Loading...
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#include "../perf.h"
12#include "../util/util.h"
13#include "../util/stat.h"
14#include <subcmd/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
24static u_int32_t futex1 = 0, futex2 = 0;
25
26/*
27 * How many tasks to requeue at a time.
28 * Default to 1 in order to make the kernel work more.
29 */
30static unsigned int nrequeue = 1;
31
32static pthread_t *worker;
33static bool done = false, silent = false, fshared = false;
34static pthread_mutex_t thread_lock;
35static pthread_cond_t thread_parent, thread_worker;
36static struct stats requeuetime_stats, requeued_stats;
37static unsigned int ncpus, threads_starting, nthreads = 0;
38static int futex_flag = 0;
39
40static const struct option options[] = {
41 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
42 OPT_UINTEGER('q', "nrequeue", &nrequeue, "Specify amount of threads to requeue at once"),
43 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"),
44 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"),
45 OPT_END()
46};
47
48static const char * const bench_futex_requeue_usage[] = {
49 "perf bench futex requeue <options>",
50 NULL
51};
52
53static void print_summary(void)
54{
55 double requeuetime_avg = avg_stats(&requeuetime_stats);
56 double requeuetime_stddev = stddev_stats(&requeuetime_stats);
57 unsigned int requeued_avg = avg_stats(&requeued_stats);
58
59 printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n",
60 requeued_avg,
61 nthreads,
62 requeuetime_avg/1e3,
63 rel_stddev_stats(requeuetime_stddev, requeuetime_avg));
64}
65
66static void *workerfn(void *arg __maybe_unused)
67{
68 pthread_mutex_lock(&thread_lock);
69 threads_starting--;
70 if (!threads_starting)
71 pthread_cond_signal(&thread_parent);
72 pthread_cond_wait(&thread_worker, &thread_lock);
73 pthread_mutex_unlock(&thread_lock);
74
75 futex_wait(&futex1, 0, NULL, futex_flag);
76 return NULL;
77}
78
79static void block_threads(pthread_t *w,
80 pthread_attr_t thread_attr)
81{
82 cpu_set_t cpu;
83 unsigned int i;
84
85 threads_starting = nthreads;
86
87 /* create and block all threads */
88 for (i = 0; i < nthreads; i++) {
89 CPU_ZERO(&cpu);
90 CPU_SET(i % ncpus, &cpu);
91
92 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu))
93 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
94
95 if (pthread_create(&w[i], &thread_attr, workerfn, NULL))
96 err(EXIT_FAILURE, "pthread_create");
97 }
98}
99
100static void toggle_done(int sig __maybe_unused,
101 siginfo_t *info __maybe_unused,
102 void *uc __maybe_unused)
103{
104 done = true;
105}
106
107int bench_futex_requeue(int argc, const char **argv,
108 const char *prefix __maybe_unused)
109{
110 int ret = 0;
111 unsigned int i, j;
112 struct sigaction act;
113 pthread_attr_t thread_attr;
114
115 argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0);
116 if (argc)
117 goto err;
118
119 ncpus = sysconf(_SC_NPROCESSORS_ONLN);
120
121 sigfillset(&act.sa_mask);
122 act.sa_sigaction = toggle_done;
123 sigaction(SIGINT, &act, NULL);
124
125 if (!nthreads)
126 nthreads = ncpus;
127
128 worker = calloc(nthreads, sizeof(*worker));
129 if (!worker)
130 err(EXIT_FAILURE, "calloc");
131
132 if (!fshared)
133 futex_flag = FUTEX_PRIVATE_FLAG;
134
135 if (nrequeue > nthreads)
136 nrequeue = nthreads;
137
138 printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %p), "
139 "%d at a time.\n\n", getpid(), nthreads,
140 fshared ? "shared":"private", &futex1, &futex2, nrequeue);
141
142 init_stats(&requeued_stats);
143 init_stats(&requeuetime_stats);
144 pthread_attr_init(&thread_attr);
145 pthread_mutex_init(&thread_lock, NULL);
146 pthread_cond_init(&thread_parent, NULL);
147 pthread_cond_init(&thread_worker, NULL);
148
149 for (j = 0; j < bench_repeat && !done; j++) {
150 unsigned int nrequeued = 0;
151 struct timeval start, end, runtime;
152
153 /* create, launch & block all threads */
154 block_threads(worker, thread_attr);
155
156 /* make sure all threads are already blocked */
157 pthread_mutex_lock(&thread_lock);
158 while (threads_starting)
159 pthread_cond_wait(&thread_parent, &thread_lock);
160 pthread_cond_broadcast(&thread_worker);
161 pthread_mutex_unlock(&thread_lock);
162
163 usleep(100000);
164
165 /* Ok, all threads are patiently blocked, start requeueing */
166 gettimeofday(&start, NULL);
167 while (nrequeued < nthreads) {
168 /*
169 * Do not wakeup any tasks blocked on futex1, allowing
170 * us to really measure futex_wait functionality.
171 */
172 nrequeued += futex_cmp_requeue(&futex1, 0, &futex2, 0,
173 nrequeue, futex_flag);
174 }
175
176 gettimeofday(&end, NULL);
177 timersub(&end, &start, &runtime);
178
179 update_stats(&requeued_stats, nrequeued);
180 update_stats(&requeuetime_stats, runtime.tv_usec);
181
182 if (!silent) {
183 printf("[Run %d]: Requeued %d of %d threads in %.4f ms\n",
184 j + 1, nrequeued, nthreads, runtime.tv_usec/1e3);
185 }
186
187 /* everybody should be blocked on futex2, wake'em up */
188 nrequeued = futex_wake(&futex2, nrequeued, futex_flag);
189 if (nthreads != nrequeued)
190 warnx("couldn't wakeup all tasks (%d/%d)", nrequeued, nthreads);
191
192 for (i = 0; i < nthreads; i++) {
193 ret = pthread_join(worker[i], NULL);
194 if (ret)
195 err(EXIT_FAILURE, "pthread_join");
196 }
197 }
198
199 /* cleanup & report results */
200 pthread_cond_destroy(&thread_parent);
201 pthread_cond_destroy(&thread_worker);
202 pthread_mutex_destroy(&thread_lock);
203 pthread_attr_destroy(&thread_attr);
204
205 print_summary();
206
207 free(worker);
208 return ret;
209err:
210 usage_with_options(bench_futex_requeue_usage, options);
211 exit(EXIT_FAILURE);
212}
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}