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}