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 | ||
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 | ||
24 | static u_int32_t futex1 = 0, futex2 = 0; | |
25 | ||
26 | /* | |
27 | * How many tasks to requeue at a time. | |
28 | * Default to 1 in order to make the kernel work more. | |
29 | */ | |
30 | static unsigned int nrequeue = 1; | |
31 | ||
32 | /* | |
33 | * There can be significant variance from run to run, | |
34 | * the more repeats, the more exact the overall avg and | |
35 | * the better idea of the futex latency. | |
36 | */ | |
37 | static unsigned int repeat = 10; | |
38 | ||
39 | static pthread_t *worker; | |
40 | static bool done = 0, silent = 0; | |
41 | static pthread_mutex_t thread_lock; | |
42 | static pthread_cond_t thread_parent, thread_worker; | |
43 | static struct stats requeuetime_stats, requeued_stats; | |
44 | static unsigned int ncpus, threads_starting, nthreads = 0; | |
45 | ||
46 | static const struct option options[] = { | |
47 | OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"), | |
48 | OPT_UINTEGER('q', "nrequeue", &nrequeue, "Specify amount of threads to requeue at once"), | |
49 | OPT_UINTEGER('r', "repeat", &repeat, "Specify amount of times to repeat the run"), | |
50 | OPT_BOOLEAN( 's', "silent", &silent, "Silent mode: do not display data/details"), | |
51 | OPT_END() | |
52 | }; | |
53 | ||
54 | static const char * const bench_futex_requeue_usage[] = { | |
55 | "perf bench futex requeue <options>", | |
56 | NULL | |
57 | }; | |
58 | ||
59 | static void print_summary(void) | |
60 | { | |
61 | double requeuetime_avg = avg_stats(&requeuetime_stats); | |
62 | double requeuetime_stddev = stddev_stats(&requeuetime_stats); | |
63 | unsigned int requeued_avg = avg_stats(&requeued_stats); | |
64 | ||
65 | printf("Requeued %d of %d threads in %.4f ms (+-%.2f%%)\n", | |
66 | requeued_avg, | |
67 | nthreads, | |
68 | requeuetime_avg/1e3, | |
69 | rel_stddev_stats(requeuetime_stddev, requeuetime_avg)); | |
70 | } | |
71 | ||
72 | static void *workerfn(void *arg __maybe_unused) | |
73 | { | |
74 | pthread_mutex_lock(&thread_lock); | |
75 | threads_starting--; | |
76 | if (!threads_starting) | |
77 | pthread_cond_signal(&thread_parent); | |
78 | pthread_cond_wait(&thread_worker, &thread_lock); | |
79 | pthread_mutex_unlock(&thread_lock); | |
80 | ||
81 | futex_wait(&futex1, 0, NULL, FUTEX_PRIVATE_FLAG); | |
82 | return NULL; | |
83 | } | |
84 | ||
85 | static void block_threads(pthread_t *w, | |
86 | pthread_attr_t thread_attr) | |
87 | { | |
88 | cpu_set_t cpu; | |
89 | unsigned int i; | |
90 | ||
91 | threads_starting = nthreads; | |
92 | ||
93 | /* create and block all threads */ | |
94 | for (i = 0; i < nthreads; i++) { | |
95 | CPU_ZERO(&cpu); | |
96 | CPU_SET(i % ncpus, &cpu); | |
97 | ||
98 | if (pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu)) | |
99 | err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); | |
100 | ||
101 | if (pthread_create(&w[i], &thread_attr, workerfn, NULL)) | |
102 | err(EXIT_FAILURE, "pthread_create"); | |
103 | } | |
104 | } | |
105 | ||
106 | static void toggle_done(int sig __maybe_unused, | |
107 | siginfo_t *info __maybe_unused, | |
108 | void *uc __maybe_unused) | |
109 | { | |
110 | done = true; | |
111 | } | |
112 | ||
113 | int bench_futex_requeue(int argc, const char **argv, | |
114 | const char *prefix __maybe_unused) | |
115 | { | |
116 | int ret = 0; | |
117 | unsigned int i, j; | |
118 | struct sigaction act; | |
119 | pthread_attr_t thread_attr; | |
120 | ||
121 | argc = parse_options(argc, argv, options, bench_futex_requeue_usage, 0); | |
122 | if (argc) | |
123 | goto err; | |
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) | |
132 | nthreads = ncpus; | |
133 | ||
134 | worker = calloc(nthreads, sizeof(*worker)); | |
135 | if (!worker) | |
136 | err(EXIT_FAILURE, "calloc"); | |
137 | ||
138 | printf("Run summary [PID %d]: Requeuing %d threads (from %p to %p), " | |
139 | "%d at a time.\n\n", | |
140 | getpid(), nthreads, &futex1, &futex2, nrequeue); | |
141 | ||
142 | init_stats(&requeued_stats); | |
143 | init_stats(&requeuetime_stats); | |
144 | pthread_attr_init(&thread_attr); | |
145 | pthread_mutex_init(&thread_lock, NULL); | |
146 | pthread_cond_init(&thread_parent, NULL); | |
147 | pthread_cond_init(&thread_worker, NULL); | |
148 | ||
149 | for (j = 0; j < repeat && !done; j++) { | |
150 | unsigned int nrequeued = 0; | |
151 | struct timeval start, end, runtime; | |
152 | ||
153 | /* create, launch & block all threads */ | |
154 | block_threads(worker, thread_attr); | |
155 | ||
156 | /* make sure all threads are already blocked */ | |
157 | pthread_mutex_lock(&thread_lock); | |
158 | while (threads_starting) | |
159 | pthread_cond_wait(&thread_parent, &thread_lock); | |
160 | pthread_cond_broadcast(&thread_worker); | |
161 | pthread_mutex_unlock(&thread_lock); | |
162 | ||
163 | usleep(100000); | |
164 | ||
165 | /* Ok, all threads are patiently blocked, start requeueing */ | |
166 | gettimeofday(&start, NULL); | |
167 | for (nrequeued = 0; nrequeued < nthreads; nrequeued += nrequeue) | |
168 | /* | |
169 | * Do not wakeup any tasks blocked on futex1, allowing | |
170 | * us to really measure futex_wait functionality. | |
171 | */ | |
172 | futex_cmp_requeue(&futex1, 0, &futex2, 0, nrequeue, | |
173 | FUTEX_PRIVATE_FLAG); | |
174 | gettimeofday(&end, NULL); | |
175 | timersub(&end, &start, &runtime); | |
176 | ||
177 | update_stats(&requeued_stats, nrequeued); | |
178 | update_stats(&requeuetime_stats, runtime.tv_usec); | |
179 | ||
180 | if (!silent) { | |
181 | printf("[Run %d]: Requeued %d of %d threads in %.4f ms\n", | |
182 | j + 1, nrequeued, nthreads, runtime.tv_usec/1e3); | |
183 | } | |
184 | ||
185 | /* everybody should be blocked on futex2, wake'em up */ | |
186 | nrequeued = futex_wake(&futex2, nthreads, FUTEX_PRIVATE_FLAG); | |
187 | if (nthreads != nrequeued) | |
188 | warnx("couldn't wakeup all tasks (%d/%d)", nrequeued, nthreads); | |
189 | ||
190 | for (i = 0; i < nthreads; i++) { | |
191 | ret = pthread_join(worker[i], NULL); | |
192 | if (ret) | |
193 | err(EXIT_FAILURE, "pthread_join"); | |
194 | } | |
195 | ||
196 | } | |
197 | ||
198 | /* cleanup & report results */ | |
199 | pthread_cond_destroy(&thread_parent); | |
200 | pthread_cond_destroy(&thread_worker); | |
201 | pthread_mutex_destroy(&thread_lock); | |
202 | pthread_attr_destroy(&thread_attr); | |
203 | ||
204 | print_summary(); | |
205 | ||
206 | free(worker); | |
207 | return ret; | |
208 | err: | |
209 | usage_with_options(bench_futex_requeue_usage, options); | |
210 | exit(EXIT_FAILURE); | |
211 | } |