requeued_stats);
153 init_stats(&requeuetime_stats);
154 pthread_attr_init(&thread_attr);
155 pthread_mutex_init(&thread_lock, NULL);
156 pthread_cond_init(&thread_parent, NULL);
157 pthread_cond_init(&thread_worker, NULL);
158
159 for (j = 0; j < bench_repeat && !done; j++) {
160 unsigned int nrequeued = 0;
161 struct timeval start, end, runtime;
162
163 /* create, launch & block all threads */
164 block_threads(worker, thread_attr, cpu);
165
166 /* make sure all threads are already blocked */
167 pthread_mutex_lock(&thread_lock);
168 while (threads_starting)
169 pthread_cond_wait(&thread_parent, &thread_lock);
170 pthread_cond_broadcast(&thread_worker);
171 pthread_mutex_unlock(&thread_lock);
172
173 usleep(100000);
174
175 /* Ok, all threads are patiently blocked, start requeueing */
176 gettimeofday(&start, NULL);
177 while (nrequeued < nthreads) {
178 /*
179 * Do not wakeup any tasks blocked on futex1, allowing
180 * us to really measure futex_wait functionality.
181 */
182 nrequeued += futex_cmp_requeue(&futex1, 0, &futex2, 0,
183 nrequeue, futex_flag);
184 }
185
186 gettimeofday(&end, NULL);
187 timersub(&end, &start, &runtime);
188
189 update_stats(&requeued_stats, nrequeued);
190 update_stats(&requeuetime_stats, runtime.tv_usec);
191
192 if (!silent) {
193 printf("[Run %d]: Requeued %d of %d threads in %.4f ms\n",
194 j + 1, nrequeued, nthreads, runtime.tv_usec / (double)USEC_PER_MSEC);
195 }
196
197 /* everybody should be blocked on futex2, wake'em up */
198 nrequeued = futex_wake(&futex2, nrequeued, futex_flag);
199 if (nthreads != nrequeued)
200 warnx("couldn't wakeup all tasks (%d/%d)", nrequeued, nthreads);
201
202 for (i = 0; i < nthreads; i++) {
203 ret = pthread_join(worker[i], NULL);
204 if (ret)
205 err(EXIT_FAILURE, "pthread_join");
206 }
207 }
208
209 /* cleanup & report results */
210 pthread_cond_destroy(&thread_parent);
211 pthread_cond_destroy(&thread_worker);
212 pthread_mutex_destroy(&thread_lock);
213 pthread_attr_destroy(&thread_attr);
214
215 print_summary();
216
217 free(worker);
218 return ret;
219err:
220 usage_with_options(bench_futex_requeue_usage, options);
221 exit(EXIT_FAILURE);
222}
1/*
2 * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com>
3 *
4 * futex-requeue: Block a bunch of threads on futex1 and requeue them
5 * on futex2, N at a time.
6 *
7 * This program is particularly useful to measure the latency of nthread
8 * requeues without waking up any tasks -- thus mimicking a regular futex_wait.
9 */
10
11/* For the CLR_() macros */
12#include <pthread.h>
13
14#include <signal.h>
15#include "../util/stat.h"
16#include <subcmd/parse-options.h>
17#include <linux/compiler.h>
18#include <linux/kernel.h>
19#include <linux/time64.h>
20#include <errno.h>
21#include "bench.h"
22#include "futex.h"
23
24#include <err.h>
25#include <stdlib.h>
26#include <sys/time.h>
27
28static u_int32_t futex1 = 0, futex2 = 0;
29
30/*
31 * How many tasks to requeue at a time.
32 * Default to 1 in order to make the kernel work more.
33 */
34static unsigned int nrequeue = 1;
35
36static pthread_t *worker;
37static bool done = false, silent = false, fshared = false;
38static pthread_mutex_t thread_lock;
39static pthread_cond_t thread_parent, thread_worker;
40static struct stats requeuetime_stats, requeued_stats;
41static unsigned int ncpus, threads_starting, nthreads = 0;
42static int futex_flag = 0;
43
44static const struct option options[] = {
45 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
46 OPT_UINTEGER('q', "nrequeue", &nrequeue, "Specify amount of threads to requeue at once"),
47 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"),
48 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"),
49 OPT_END()
50};
51
52static const char * const bench_futex_requeue_usage[] = {
53 "perf bench futex requeue <options>",
54 NULL
55};
56
57static void print_summary(void)
58{
59 double requeuetime_avg = avg_stats(&requeuetime_stats);
60 double requeuetime_stddev = stddev_stats(&requeuetime_stats);
61 unsigned int requeued_avg = avg_stats(&requeued_stats);
62
63 printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n",
64 requeued_avg,
65 nthreads,
66 requeuetime_avg / USEC_PER_MSEC,
67 rel_stddev_stats(requeuetime_stddev, requeuetime_avg));
68}
69
70static void *workerfn(void *arg __maybe_unused)
71{
72 pthread_mutex_lock(&thread_lock);
73 threads_starting--;
74 if (!threads_starting)
75 pthread_cond_signal(&thread_parent);
76 pthread_cond_wait(&thread_worker, &thread_lock);
77 pthread_mutex_unlock(&thread_lock);
78
79 futex_wait(&futex1, 0, NULL, futex_flag);
80 return NULL;
81}
82
83static void block_threads(pthread_t *w,
84 pthread_attr_t thread_attr)
85{
86 cpu_set_t cpu;
87 unsigned int i;
88
89 threads_starting = nthreads;
90
91 /* create and block all threads */
92 for (i = 0; i < nthreads; i++) {
93 CPU_ZERO(&cpu);
94 CPU_SET(i % ncpus, &cpu);
95
96 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu))
97 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
98
99 if (pthread_create(&w[i], &thread_attr, workerfn, NULL))
100 err(EXIT_FAILURE, "pthread_create");
101 }
102}
103
104static void toggle_done(int sig __maybe_unused,
105 siginfo_t *info __maybe_unused,
106 void *uc __maybe_unused)
107{
108 done = true;
109}
110
111int bench_futex_requeue(int argc, const char **argv,
112 const char *prefix __maybe_unused)
113{
114 int ret = 0;
115 unsigned int i, j;
116 struct sigaction act;
117 pthread_attr_t thread_attr;
118
119 argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0);
120 if (argc)
121 goto err;
122
123 ncpus = sysconf(_SC_NPROCESSORS_ONLN);
124
125 sigfillset(&act.sa_mask);
126 act.sa_sigaction = toggle_done;
127 sigaction(SIGINT, &act, NULL);
128
129 if (!nthreads)
130 nthreads = ncpus;
131 else
132 nthreads = futexbench_sanitize_numeric(nthreads);
133
134 worker = calloc(nthreads, sizeof(*worker));
135 if (!worker)
136 err(EXIT_FAILURE, "calloc");
137
138 if (!fshared)
139 futex_flag = FUTEX_PRIVATE_FLAG;
140
141 if (nrequeue > nthreads)
142 nrequeue = nthreads;
143
144 printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %p), "
145 "%d at a time.\n\n", getpid(), nthreads,
146 fshared ? "shared":"private", &futex1, &futex2, nrequeue);
147
148 init_stats(&requeued_stats);
149 init_stats(&requeuetime_stats);
150 pthread_attr_init(&thread_attr);
151 pthread_mutex_init(&thread_lock, NULL);
152 pthread_cond_init(&thread_parent, NULL);
153 pthread_cond_init(&thread_worker, NULL);
154
155 for (j = 0; j < bench_repeat && !done; j++) {
156 unsigned int nrequeued = 0;
157 struct timeval start, end, runtime;
158
159 /* create, launch & block all threads */
160 block_threads(worker, thread_attr);
161
162 /* make sure all threads are already blocked */
163 pthread_mutex_lock(&thread_lock);
164 while (threads_starting)
165 pthread_cond_wait(&thread_parent, &thread_lock);
166 pthread_cond_broadcast(&thread_worker);
167 pthread_mutex_unlock(&thread_lock);
168
169 usleep(100000);
170
171 /* Ok, all threads are patiently blocked, start requeueing */
172 gettimeofday(&start, NULL);
173 while (nrequeued < nthreads) {
174 /*
175 * Do not wakeup any tasks blocked on futex1, allowing
176 * us to really measure futex_wait functionality.
177 */
178 nrequeued += futex_cmp_requeue(&futex1, 0, &futex2, 0,
179 nrequeue, futex_flag);
180 }
181
182 gettimeofday(&end, NULL);
183 timersub(&end, &start, &runtime);
184
185 update_stats(&requeued_stats, nrequeued);
186 update_stats(&requeuetime_stats, runtime.tv_usec);
187
188 if (!silent) {
189 printf("[Run %d]: Requeued %d of %d threads in %.4f ms\n",
190 j + 1, nrequeued, nthreads, runtime.tv_usec / (double)USEC_PER_MSEC);
191 }
192
193 /* everybody should be blocked on futex2, wake'em up */
194 nrequeued = futex_wake(&futex2, nrequeued, futex_flag);
195 if (nthreads != nrequeued)
196 warnx("couldn't wakeup all tasks (%d/%d)", nrequeued, nthreads);
197
198 for (i = 0; i < nthreads; i++) {
199 ret = pthread_join(worker[i], NULL);
200 if (ret)
201 err(EXIT_FAILURE, "pthread_join");
202 }
203 }
204
205 /* cleanup & report results */
206 pthread_cond_destroy(&thread_parent);
207 pthread_cond_destroy(&thread_worker);
208 pthread_mutex_destroy(&thread_lock);
209 pthread_attr_destroy(&thread_attr);
210
211 print_summary();
212
213 free(worker);
214 return ret;
215err:
216 usage_with_options(bench_futex_requeue_usage, options);
217 exit(EXIT_FAILURE);
218}