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;
29static int futex_flag = 0;
30
31struct timeval start, end, runtime;
32static pthread_mutex_t thread_lock;
33static unsigned int threads_starting;
34static struct stats throughput_stats;
35static pthread_cond_t thread_parent, thread_worker;
36
37struct worker {
38	int tid;
39	u_int32_t *futex;
40	pthread_t thread;
41	unsigned long ops;
42};
43
44static 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_UINTEGER('f', "futexes", &nfutexes, "Specify amount of futexes per threads"),
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
53static const char * const bench_futex_hash_usage[] = {
54	"perf bench futex hash <options>",
55	NULL
56};
57
58static void *workerfn(void *arg)
59{
60	int ret;
61	unsigned int i;
62	struct worker *w = (struct worker *) arg;
63
64	pthread_mutex_lock(&thread_lock);
65	threads_starting--;
66	if (!threads_starting)
67		pthread_cond_signal(&thread_parent);
68	pthread_cond_wait(&thread_worker, &thread_lock);
69	pthread_mutex_unlock(&thread_lock);
70
71	do {
72		for (i = 0; i < nfutexes; i++, w->ops++) {
73			/*
74			 * We want the futex calls to fail in order to stress
75			 * the hashing of uaddr and not measure other steps,
76			 * such as internal waitqueue handling, thus enlarging
77			 * the critical region protected by hb->lock.
78			 */
79			ret = futex_wait(&w->futex[i], 1234, NULL, futex_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	if (!fshared)
139		futex_flag = FUTEX_PRIVATE_FLAG;
140
141	printf("Run summary [PID %d]: %d threads, each operating on %d [%s] futexes for %d secs.\n\n",
142	       getpid(), nthreads, nfutexes, fshared ? "shared":"private", nsecs);
143
144	init_stats(&throughput_stats);
145	pthread_mutex_init(&thread_lock, NULL);
146	pthread_cond_init(&thread_parent, NULL);
147	pthread_cond_init(&thread_worker, NULL);
148
149	threads_starting = nthreads;
150	pthread_attr_init(&thread_attr);
151	gettimeofday(&start, NULL);
152	for (i = 0; i < nthreads; i++) {
153		worker[i].tid = i;
154		worker[i].futex = calloc(nfutexes, sizeof(*worker[i].futex));
155		if (!worker[i].futex)
156			goto errmem;
157
158		CPU_ZERO(&cpu);
159		CPU_SET(i % ncpus, &cpu);
160
161		ret = pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu);
162		if (ret)
163			err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
164
165		ret = pthread_create(&worker[i].thread, &thread_attr, workerfn,
166				     (void *)(struct worker *) &worker[i]);
167		if (ret)
168			err(EXIT_FAILURE, "pthread_create");
169
170	}
171	pthread_attr_destroy(&thread_attr);
172
173	pthread_mutex_lock(&thread_lock);
174	while (threads_starting)
175		pthread_cond_wait(&thread_parent, &thread_lock);
176	pthread_cond_broadcast(&thread_worker);
177	pthread_mutex_unlock(&thread_lock);
178
179	sleep(nsecs);
180	toggle_done(0, NULL, NULL);
181
182	for (i = 0; i < nthreads; i++) {
183		ret = pthread_join(worker[i].thread, NULL);
184		if (ret)
185			err(EXIT_FAILURE, "pthread_join");
186	}
187
188	/* cleanup & report results */
189	pthread_cond_destroy(&thread_parent);
190	pthread_cond_destroy(&thread_worker);
191	pthread_mutex_destroy(&thread_lock);
192
193	for (i = 0; i < nthreads; i++) {
194		unsigned long t = worker[i].ops/runtime.tv_sec;
195		update_stats(&throughput_stats, t);
196		if (!silent) {
197			if (nfutexes == 1)
198				printf("[thread %2d] futex: %p [ %ld ops/sec ]\n",
199				       worker[i].tid, &worker[i].futex[0], t);
200			else
201				printf("[thread %2d] futexes: %p ... %p [ %ld ops/sec ]\n",
202				       worker[i].tid, &worker[i].futex[0],
203				       &worker[i].futex[nfutexes-1], t);
204		}
205
206		free(worker[i].futex);
207	}
208
209	print_summary();
210
211	free(worker);
212	return ret;
213errmem:
214	err(EXIT_FAILURE, "calloc");
215}
216