xref: /openbmc/linux/tools/perf/bench/futex-lock-pi.c (revision 29d97219)
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 <perf/cpumap.h>
18 #include "bench.h"
19 #include "futex.h"
20 
21 #include <err.h>
22 #include <stdlib.h>
23 #include <sys/time.h>
24 
25 struct worker {
26 	int tid;
27 	u_int32_t *futex;
28 	pthread_t thread;
29 	unsigned long ops;
30 };
31 
32 static u_int32_t global_futex = 0;
33 static struct worker *worker;
34 static unsigned int nsecs = 10;
35 static bool silent = false, multi = false;
36 static bool done = false, fshared = false;
37 static unsigned int nthreads = 0;
38 static int futex_flag = 0;
39 static pthread_mutex_t thread_lock;
40 static unsigned int threads_starting;
41 static struct stats throughput_stats;
42 static pthread_cond_t thread_parent, thread_worker;
43 
44 static const struct option options[] = {
45 	OPT_UINTEGER('t', "threads",  &nthreads, "Specify amount of threads"),
46 	OPT_UINTEGER('r', "runtime", &nsecs,     "Specify runtime (in seconds)"),
47 	OPT_BOOLEAN( 'M', "multi",   &multi,     "Use multiple futexes"),
48 	OPT_BOOLEAN( 's', "silent",  &silent,    "Silent mode: do not display data/details"),
49 	OPT_BOOLEAN( 'S', "shared",  &fshared,   "Use shared futexes instead of private ones"),
50 	OPT_END()
51 };
52 
53 static const char * const bench_futex_lock_pi_usage[] = {
54 	"perf bench futex lock-pi <options>",
55 	NULL
56 };
57 
58 static void print_summary(void)
59 {
60 	unsigned long avg = avg_stats(&throughput_stats);
61 	double stddev = stddev_stats(&throughput_stats);
62 
63 	printf("%sAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
64 	       !silent ? "\n" : "", avg, rel_stddev_stats(stddev, avg),
65 	       (int)bench__runtime.tv_sec);
66 }
67 
68 static void toggle_done(int sig __maybe_unused,
69 			siginfo_t *info __maybe_unused,
70 			void *uc __maybe_unused)
71 {
72 	/* inform all threads that we're done for the day */
73 	done = true;
74 	gettimeofday(&bench__end, NULL);
75 	timersub(&bench__end, &bench__start, &bench__runtime);
76 }
77 
78 static void *workerfn(void *arg)
79 {
80 	struct worker *w = (struct worker *) arg;
81 	unsigned long ops = w->ops;
82 
83 	pthread_mutex_lock(&thread_lock);
84 	threads_starting--;
85 	if (!threads_starting)
86 		pthread_cond_signal(&thread_parent);
87 	pthread_cond_wait(&thread_worker, &thread_lock);
88 	pthread_mutex_unlock(&thread_lock);
89 
90 	do {
91 		int ret;
92 	again:
93 		ret = futex_lock_pi(w->futex, NULL, futex_flag);
94 
95 		if (ret) { /* handle lock acquisition */
96 			if (!silent)
97 				warn("thread %d: Could not lock pi-lock for %p (%d)",
98 				     w->tid, w->futex, ret);
99 			if (done)
100 				break;
101 
102 			goto again;
103 		}
104 
105 		usleep(1);
106 		ret = futex_unlock_pi(w->futex, futex_flag);
107 		if (ret && !silent)
108 			warn("thread %d: Could not unlock pi-lock for %p (%d)",
109 			     w->tid, w->futex, ret);
110 		ops++; /* account for thread's share of work */
111 	}  while (!done);
112 
113 	w->ops = ops;
114 	return NULL;
115 }
116 
117 static void create_threads(struct worker *w, pthread_attr_t thread_attr,
118 			   struct perf_cpu_map *cpu)
119 {
120 	cpu_set_t cpuset;
121 	unsigned int i;
122 
123 	threads_starting = nthreads;
124 
125 	for (i = 0; i < nthreads; i++) {
126 		worker[i].tid = i;
127 
128 		if (multi) {
129 			worker[i].futex = calloc(1, sizeof(u_int32_t));
130 			if (!worker[i].futex)
131 				err(EXIT_FAILURE, "calloc");
132 		} else
133 			worker[i].futex = &global_futex;
134 
135 		CPU_ZERO(&cpuset);
136 		CPU_SET(cpu->map[i % cpu->nr], &cpuset);
137 
138 		if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset))
139 			err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
140 
141 		if (pthread_create(&w[i].thread, &thread_attr, workerfn, &worker[i]))
142 			err(EXIT_FAILURE, "pthread_create");
143 	}
144 }
145 
146 int bench_futex_lock_pi(int argc, const char **argv)
147 {
148 	int ret = 0;
149 	unsigned int i;
150 	struct sigaction act;
151 	pthread_attr_t thread_attr;
152 	struct perf_cpu_map *cpu;
153 
154 	argc = parse_options(argc, argv, options, bench_futex_lock_pi_usage, 0);
155 	if (argc)
156 		goto err;
157 
158 	cpu = perf_cpu_map__new(NULL);
159 	if (!cpu)
160 		err(EXIT_FAILURE, "calloc");
161 
162 	memset(&act, 0, sizeof(act));
163 	sigfillset(&act.sa_mask);
164 	act.sa_sigaction = toggle_done;
165 	sigaction(SIGINT, &act, NULL);
166 
167 	if (!nthreads)
168 		nthreads = cpu->nr;
169 
170 	worker = calloc(nthreads, sizeof(*worker));
171 	if (!worker)
172 		err(EXIT_FAILURE, "calloc");
173 
174 	if (!fshared)
175 		futex_flag = FUTEX_PRIVATE_FLAG;
176 
177 	printf("Run summary [PID %d]: %d threads doing pi lock/unlock pairing for %d secs.\n\n",
178 	       getpid(), nthreads, nsecs);
179 
180 	init_stats(&throughput_stats);
181 	pthread_mutex_init(&thread_lock, NULL);
182 	pthread_cond_init(&thread_parent, NULL);
183 	pthread_cond_init(&thread_worker, NULL);
184 
185 	threads_starting = nthreads;
186 	pthread_attr_init(&thread_attr);
187 	gettimeofday(&bench__start, NULL);
188 
189 	create_threads(worker, thread_attr, cpu);
190 	pthread_attr_destroy(&thread_attr);
191 
192 	pthread_mutex_lock(&thread_lock);
193 	while (threads_starting)
194 		pthread_cond_wait(&thread_parent, &thread_lock);
195 	pthread_cond_broadcast(&thread_worker);
196 	pthread_mutex_unlock(&thread_lock);
197 
198 	sleep(nsecs);
199 	toggle_done(0, NULL, NULL);
200 
201 	for (i = 0; i < nthreads; i++) {
202 		ret = pthread_join(worker[i].thread, NULL);
203 		if (ret)
204 			err(EXIT_FAILURE, "pthread_join");
205 	}
206 
207 	/* cleanup & report results */
208 	pthread_cond_destroy(&thread_parent);
209 	pthread_cond_destroy(&thread_worker);
210 	pthread_mutex_destroy(&thread_lock);
211 
212 	for (i = 0; i < nthreads; i++) {
213 		unsigned long t = bench__runtime.tv_sec > 0 ?
214 			worker[i].ops / bench__runtime.tv_sec : 0;
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