]>
Commit | Line | Data |
---|---|---|
0fb298cf DB |
1 | /* |
2 | * Copyright (C) 2013 Davidlohr Bueso <davidlohr@hp.com> | |
3 | * | |
4 | * futex-requeue: Block a bunch of threads on futex1 and requeue them | |
5 | * on futex2, N at a time. | |
6 | * | |
7 | * This program is particularly useful to measure the latency of nthread | |
8 | * requeues without waking up any tasks -- thus mimicking a regular futex_wait. | |
9 | */ | |
10 | ||
8a158589 ACM |
11 | /* For the CLR_() macros */ |
12 | #include <pthread.h> | |
13 | ||
9c304f6c | 14 | #include <signal.h> |
0fb298cf | 15 | #include "../util/stat.h" |
4b6ab94e | 16 | #include <subcmd/parse-options.h> |
86695f59 | 17 | #include <linux/compiler.h> |
9c304f6c | 18 | #include <linux/kernel.h> |
565e6911 | 19 | #include <linux/time64.h> |
9c304f6c | 20 | #include <errno.h> |
0fb298cf DB |
21 | #include "bench.h" |
22 | #include "futex.h" | |
23 | ||
24 | #include <err.h> | |
25 | #include <stdlib.h> | |
26 | #include <sys/time.h> | |
0fb298cf DB |
27 | |
28 | static u_int32_t futex1 = 0, futex2 = 0; | |
29 | ||
30 | /* | |
31 | * How many tasks to requeue at a time. | |
32 | * Default to 1 in order to make the kernel work more. | |
33 | */ | |
34 | static unsigned int nrequeue = 1; | |
35 | ||
0fb298cf | 36 | static pthread_t *worker; |
86c87e13 | 37 | static bool done = false, silent = false, fshared = false; |
0fb298cf DB |
38 | static pthread_mutex_t thread_lock; |
39 | static pthread_cond_t thread_parent, thread_worker; | |
40 | static struct stats requeuetime_stats, requeued_stats; | |
41 | static unsigned int ncpus, threads_starting, nthreads = 0; | |
86c87e13 | 42 | static int futex_flag = 0; |
0fb298cf DB |
43 | |
44 | static const struct option options[] = { | |
45 | OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"), | |
46 | OPT_UINTEGER('q', "nrequeue", &nrequeue, "Specify amount of threads to requeue at once"), | |
0fb298cf | 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"), |
0fb298cf DB |
49 | OPT_END() |
50 | }; | |
51 | ||
52 | static const char * const bench_futex_requeue_usage[] = { | |
53 | "perf bench futex requeue <options>", | |
54 | NULL | |
55 | }; | |
56 | ||
57 | static void print_summary(void) | |
58 | { | |
59 | double requeuetime_avg = avg_stats(&requeuetime_stats); | |
60 | double requeuetime_stddev = stddev_stats(&requeuetime_stats); | |
61 | unsigned int requeued_avg = avg_stats(&requeued_stats); | |
62 | ||
63 | printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n", | |
64 | requeued_avg, | |
65 | nthreads, | |
565e6911 | 66 | requeuetime_avg / USEC_PER_MSEC, |
0fb298cf DB |
67 | rel_stddev_stats(requeuetime_stddev, requeuetime_avg)); |
68 | } | |
69 | ||
70 | static void *workerfn(void *arg __maybe_unused) | |
71 | { | |
72 | pthread_mutex_lock(&thread_lock); | |
73 | threads_starting--; | |
74 | if (!threads_starting) | |
75 | pthread_cond_signal(&thread_parent); | |
76 | pthread_cond_wait(&thread_worker, &thread_lock); | |
77 | pthread_mutex_unlock(&thread_lock); | |
78 | ||
86c87e13 | 79 | futex_wait(&futex1, 0, NULL, futex_flag); |
0fb298cf DB |
80 | return NULL; |
81 | } | |
82 | ||
83 | static void block_threads(pthread_t *w, | |
84 | pthread_attr_t thread_attr) | |
85 | { | |
86 | cpu_set_t cpu; | |
87 | unsigned int i; | |
88 | ||
89 | threads_starting = nthreads; | |
90 | ||
91 | /* create and block all threads */ | |
92 | for (i = 0; i < nthreads; i++) { | |
93 | CPU_ZERO(&cpu); | |
94 | CPU_SET(i % ncpus, &cpu); | |
95 | ||
96 | if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu)) | |
97 | err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); | |
98 | ||
99 | if (pthread_create(&w[i], &thread_attr, workerfn, NULL)) | |
100 | err(EXIT_FAILURE, "pthread_create"); | |
101 | } | |
102 | } | |
103 | ||
104 | static void toggle_done(int sig __maybe_unused, | |
105 | siginfo_t *info __maybe_unused, | |
106 | void *uc __maybe_unused) | |
107 | { | |
108 | done = true; | |
109 | } | |
110 | ||
111 | int bench_futex_requeue(int argc, const char **argv, | |
112 | const char *prefix __maybe_unused) | |
113 | { | |
114 | int ret = 0; | |
115 | unsigned int i, j; | |
116 | struct sigaction act; | |
117 | pthread_attr_t thread_attr; | |
118 | ||
119 | argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0); | |
120 | if (argc) | |
121 | goto err; | |
122 | ||
123 | ncpus = sysconf(_SC_NPROCESSORS_ONLN); | |
124 | ||
125 | sigfillset(&act.sa_mask); | |
126 | act.sa_sigaction = toggle_done; | |
127 | sigaction(SIGINT, &act, NULL); | |
128 | ||
129 | if (!nthreads) | |
130 | nthreads = ncpus; | |
131 | ||
132 | worker = calloc(nthreads, sizeof(*worker)); | |
133 | if (!worker) | |
134 | err(EXIT_FAILURE, "calloc"); | |
135 | ||
86c87e13 DB |
136 | if (!fshared) |
137 | futex_flag = FUTEX_PRIVATE_FLAG; | |
138 | ||
052b0f6e DB |
139 | if (nrequeue > nthreads) |
140 | nrequeue = nthreads; | |
141 | ||
86c87e13 DB |
142 | printf("Run summary [PID %d]: Requeuing %d threads (from [%s] %p to %p), " |
143 | "%d at a time.\n\n", getpid(), nthreads, | |
144 | fshared ? "shared":"private", &futex1, &futex2, nrequeue); | |
0fb298cf DB |
145 | |
146 | init_stats(&requeued_stats); | |
147 | init_stats(&requeuetime_stats); | |
148 | pthread_attr_init(&thread_attr); | |
149 | pthread_mutex_init(&thread_lock, NULL); | |
150 | pthread_cond_init(&thread_parent, NULL); | |
151 | pthread_cond_init(&thread_worker, NULL); | |
152 | ||
d9de84af | 153 | for (j = 0; j < bench_repeat && !done; j++) { |
0fb298cf DB |
154 | unsigned int nrequeued = 0; |
155 | struct timeval start, end, runtime; | |
156 | ||
157 | /* create, launch & block all threads */ | |
158 | block_threads(worker, thread_attr); | |
159 | ||
160 | /* make sure all threads are already blocked */ | |
161 | pthread_mutex_lock(&thread_lock); | |
162 | while (threads_starting) | |
163 | pthread_cond_wait(&thread_parent, &thread_lock); | |
164 | pthread_cond_broadcast(&thread_worker); | |
165 | pthread_mutex_unlock(&thread_lock); | |
166 | ||
167 | usleep(100000); | |
168 | ||
169 | /* Ok, all threads are patiently blocked, start requeueing */ | |
170 | gettimeofday(&start, NULL); | |
052b0f6e | 171 | while (nrequeued < nthreads) { |
0fb298cf DB |
172 | /* |
173 | * Do not wakeup any tasks blocked on futex1, allowing | |
174 | * us to really measure futex_wait functionality. | |
175 | */ | |
052b0f6e DB |
176 | nrequeued += futex_cmp_requeue(&futex1, 0, &futex2, 0, |
177 | nrequeue, futex_flag); | |
86c87e13 | 178 | } |
052b0f6e | 179 | |
0fb298cf DB |
180 | gettimeofday(&end, NULL); |
181 | timersub(&end, &start, &runtime); | |
182 | ||
183 | update_stats(&requeued_stats, nrequeued); | |
184 | update_stats(&requeuetime_stats, runtime.tv_usec); | |
185 | ||
186 | if (!silent) { | |
187 | printf("[Run %d]: Requeued %d of %d threads in %.4f ms\n", | |
565e6911 | 188 | j + 1, nrequeued, nthreads, runtime.tv_usec / (double)USEC_PER_MSEC); |
0fb298cf DB |
189 | } |
190 | ||
191 | /* everybody should be blocked on futex2, wake'em up */ | |
052b0f6e | 192 | nrequeued = futex_wake(&futex2, nrequeued, futex_flag); |
0fb298cf DB |
193 | if (nthreads != nrequeued) |
194 | warnx("couldn't wakeup all tasks (%d/%d)", nrequeued, nthreads); | |
195 | ||
196 | for (i = 0; i < nthreads; i++) { | |
197 | ret = pthread_join(worker[i], NULL); | |
198 | if (ret) | |
199 | err(EXIT_FAILURE, "pthread_join"); | |
200 | } | |
0fb298cf DB |
201 | } |
202 | ||
203 | /* cleanup & report results */ | |
204 | pthread_cond_destroy(&thread_parent); | |
205 | pthread_cond_destroy(&thread_worker); | |
206 | pthread_mutex_destroy(&thread_lock); | |
207 | pthread_attr_destroy(&thread_attr); | |
208 | ||
209 | print_summary(); | |
210 | ||
211 | free(worker); | |
212 | return ret; | |
213 | err: | |
214 | usage_with_options(bench_futex_requeue_usage, options); | |
215 | exit(EXIT_FAILURE); | |
216 | } |