1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * Copyright (C) 2013  Davidlohr Bueso <davidlohr@hp.com>
  4 *
  5 * futex-hash: Stress the hell out of the Linux kernel futex uaddr hashing.
  6 *
  7 * This program is particularly useful for measuring the kernel's futex hash
  8 * table/function implementation. In order for it to make sense, use with as
  9 * many threads and futexes as possible.
 10 */
 11
 12/* For the CLR_() macros */
 13#include <string.h>
 14#include <pthread.h>
 15
 16#include <errno.h>
 17#include <signal.h>
 18#include <stdlib.h>
 19#include <linux/compiler.h>
 20#include <linux/kernel.h>
 21#include <linux/zalloc.h>
 22#include <sys/time.h>
 23#include <internal/cpumap.h>
 24#include <perf/cpumap.h>
 25
 26#include "../util/stat.h"
 27#include <subcmd/parse-options.h>
 28#include "bench.h"
 29#include "futex.h"
 30
 31#include <err.h>
 32
 33static unsigned int nthreads = 0;
 34static unsigned int nsecs    = 10;
 35/* amount of futexes per thread */
 36static unsigned int nfutexes = 1024;
 37static bool fshared = false, done = false, silent = false;
 38static int futex_flag = 0;
 39
 40struct timeval bench__start, bench__end, bench__runtime;
 41static pthread_mutex_t thread_lock;
 42static unsigned int threads_starting;
 43static struct stats throughput_stats;
 44static pthread_cond_t thread_parent, thread_worker;
 45
 46struct worker {
 47	int tid;
 48	u_int32_t *futex;
 49	pthread_t thread;
 50	unsigned long ops;
 51};
 52
 53static const struct option options[] = {
 54	OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
 55	OPT_UINTEGER('r', "runtime", &nsecs,    "Specify runtime (in seconds)"),
 56	OPT_UINTEGER('f', "futexes", &nfutexes, "Specify amount of futexes per threads"),
 57	OPT_BOOLEAN( 's', "silent",  &silent,   "Silent mode: do not display data/details"),
 58	OPT_BOOLEAN( 'S', "shared",  &fshared,  "Use shared futexes instead of private ones"),
 59	OPT_END()
 60};
 61
 62static const char * const bench_futex_hash_usage[] = {
 63	"perf bench futex hash <options>",
 64	NULL
 65};
 66
 67static void *workerfn(void *arg)
 68{
 69	int ret;
 70	struct worker *w = (struct worker *) arg;
 71	unsigned int i;
 72	unsigned long ops = w->ops; /* avoid cacheline bouncing */
 73
 74	pthread_mutex_lock(&thread_lock);
 75	threads_starting--;
 76	if (!threads_starting)
 77		pthread_cond_signal(&thread_parent);
 78	pthread_cond_wait(&thread_worker, &thread_lock);
 79	pthread_mutex_unlock(&thread_lock);
 80
 81	do {
 82		for (i = 0; i < nfutexes; i++, ops++) {
 83			/*
 84			 * We want the futex calls to fail in order to stress
 85			 * the hashing of uaddr and not measure other steps,
 86			 * such as internal waitqueue handling, thus enlarging
 87			 * the critical region protected by hb->lock.
 88			 */
 89			ret = futex_wait(&w->futex[i], 1234, NULL, futex_flag);
 90			if (!silent &&
 91			    (!ret || errno != EAGAIN || errno != EWOULDBLOCK))
 92				warn("Non-expected futex return call");
 93		}
 94	}  while (!done);
 95
 96	w->ops = ops;
 97	return NULL;
 98}
 99
100static void toggle_done(int sig __maybe_unused,
101			siginfo_t *info __maybe_unused,
102			void *uc __maybe_unused)
103{
104	/* inform all threads that we're done for the day */
105	done = true;
106	gettimeofday(&bench__end, NULL);
107	timersub(&bench__end, &bench__start, &bench__runtime);
108}
109
110static void print_summary(void)
111{
112	unsigned long avg = avg_stats(&throughput_stats);
113	double stddev = stddev_stats(&throughput_stats);
114
115	printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
116	       !silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg),
117	       (int)bench__runtime.tv_sec);
118}
119
120int bench_futex_hash(int argc, const char **argv)
121{
122	int ret = 0;
123	cpu_set_t cpuset;
124	struct sigaction act;
125	unsigned int i;
126	pthread_attr_t thread_attr;
127	struct worker *worker = NULL;
128	struct perf_cpu_map *cpu;
129
130	argc = parse_options(argc, argv, options, bench_futex_hash_usage, 0);
131	if (argc) {
132		usage_with_options(bench_futex_hash_usage, options);
133		exit(EXIT_FAILURE);
134	}
135
136	cpu = perf_cpu_map__new(NULL);
137	if (!cpu)
138		goto errmem;
139
140	memset(&act, 0, sizeof(act));
141	sigfillset(&act.sa_mask);
142	act.sa_sigaction = toggle_done;
143	sigaction(SIGINT, &act, NULL);
144
145	if (!nthreads) /* default to the number of CPUs */
146		nthreads = cpu->nr;
147
148	worker = calloc(nthreads, sizeof(*worker));
149	if (!worker)
150		goto errmem;
151
152	if (!fshared)
153		futex_flag = FUTEX_PRIVATE_FLAG;
154
155	printf("Run summary [PID %d]: %d threads, each operating on %d [%s] futexes for %d secs.\n\n",
156	       getpid(), nthreads, nfutexes, fshared ? "shared":"private", nsecs);
157
158	init_stats(&throughput_stats);
159	pthread_mutex_init(&thread_lock, NULL);
160	pthread_cond_init(&thread_parent, NULL);
161	pthread_cond_init(&thread_worker, NULL);
162
163	threads_starting = nthreads;
164	pthread_attr_init(&thread_attr);
165	gettimeofday(&bench__start, NULL);
166	for (i = 0; i < nthreads; i++) {
167		worker[i].tid = i;
168		worker[i].futex = calloc(nfutexes, sizeof(*worker[i].futex));
169		if (!worker[i].futex)
170			goto errmem;
171
172		CPU_ZERO(&cpuset);
173		CPU_SET(cpu->map[i % cpu->nr], &cpuset);
174
175		ret = pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset);
176		if (ret)
177			err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
178
179		ret = pthread_create(&worker[i].thread, &thread_attr, workerfn,
180				     (void *)(struct worker *) &worker[i]);
181		if (ret)
182			err(EXIT_FAILURE, "pthread_create");
183
184	}
185	pthread_attr_destroy(&thread_attr);
186
187	pthread_mutex_lock(&thread_lock);
188	while (threads_starting)
189		pthread_cond_wait(&thread_parent, &thread_lock);
190	pthread_cond_broadcast(&thread_worker);
191	pthread_mutex_unlock(&thread_lock);
192
193	sleep(nsecs);
194	toggle_done(0, NULL, NULL);
195
196	for (i = 0; i < nthreads; i++) {
197		ret = pthread_join(worker[i].thread, NULL);
198		if (ret)
199			err(EXIT_FAILURE, "pthread_join");
200	}
201
202	/* cleanup & report results */
203	pthread_cond_destroy(&thread_parent);
204	pthread_cond_destroy(&thread_worker);
205	pthread_mutex_destroy(&thread_lock);
206
207	for (i = 0; i < nthreads; i++) {
208		unsigned long t = bench__runtime.tv_sec > 0 ?
209			worker[i].ops / bench__runtime.tv_sec : 0;
210		update_stats(&throughput_stats, t);
211		if (!silent) {
212			if (nfutexes == 1)
213				printf("[thread %2d] futex: %p [ %ld ops/sec ]\n",
214				       worker[i].tid, &worker[i].futex[0], t);
215			else
216				printf("[thread %2d] futexes: %p ... %p [ %ld ops/sec ]\n",
217				       worker[i].tid, &worker[i].futex[0],
218				       &worker[i].futex[nfutexes-1], t);
219		}
220
221		zfree(&worker[i].futex);
222	}
223
224	print_summary();
225
226	free(worker);
227	free(cpu);
228	return ret;
229errmem:
230	err(EXIT_FAILURE, "calloc");
231}