xref: /openbmc/linux/tools/perf/bench/futex-lock-pi.c (revision 7d94ab16)
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/stat.h"
12 #include <subcmd/parse-options.h>
13 #include <linux/compiler.h>
14 #include <linux/kernel.h>
15 #include <linux/zalloc.h>
16 #include <errno.h>
17 #include <internal/cpumap.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 
26 struct worker {
27 	int tid;
28 	u_int32_t *futex;
29 	pthread_t thread;
30 	unsigned long ops;
31 };
32 
33 static u_int32_t global_futex = 0;
34 static struct worker *worker;
35 static unsigned int nsecs = 10;
36 static bool silent = false, multi = false;
37 static bool done = false, fshared = false;
38 static unsigned int nthreads = 0;
39 static int futex_flag = 0;
40 struct timeval start, end, runtime;
41 static pthread_mutex_t thread_lock;
42 static unsigned int threads_starting;
43 static struct stats throughput_stats;
44 static pthread_cond_t thread_parent, thread_worker;
45 
46 static const struct option options[] = {
47 	OPT_UINTEGER('t', "threads",  &nthreads, "Specify amount of threads"),
48 	OPT_UINTEGER('r', "runtime", &nsecs,     "Specify runtime (in seconds)"),
49 	OPT_BOOLEAN( 'M', "multi",   &multi,     "Use multiple futexes"),
50 	OPT_BOOLEAN( 's', "silent",  &silent,    "Silent mode: do not display data/details"),
51 	OPT_BOOLEAN( 'S', "shared",  &fshared,   "Use shared futexes instead of private ones"),
52 	OPT_END()
53 };
54 
55 static const char * const bench_futex_lock_pi_usage[] = {
56 	"perf bench futex lock-pi <options>",
57 	NULL
58 };
59 
60 static void print_summary(void)
61 {
62 	unsigned long avg = avg_stats(&throughput_stats);
63 	double stddev = stddev_stats(&throughput_stats);
64 
65 	printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
66 	       !silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg),
67 	       (int) runtime.tv_sec);
68 }
69 
70 static void toggle_done(int sig __maybe_unused,
71 			siginfo_t *info __maybe_unused,
72 			void *uc __maybe_unused)
73 {
74 	/* inform all threads that we're done for the day */
75 	done = true;
76 	gettimeofday(&end, NULL);
77 	timersub(&end, &start, &runtime);
78 }
79 
80 static void *workerfn(void *arg)
81 {
82 	struct worker *w = (struct worker *) arg;
83 	unsigned long ops = w->ops;
84 
85 	pthread_mutex_lock(&thread_lock);
86 	threads_starting--;
87 	if (!threads_starting)
88 		pthread_cond_signal(&thread_parent);
89 	pthread_cond_wait(&thread_worker, &thread_lock);
90 	pthread_mutex_unlock(&thread_lock);
91 
92 	do {
93 		int ret;
94 	again:
95 		ret = futex_lock_pi(w->futex, NULL, futex_flag);
96 
97 		if (ret) { /* handle lock acquisition */
98 			if (!silent)
99 				warn("thread %d: Could not lock pi-lock for %p (%d)",
100 				     w->tid, w->futex, ret);
101 			if (done)
102 				break;
103 
104 			goto again;
105 		}
106 
107 		usleep(1);
108 		ret = futex_unlock_pi(w->futex, futex_flag);
109 		if (ret && !silent)
110 			warn("thread %d: Could not unlock pi-lock for %p (%d)",
111 			     w->tid, w->futex, ret);
112 		ops++; /* account for thread's share of work */
113 	}  while (!done);
114 
115 	w->ops = ops;
116 	return NULL;
117 }
118 
119 static void create_threads(struct worker *w, pthread_attr_t thread_attr,
120 			   struct perf_cpu_map *cpu)
121 {
122 	cpu_set_t cpuset;
123 	unsigned int i;
124 
125 	threads_starting = nthreads;
126 
127 	for (i = 0; i < nthreads; i++) {
128 		worker[i].tid = i;
129 
130 		if (multi) {
131 			worker[i].futex = calloc(1, sizeof(u_int32_t));
132 			if (!worker[i].futex)
133 				err(EXIT_FAILURE, "calloc");
134 		} else
135 			worker[i].futex = &global_futex;
136 
137 		CPU_ZERO(&cpuset);
138 		CPU_SET(cpu->map[i % cpu->nr], &cpuset);
139 
140 		if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset))
141 			err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
142 
143 		if (pthread_create(&w[i].thread, &thread_attr, workerfn, &worker[i]))
144 			err(EXIT_FAILURE, "pthread_create");
145 	}
146 }
147 
148 int bench_futex_lock_pi(int argc, const char **argv)
149 {
150 	int ret = 0;
151 	unsigned int i;
152 	struct sigaction act;
153 	pthread_attr_t thread_attr;
154 	struct perf_cpu_map *cpu;
155 
156 	argc = parse_options(argc, argv, options, bench_futex_lock_pi_usage, 0);
157 	if (argc)
158 		goto err;
159 
160 	cpu = perf_cpu_map__new(NULL);
161 	if (!cpu)
162 		err(EXIT_FAILURE, "calloc");
163 
164 	sigfillset(&act.sa_mask);
165 	act.sa_sigaction = toggle_done;
166 	sigaction(SIGINT, &act, NULL);
167 
168 	if (!nthreads)
169 		nthreads = cpu->nr;
170 
171 	worker = calloc(nthreads, sizeof(*worker));
172 	if (!worker)
173 		err(EXIT_FAILURE, "calloc");
174 
175 	if (!fshared)
176 		futex_flag = FUTEX_PRIVATE_FLAG;
177 
178 	printf("Run summary [PID %d]: %d threads doing pi lock/unlock pairing for %d secs.\n\n",
179 	       getpid(), nthreads, nsecs);
180 
181 	init_stats(&throughput_stats);
182 	pthread_mutex_init(&thread_lock, NULL);
183 	pthread_cond_init(&thread_parent, NULL);
184 	pthread_cond_init(&thread_worker, NULL);
185 
186 	threads_starting = nthreads;
187 	pthread_attr_init(&thread_attr);
188 	gettimeofday(&start, NULL);
189 
190 	create_threads(worker, thread_attr, cpu);
191 	pthread_attr_destroy(&thread_attr);
192 
193 	pthread_mutex_lock(&thread_lock);
194 	while (threads_starting)
195 		pthread_cond_wait(&thread_parent, &thread_lock);
196 	pthread_cond_broadcast(&thread_worker);
197 	pthread_mutex_unlock(&thread_lock);
198 
199 	sleep(nsecs);
200 	toggle_done(0, NULL, NULL);
201 
202 	for (i = 0; i < nthreads; i++) {
203 		ret = pthread_join(worker[i].thread, NULL);
204 		if (ret)
205 			err(EXIT_FAILURE, "pthread_join");
206 	}
207 
208 	/* cleanup & report results */
209 	pthread_cond_destroy(&thread_parent);
210 	pthread_cond_destroy(&thread_worker);
211 	pthread_mutex_destroy(&thread_lock);
212 
213 	for (i = 0; i < nthreads; i++) {
214 		unsigned long t = worker[i].ops/runtime.tv_sec;
215 
216 		update_stats(&throughput_stats, t);
217 		if (!silent)
218 			printf("[thread %3d] futex: %p [ %ld ops/sec ]\n",
219 			       worker[i].tid, worker[i].futex, t);
220 
221 		if (multi)
222 			zfree(&worker[i].futex);
223 	}
224 
225 	print_summary();
226 
227 	free(worker);
228 	return ret;
229 err:
230 	usage_with_options(bench_futex_lock_pi_usage, options);
231 	exit(EXIT_FAILURE);
232 }
233