1 // SPDX-License-Identifier: GPL-2.0
6 * messaging: Benchmark for scheduler and IPC mechanisms
8 * Based on hackbench by Rusty Russell <rusty@rustcorp.com.au>
9 * Ported to perf by Hitoshi Mitake <mitake@dcl.info.waseda.ac.jp>
14 #include "../util/util.h"
15 #include <subcmd/parse-options.h>
16 #include "../builtin.h"
19 /* Test groups of 20 processes spraying to 20 receivers */
26 #include <sys/types.h>
27 #include <sys/socket.h>
33 #include <linux/time64.h>
37 static bool use_pipes = false;
38 static unsigned int nr_loops = 100;
39 static bool thread_mode = false;
40 static unsigned int num_groups = 10;
42 struct sender_context {
49 struct receiver_context {
50 unsigned int num_packets;
56 static void fdpair(int fds[2])
62 if (socketpair(AF_UNIX, SOCK_STREAM, 0, fds) == 0)
66 err(EXIT_FAILURE, use_pipes ? "pipe()" : "socketpair()");
69 /* Block until we're ready to go */
70 static void ready(int ready_out, int wakefd)
73 struct pollfd pollfd = { .fd = wakefd, .events = POLLIN };
75 /* Tell them we're ready. */
76 if (write(ready_out, &dummy, 1) != 1)
77 err(EXIT_FAILURE, "CLIENT: ready write");
79 /* Wait for "GO" signal */
80 if (poll(&pollfd, 1, -1) != 1)
81 err(EXIT_FAILURE, "poll");
84 /* Sender sprays nr_loops messages down each file descriptor */
85 static void *sender(struct sender_context *ctx)
90 ready(ctx->ready_out, ctx->wakefd);
92 /* Now pump to every receiver. */
93 for (i = 0; i < nr_loops; i++) {
94 for (j = 0; j < ctx->num_fds; j++) {
98 ret = write(ctx->out_fds[j], data + done,
101 err(EXIT_FAILURE, "SENDER: write");
112 /* One receiver per fd */
113 static void *receiver(struct receiver_context* ctx)
118 close(ctx->in_fds[1]);
120 /* Wait for start... */
121 ready(ctx->ready_out, ctx->wakefd);
123 /* Receive them all */
124 for (i = 0; i < ctx->num_packets; i++) {
129 ret = read(ctx->in_fds[0], data + done, DATASIZE - done);
131 err(EXIT_FAILURE, "SERVER: read");
140 static pthread_t create_worker(void *ctx, void *(*func)(void *))
148 /* Fork the receiver. */
151 err(EXIT_FAILURE, "fork()");
164 if (pthread_attr_init(&attr) != 0)
165 err(EXIT_FAILURE, "pthread_attr_init:");
168 if (pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN) != 0)
169 err(EXIT_FAILURE, "pthread_attr_setstacksize");
172 ret = pthread_create(&childid, &attr, func, ctx);
174 err(EXIT_FAILURE, "pthread_create failed");
179 static void reap_worker(pthread_t id)
187 if (!WIFEXITED(proc_status))
190 pthread_join(id, &thread_status);
194 /* One group of senders and receivers */
195 static unsigned int group(pthread_t *pth,
196 unsigned int num_fds,
201 struct sender_context *snd_ctx = malloc(sizeof(struct sender_context)
202 + num_fds * sizeof(int));
205 err(EXIT_FAILURE, "malloc()");
207 for (i = 0; i < num_fds; i++) {
209 struct receiver_context *ctx = malloc(sizeof(*ctx));
212 err(EXIT_FAILURE, "malloc()");
215 /* Create the pipe between client and server */
218 ctx->num_packets = num_fds * nr_loops;
219 ctx->in_fds[0] = fds[0];
220 ctx->in_fds[1] = fds[1];
221 ctx->ready_out = ready_out;
222 ctx->wakefd = wakefd;
224 pth[i] = create_worker(ctx, (void *)receiver);
226 snd_ctx->out_fds[i] = fds[1];
231 /* Now we have all the fds, fork the senders */
232 for (i = 0; i < num_fds; i++) {
233 snd_ctx->ready_out = ready_out;
234 snd_ctx->wakefd = wakefd;
235 snd_ctx->num_fds = num_fds;
237 pth[num_fds+i] = create_worker(snd_ctx, (void *)sender);
240 /* Close the fds we have left */
242 for (i = 0; i < num_fds; i++)
243 close(snd_ctx->out_fds[i]);
245 /* Return number of children to reap */
249 static const struct option options[] = {
250 OPT_BOOLEAN('p', "pipe", &use_pipes,
251 "Use pipe() instead of socketpair()"),
252 OPT_BOOLEAN('t', "thread", &thread_mode,
253 "Be multi thread instead of multi process"),
254 OPT_UINTEGER('g', "group", &num_groups, "Specify number of groups"),
255 OPT_UINTEGER('l', "nr_loops", &nr_loops, "Specify the number of loops to run (default: 100)"),
259 static const char * const bench_sched_message_usage[] = {
260 "perf bench sched messaging <options>",
264 int bench_sched_messaging(int argc, const char **argv)
266 unsigned int i, total_children;
267 struct timeval start, stop, diff;
268 unsigned int num_fds = 20;
269 int readyfds[2], wakefds[2];
273 argc = parse_options(argc, argv, options,
274 bench_sched_message_usage, 0);
276 pth_tab = malloc(num_fds * 2 * num_groups * sizeof(pthread_t));
278 err(EXIT_FAILURE, "main:malloc()");
284 for (i = 0; i < num_groups; i++)
285 total_children += group(pth_tab+total_children, num_fds,
286 readyfds[1], wakefds[0]);
288 /* Wait for everyone to be ready */
289 for (i = 0; i < total_children; i++)
290 if (read(readyfds[0], &dummy, 1) != 1)
291 err(EXIT_FAILURE, "Reading for readyfds");
293 gettimeofday(&start, NULL);
296 if (write(wakefds[1], &dummy, 1) != 1)
297 err(EXIT_FAILURE, "Writing to start them");
300 for (i = 0; i < total_children; i++)
301 reap_worker(pth_tab[i]);
303 gettimeofday(&stop, NULL);
305 timersub(&stop, &start, &diff);
307 switch (bench_format) {
308 case BENCH_FORMAT_DEFAULT:
309 printf("# %d sender and receiver %s per group\n",
310 num_fds, thread_mode ? "threads" : "processes");
311 printf("# %d groups == %d %s run\n\n",
312 num_groups, num_groups * 2 * num_fds,
313 thread_mode ? "threads" : "processes");
314 printf(" %14s: %lu.%03lu [sec]\n", "Total time",
316 (unsigned long) (diff.tv_usec / USEC_PER_MSEC));
318 case BENCH_FORMAT_SIMPLE:
319 printf("%lu.%03lu\n", diff.tv_sec,
320 (unsigned long) (diff.tv_usec / USEC_PER_MSEC));
323 /* reaching here is something disaster */
324 fprintf(stderr, "Unknown format:%d\n", bench_format);