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 #include "../util/mutex.h"
14 
15 #ifndef HAVE_PTHREAD_BARRIER
16 int bench_futex_wake_parallel(int argc __maybe_unused, const char **argv __maybe_unused)
17 {
18 	pr_err("%s: pthread_barrier_t unavailable, disabling this test...\n", __func__);
19 	return 0;
20 }
21 #else /* HAVE_PTHREAD_BARRIER */
22 /* For the CLR_() macros */
23 #include <string.h>
24 #include <pthread.h>
25 
26 #include <signal.h>
27 #include "../util/stat.h"
28 #include <subcmd/parse-options.h>
29 #include <linux/kernel.h>
30 #include <linux/time64.h>
31 #include <errno.h>
32 #include "futex.h"
33 #include <perf/cpumap.h>
34 
35 #include <err.h>
36 #include <stdlib.h>
37 #include <sys/time.h>
38 #include <sys/mman.h>
39 
40 struct thread_data {
41 	pthread_t worker;
42 	unsigned int nwoken;
43 	struct timeval runtime;
44 };
45 
46 static unsigned int nwakes = 1;
47 
48 /* all threads will block on the same futex -- hash bucket chaos ;) */
49 static u_int32_t futex = 0;
50 
51 static pthread_t *blocked_worker;
52 static bool done = false;
53 static struct mutex thread_lock;
54 static struct cond thread_parent, thread_worker;
55 static pthread_barrier_t barrier;
56 static struct stats waketime_stats, wakeup_stats;
57 static unsigned int threads_starting;
58 static int futex_flag = 0;
59 
60 static struct bench_futex_parameters params;
61 
62 static const struct option options[] = {
63 	OPT_UINTEGER('t', "threads", &params.nthreads, "Specify amount of threads"),
64 	OPT_UINTEGER('w', "nwakers", &params.nwakes, "Specify amount of waking threads"),
65 	OPT_BOOLEAN( 's', "silent",  &params.silent, "Silent mode: do not display data/details"),
66 	OPT_BOOLEAN( 'S', "shared",  &params.fshared, "Use shared futexes instead of private ones"),
67 	OPT_BOOLEAN( 'm', "mlockall", &params.mlockall, "Lock all current and future memory"),
68 
69 	OPT_END()
70 };
71 
72 static const char * const bench_futex_wake_parallel_usage[] = {
73 	"perf bench futex wake-parallel <options>",
74 	NULL
75 };
76 
77 static void *waking_workerfn(void *arg)
78 {
79 	struct thread_data *waker = (struct thread_data *) arg;
80 	struct timeval start, end;
81 
82 	pthread_barrier_wait(&barrier);
83 
84 	gettimeofday(&start, NULL);
85 
86 	waker->nwoken = futex_wake(&futex, nwakes, futex_flag);
87 	if (waker->nwoken != nwakes)
88 		warnx("couldn't wakeup all tasks (%d/%d)",
89 		      waker->nwoken, nwakes);
90 
91 	gettimeofday(&end, NULL);
92 	timersub(&end, &start, &waker->runtime);
93 
94 	pthread_exit(NULL);
95 	return NULL;
96 }
97 
98 static void wakeup_threads(struct thread_data *td, pthread_attr_t thread_attr)
99 {
100 	unsigned int i;
101 
102 	pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
103 
104 	pthread_barrier_init(&barrier, NULL, params.nwakes + 1);
105 
106 	/* create and block all threads */
107 	for (i = 0; i < params.nwakes; i++) {
108 		/*
109 		 * Thread creation order will impact per-thread latency
110 		 * as it will affect the order to acquire the hb spinlock.
111 		 * For now let the scheduler decide.
112 		 */
113 		if (pthread_create(&td[i].worker, &thread_attr,
114 				   waking_workerfn, (void *)&td[i]))
115 			err(EXIT_FAILURE, "pthread_create");
116 	}
117 
118 	pthread_barrier_wait(&barrier);
119 
120 	for (i = 0; i < params.nwakes; i++)
121 		if (pthread_join(td[i].worker, NULL))
122 			err(EXIT_FAILURE, "pthread_join");
123 
124 	pthread_barrier_destroy(&barrier);
125 }
126 
127 static void *blocked_workerfn(void *arg __maybe_unused)
128 {
129 	mutex_lock(&thread_lock);
130 	threads_starting--;
131 	if (!threads_starting)
132 		cond_signal(&thread_parent);
133 	cond_wait(&thread_worker, &thread_lock);
134 	mutex_unlock(&thread_lock);
135 
136 	while (1) { /* handle spurious wakeups */
137 		if (futex_wait(&futex, 0, NULL, futex_flag) != EINTR)
138 			break;
139 	}
140 
141 	pthread_exit(NULL);
142 	return NULL;
143 }
144 
145 static void block_threads(pthread_t *w, pthread_attr_t thread_attr,
146 			  struct perf_cpu_map *cpu)
147 {
148 	cpu_set_t *cpuset;
149 	unsigned int i;
150 	int nrcpus = perf_cpu_map__nr(cpu);
151 	size_t size;
152 
153 	threads_starting = params.nthreads;
154 
155 	cpuset = CPU_ALLOC(nrcpus);
156 	BUG_ON(!cpuset);
157 	size = CPU_ALLOC_SIZE(nrcpus);
158 
159 	/* create and block all threads */
160 	for (i = 0; i < params.nthreads; i++) {
161 		CPU_ZERO_S(size, cpuset);
162 		CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset);
163 
164 		if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) {
165 			CPU_FREE(cpuset);
166 			err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
167 		}
168 
169 		if (pthread_create(&w[i], &thread_attr, blocked_workerfn, NULL)) {
170 			CPU_FREE(cpuset);
171 			err(EXIT_FAILURE, "pthread_create");
172 		}
173 	}
174 	CPU_FREE(cpuset);
175 }
176 
177 static void print_run(struct thread_data *waking_worker, unsigned int run_num)
178 {
179 	unsigned int i, wakeup_avg;
180 	double waketime_avg, waketime_stddev;
181 	struct stats __waketime_stats, __wakeup_stats;
182 
183 	init_stats(&__wakeup_stats);
184 	init_stats(&__waketime_stats);
185 
186 	for (i = 0; i < params.nwakes; i++) {
187 		update_stats(&__waketime_stats, waking_worker[i].runtime.tv_usec);
188 		update_stats(&__wakeup_stats, waking_worker[i].nwoken);
189 	}
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("[Run %d]: Avg per-thread latency (waking %d/%d threads) "
196 	       "in %.4f ms (+-%.2f%%)\n", run_num + 1, wakeup_avg,
197 	       params.nthreads, waketime_avg / USEC_PER_MSEC,
198 	       rel_stddev_stats(waketime_stddev, waketime_avg));
199 }
200 
201 static void print_summary(void)
202 {
203 	unsigned int wakeup_avg;
204 	double waketime_avg, waketime_stddev;
205 
206 	waketime_avg = avg_stats(&waketime_stats);
207 	waketime_stddev = stddev_stats(&waketime_stats);
208 	wakeup_avg = avg_stats(&wakeup_stats);
209 
210 	printf("Avg per-thread latency (waking %d/%d threads) in %.4f ms (+-%.2f%%)\n",
211 	       wakeup_avg,
212 	       params.nthreads,
213 	       waketime_avg / USEC_PER_MSEC,
214 	       rel_stddev_stats(waketime_stddev, waketime_avg));
215 }
216 
217 
218 static void do_run_stats(struct thread_data *waking_worker)
219 {
220 	unsigned int i;
221 
222 	for (i = 0; i < params.nwakes; i++) {
223 		update_stats(&waketime_stats, waking_worker[i].runtime.tv_usec);
224 		update_stats(&wakeup_stats, waking_worker[i].nwoken);
225 	}
226 
227 }
228 
229 static void toggle_done(int sig __maybe_unused,
230 			siginfo_t *info __maybe_unused,
231 			void *uc __maybe_unused)
232 {
233 	done = true;
234 }
235 
236 int bench_futex_wake_parallel(int argc, const char **argv)
237 {
238 	int ret = 0;
239 	unsigned int i, j;
240 	struct sigaction act;
241 	pthread_attr_t thread_attr;
242 	struct thread_data *waking_worker;
243 	struct perf_cpu_map *cpu;
244 
245 	argc = parse_options(argc, argv, options,
246 			     bench_futex_wake_parallel_usage, 0);
247 	if (argc) {
248 		usage_with_options(bench_futex_wake_parallel_usage, options);
249 		exit(EXIT_FAILURE);
250 	}
251 
252 	memset(&act, 0, sizeof(act));
253 	sigfillset(&act.sa_mask);
254 	act.sa_sigaction = toggle_done;
255 	sigaction(SIGINT, &act, NULL);
256 
257 	if (params.mlockall) {
258 		if (mlockall(MCL_CURRENT | MCL_FUTURE))
259 			err(EXIT_FAILURE, "mlockall");
260 	}
261 
262 	cpu = perf_cpu_map__new(NULL);
263 	if (!cpu)
264 		err(EXIT_FAILURE, "calloc");
265 
266 	if (!params.nthreads)
267 		params.nthreads = perf_cpu_map__nr(cpu);
268 
269 	/* some sanity checks */
270 	if (params.nwakes > params.nthreads ||
271 	    !params.nwakes)
272 		params.nwakes = params.nthreads;
273 
274 	if (params.nthreads % params.nwakes)
275 		errx(EXIT_FAILURE, "Must be perfectly divisible");
276 	/*
277 	 * Each thread will wakeup nwakes tasks in
278 	 * a single futex_wait call.
279 	 */
280 	nwakes = params.nthreads/params.nwakes;
281 
282 	blocked_worker = calloc(params.nthreads, sizeof(*blocked_worker));
283 	if (!blocked_worker)
284 		err(EXIT_FAILURE, "calloc");
285 
286 	if (!params.fshared)
287 		futex_flag = FUTEX_PRIVATE_FLAG;
288 
289 	printf("Run summary [PID %d]: blocking on %d threads (at [%s] "
290 	       "futex %p), %d threads waking up %d at a time.\n\n",
291 	       getpid(), params.nthreads, params.fshared ? "shared":"private",
292 	       &futex, params.nwakes, nwakes);
293 
294 	init_stats(&wakeup_stats);
295 	init_stats(&waketime_stats);
296 
297 	pthread_attr_init(&thread_attr);
298 	mutex_init(&thread_lock);
299 	cond_init(&thread_parent);
300 	cond_init(&thread_worker);
301 
302 	for (j = 0; j < bench_repeat && !done; j++) {
303 		waking_worker = calloc(params.nwakes, sizeof(*waking_worker));
304 		if (!waking_worker)
305 			err(EXIT_FAILURE, "calloc");
306 
307 		/* create, launch & block all threads */
308 		block_threads(blocked_worker, thread_attr, cpu);
309 
310 		/* make sure all threads are already blocked */
311 		mutex_lock(&thread_lock);
312 		while (threads_starting)
313 			cond_wait(&thread_parent, &thread_lock);
314 		cond_broadcast(&thread_worker);
315 		mutex_unlock(&thread_lock);
316 
317 		usleep(100000);
318 
319 		/* Ok, all threads are patiently blocked, start waking folks up */
320 		wakeup_threads(waking_worker, thread_attr);
321 
322 		for (i = 0; i < params.nthreads; i++) {
323 			ret = pthread_join(blocked_worker[i], NULL);
324 			if (ret)
325 				err(EXIT_FAILURE, "pthread_join");
326 		}
327 
328 		do_run_stats(waking_worker);
329 		if (!params.silent)
330 			print_run(waking_worker, j);
331 
332 		free(waking_worker);
333 	}
334 
335 	/* cleanup & report results */
336 	cond_destroy(&thread_parent);
337 	cond_destroy(&thread_worker);
338 	mutex_destroy(&thread_lock);
339 	pthread_attr_destroy(&thread_attr);
340 
341 	print_summary();
342 
343 	free(blocked_worker);
344 	perf_cpu_map__put(cpu);
345 	return ret;
346 }
347 #endif /* HAVE_PTHREAD_BARRIER */
348