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

  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 <sys/mman.h>
 24#include <perf/cpumap.h>
 25
 26#include "../util/mutex.h"
 27#include "../util/stat.h"
 28#include <subcmd/parse-options.h>
 
 29#include "bench.h"
 30#include "futex.h"
 31
 32#include <err.h>
 
 
 
 33
 34static bool done = false;
 35static int futex_flag = 0;
 
 
 
 36
 37struct timeval bench__start, bench__end, bench__runtime;
 38static struct mutex thread_lock;
 39static unsigned int threads_starting;
 40static struct stats throughput_stats;
 41static struct cond thread_parent, thread_worker;
 42
 43struct worker {
 44	int tid;
 45	u_int32_t *futex;
 46	pthread_t thread;
 47	unsigned long ops;
 48};
 49
 50static struct bench_futex_parameters params = {
 51	.nfutexes = 1024,
 52	.runtime  = 10,
 53};
 54
 55static const struct option options[] = {
 56	OPT_UINTEGER('t', "threads", &params.nthreads, "Specify amount of threads"),
 57	OPT_UINTEGER('r', "runtime", &params.runtime, "Specify runtime (in seconds)"),
 58	OPT_UINTEGER('f', "futexes", &params.nfutexes, "Specify amount of futexes per threads"),
 59	OPT_BOOLEAN( 's', "silent",  &params.silent, "Silent mode: do not display data/details"),
 60	OPT_BOOLEAN( 'S', "shared",  &params.fshared, "Use shared futexes instead of private ones"),
 61	OPT_BOOLEAN( 'm', "mlockall", &params.mlockall, "Lock all current and future memory"),
 62	OPT_END()
 63};
 64
 65static const char * const bench_futex_hash_usage[] = {
 66	"perf bench futex hash <options>",
 67	NULL
 68};
 69
 70static void *workerfn(void *arg)
 71{
 72	int ret;
 
 73	struct worker *w = (struct worker *) arg;
 74	unsigned int i;
 75	unsigned long ops = w->ops; /* avoid cacheline bouncing */
 76
 77	mutex_lock(&thread_lock);
 78	threads_starting--;
 79	if (!threads_starting)
 80		cond_signal(&thread_parent);
 81	cond_wait(&thread_worker, &thread_lock);
 82	mutex_unlock(&thread_lock);
 83
 84	do {
 85		for (i = 0; i < params.nfutexes; i++, ops++) {
 86			/*
 87			 * We want the futex calls to fail in order to stress
 88			 * the hashing of uaddr and not measure other steps,
 89			 * such as internal waitqueue handling, thus enlarging
 90			 * the critical region protected by hb->lock.
 91			 */
 92			ret = futex_wait(&w->futex[i], 1234, NULL, futex_flag);
 93			if (!params.silent &&
 
 94			    (!ret || errno != EAGAIN || errno != EWOULDBLOCK))
 95				warn("Non-expected futex return call");
 96		}
 97	}  while (!done);
 98
 99	w->ops = ops;
100	return NULL;
101}
102
103static void toggle_done(int sig __maybe_unused,
104			siginfo_t *info __maybe_unused,
105			void *uc __maybe_unused)
106{
107	/* inform all threads that we're done for the day */
108	done = true;
109	gettimeofday(&bench__end, NULL);
110	timersub(&bench__end, &bench__start, &bench__runtime);
111}
112
113static void print_summary(void)
114{
115	unsigned long avg = avg_stats(&throughput_stats);
116	double stddev = stddev_stats(&throughput_stats);
117
118	printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
119	       !params.silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg),
120	       (int)bench__runtime.tv_sec);
121}
122
123int bench_futex_hash(int argc, const char **argv)
 
124{
125	int ret = 0;
126	cpu_set_t *cpuset;
127	struct sigaction act;
128	unsigned int i;
129	pthread_attr_t thread_attr;
130	struct worker *worker = NULL;
131	struct perf_cpu_map *cpu;
132	int nrcpus;
133	size_t size;
134
135	argc = parse_options(argc, argv, options, bench_futex_hash_usage, 0);
136	if (argc) {
137		usage_with_options(bench_futex_hash_usage, options);
138		exit(EXIT_FAILURE);
139	}
140
141	cpu = perf_cpu_map__new(NULL);
142	if (!cpu)
143		goto errmem;
144
145	memset(&act, 0, sizeof(act));
146	sigfillset(&act.sa_mask);
147	act.sa_sigaction = toggle_done;
148	sigaction(SIGINT, &act, NULL);
149
150	if (params.mlockall) {
151		if (mlockall(MCL_CURRENT | MCL_FUTURE))
152			err(EXIT_FAILURE, "mlockall");
153	}
154
155	if (!params.nthreads) /* default to the number of CPUs */
156		params.nthreads = perf_cpu_map__nr(cpu);
157
158	worker = calloc(params.nthreads, sizeof(*worker));
159	if (!worker)
160		goto errmem;
161
162	if (!params.fshared)
163		futex_flag = FUTEX_PRIVATE_FLAG;
164
165	printf("Run summary [PID %d]: %d threads, each operating on %d [%s] futexes for %d secs.\n\n",
166	       getpid(), params.nthreads, params.nfutexes, params.fshared ? "shared":"private", params.runtime);
167
168	init_stats(&throughput_stats);
169	mutex_init(&thread_lock);
170	cond_init(&thread_parent);
171	cond_init(&thread_worker);
172
173	threads_starting = params.nthreads;
174	pthread_attr_init(&thread_attr);
175	gettimeofday(&bench__start, NULL);
176
177	nrcpus = perf_cpu_map__nr(cpu);
178	cpuset = CPU_ALLOC(nrcpus);
179	BUG_ON(!cpuset);
180	size = CPU_ALLOC_SIZE(nrcpus);
181
182	for (i = 0; i < params.nthreads; i++) {
183		worker[i].tid = i;
184		worker[i].futex = calloc(params.nfutexes, sizeof(*worker[i].futex));
185		if (!worker[i].futex)
186			goto errmem;
187
188		CPU_ZERO_S(size, cpuset);
 
189
190		CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset);
191		ret = pthread_attr_setaffinity_np(&thread_attr, size, cpuset);
192		if (ret) {
193			CPU_FREE(cpuset);
194			err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
195		}
196		ret = pthread_create(&worker[i].thread, &thread_attr, workerfn,
197				     (void *)(struct worker *) &worker[i]);
198		if (ret) {
199			CPU_FREE(cpuset);
200			err(EXIT_FAILURE, "pthread_create");
201		}
202
203	}
204	CPU_FREE(cpuset);
205	pthread_attr_destroy(&thread_attr);
206
207	mutex_lock(&thread_lock);
208	while (threads_starting)
209		cond_wait(&thread_parent, &thread_lock);
210	cond_broadcast(&thread_worker);
211	mutex_unlock(&thread_lock);
212
213	sleep(params.runtime);
214	toggle_done(0, NULL, NULL);
215
216	for (i = 0; i < params.nthreads; i++) {
217		ret = pthread_join(worker[i].thread, NULL);
218		if (ret)
219			err(EXIT_FAILURE, "pthread_join");
220	}
221
222	/* cleanup & report results */
223	cond_destroy(&thread_parent);
224	cond_destroy(&thread_worker);
225	mutex_destroy(&thread_lock);
226
227	for (i = 0; i < params.nthreads; i++) {
228		unsigned long t = bench__runtime.tv_sec > 0 ?
229			worker[i].ops / bench__runtime.tv_sec : 0;
230		update_stats(&throughput_stats, t);
231		if (!params.silent) {
232			if (params.nfutexes == 1)
233				printf("[thread %2d] futex: %p [ %ld ops/sec ]\n",
234				       worker[i].tid, &worker[i].futex[0], t);
235			else
236				printf("[thread %2d] futexes: %p ... %p [ %ld ops/sec ]\n",
237				       worker[i].tid, &worker[i].futex[0],
238				       &worker[i].futex[params.nfutexes-1], t);
239		}
240
241		zfree(&worker[i].futex);
242	}
243
244	print_summary();
245
246	free(worker);
247	free(cpu);
248	return ret;
249errmem:
250	err(EXIT_FAILURE, "calloc");
251}