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 (in the non-pi case) -- thus
10 * mimicking a regular futex_wait.
11 */
12
13 /* For the CLR_() macros */
14 #include <string.h>
15 #include <pthread.h>
16
17 #include <signal.h>
18 #include "../util/mutex.h"
19 #include "../util/stat.h"
20 #include <subcmd/parse-options.h>
21 #include <linux/compiler.h>
22 #include <linux/kernel.h>
23 #include <linux/time64.h>
24 #include <errno.h>
25 #include <perf/cpumap.h>
26 #include "bench.h"
27 #include "futex.h"
28
29 #include <err.h>
30 #include <stdlib.h>
31 #include <sys/time.h>
32 #include <sys/mman.h>
33
34 static u_int32_t futex1 = 0, futex2 = 0;
35
36 static pthread_t *worker;
37 static bool done = false;
38 static struct mutex thread_lock;
39 static struct cond thread_parent, thread_worker;
40 static struct stats requeuetime_stats, requeued_stats;
41 static unsigned int threads_starting;
42 static int futex_flag = 0;
43
44 static struct bench_futex_parameters params = {
45 /*
46 * How many tasks to requeue at a time.
47 * Default to 1 in order to make the kernel work more.
48 */
49 .nrequeue = 1,
50 };
51
52 static const struct option options[] = {
53 OPT_UINTEGER('t', "threads", ¶ms.nthreads, "Specify amount of threads"),
54 OPT_UINTEGER('q', "nrequeue", ¶ms.nrequeue, "Specify amount of threads to requeue 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 OPT_BOOLEAN( 'B', "broadcast", ¶ms.broadcast, "Requeue all threads at once"),
59 OPT_BOOLEAN( 'p', "pi", ¶ms.pi, "Use PI-aware variants of FUTEX_CMP_REQUEUE"),
60
61 OPT_END()
62 };
63
64 static const char * const bench_futex_requeue_usage[] = {
65 "perf bench futex requeue <options>",
66 NULL
67 };
68
print_summary(void)69 static void print_summary(void)
70 {
71 double requeuetime_avg = avg_stats(&requeuetime_stats);
72 double requeuetime_stddev = stddev_stats(&requeuetime_stats);
73 unsigned int requeued_avg = avg_stats(&requeued_stats);
74
75 printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n",
76 requeued_avg,
77 params.nthreads,
78 requeuetime_avg / USEC_PER_MSEC,
79 rel_stddev_stats(requeuetime_stddev, requeuetime_avg));
80 }
81
workerfn(void * arg __maybe_unused)82 static void *workerfn(void *arg __maybe_unused)
83 {
84 int ret;
85
86 mutex_lock(&thread_lock);
87 threads_starting--;
88 if (!threads_starting)
89 cond_signal(&thread_parent);
90 cond_wait(&thread_worker, &thread_lock);
91 mutex_unlock(&thread_lock);
92
93 while (1) {
94 if (!params.pi) {
95 ret = futex_wait(&futex1, 0, NULL, futex_flag);
96 if (!ret)
97 break;
98
99 if (ret && errno != EAGAIN) {
100 if (!params.silent)
101 warnx("futex_wait");
102 break;
103 }
104 } else {
105 ret = futex_wait_requeue_pi(&futex1, 0, &futex2,
106 NULL, futex_flag);
107 if (!ret) {
108 /* got the lock at futex2 */
109 futex_unlock_pi(&futex2, futex_flag);
110 break;
111 }
112
113 if (ret && errno != EAGAIN) {
114 if (!params.silent)
115 warnx("futex_wait_requeue_pi");
116 break;
117 }
118 }
119 }
120
121 return NULL;
122 }
123
block_threads(pthread_t * w,pthread_attr_t thread_attr,struct perf_cpu_map * cpu)124 static void block_threads(pthread_t *w,
125 pthread_attr_t thread_attr, struct perf_cpu_map *cpu)
126 {
127 cpu_set_t *cpuset;
128 unsigned int i;
129 int nrcpus = perf_cpu_map__nr(cpu);
130 size_t size;
131
132 threads_starting = params.nthreads;
133
134 cpuset = CPU_ALLOC(nrcpus);
135 BUG_ON(!cpuset);
136 size = CPU_ALLOC_SIZE(nrcpus);
137
138 /* create and block all threads */
139 for (i = 0; i < params.nthreads; i++) {
140 CPU_ZERO_S(size, cpuset);
141 CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset);
142
143 if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) {
144 CPU_FREE(cpuset);
145 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
146 }
147
148 if (pthread_create(&w[i], &thread_attr, workerfn, NULL)) {
149 CPU_FREE(cpuset);
150 err(EXIT_FAILURE, "pthread_create");
151 }
152 }
153 CPU_FREE(cpuset);
154 }
155
toggle_done(int sig __maybe_unused,siginfo_t * info __maybe_unused,void * uc __maybe_unused)156 static void toggle_done(int sig __maybe_unused,
157 siginfo_t *info __maybe_unused,
158 void *uc __maybe_unused)
159 {
160 done = true;
161 }
162
bench_futex_requeue(int argc,const char ** argv)163 int bench_futex_requeue(int argc, const char **argv)
164 {
165 int ret = 0;
166 unsigned int i, j;
167 struct sigaction act;
168 pthread_attr_t thread_attr;
169 struct perf_cpu_map *cpu;
170
171 argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0);
172 if (argc)
173 goto err;
174
175 cpu = perf_cpu_map__new(NULL);
176 if (!cpu)
177 err(EXIT_FAILURE, "cpu_map__new");
178
179 memset(&act, 0, sizeof(act));
180 sigfillset(&act.sa_mask);
181 act.sa_sigaction = toggle_done;
182 sigaction(SIGINT, &act, NULL);
183
184 if (params.mlockall) {
185 if (mlockall(MCL_CURRENT | MCL_FUTURE))
186 err(EXIT_FAILURE, "mlockall");
187 }
188
189 if (!params.nthreads)
190 params.nthreads = perf_cpu_map__nr(cpu);
191
192 worker = calloc(params.nthreads, sizeof(*worker));
193 if (!worker)
194 err(EXIT_FAILURE, "calloc");
195
196 if (!params.fshared)
197 futex_flag = FUTEX_PRIVATE_FLAG;
198
199 if (params.nrequeue > params.nthreads)
200 params.nrequeue = params.nthreads;
201
202 if (params.broadcast)
203 params.nrequeue = params.nthreads;
204
205 printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %s%p), "
206 "%d at a time.\n\n", getpid(), params.nthreads,
207 params.fshared ? "shared":"private", &futex1,
208 params.pi ? "PI ": "", &futex2, params.nrequeue);
209
210 init_stats(&requeued_stats);
211 init_stats(&requeuetime_stats);
212 pthread_attr_init(&thread_attr);
213 mutex_init(&thread_lock);
214 cond_init(&thread_parent);
215 cond_init(&thread_worker);
216
217 for (j = 0; j < bench_repeat && !done; j++) {
218 unsigned int nrequeued = 0, wakeups = 0;
219 struct timeval start, end, runtime;
220
221 /* create, launch & block all threads */
222 block_threads(worker, thread_attr, cpu);
223
224 /* make sure all threads are already blocked */
225 mutex_lock(&thread_lock);
226 while (threads_starting)
227 cond_wait(&thread_parent, &thread_lock);
228 cond_broadcast(&thread_worker);
229 mutex_unlock(&thread_lock);
230
231 usleep(100000);
232
233 /* Ok, all threads are patiently blocked, start requeueing */
234 gettimeofday(&start, NULL);
235 while (nrequeued < params.nthreads) {
236 int r;
237
238 /*
239 * For the regular non-pi case, do not wakeup any tasks
240 * blocked on futex1, allowing us to really measure
241 * futex_wait functionality. For the PI case the first
242 * waiter is always awoken.
243 */
244 if (!params.pi) {
245 r = futex_cmp_requeue(&futex1, 0, &futex2, 0,
246 params.nrequeue,
247 futex_flag);
248 } else {
249 r = futex_cmp_requeue_pi(&futex1, 0, &futex2,
250 params.nrequeue,
251 futex_flag);
252 wakeups++; /* assume no error */
253 }
254
255 if (r < 0)
256 err(EXIT_FAILURE, "couldn't requeue from %p to %p",
257 &futex1, &futex2);
258
259 nrequeued += r;
260 }
261
262 gettimeofday(&end, NULL);
263 timersub(&end, &start, &runtime);
264
265 update_stats(&requeued_stats, nrequeued);
266 update_stats(&requeuetime_stats, runtime.tv_usec);
267
268 if (!params.silent) {
269 if (!params.pi)
270 printf("[Run %d]: Requeued %d of %d threads in "
271 "%.4f ms\n", j + 1, nrequeued,
272 params.nthreads,
273 runtime.tv_usec / (double)USEC_PER_MSEC);
274 else {
275 nrequeued -= wakeups;
276 printf("[Run %d]: Awoke and Requeued (%d+%d) of "
277 "%d threads in %.4f ms\n",
278 j + 1, wakeups, nrequeued,
279 params.nthreads,
280 runtime.tv_usec / (double)USEC_PER_MSEC);
281 }
282
283 }
284
285 if (!params.pi) {
286 /* everybody should be blocked on futex2, wake'em up */
287 nrequeued = futex_wake(&futex2, nrequeued, futex_flag);
288 if (params.nthreads != nrequeued)
289 warnx("couldn't wakeup all tasks (%d/%d)",
290 nrequeued, params.nthreads);
291 }
292
293 for (i = 0; i < params.nthreads; i++) {
294 ret = pthread_join(worker[i], NULL);
295 if (ret)
296 err(EXIT_FAILURE, "pthread_join");
297 }
298 }
299
300 /* cleanup & report results */
301 cond_destroy(&thread_parent);
302 cond_destroy(&thread_worker);
303 mutex_destroy(&thread_lock);
304 pthread_attr_destroy(&thread_attr);
305
306 print_summary();
307
308 free(worker);
309 perf_cpu_map__put(cpu);
310 return ret;
311 err:
312 usage_with_options(bench_futex_requeue_usage, options);
313 exit(EXIT_FAILURE);
314 }
315