futex-wake-parallel.c

  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
 16int 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
 40struct thread_data {
 41	pthread_t worker;
 42	unsigned int nwoken;
 43	struct timeval runtime;
 44};
 45
 46static unsigned int nwakes = 1;
 47
 48/* all threads will block on the same futex -- hash bucket chaos ;) */
 49static u_int32_t futex = 0;
 50
 51static pthread_t *blocked_worker;
 52static bool done = false;
 53static struct mutex thread_lock;
 54static struct cond 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 struct bench_futex_parameters params = {
 61	.nbuckets = -1,
 62};
 63
 64static const struct option options[] = {
 65	OPT_INTEGER( 'b', "buckets", &params.nbuckets, "Specify amount of hash buckets"),
 66	OPT_UINTEGER('t', "threads", &params.nthreads, "Specify amount of threads"),
 67	OPT_UINTEGER('w', "nwakers", &params.nwakes, "Specify amount of waking threads"),
 68	OPT_BOOLEAN( 's', "silent",  &params.silent, "Silent mode: do not display data/details"),
 69	OPT_BOOLEAN( 'S', "shared",  &params.fshared, "Use shared futexes instead of private ones"),
 70	OPT_BOOLEAN( 'm', "mlockall", &params.mlockall, "Lock all current and future memory"),
 71
 72	OPT_END()
 73};
 74
 75static const char * const bench_futex_wake_parallel_usage[] = {
 76	"perf bench futex wake-parallel <options>",
 77	NULL
 78};
 79
 80static void *waking_workerfn(void *arg)
 81{
 82	struct thread_data *waker = (struct thread_data *) arg;
 83	struct timeval start, end;
 84
 85	pthread_barrier_wait(&barrier);
 86
 87	gettimeofday(&start, NULL);
 88
 89	waker->nwoken = futex_wake(&futex, nwakes, futex_flag);
 90	if (waker->nwoken != nwakes)
 91		warnx("couldn't wakeup all tasks (%d/%d)",
 92		      waker->nwoken, nwakes);
 93
 94	gettimeofday(&end, NULL);
 95	timersub(&end, &start, &waker->runtime);
 96
 97	pthread_exit(NULL);
 98	return NULL;
 99}
100
101static void wakeup_threads(struct thread_data *td)
102{
103	unsigned int i;
104	pthread_attr_t thread_attr;
105
106	pthread_attr_init(&thread_attr);
107	pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
108
109	pthread_barrier_init(&barrier, NULL, params.nwakes + 1);
110
111	/* create and block all threads */
112	for (i = 0; i < params.nwakes; i++) {
113		/*
114		 * Thread creation order will impact per-thread latency
115		 * as it will affect the order to acquire the hb spinlock.
116		 * For now let the scheduler decide.
117		 */
118		if (pthread_create(&td[i].worker, &thread_attr,
119				   waking_workerfn, (void *)&td[i]))
120			err(EXIT_FAILURE, "pthread_create");
121	}
122
123	pthread_barrier_wait(&barrier);
124
125	for (i = 0; i < params.nwakes; i++)
126		if (pthread_join(td[i].worker, NULL))
127			err(EXIT_FAILURE, "pthread_join");
128
129	pthread_barrier_destroy(&barrier);
130	pthread_attr_destroy(&thread_attr);
131}
132
133static void *blocked_workerfn(void *arg __maybe_unused)
134{
135	mutex_lock(&thread_lock);
136	threads_starting--;
137	if (!threads_starting)
138		cond_signal(&thread_parent);
139	cond_wait(&thread_worker, &thread_lock);
140	mutex_unlock(&thread_lock);
141
142	while (1) { /* handle spurious wakeups */
143		if (futex_wait(&futex, 0, NULL, futex_flag) != EINTR)
144			break;
145	}
146
147	pthread_exit(NULL);
148	return NULL;
149}
150
151static void block_threads(pthread_t *w, struct perf_cpu_map *cpu)
152{
153	cpu_set_t *cpuset;
154	unsigned int i;
155	int nrcpus = cpu__max_cpu().cpu;
156	size_t size;
157
158	threads_starting = params.nthreads;
159
160	cpuset = CPU_ALLOC(nrcpus);
161	BUG_ON(!cpuset);
162	size = CPU_ALLOC_SIZE(nrcpus);
163
164	/* create and block all threads */
165	for (i = 0; i < params.nthreads; i++) {
166		pthread_attr_t thread_attr;
167
168		pthread_attr_init(&thread_attr);
169		CPU_ZERO_S(size, cpuset);
170		CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset);
171
172		if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) {
173			CPU_FREE(cpuset);
174			err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
175		}
176
177		if (pthread_create(&w[i], &thread_attr, blocked_workerfn, NULL)) {
178			CPU_FREE(cpuset);
179			err(EXIT_FAILURE, "pthread_create");
180		}
181		pthread_attr_destroy(&thread_attr);
182	}
183	CPU_FREE(cpuset);
184}
185
186static void print_run(struct thread_data *waking_worker, unsigned int run_num)
187{
188	unsigned int i, wakeup_avg;
189	double waketime_avg, waketime_stddev;
190	struct stats __waketime_stats, __wakeup_stats;
191
192	init_stats(&__wakeup_stats);
193	init_stats(&__waketime_stats);
194
195	for (i = 0; i < params.nwakes; i++) {
196		update_stats(&__waketime_stats, waking_worker[i].runtime.tv_usec);
197		update_stats(&__wakeup_stats, waking_worker[i].nwoken);
198	}
199
200	waketime_avg = avg_stats(&__waketime_stats);
201	waketime_stddev = stddev_stats(&__waketime_stats);
202	wakeup_avg = avg_stats(&__wakeup_stats);
203
204	printf("[Run %d]: Avg per-thread latency (waking %d/%d threads) "
205	       "in %.4f ms (+-%.2f%%)\n", run_num + 1, wakeup_avg,
206	       params.nthreads, waketime_avg / USEC_PER_MSEC,
207	       rel_stddev_stats(waketime_stddev, waketime_avg));
208}
209
210static void print_summary(void)
211{
212	unsigned int wakeup_avg;
213	double waketime_avg, waketime_stddev;
214
215	waketime_avg = avg_stats(&waketime_stats);
216	waketime_stddev = stddev_stats(&waketime_stats);
217	wakeup_avg = avg_stats(&wakeup_stats);
218
219	printf("Avg per-thread latency (waking %d/%d threads) in %.4f ms (+-%.2f%%)\n",
220	       wakeup_avg,
221	       params.nthreads,
222	       waketime_avg / USEC_PER_MSEC,
223	       rel_stddev_stats(waketime_stddev, waketime_avg));
224	futex_print_nbuckets(&params);
225}
226
227
228static void do_run_stats(struct thread_data *waking_worker)
229{
230	unsigned int i;
231
232	for (i = 0; i < params.nwakes; i++) {
233		update_stats(&waketime_stats, waking_worker[i].runtime.tv_usec);
234		update_stats(&wakeup_stats, waking_worker[i].nwoken);
235	}
236
237}
238
239static void toggle_done(int sig __maybe_unused,
240			siginfo_t *info __maybe_unused,
241			void *uc __maybe_unused)
242{
243	done = true;
244}
245
246int bench_futex_wake_parallel(int argc, const char **argv)
247{
248	int ret = 0;
249	unsigned int i, j;
250	struct sigaction act;
251	struct thread_data *waking_worker;
252	struct perf_cpu_map *cpu;
253
254	argc = parse_options(argc, argv, options,
255			     bench_futex_wake_parallel_usage, 0);
256	if (argc) {
257		usage_with_options(bench_futex_wake_parallel_usage, options);
258		exit(EXIT_FAILURE);
259	}
260
261	memset(&act, 0, sizeof(act));
262	sigfillset(&act.sa_mask);
263	act.sa_sigaction = toggle_done;
264	sigaction(SIGINT, &act, NULL);
265
266	if (params.mlockall) {
267		if (mlockall(MCL_CURRENT | MCL_FUTURE))
268			err(EXIT_FAILURE, "mlockall");
269	}
270
271	cpu = perf_cpu_map__new_online_cpus();
272	if (!cpu)
273		err(EXIT_FAILURE, "calloc");
274
275	if (!params.nthreads)
276		params.nthreads = perf_cpu_map__nr(cpu);
277
278	/* some sanity checks */
279	if (params.nwakes > params.nthreads ||
280	    !params.nwakes)
281		params.nwakes = params.nthreads;
282
283	if (params.nthreads % params.nwakes)
284		errx(EXIT_FAILURE, "Must be perfectly divisible");
285	/*
286	 * Each thread will wakeup nwakes tasks in
287	 * a single futex_wait call.
288	 */
289	nwakes = params.nthreads/params.nwakes;
290
291	blocked_worker = calloc(params.nthreads, sizeof(*blocked_worker));
292	if (!blocked_worker)
293		err(EXIT_FAILURE, "calloc");
294
295	if (!params.fshared)
296		futex_flag = FUTEX_PRIVATE_FLAG;
297
298	futex_set_nbuckets_param(&params);
299
300	printf("Run summary [PID %d]: blocking on %d threads (at [%s] "
301	       "futex %p), %d threads waking up %d at a time.\n\n",
302	       getpid(), params.nthreads, params.fshared ? "shared":"private",
303	       &futex, params.nwakes, nwakes);
304
305	init_stats(&wakeup_stats);
306	init_stats(&waketime_stats);
307
308	mutex_init(&thread_lock);
309	cond_init(&thread_parent);
310	cond_init(&thread_worker);
311
312	for (j = 0; j < bench_repeat && !done; j++) {
313		waking_worker = calloc(params.nwakes, sizeof(*waking_worker));
314		if (!waking_worker)
315			err(EXIT_FAILURE, "calloc");
316
317		/* create, launch & block all threads */
318		block_threads(blocked_worker, cpu);
319
320		/* make sure all threads are already blocked */
321		mutex_lock(&thread_lock);
322		while (threads_starting)
323			cond_wait(&thread_parent, &thread_lock);
324		cond_broadcast(&thread_worker);
325		mutex_unlock(&thread_lock);
326
327		usleep(200000);
328
329		/* Ok, all threads are patiently blocked, start waking folks up */
330		wakeup_threads(waking_worker);
331
332		for (i = 0; i < params.nthreads; i++) {
333			ret = pthread_join(blocked_worker[i], NULL);
334			if (ret)
335				err(EXIT_FAILURE, "pthread_join");
336		}
337
338		do_run_stats(waking_worker);
339		if (!params.silent)
340			print_run(waking_worker, j);
341
342		free(waking_worker);
343	}
344
345	/* cleanup & report results */
346	cond_destroy(&thread_parent);
347	cond_destroy(&thread_worker);
348	mutex_destroy(&thread_lock);
349
350	print_summary();
351
352	free(blocked_worker);
353	perf_cpu_map__put(cpu);
354	return ret;
355}
356#endif /* HAVE_PTHREAD_BARRIER */