Loading...
Note: File does not exist in v3.15.
1// SPDX-License-Identifier: GPL-2.0
2#ifdef HAVE_EVENTFD_SUPPORT
3/*
4 * Copyright (C) 2018 Davidlohr Bueso.
5 *
6 * This program benchmarks concurrent epoll_wait(2) monitoring multiple
7 * file descriptors under one or two load balancing models. The first,
8 * and default, is the single/combined queueing (which refers to a single
9 * epoll instance for N worker threads):
10 *
11 * |---> [worker A]
12 * |---> [worker B]
13 * [combined queue] .---> [worker C]
14 * |---> [worker D]
15 * |---> [worker E]
16 *
17 * While the second model, enabled via --multiq option, uses multiple
18 * queueing (which refers to one epoll instance per worker). For example,
19 * short lived tcp connections in a high throughput httpd server will
20 * ditribute the accept()'ing connections across CPUs. In this case each
21 * worker does a limited amount of processing.
22 *
23 * [queue A] ---> [worker]
24 * [queue B] ---> [worker]
25 * [queue C] ---> [worker]
26 * [queue D] ---> [worker]
27 * [queue E] ---> [worker]
28 *
29 * Naturally, the single queue will enforce more concurrency on the epoll
30 * instance, and can therefore scale poorly compared to multiple queues.
31 * However, this is a benchmark raw data and must be taken with a grain of
32 * salt when choosing how to make use of sys_epoll.
33
34 * Each thread has a number of private, nonblocking file descriptors,
35 * referred to as fdmap. A writer thread will constantly be writing to
36 * the fdmaps of all threads, minimizing each threads's chances of
37 * epoll_wait not finding any ready read events and blocking as this
38 * is not what we want to stress. The size of the fdmap can be adjusted
39 * by the user; enlarging the value will increase the chances of
40 * epoll_wait(2) blocking as the lineal writer thread will take "longer",
41 * at least at a high level.
42 *
43 * Note that because fds are private to each thread, this workload does
44 * not stress scenarios where multiple tasks are awoken per ready IO; ie:
45 * EPOLLEXCLUSIVE semantics.
46 *
47 * The end result/metric is throughput: number of ops/second where an
48 * operation consists of:
49 *
50 * epoll_wait(2) + [others]
51 *
52 * ... where [others] is the cost of re-adding the fd (EPOLLET),
53 * or rearming it (EPOLLONESHOT).
54 *
55 *
56 * The purpose of this is program is that it be useful for measuring
57 * kernel related changes to the sys_epoll, and not comparing different
58 * IO polling methods, for example. Hence everything is very adhoc and
59 * outputs raw microbenchmark numbers. Also this uses eventfd, similar
60 * tools tend to use pipes or sockets, but the result is the same.
61 */
62
63/* For the CLR_() macros */
64#include <string.h>
65#include <pthread.h>
66#include <unistd.h>
67
68#include <errno.h>
69#include <inttypes.h>
70#include <signal.h>
71#include <stdlib.h>
72#include <linux/compiler.h>
73#include <linux/kernel.h>
74#include <sys/time.h>
75#include <sys/resource.h>
76#include <sys/epoll.h>
77#include <sys/eventfd.h>
78#include <sys/types.h>
79#include <internal/cpumap.h>
80#include <perf/cpumap.h>
81
82#include "../util/stat.h"
83#include <subcmd/parse-options.h>
84#include "bench.h"
85
86#include <err.h>
87
88#define printinfo(fmt, arg...) \
89 do { if (__verbose) { printf(fmt, ## arg); fflush(stdout); } } while (0)
90
91static unsigned int nthreads = 0;
92static unsigned int nsecs = 8;
93static bool wdone, done, __verbose, randomize, nonblocking;
94
95/*
96 * epoll related shared variables.
97 */
98
99/* Maximum number of nesting allowed inside epoll sets */
100#define EPOLL_MAXNESTS 4
101
102static int epollfd;
103static int *epollfdp;
104static bool noaffinity;
105static unsigned int nested = 0;
106static bool et; /* edge-trigger */
107static bool oneshot;
108static bool multiq; /* use an epoll instance per thread */
109
110/* amount of fds to monitor, per thread */
111static unsigned int nfds = 64;
112
113static pthread_mutex_t thread_lock;
114static unsigned int threads_starting;
115static struct stats throughput_stats;
116static pthread_cond_t thread_parent, thread_worker;
117
118struct worker {
119 int tid;
120 int epollfd; /* for --multiq */
121 pthread_t thread;
122 unsigned long ops;
123 int *fdmap;
124};
125
126static const struct option options[] = {
127 /* general benchmark options */
128 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
129 OPT_UINTEGER('r', "runtime", &nsecs, "Specify runtime (in seconds)"),
130 OPT_UINTEGER('f', "nfds", &nfds, "Specify amount of file descriptors to monitor for each thread"),
131 OPT_BOOLEAN( 'n', "noaffinity", &noaffinity, "Disables CPU affinity"),
132 OPT_BOOLEAN('R', "randomize", &randomize, "Enable random write behaviour (default is lineal)"),
133 OPT_BOOLEAN( 'v', "verbose", &__verbose, "Verbose mode"),
134
135 /* epoll specific options */
136 OPT_BOOLEAN( 'm', "multiq", &multiq, "Use multiple epoll instances (one per thread)"),
137 OPT_BOOLEAN( 'B', "nonblocking", &nonblocking, "Nonblocking epoll_wait(2) behaviour"),
138 OPT_UINTEGER( 'N', "nested", &nested, "Nesting level epoll hierarchy (default is 0, no nesting)"),
139 OPT_BOOLEAN( 'S', "oneshot", &oneshot, "Use EPOLLONESHOT semantics"),
140 OPT_BOOLEAN( 'E', "edge", &et, "Use Edge-triggered interface (default is LT)"),
141
142 OPT_END()
143};
144
145static const char * const bench_epoll_wait_usage[] = {
146 "perf bench epoll wait <options>",
147 NULL
148};
149
150
151/*
152 * Arrange the N elements of ARRAY in random order.
153 * Only effective if N is much smaller than RAND_MAX;
154 * if this may not be the case, use a better random
155 * number generator. -- Ben Pfaff.
156 */
157static void shuffle(void *array, size_t n, size_t size)
158{
159 char *carray = array;
160 void *aux;
161 size_t i;
162
163 if (n <= 1)
164 return;
165
166 aux = calloc(1, size);
167 if (!aux)
168 err(EXIT_FAILURE, "calloc");
169
170 for (i = 1; i < n; ++i) {
171 size_t j = i + rand() / (RAND_MAX / (n - i) + 1);
172 j *= size;
173
174 memcpy(aux, &carray[j], size);
175 memcpy(&carray[j], &carray[i*size], size);
176 memcpy(&carray[i*size], aux, size);
177 }
178
179 free(aux);
180}
181
182
183static void *workerfn(void *arg)
184{
185 int fd, ret, r;
186 struct worker *w = (struct worker *) arg;
187 unsigned long ops = w->ops;
188 struct epoll_event ev;
189 uint64_t val;
190 int to = nonblocking? 0 : -1;
191 int efd = multiq ? w->epollfd : epollfd;
192
193 pthread_mutex_lock(&thread_lock);
194 threads_starting--;
195 if (!threads_starting)
196 pthread_cond_signal(&thread_parent);
197 pthread_cond_wait(&thread_worker, &thread_lock);
198 pthread_mutex_unlock(&thread_lock);
199
200 do {
201 /*
202 * Block undefinitely waiting for the IN event.
203 * In order to stress the epoll_wait(2) syscall,
204 * call it event per event, instead of a larger
205 * batch (max)limit.
206 */
207 do {
208 ret = epoll_wait(efd, &ev, 1, to);
209 } while (ret < 0 && errno == EINTR);
210 if (ret < 0)
211 err(EXIT_FAILURE, "epoll_wait");
212
213 fd = ev.data.fd;
214
215 do {
216 r = read(fd, &val, sizeof(val));
217 } while (!done && (r < 0 && errno == EAGAIN));
218
219 if (et) {
220 ev.events = EPOLLIN | EPOLLET;
221 ret = epoll_ctl(efd, EPOLL_CTL_ADD, fd, &ev);
222 }
223
224 if (oneshot) {
225 /* rearm the file descriptor with a new event mask */
226 ev.events |= EPOLLIN | EPOLLONESHOT;
227 ret = epoll_ctl(efd, EPOLL_CTL_MOD, fd, &ev);
228 }
229
230 ops++;
231 } while (!done);
232
233 if (multiq)
234 close(w->epollfd);
235
236 w->ops = ops;
237 return NULL;
238}
239
240static void nest_epollfd(struct worker *w)
241{
242 unsigned int i;
243 struct epoll_event ev;
244 int efd = multiq ? w->epollfd : epollfd;
245
246 if (nested > EPOLL_MAXNESTS)
247 nested = EPOLL_MAXNESTS;
248
249 epollfdp = calloc(nested, sizeof(*epollfdp));
250 if (!epollfdp)
251 err(EXIT_FAILURE, "calloc");
252
253 for (i = 0; i < nested; i++) {
254 epollfdp[i] = epoll_create(1);
255 if (epollfdp[i] < 0)
256 err(EXIT_FAILURE, "epoll_create");
257 }
258
259 ev.events = EPOLLHUP; /* anything */
260 ev.data.u64 = i; /* any number */
261
262 for (i = nested - 1; i; i--) {
263 if (epoll_ctl(epollfdp[i - 1], EPOLL_CTL_ADD,
264 epollfdp[i], &ev) < 0)
265 err(EXIT_FAILURE, "epoll_ctl");
266 }
267
268 if (epoll_ctl(efd, EPOLL_CTL_ADD, *epollfdp, &ev) < 0)
269 err(EXIT_FAILURE, "epoll_ctl");
270}
271
272static void toggle_done(int sig __maybe_unused,
273 siginfo_t *info __maybe_unused,
274 void *uc __maybe_unused)
275{
276 /* inform all threads that we're done for the day */
277 done = true;
278 gettimeofday(&bench__end, NULL);
279 timersub(&bench__end, &bench__start, &bench__runtime);
280}
281
282static void print_summary(void)
283{
284 unsigned long avg = avg_stats(&throughput_stats);
285 double stddev = stddev_stats(&throughput_stats);
286
287 printf("\nAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
288 avg, rel_stddev_stats(stddev, avg),
289 (int)bench__runtime.tv_sec);
290}
291
292static int do_threads(struct worker *worker, struct perf_cpu_map *cpu)
293{
294 pthread_attr_t thread_attr, *attrp = NULL;
295 cpu_set_t cpuset;
296 unsigned int i, j;
297 int ret = 0, events = EPOLLIN;
298
299 if (oneshot)
300 events |= EPOLLONESHOT;
301 if (et)
302 events |= EPOLLET;
303
304 printinfo("starting worker/consumer %sthreads%s\n",
305 noaffinity ? "":"CPU affinity ",
306 nonblocking ? " (nonblocking)":"");
307 if (!noaffinity)
308 pthread_attr_init(&thread_attr);
309
310 for (i = 0; i < nthreads; i++) {
311 struct worker *w = &worker[i];
312
313 if (multiq) {
314 w->epollfd = epoll_create(1);
315 if (w->epollfd < 0)
316 err(EXIT_FAILURE, "epoll_create");
317
318 if (nested)
319 nest_epollfd(w);
320 }
321
322 w->tid = i;
323 w->fdmap = calloc(nfds, sizeof(int));
324 if (!w->fdmap)
325 return 1;
326
327 for (j = 0; j < nfds; j++) {
328 int efd = multiq ? w->epollfd : epollfd;
329 struct epoll_event ev;
330
331 w->fdmap[j] = eventfd(0, EFD_NONBLOCK);
332 if (w->fdmap[j] < 0)
333 err(EXIT_FAILURE, "eventfd");
334
335 ev.data.fd = w->fdmap[j];
336 ev.events = events;
337
338 ret = epoll_ctl(efd, EPOLL_CTL_ADD,
339 w->fdmap[j], &ev);
340 if (ret < 0)
341 err(EXIT_FAILURE, "epoll_ctl");
342 }
343
344 if (!noaffinity) {
345 CPU_ZERO(&cpuset);
346 CPU_SET(cpu->map[i % cpu->nr], &cpuset);
347
348 ret = pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset);
349 if (ret)
350 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
351
352 attrp = &thread_attr;
353 }
354
355 ret = pthread_create(&w->thread, attrp, workerfn,
356 (void *)(struct worker *) w);
357 if (ret)
358 err(EXIT_FAILURE, "pthread_create");
359 }
360
361 if (!noaffinity)
362 pthread_attr_destroy(&thread_attr);
363
364 return ret;
365}
366
367static void *writerfn(void *p)
368{
369 struct worker *worker = p;
370 size_t i, j, iter;
371 const uint64_t val = 1;
372 ssize_t sz;
373 struct timespec ts = { .tv_sec = 0,
374 .tv_nsec = 500 };
375
376 printinfo("starting writer-thread: doing %s writes ...\n",
377 randomize? "random":"lineal");
378
379 for (iter = 0; !wdone; iter++) {
380 if (randomize) {
381 shuffle((void *)worker, nthreads, sizeof(*worker));
382 }
383
384 for (i = 0; i < nthreads; i++) {
385 struct worker *w = &worker[i];
386
387 if (randomize) {
388 shuffle((void *)w->fdmap, nfds, sizeof(int));
389 }
390
391 for (j = 0; j < nfds; j++) {
392 do {
393 sz = write(w->fdmap[j], &val, sizeof(val));
394 } while (!wdone && (sz < 0 && errno == EAGAIN));
395 }
396 }
397
398 nanosleep(&ts, NULL);
399 }
400
401 printinfo("exiting writer-thread (total full-loops: %zd)\n", iter);
402 return NULL;
403}
404
405static int cmpworker(const void *p1, const void *p2)
406{
407
408 struct worker *w1 = (struct worker *) p1;
409 struct worker *w2 = (struct worker *) p2;
410 return w1->tid > w2->tid;
411}
412
413int bench_epoll_wait(int argc, const char **argv)
414{
415 int ret = 0;
416 struct sigaction act;
417 unsigned int i;
418 struct worker *worker = NULL;
419 struct perf_cpu_map *cpu;
420 pthread_t wthread;
421 struct rlimit rl, prevrl;
422
423 argc = parse_options(argc, argv, options, bench_epoll_wait_usage, 0);
424 if (argc) {
425 usage_with_options(bench_epoll_wait_usage, options);
426 exit(EXIT_FAILURE);
427 }
428
429 memset(&act, 0, sizeof(act));
430 sigfillset(&act.sa_mask);
431 act.sa_sigaction = toggle_done;
432 sigaction(SIGINT, &act, NULL);
433
434 cpu = perf_cpu_map__new(NULL);
435 if (!cpu)
436 goto errmem;
437
438 /* a single, main epoll instance */
439 if (!multiq) {
440 epollfd = epoll_create(1);
441 if (epollfd < 0)
442 err(EXIT_FAILURE, "epoll_create");
443
444 /*
445 * Deal with nested epolls, if any.
446 */
447 if (nested)
448 nest_epollfd(NULL);
449 }
450
451 printinfo("Using %s queue model\n", multiq ? "multi" : "single");
452 printinfo("Nesting level(s): %d\n", nested);
453
454 /* default to the number of CPUs and leave one for the writer pthread */
455 if (!nthreads)
456 nthreads = cpu->nr - 1;
457
458 worker = calloc(nthreads, sizeof(*worker));
459 if (!worker) {
460 goto errmem;
461 }
462
463 if (getrlimit(RLIMIT_NOFILE, &prevrl))
464 err(EXIT_FAILURE, "getrlimit");
465 rl.rlim_cur = rl.rlim_max = nfds * nthreads * 2 + 50;
466 printinfo("Setting RLIMIT_NOFILE rlimit from %" PRIu64 " to: %" PRIu64 "\n",
467 (uint64_t)prevrl.rlim_max, (uint64_t)rl.rlim_max);
468 if (setrlimit(RLIMIT_NOFILE, &rl) < 0)
469 err(EXIT_FAILURE, "setrlimit");
470
471 printf("Run summary [PID %d]: %d threads monitoring%s on "
472 "%d file-descriptors for %d secs.\n\n",
473 getpid(), nthreads, oneshot ? " (EPOLLONESHOT semantics)": "", nfds, nsecs);
474
475 init_stats(&throughput_stats);
476 pthread_mutex_init(&thread_lock, NULL);
477 pthread_cond_init(&thread_parent, NULL);
478 pthread_cond_init(&thread_worker, NULL);
479
480 threads_starting = nthreads;
481
482 gettimeofday(&bench__start, NULL);
483
484 do_threads(worker, cpu);
485
486 pthread_mutex_lock(&thread_lock);
487 while (threads_starting)
488 pthread_cond_wait(&thread_parent, &thread_lock);
489 pthread_cond_broadcast(&thread_worker);
490 pthread_mutex_unlock(&thread_lock);
491
492 /*
493 * At this point the workers should be blocked waiting for read events
494 * to become ready. Launch the writer which will constantly be writing
495 * to each thread's fdmap.
496 */
497 ret = pthread_create(&wthread, NULL, writerfn,
498 (void *)(struct worker *) worker);
499 if (ret)
500 err(EXIT_FAILURE, "pthread_create");
501
502 sleep(nsecs);
503 toggle_done(0, NULL, NULL);
504 printinfo("main thread: toggling done\n");
505
506 sleep(1); /* meh */
507 wdone = true;
508 ret = pthread_join(wthread, NULL);
509 if (ret)
510 err(EXIT_FAILURE, "pthread_join");
511
512 /* cleanup & report results */
513 pthread_cond_destroy(&thread_parent);
514 pthread_cond_destroy(&thread_worker);
515 pthread_mutex_destroy(&thread_lock);
516
517 /* sort the array back before reporting */
518 if (randomize)
519 qsort(worker, nthreads, sizeof(struct worker), cmpworker);
520
521 for (i = 0; i < nthreads; i++) {
522 unsigned long t = bench__runtime.tv_sec > 0 ?
523 worker[i].ops / bench__runtime.tv_sec : 0;
524
525 update_stats(&throughput_stats, t);
526
527 if (nfds == 1)
528 printf("[thread %2d] fdmap: %p [ %04ld ops/sec ]\n",
529 worker[i].tid, &worker[i].fdmap[0], t);
530 else
531 printf("[thread %2d] fdmap: %p ... %p [ %04ld ops/sec ]\n",
532 worker[i].tid, &worker[i].fdmap[0],
533 &worker[i].fdmap[nfds-1], t);
534 }
535
536 print_summary();
537
538 close(epollfd);
539 return ret;
540errmem:
541 err(EXIT_FAILURE, "calloc");
542}
543#endif // HAVE_EVENTFD_SUPPORT