1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * Copyright (C) 2015 Davidlohr Bueso.
  4 */
  5
  6/* For the CLR_() macros */
  7#include <string.h>
  8#include <pthread.h>
  9
 10#include <signal.h>
 11#include "../util/mutex.h"
 12#include "../util/stat.h"
 13#include <subcmd/parse-options.h>
 14#include <linux/compiler.h>
 15#include <linux/kernel.h>
 16#include <linux/zalloc.h>
 17#include <errno.h>
 18#include <perf/cpumap.h>
 19#include "bench.h"
 20#include "futex.h"
 21
 22#include <err.h>
 23#include <stdlib.h>
 24#include <sys/time.h>
 25#include <sys/mman.h>
 26
 27struct worker {
 28	int tid;
 29	u_int32_t *futex;
 30	pthread_t thread;
 31	unsigned long ops;
 32};
 33
 34static u_int32_t global_futex = 0;
 35static struct worker *worker;
 36static bool done = false;
 37static int futex_flag = 0;
 38static struct mutex thread_lock;
 39static unsigned int threads_starting;
 40static struct stats throughput_stats;
 41static struct cond thread_parent, thread_worker;
 42
 43static struct bench_futex_parameters params = {
 44	.runtime  = 10,
 45};
 46
 47static const struct option options[] = {
 48	OPT_UINTEGER('t', "threads", &params.nthreads, "Specify amount of threads"),
 49	OPT_UINTEGER('r', "runtime", &params.runtime, "Specify runtime (in seconds)"),
 50	OPT_BOOLEAN( 'M', "multi",   &params.multi, "Use multiple futexes"),
 51	OPT_BOOLEAN( 's', "silent",  &params.silent, "Silent mode: do not display data/details"),
 52	OPT_BOOLEAN( 'S', "shared",  &params.fshared, "Use shared futexes instead of private ones"),
 53	OPT_BOOLEAN( 'm', "mlockall", &params.mlockall, "Lock all current and future memory"),
 54	OPT_END()
 55};
 56
 57static const char * const bench_futex_lock_pi_usage[] = {
 58	"perf bench futex lock-pi <options>",
 59	NULL
 60};
 61
 62static void print_summary(void)
 63{
 64	unsigned long avg = avg_stats(&throughput_stats);
 65	double stddev = stddev_stats(&throughput_stats);
 66
 67	printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
 68	       !params.silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg),
 69	       (int)bench__runtime.tv_sec);
 70}
 71
 72static void toggle_done(int sig __maybe_unused,
 73			siginfo_t *info __maybe_unused,
 74			void *uc __maybe_unused)
 75{
 76	/* inform all threads that we're done for the day */
 77	done = true;
 78	gettimeofday(&bench__end, NULL);
 79	timersub(&bench__end, &bench__start, &bench__runtime);
 80}
 81
 82static void *workerfn(void *arg)
 83{
 84	struct worker *w = (struct worker *) arg;
 85	unsigned long ops = w->ops;
 86
 87	mutex_lock(&thread_lock);
 88	threads_starting--;
 89	if (!threads_starting)
 90		cond_signal(&thread_parent);
 91	cond_wait(&thread_worker, &thread_lock);
 92	mutex_unlock(&thread_lock);
 93
 94	do {
 95		int ret;
 96	again:
 97		ret = futex_lock_pi(w->futex, NULL, futex_flag);
 98
 99		if (ret) { /* handle lock acquisition */
100			if (!params.silent)
101				warn("thread %d: Could not lock pi-lock for %p (%d)",
102				     w->tid, w->futex, ret);
103			if (done)
104				break;
105
106			goto again;
107		}
108
109		usleep(1);
110		ret = futex_unlock_pi(w->futex, futex_flag);
111		if (ret && !params.silent)
112			warn("thread %d: Could not unlock pi-lock for %p (%d)",
113			     w->tid, w->futex, ret);
114		ops++; /* account for thread's share of work */
115	}  while (!done);
116
117	w->ops = ops;
118	return NULL;
119}
120
121static void create_threads(struct worker *w, struct perf_cpu_map *cpu)
122{
123	cpu_set_t *cpuset;
124	unsigned int i;
125	int nrcpus =  perf_cpu_map__nr(cpu);
126	size_t size;
127
128	threads_starting = params.nthreads;
129
130	cpuset = CPU_ALLOC(nrcpus);
131	BUG_ON(!cpuset);
132	size = CPU_ALLOC_SIZE(nrcpus);
133
134	for (i = 0; i < params.nthreads; i++) {
135		pthread_attr_t thread_attr;
136
137		pthread_attr_init(&thread_attr);
138		worker[i].tid = i;
139
140		if (params.multi) {
141			worker[i].futex = calloc(1, sizeof(u_int32_t));
142			if (!worker[i].futex)
143				err(EXIT_FAILURE, "calloc");
144		} else
145			worker[i].futex = &global_futex;
146
147		CPU_ZERO_S(size, cpuset);
148		CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset);
149
150		if (pthread_attr_setaffinity_np(&thread_attr, size, cpuset)) {
151			CPU_FREE(cpuset);
152			err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
153		}
154
155		if (pthread_create(&w[i].thread, &thread_attr, workerfn, &worker[i])) {
156			CPU_FREE(cpuset);
157			err(EXIT_FAILURE, "pthread_create");
158		}
159		pthread_attr_destroy(&thread_attr);
160	}
161	CPU_FREE(cpuset);
162}
163
164int bench_futex_lock_pi(int argc, const char **argv)
165{
166	int ret = 0;
167	unsigned int i;
168	struct sigaction act;
169	struct perf_cpu_map *cpu;
170
171	argc = parse_options(argc, argv, options, bench_futex_lock_pi_usage, 0);
172	if (argc)
173		goto err;
174
175	cpu = perf_cpu_map__new_online_cpus();
176	if (!cpu)
177		err(EXIT_FAILURE, "calloc");
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	printf("Run summary [PID %d]: %d threads doing pi lock/unlock pairing for %d secs.\n\n",
200	       getpid(), params.nthreads, params.runtime);
201
202	init_stats(&throughput_stats);
203	mutex_init(&thread_lock);
204	cond_init(&thread_parent);
205	cond_init(&thread_worker);
206
207	threads_starting = params.nthreads;
208	gettimeofday(&bench__start, NULL);
209
210	create_threads(worker, cpu);
211
212	mutex_lock(&thread_lock);
213	while (threads_starting)
214		cond_wait(&thread_parent, &thread_lock);
215	cond_broadcast(&thread_worker);
216	mutex_unlock(&thread_lock);
217
218	sleep(params.runtime);
219	toggle_done(0, NULL, NULL);
220
221	for (i = 0; i < params.nthreads; i++) {
222		ret = pthread_join(worker[i].thread, NULL);
223		if (ret)
224			err(EXIT_FAILURE, "pthread_join");
225	}
226
227	/* cleanup & report results */
228	cond_destroy(&thread_parent);
229	cond_destroy(&thread_worker);
230	mutex_destroy(&thread_lock);
231
232	for (i = 0; i < params.nthreads; i++) {
233		unsigned long t = bench__runtime.tv_sec > 0 ?
234			worker[i].ops / bench__runtime.tv_sec : 0;
235
236		update_stats(&throughput_stats, t);
237		if (!params.silent)
238			printf("[thread %3d] futex: %p [ %ld ops/sec ]\n",
239			       worker[i].tid, worker[i].futex, t);
240
241		if (params.multi)
242			zfree(&worker[i].futex);
243	}
244
245	print_summary();
246
247	free(worker);
248	perf_cpu_map__put(cpu);
249	return ret;
250err:
251	usage_with_options(bench_futex_lock_pi_usage, options);
252	exit(EXIT_FAILURE);
253}