Commit | Line | Data |
---|---|---|
27db7830 DB |
1 | /* |
2 | * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com> | |
3 | * | |
4 | * futex-wake: Block a bunch of threads on a futex and wake'em up, N at a time. | |
5 | * | |
6 | * This program is particularly useful to measure the latency of nthread wakeups | |
7 | * in non-error situations: all waiters are queued and all wake calls wakeup | |
8 | * one or more tasks, and thus the waitqueue is never empty. | |
9 | */ | |
10 | ||
8a158589 ACM |
11 | /* For the CLR_() macros */ |
12 | #include <pthread.h> | |
13 | ||
9c304f6c | 14 | #include <signal.h> |
27db7830 | 15 | #include "../util/stat.h" |
4b6ab94e | 16 | #include <subcmd/parse-options.h> |
86695f59 | 17 | #include <linux/compiler.h> |
9c304f6c ACM |
18 | #include <linux/kernel.h> |
19 | #include <errno.h> | |
27db7830 DB |
20 | #include "bench.h" |
21 | #include "futex.h" | |
22 | ||
23 | #include <err.h> | |
24 | #include <stdlib.h> | |
25 | #include <sys/time.h> | |
27db7830 DB |
26 | |
27 | /* all threads will block on the same futex */ | |
28 | static u_int32_t futex1 = 0; | |
29 | ||
30 | /* | |
31 | * How many wakeups to do at a time. | |
32 | * Default to 1 in order to make the kernel work more. | |
33 | */ | |
34 | static unsigned int nwakes = 1; | |
35 | ||
27db7830 | 36 | pthread_t *worker; |
86c87e13 | 37 | static bool done = false, silent = false, fshared = false; |
27db7830 DB |
38 | static pthread_mutex_t thread_lock; |
39 | static pthread_cond_t thread_parent, thread_worker; | |
40 | static struct stats waketime_stats, wakeup_stats; | |
41 | static unsigned int ncpus, threads_starting, nthreads = 0; | |
86c87e13 | 42 | static int futex_flag = 0; |
27db7830 DB |
43 | |
44 | static const struct option options[] = { | |
45 | OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"), | |
46 | OPT_UINTEGER('w', "nwakes", &nwakes, "Specify amount of threads to wake at once"), | |
27db7830 | 47 | OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"), |
86c87e13 | 48 | OPT_BOOLEAN( 'S', "shared", &fshared, "Use shared futexes instead of private ones"), |
27db7830 DB |
49 | OPT_END() |
50 | }; | |
51 | ||
52 | static const char * const bench_futex_wake_usage[] = { | |
53 | "perf bench futex wake <options>", | |
54 | NULL | |
55 | }; | |
56 | ||
57 | static void *workerfn(void *arg __maybe_unused) | |
58 | { | |
59 | pthread_mutex_lock(&thread_lock); | |
60 | threads_starting--; | |
61 | if (!threads_starting) | |
62 | pthread_cond_signal(&thread_parent); | |
63 | pthread_cond_wait(&thread_worker, &thread_lock); | |
64 | pthread_mutex_unlock(&thread_lock); | |
65 | ||
598adc5c DB |
66 | while (1) { |
67 | if (futex_wait(&futex1, 0, NULL, futex_flag) != EINTR) | |
68 | break; | |
69 | } | |
70 | ||
71 | pthread_exit(NULL); | |
27db7830 DB |
72 | return NULL; |
73 | } | |
74 | ||
75 | static void print_summary(void) | |
76 | { | |
77 | double waketime_avg = avg_stats(&waketime_stats); | |
78 | double waketime_stddev = stddev_stats(&waketime_stats); | |
79 | unsigned int wakeup_avg = avg_stats(&wakeup_stats); | |
80 | ||
81 | printf("Wokeup %d of %d threads in %.4f ms (+-%.2f%%)\n", | |
82 | wakeup_avg, | |
83 | nthreads, | |
84 | waketime_avg/1e3, | |
85 | rel_stddev_stats(waketime_stddev, waketime_avg)); | |
86 | } | |
87 | ||
88 | static void block_threads(pthread_t *w, | |
89 | pthread_attr_t thread_attr) | |
90 | { | |
91 | cpu_set_t cpu; | |
92 | unsigned int i; | |
93 | ||
94 | threads_starting = nthreads; | |
95 | ||
96 | /* create and block all threads */ | |
97 | for (i = 0; i < nthreads; i++) { | |
98 | CPU_ZERO(&cpu); | |
99 | CPU_SET(i % ncpus, &cpu); | |
100 | ||
101 | if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu)) | |
102 | err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); | |
103 | ||
104 | if (pthread_create(&w[i], &thread_attr, workerfn, NULL)) | |
105 | err(EXIT_FAILURE, "pthread_create"); | |
106 | } | |
107 | } | |
108 | ||
109 | static void toggle_done(int sig __maybe_unused, | |
110 | siginfo_t *info __maybe_unused, | |
111 | void *uc __maybe_unused) | |
112 | { | |
113 | done = true; | |
114 | } | |
115 | ||
116 | int bench_futex_wake(int argc, const char **argv, | |
117 | const char *prefix __maybe_unused) | |
118 | { | |
119 | int ret = 0; | |
120 | unsigned int i, j; | |
121 | struct sigaction act; | |
122 | pthread_attr_t thread_attr; | |
123 | ||
124 | argc = parse_options(argc, argv, options, bench_futex_wake_usage, 0); | |
125 | if (argc) { | |
126 | usage_with_options(bench_futex_wake_usage, options); | |
127 | exit(EXIT_FAILURE); | |
128 | } | |
129 | ||
130 | ncpus = sysconf(_SC_NPROCESSORS_ONLN); | |
131 | ||
132 | sigfillset(&act.sa_mask); | |
133 | act.sa_sigaction = toggle_done; | |
134 | sigaction(SIGINT, &act, NULL); | |
135 | ||
136 | if (!nthreads) | |
137 | nthreads = ncpus; | |
138 | ||
139 | worker = calloc(nthreads, sizeof(*worker)); | |
140 | if (!worker) | |
141 | err(EXIT_FAILURE, "calloc"); | |
142 | ||
86c87e13 DB |
143 | if (!fshared) |
144 | futex_flag = FUTEX_PRIVATE_FLAG; | |
145 | ||
146 | printf("Run summary [PID %d]: blocking on %d threads (at [%s] futex %p), " | |
27db7830 | 147 | "waking up %d at a time.\n\n", |
86c87e13 | 148 | getpid(), nthreads, fshared ? "shared":"private", &futex1, nwakes); |
27db7830 DB |
149 | |
150 | init_stats(&wakeup_stats); | |
151 | init_stats(&waketime_stats); | |
152 | pthread_attr_init(&thread_attr); | |
153 | pthread_mutex_init(&thread_lock, NULL); | |
154 | pthread_cond_init(&thread_parent, NULL); | |
155 | pthread_cond_init(&thread_worker, NULL); | |
156 | ||
d9de84af | 157 | for (j = 0; j < bench_repeat && !done; j++) { |
27db7830 DB |
158 | unsigned int nwoken = 0; |
159 | struct timeval start, end, runtime; | |
160 | ||
161 | /* create, launch & block all threads */ | |
162 | block_threads(worker, thread_attr); | |
163 | ||
164 | /* make sure all threads are already blocked */ | |
165 | pthread_mutex_lock(&thread_lock); | |
166 | while (threads_starting) | |
167 | pthread_cond_wait(&thread_parent, &thread_lock); | |
168 | pthread_cond_broadcast(&thread_worker); | |
169 | pthread_mutex_unlock(&thread_lock); | |
170 | ||
171 | usleep(100000); | |
172 | ||
173 | /* Ok, all threads are patiently blocked, start waking folks up */ | |
174 | gettimeofday(&start, NULL); | |
175 | while (nwoken != nthreads) | |
86c87e13 | 176 | nwoken += futex_wake(&futex1, nwakes, futex_flag); |
27db7830 DB |
177 | gettimeofday(&end, NULL); |
178 | timersub(&end, &start, &runtime); | |
179 | ||
180 | update_stats(&wakeup_stats, nwoken); | |
181 | update_stats(&waketime_stats, runtime.tv_usec); | |
182 | ||
183 | if (!silent) { | |
184 | printf("[Run %d]: Wokeup %d of %d threads in %.4f ms\n", | |
185 | j + 1, nwoken, nthreads, runtime.tv_usec/1e3); | |
186 | } | |
187 | ||
188 | for (i = 0; i < nthreads; i++) { | |
189 | ret = pthread_join(worker[i], NULL); | |
190 | if (ret) | |
191 | err(EXIT_FAILURE, "pthread_join"); | |
192 | } | |
193 | ||
194 | } | |
195 | ||
196 | /* cleanup & report results */ | |
197 | pthread_cond_destroy(&thread_parent); | |
198 | pthread_cond_destroy(&thread_worker); | |
199 | pthread_mutex_destroy(&thread_lock); | |
200 | pthread_attr_destroy(&thread_attr); | |
201 | ||
202 | print_summary(); | |
203 | ||
204 | free(worker); | |
205 | return ret; | |
206 | } |