Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2015 Davidlohr Bueso.
4 *
5 * Block a bunch of threads and let parallel waker threads wakeup an
6 * equal amount of them. The program output reflects the avg latency
7 * for each individual thread to service its share of work. Ultimately
8 * it can be used to measure futex_wake() changes.
9 */
10#include "bench.h"
11#include <linux/compiler.h>
12#include "../util/debug.h"
13
14#ifndef HAVE_PTHREAD_BARRIER
15int bench_futex_wake_parallel(int argc __maybe_unused, const char **argv __maybe_unused)
16{
17 pr_err("%s: pthread_barrier_t unavailable, disabling this test...\n", __func__);
18 return 0;
19}
20#else /* HAVE_PTHREAD_BARRIER */
21/* For the CLR_() macros */
22#include <string.h>
23#include <pthread.h>
24
25#include <signal.h>
26#include "../util/stat.h"
27#include <subcmd/parse-options.h>
28#include <linux/kernel.h>
29#include <linux/time64.h>
30#include <errno.h>
31#include "futex.h"
32#include <internal/cpumap.h>
33#include <perf/cpumap.h>
34
35#include <err.h>
36#include <stdlib.h>
37#include <sys/time.h>
38
39struct thread_data {
40 pthread_t worker;
41 unsigned int nwoken;
42 struct timeval runtime;
43};
44
45static unsigned int nwakes = 1;
46
47/* all threads will block on the same futex -- hash bucket chaos ;) */
48static u_int32_t futex = 0;
49
50static pthread_t *blocked_worker;
51static bool done = false, silent = false, fshared = false;
52static unsigned int nblocked_threads = 0, nwaking_threads = 0;
53static pthread_mutex_t thread_lock;
54static pthread_cond_t thread_parent, thread_worker;
55static pthread_barrier_t barrier;
56static struct stats waketime_stats, wakeup_stats;
57static unsigned int threads_starting;
58static int futex_flag = 0;
59
60static const struct option options[] = {
61 OPT_UINTEGER('t', "threads", &nblocked_threads, "Specify amount of threads"),
62 OPT_UINTEGER('w', "nwakers", &nwaking_threads, "Specify amount of waking threads"),
63 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"),
64 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"),
65 OPT_END()
66};
67
68static const char * const bench_futex_wake_parallel_usage[] = {
69 "perf bench futex wake-parallel <options>",
70 NULL
71};
72
73static void *waking_workerfn(void *arg)
74{
75 struct thread_data *waker = (struct thread_data *) arg;
76 struct timeval start, end;
77
78 pthread_barrier_wait(&barrier);
79
80 gettimeofday(&start, NULL);
81
82 waker->nwoken = futex_wake(&futex, nwakes, futex_flag);
83 if (waker->nwoken != nwakes)
84 warnx("couldn't wakeup all tasks (%d/%d)",
85 waker->nwoken, nwakes);
86
87 gettimeofday(&end, NULL);
88 timersub(&end, &start, &waker->runtime);
89
90 pthread_exit(NULL);
91 return NULL;
92}
93
94static void wakeup_threads(struct thread_data *td, pthread_attr_t thread_attr)
95{
96 unsigned int i;
97
98 pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
99
100 pthread_barrier_init(&barrier, NULL, nwaking_threads + 1);
101
102 /* create and block all threads */
103 for (i = 0; i < nwaking_threads; i++) {
104 /*
105 * Thread creation order will impact per-thread latency
106 * as it will affect the order to acquire the hb spinlock.
107 * For now let the scheduler decide.
108 */
109 if (pthread_create(&td[i].worker, &thread_attr,
110 waking_workerfn, (void *)&td[i]))
111 err(EXIT_FAILURE, "pthread_create");
112 }
113
114 pthread_barrier_wait(&barrier);
115
116 for (i = 0; i < nwaking_threads; i++)
117 if (pthread_join(td[i].worker, NULL))
118 err(EXIT_FAILURE, "pthread_join");
119
120 pthread_barrier_destroy(&barrier);
121}
122
123static void *blocked_workerfn(void *arg __maybe_unused)
124{
125 pthread_mutex_lock(&thread_lock);
126 threads_starting--;
127 if (!threads_starting)
128 pthread_cond_signal(&thread_parent);
129 pthread_cond_wait(&thread_worker, &thread_lock);
130 pthread_mutex_unlock(&thread_lock);
131
132 while (1) { /* handle spurious wakeups */
133 if (futex_wait(&futex, 0, NULL, futex_flag) != EINTR)
134 break;
135 }
136
137 pthread_exit(NULL);
138 return NULL;
139}
140
141static void block_threads(pthread_t *w, pthread_attr_t thread_attr,
142 struct perf_cpu_map *cpu)
143{
144 cpu_set_t cpuset;
145 unsigned int i;
146
147 threads_starting = nblocked_threads;
148
149 /* create and block all threads */
150 for (i = 0; i < nblocked_threads; i++) {
151 CPU_ZERO(&cpuset);
152 CPU_SET(cpu->map[i % cpu->nr], &cpuset);
153
154 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset))
155 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
156
157 if (pthread_create(&w[i], &thread_attr, blocked_workerfn, NULL))
158 err(EXIT_FAILURE, "pthread_create");
159 }
160}
161
162static void print_run(struct thread_data *waking_worker, unsigned int run_num)
163{
164 unsigned int i, wakeup_avg;
165 double waketime_avg, waketime_stddev;
166 struct stats __waketime_stats, __wakeup_stats;
167
168 init_stats(&__wakeup_stats);
169 init_stats(&__waketime_stats);
170
171 for (i = 0; i < nwaking_threads; i++) {
172 update_stats(&__waketime_stats, waking_worker[i].runtime.tv_usec);
173 update_stats(&__wakeup_stats, waking_worker[i].nwoken);
174 }
175
176 waketime_avg = avg_stats(&__waketime_stats);
177 waketime_stddev = stddev_stats(&__waketime_stats);
178 wakeup_avg = avg_stats(&__wakeup_stats);
179
180 printf("[Run %d]: Avg per-thread latency (waking %d/%d threads) "
181 "in %.4f ms (+-%.2f%%)\n", run_num + 1, wakeup_avg,
182 nblocked_threads, waketime_avg / USEC_PER_MSEC,
183 rel_stddev_stats(waketime_stddev, waketime_avg));
184}
185
186static void print_summary(void)
187{
188 unsigned int wakeup_avg;
189 double waketime_avg, waketime_stddev;
190
191 waketime_avg = avg_stats(&waketime_stats);
192 waketime_stddev = stddev_stats(&waketime_stats);
193 wakeup_avg = avg_stats(&wakeup_stats);
194
195 printf("Avg per-thread latency (waking %d/%d threads) in %.4f ms (+-%.2f%%)\n",
196 wakeup_avg,
197 nblocked_threads,
198 waketime_avg / USEC_PER_MSEC,
199 rel_stddev_stats(waketime_stddev, waketime_avg));
200}
201
202
203static void do_run_stats(struct thread_data *waking_worker)
204{
205 unsigned int i;
206
207 for (i = 0; i < nwaking_threads; i++) {
208 update_stats(&waketime_stats, waking_worker[i].runtime.tv_usec);
209 update_stats(&wakeup_stats, waking_worker[i].nwoken);
210 }
211
212}
213
214static void toggle_done(int sig __maybe_unused,
215 siginfo_t *info __maybe_unused,
216 void *uc __maybe_unused)
217{
218 done = true;
219}
220
221int bench_futex_wake_parallel(int argc, const char **argv)
222{
223 int ret = 0;
224 unsigned int i, j;
225 struct sigaction act;
226 pthread_attr_t thread_attr;
227 struct thread_data *waking_worker;
228 struct perf_cpu_map *cpu;
229
230 argc = parse_options(argc, argv, options,
231 bench_futex_wake_parallel_usage, 0);
232 if (argc) {
233 usage_with_options(bench_futex_wake_parallel_usage, options);
234 exit(EXIT_FAILURE);
235 }
236
237 sigfillset(&act.sa_mask);
238 act.sa_sigaction = toggle_done;
239 sigaction(SIGINT, &act, NULL);
240
241 cpu = perf_cpu_map__new(NULL);
242 if (!cpu)
243 err(EXIT_FAILURE, "calloc");
244
245 if (!nblocked_threads)
246 nblocked_threads = cpu->nr;
247
248 /* some sanity checks */
249 if (nwaking_threads > nblocked_threads || !nwaking_threads)
250 nwaking_threads = nblocked_threads;
251
252 if (nblocked_threads % nwaking_threads)
253 errx(EXIT_FAILURE, "Must be perfectly divisible");
254 /*
255 * Each thread will wakeup nwakes tasks in
256 * a single futex_wait call.
257 */
258 nwakes = nblocked_threads/nwaking_threads;
259
260 blocked_worker = calloc(nblocked_threads, sizeof(*blocked_worker));
261 if (!blocked_worker)
262 err(EXIT_FAILURE, "calloc");
263
264 if (!fshared)
265 futex_flag = FUTEX_PRIVATE_FLAG;
266
267 printf("Run summary [PID %d]: blocking on %d threads (at [%s] "
268 "futex %p), %d threads waking up %d at a time.\n\n",
269 getpid(), nblocked_threads, fshared ? "shared":"private",
270 &futex, nwaking_threads, nwakes);
271
272 init_stats(&wakeup_stats);
273 init_stats(&waketime_stats);
274
275 pthread_attr_init(&thread_attr);
276 pthread_mutex_init(&thread_lock, NULL);
277 pthread_cond_init(&thread_parent, NULL);
278 pthread_cond_init(&thread_worker, NULL);
279
280 for (j = 0; j < bench_repeat && !done; j++) {
281 waking_worker = calloc(nwaking_threads, sizeof(*waking_worker));
282 if (!waking_worker)
283 err(EXIT_FAILURE, "calloc");
284
285 /* create, launch & block all threads */
286 block_threads(blocked_worker, thread_attr, cpu);
287
288 /* make sure all threads are already blocked */
289 pthread_mutex_lock(&thread_lock);
290 while (threads_starting)
291 pthread_cond_wait(&thread_parent, &thread_lock);
292 pthread_cond_broadcast(&thread_worker);
293 pthread_mutex_unlock(&thread_lock);
294
295 usleep(100000);
296
297 /* Ok, all threads are patiently blocked, start waking folks up */
298 wakeup_threads(waking_worker, thread_attr);
299
300 for (i = 0; i < nblocked_threads; i++) {
301 ret = pthread_join(blocked_worker[i], NULL);
302 if (ret)
303 err(EXIT_FAILURE, "pthread_join");
304 }
305
306 do_run_stats(waking_worker);
307 if (!silent)
308 print_run(waking_worker, j);
309
310 free(waking_worker);
311 }
312
313 /* cleanup & report results */
314 pthread_cond_destroy(&thread_parent);
315 pthread_cond_destroy(&thread_worker);
316 pthread_mutex_destroy(&thread_lock);
317 pthread_attr_destroy(&thread_attr);
318
319 print_summary();
320
321 free(blocked_worker);
322 return ret;
323}
324#endif /* HAVE_PTHREAD_BARRIER */
1/*
2 * Copyright (C) 2015 Davidlohr Bueso.
3 *
4 * Block a bunch of threads and let parallel waker threads wakeup an
5 * equal amount of them. The program output reflects the avg latency
6 * for each individual thread to service its share of work. Ultimately
7 * it can be used to measure futex_wake() changes.
8 */
9
10#include "../perf.h"
11#include "../util/util.h"
12#include "../util/stat.h"
13#include <subcmd/parse-options.h>
14#include "../util/header.h"
15#include "bench.h"
16#include "futex.h"
17
18#include <err.h>
19#include <stdlib.h>
20#include <sys/time.h>
21#include <pthread.h>
22
23struct thread_data {
24 pthread_t worker;
25 unsigned int nwoken;
26 struct timeval runtime;
27};
28
29static unsigned int nwakes = 1;
30
31/* all threads will block on the same futex -- hash bucket chaos ;) */
32static u_int32_t futex = 0;
33
34static pthread_t *blocked_worker;
35static bool done = false, silent = false, fshared = false;
36static unsigned int nblocked_threads = 0, nwaking_threads = 0;
37static pthread_mutex_t thread_lock;
38static pthread_cond_t thread_parent, thread_worker;
39static struct stats waketime_stats, wakeup_stats;
40static unsigned int ncpus, threads_starting;
41static int futex_flag = 0;
42
43static const struct option options[] = {
44 OPT_UINTEGER('t', "threads", &nblocked_threads, "Specify amount of threads"),
45 OPT_UINTEGER('w', "nwakers", &nwaking_threads, "Specify amount of waking threads"),
46 OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"),
47 OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"),
48 OPT_END()
49};
50
51static const char * const bench_futex_wake_parallel_usage[] = {
52 "perf bench futex wake-parallel <options>",
53 NULL
54};
55
56static void *waking_workerfn(void *arg)
57{
58 struct thread_data *waker = (struct thread_data *) arg;
59 struct timeval start, end;
60
61 gettimeofday(&start, NULL);
62
63 waker->nwoken = futex_wake(&futex, nwakes, futex_flag);
64 if (waker->nwoken != nwakes)
65 warnx("couldn't wakeup all tasks (%d/%d)",
66 waker->nwoken, nwakes);
67
68 gettimeofday(&end, NULL);
69 timersub(&end, &start, &waker->runtime);
70
71 pthread_exit(NULL);
72 return NULL;
73}
74
75static void wakeup_threads(struct thread_data *td, pthread_attr_t thread_attr)
76{
77 unsigned int i;
78
79 pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
80
81 /* create and block all threads */
82 for (i = 0; i < nwaking_threads; i++) {
83 /*
84 * Thread creation order will impact per-thread latency
85 * as it will affect the order to acquire the hb spinlock.
86 * For now let the scheduler decide.
87 */
88 if (pthread_create(&td[i].worker, &thread_attr,
89 waking_workerfn, (void *)&td[i]))
90 err(EXIT_FAILURE, "pthread_create");
91 }
92
93 for (i = 0; i < nwaking_threads; i++)
94 if (pthread_join(td[i].worker, NULL))
95 err(EXIT_FAILURE, "pthread_join");
96}
97
98static void *blocked_workerfn(void *arg __maybe_unused)
99{
100 pthread_mutex_lock(&thread_lock);
101 threads_starting--;
102 if (!threads_starting)
103 pthread_cond_signal(&thread_parent);
104 pthread_cond_wait(&thread_worker, &thread_lock);
105 pthread_mutex_unlock(&thread_lock);
106
107 while (1) { /* handle spurious wakeups */
108 if (futex_wait(&futex, 0, NULL, futex_flag) != EINTR)
109 break;
110 }
111
112 pthread_exit(NULL);
113 return NULL;
114}
115
116static void block_threads(pthread_t *w, pthread_attr_t thread_attr)
117{
118 cpu_set_t cpu;
119 unsigned int i;
120
121 threads_starting = nblocked_threads;
122
123 /* create and block all threads */
124 for (i = 0; i < nblocked_threads; i++) {
125 CPU_ZERO(&cpu);
126 CPU_SET(i % ncpus, &cpu);
127
128 if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu))
129 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
130
131 if (pthread_create(&w[i], &thread_attr, blocked_workerfn, NULL))
132 err(EXIT_FAILURE, "pthread_create");
133 }
134}
135
136static void print_run(struct thread_data *waking_worker, unsigned int run_num)
137{
138 unsigned int i, wakeup_avg;
139 double waketime_avg, waketime_stddev;
140 struct stats __waketime_stats, __wakeup_stats;
141
142 init_stats(&__wakeup_stats);
143 init_stats(&__waketime_stats);
144
145 for (i = 0; i < nwaking_threads; i++) {
146 update_stats(&__waketime_stats, waking_worker[i].runtime.tv_usec);
147 update_stats(&__wakeup_stats, waking_worker[i].nwoken);
148 }
149
150 waketime_avg = avg_stats(&__waketime_stats);
151 waketime_stddev = stddev_stats(&__waketime_stats);
152 wakeup_avg = avg_stats(&__wakeup_stats);
153
154 printf("[Run %d]: Avg per-thread latency (waking %d/%d threads) "
155 "in %.4f ms (+-%.2f%%)\n", run_num + 1, wakeup_avg,
156 nblocked_threads, waketime_avg/1e3,
157 rel_stddev_stats(waketime_stddev, waketime_avg));
158}
159
160static void print_summary(void)
161{
162 unsigned int wakeup_avg;
163 double waketime_avg, waketime_stddev;
164
165 waketime_avg = avg_stats(&waketime_stats);
166 waketime_stddev = stddev_stats(&waketime_stats);
167 wakeup_avg = avg_stats(&wakeup_stats);
168
169 printf("Avg per-thread latency (waking %d/%d threads) in %.4f ms (+-%.2f%%)\n",
170 wakeup_avg,
171 nblocked_threads,
172 waketime_avg/1e3,
173 rel_stddev_stats(waketime_stddev, waketime_avg));
174}
175
176
177static void do_run_stats(struct thread_data *waking_worker)
178{
179 unsigned int i;
180
181 for (i = 0; i < nwaking_threads; i++) {
182 update_stats(&waketime_stats, waking_worker[i].runtime.tv_usec);
183 update_stats(&wakeup_stats, waking_worker[i].nwoken);
184 }
185
186}
187
188static void toggle_done(int sig __maybe_unused,
189 siginfo_t *info __maybe_unused,
190 void *uc __maybe_unused)
191{
192 done = true;
193}
194
195int bench_futex_wake_parallel(int argc, const char **argv,
196 const char *prefix __maybe_unused)
197{
198 int ret = 0;
199 unsigned int i, j;
200 struct sigaction act;
201 pthread_attr_t thread_attr;
202 struct thread_data *waking_worker;
203
204 argc = parse_options(argc, argv, options,
205 bench_futex_wake_parallel_usage, 0);
206 if (argc) {
207 usage_with_options(bench_futex_wake_parallel_usage, options);
208 exit(EXIT_FAILURE);
209 }
210
211 sigfillset(&act.sa_mask);
212 act.sa_sigaction = toggle_done;
213 sigaction(SIGINT, &act, NULL);
214
215 ncpus = sysconf(_SC_NPROCESSORS_ONLN);
216 if (!nblocked_threads)
217 nblocked_threads = ncpus;
218
219 /* some sanity checks */
220 if (nwaking_threads > nblocked_threads || !nwaking_threads)
221 nwaking_threads = nblocked_threads;
222
223 if (nblocked_threads % nwaking_threads)
224 errx(EXIT_FAILURE, "Must be perfectly divisible");
225 /*
226 * Each thread will wakeup nwakes tasks in
227 * a single futex_wait call.
228 */
229 nwakes = nblocked_threads/nwaking_threads;
230
231 blocked_worker = calloc(nblocked_threads, sizeof(*blocked_worker));
232 if (!blocked_worker)
233 err(EXIT_FAILURE, "calloc");
234
235 if (!fshared)
236 futex_flag = FUTEX_PRIVATE_FLAG;
237
238 printf("Run summary [PID %d]: blocking on %d threads (at [%s] "
239 "futex %p), %d threads waking up %d at a time.\n\n",
240 getpid(), nblocked_threads, fshared ? "shared":"private",
241 &futex, nwaking_threads, nwakes);
242
243 init_stats(&wakeup_stats);
244 init_stats(&waketime_stats);
245
246 pthread_attr_init(&thread_attr);
247 pthread_mutex_init(&thread_lock, NULL);
248 pthread_cond_init(&thread_parent, NULL);
249 pthread_cond_init(&thread_worker, NULL);
250
251 for (j = 0; j < bench_repeat && !done; j++) {
252 waking_worker = calloc(nwaking_threads, sizeof(*waking_worker));
253 if (!waking_worker)
254 err(EXIT_FAILURE, "calloc");
255
256 /* create, launch & block all threads */
257 block_threads(blocked_worker, thread_attr);
258
259 /* make sure all threads are already blocked */
260 pthread_mutex_lock(&thread_lock);
261 while (threads_starting)
262 pthread_cond_wait(&thread_parent, &thread_lock);
263 pthread_cond_broadcast(&thread_worker);
264 pthread_mutex_unlock(&thread_lock);
265
266 usleep(100000);
267
268 /* Ok, all threads are patiently blocked, start waking folks up */
269 wakeup_threads(waking_worker, thread_attr);
270
271 for (i = 0; i < nblocked_threads; i++) {
272 ret = pthread_join(blocked_worker[i], NULL);
273 if (ret)
274 err(EXIT_FAILURE, "pthread_join");
275 }
276
277 do_run_stats(waking_worker);
278 if (!silent)
279 print_run(waking_worker, j);
280
281 free(waking_worker);
282 }
283
284 /* cleanup & report results */
285 pthread_cond_destroy(&thread_parent);
286 pthread_cond_destroy(&thread_worker);
287 pthread_mutex_destroy(&thread_lock);
288 pthread_attr_destroy(&thread_attr);
289
290 print_summary();
291
292 free(blocked_worker);
293 return ret;
294}