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 * distribute 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 <perf/cpumap.h>
80
81 #include "../util/stat.h"
82 #include <subcmd/parse-options.h>
83 #include "bench.h"
84
85 #include <err.h>
86
87 #define printinfo(fmt, arg...) \
88 do { if (__verbose) { printf(fmt, ## arg); fflush(stdout); } } while (0)
89
90 static unsigned int nthreads = 0;
91 static unsigned int nsecs = 8;
92 static bool wdone, done, __verbose, randomize, nonblocking;
93
94 /*
95 * epoll related shared variables.
96 */
97
98 /* Maximum number of nesting allowed inside epoll sets */
99 #define EPOLL_MAXNESTS 4
100
101 static int epollfd;
102 static int *epollfdp;
103 static bool noaffinity;
104 static unsigned int nested = 0;
105 static bool et; /* edge-trigger */
106 static bool oneshot;
107 static bool multiq; /* use an epoll instance per thread */
108
109 /* amount of fds to monitor, per thread */
110 static unsigned int nfds = 64;
111
112 static pthread_mutex_t thread_lock;
113 static unsigned int threads_starting;
114 static struct stats throughput_stats;
115 static pthread_cond_t thread_parent, thread_worker;
116
117 struct worker {
118 int tid;
119 int epollfd; /* for --multiq */
120 pthread_t thread;
121 unsigned long ops;
122 int *fdmap;
123 };
124
125 static const struct option options[] = {
126 /* general benchmark options */
127 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
128 OPT_UINTEGER('r', "runtime", &nsecs, "Specify runtime (in seconds)"),
129 OPT_UINTEGER('f', "nfds", &nfds, "Specify amount of file descriptors to monitor for each thread"),
130 OPT_BOOLEAN( 'n', "noaffinity", &noaffinity, "Disables CPU affinity"),
131 OPT_BOOLEAN('R', "randomize", &randomize, "Enable random write behaviour (default is lineal)"),
132 OPT_BOOLEAN( 'v', "verbose", &__verbose, "Verbose mode"),
133
134 /* epoll specific options */
135 OPT_BOOLEAN( 'm', "multiq", &multiq, "Use multiple epoll instances (one per thread)"),
136 OPT_BOOLEAN( 'B', "nonblocking", &nonblocking, "Nonblocking epoll_wait(2) behaviour"),
137 OPT_UINTEGER( 'N', "nested", &nested, "Nesting level epoll hierarchy (default is 0, no nesting)"),
138 OPT_BOOLEAN( 'S', "oneshot", &oneshot, "Use EPOLLONESHOT semantics"),
139 OPT_BOOLEAN( 'E', "edge", &et, "Use Edge-triggered interface (default is LT)"),
140
141 OPT_END()
142 };
143
144 static const char * const bench_epoll_wait_usage[] = {
145 "perf bench epoll wait <options>",
146 NULL
147 };
148
149
150 /*
151 * Arrange the N elements of ARRAY in random order.
152 * Only effective if N is much smaller than RAND_MAX;
153 * if this may not be the case, use a better random
154 * number generator. -- Ben Pfaff.
155 */
shuffle(void * array,size_t n,size_t size)156 static void shuffle(void *array, size_t n, size_t size)
157 {
158 char *carray = array;
159 void *aux;
160 size_t i;
161
162 if (n <= 1)
163 return;
164
165 aux = calloc(1, size);
166 if (!aux)
167 err(EXIT_FAILURE, "calloc");
168
169 for (i = 1; i < n; ++i) {
170 size_t j = i + rand() / (RAND_MAX / (n - i) + 1);
171 j *= size;
172
173 memcpy(aux, &carray[j], size);
174 memcpy(&carray[j], &carray[i*size], size);
175 memcpy(&carray[i*size], aux, size);
176 }
177
178 free(aux);
179 }
180
181
workerfn(void * arg)182 static void *workerfn(void *arg)
183 {
184 int fd, ret, r;
185 struct worker *w = (struct worker *) arg;
186 unsigned long ops = w->ops;
187 struct epoll_event ev;
188 uint64_t val;
189 int to = nonblocking? 0 : -1;
190 int efd = multiq ? w->epollfd : epollfd;
191
192 pthread_mutex_lock(&thread_lock);
193 threads_starting--;
194 if (!threads_starting)
195 pthread_cond_signal(&thread_parent);
196 pthread_cond_wait(&thread_worker, &thread_lock);
197 pthread_mutex_unlock(&thread_lock);
198
199 do {
200 /*
201 * Block indefinitely waiting for the IN event.
202 * In order to stress the epoll_wait(2) syscall,
203 * call it event per event, instead of a larger
204 * batch (max)limit.
205 */
206 do {
207 ret = epoll_wait(efd, &ev, 1, to);
208 } while (ret < 0 && errno == EINTR);
209 if (ret < 0)
210 err(EXIT_FAILURE, "epoll_wait");
211
212 fd = ev.data.fd;
213
214 do {
215 r = read(fd, &val, sizeof(val));
216 } while (!done && (r < 0 && errno == EAGAIN));
217
218 if (et) {
219 ev.events = EPOLLIN | EPOLLET;
220 ret = epoll_ctl(efd, EPOLL_CTL_ADD, fd, &ev);
221 }
222
223 if (oneshot) {
224 /* rearm the file descriptor with a new event mask */
225 ev.events |= EPOLLIN | EPOLLONESHOT;
226 ret = epoll_ctl(efd, EPOLL_CTL_MOD, fd, &ev);
227 }
228
229 ops++;
230 } while (!done);
231
232 if (multiq)
233 close(w->epollfd);
234
235 w->ops = ops;
236 return NULL;
237 }
238
nest_epollfd(struct worker * w)239 static void nest_epollfd(struct worker *w)
240 {
241 unsigned int i;
242 struct epoll_event ev;
243 int efd = multiq ? w->epollfd : epollfd;
244
245 if (nested > EPOLL_MAXNESTS)
246 nested = EPOLL_MAXNESTS;
247
248 epollfdp = calloc(nested, sizeof(*epollfdp));
249 if (!epollfdp)
250 err(EXIT_FAILURE, "calloc");
251
252 for (i = 0; i < nested; i++) {
253 epollfdp[i] = epoll_create(1);
254 if (epollfdp[i] < 0)
255 err(EXIT_FAILURE, "epoll_create");
256 }
257
258 ev.events = EPOLLHUP; /* anything */
259 ev.data.u64 = i; /* any number */
260
261 for (i = nested - 1; i; i--) {
262 if (epoll_ctl(epollfdp[i - 1], EPOLL_CTL_ADD,
263 epollfdp[i], &ev) < 0)
264 err(EXIT_FAILURE, "epoll_ctl");
265 }
266
267 if (epoll_ctl(efd, EPOLL_CTL_ADD, *epollfdp, &ev) < 0)
268 err(EXIT_FAILURE, "epoll_ctl");
269 }
270
toggle_done(int sig __maybe_unused,siginfo_t * info __maybe_unused,void * uc __maybe_unused)271 static void toggle_done(int sig __maybe_unused,
272 siginfo_t *info __maybe_unused,
273 void *uc __maybe_unused)
274 {
275 /* inform all threads that we're done for the day */
276 done = true;
277 gettimeofday(&bench__end, NULL);
278 timersub(&bench__end, &bench__start, &bench__runtime);
279 }
280
print_summary(void)281 static void print_summary(void)
282 {
283 unsigned long avg = avg_stats(&throughput_stats);
284 double stddev = stddev_stats(&throughput_stats);
285
286 printf("\nAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
287 avg, rel_stddev_stats(stddev, avg),
288 (int)bench__runtime.tv_sec);
289 }
290
do_threads(struct worker * worker,struct perf_cpu_map * cpu)291 static int do_threads(struct worker *worker, struct perf_cpu_map *cpu)
292 {
293 pthread_attr_t thread_attr, *attrp = NULL;
294 cpu_set_t cpuset;
295 unsigned int i, j;
296 int ret = 0, events = EPOLLIN;
297
298 if (oneshot)
299 events |= EPOLLONESHOT;
300 if (et)
301 events |= EPOLLET;
302
303 printinfo("starting worker/consumer %sthreads%s\n",
304 noaffinity ? "":"CPU affinity ",
305 nonblocking ? " (nonblocking)":"");
306 if (!noaffinity)
307 pthread_attr_init(&thread_attr);
308
309 for (i = 0; i < nthreads; i++) {
310 struct worker *w = &worker[i];
311
312 if (multiq) {
313 w->epollfd = epoll_create(1);
314 if (w->epollfd < 0)
315 err(EXIT_FAILURE, "epoll_create");
316
317 if (nested)
318 nest_epollfd(w);
319 }
320
321 w->tid = i;
322 w->fdmap = calloc(nfds, sizeof(int));
323 if (!w->fdmap)
324 return 1;
325
326 for (j = 0; j < nfds; j++) {
327 int efd = multiq ? w->epollfd : epollfd;
328 struct epoll_event ev;
329
330 w->fdmap[j] = eventfd(0, EFD_NONBLOCK);
331 if (w->fdmap[j] < 0)
332 err(EXIT_FAILURE, "eventfd");
333
334 ev.data.fd = w->fdmap[j];
335 ev.events = events;
336
337 ret = epoll_ctl(efd, EPOLL_CTL_ADD,
338 w->fdmap[j], &ev);
339 if (ret < 0)
340 err(EXIT_FAILURE, "epoll_ctl");
341 }
342
343 if (!noaffinity) {
344 CPU_ZERO(&cpuset);
345 CPU_SET(cpu->map[i % cpu->nr], &cpuset);
346
347 ret = pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset);
348 if (ret)
349 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
350
351 attrp = &thread_attr;
352 }
353
354 ret = pthread_create(&w->thread, attrp, workerfn,
355 (void *)(struct worker *) w);
356 if (ret)
357 err(EXIT_FAILURE, "pthread_create");
358 }
359
360 if (!noaffinity)
361 pthread_attr_destroy(&thread_attr);
362
363 return ret;
364 }
365
writerfn(void * p)366 static void *writerfn(void *p)
367 {
368 struct worker *worker = p;
369 size_t i, j, iter;
370 const uint64_t val = 1;
371 ssize_t sz;
372 struct timespec ts = { .tv_sec = 0,
373 .tv_nsec = 500 };
374
375 printinfo("starting writer-thread: doing %s writes ...\n",
376 randomize? "random":"lineal");
377
378 for (iter = 0; !wdone; iter++) {
379 if (randomize) {
380 shuffle((void *)worker, nthreads, sizeof(*worker));
381 }
382
383 for (i = 0; i < nthreads; i++) {
384 struct worker *w = &worker[i];
385
386 if (randomize) {
387 shuffle((void *)w->fdmap, nfds, sizeof(int));
388 }
389
390 for (j = 0; j < nfds; j++) {
391 do {
392 sz = write(w->fdmap[j], &val, sizeof(val));
393 } while (!wdone && (sz < 0 && errno == EAGAIN));
394 }
395 }
396
397 nanosleep(&ts, NULL);
398 }
399
400 printinfo("exiting writer-thread (total full-loops: %zd)\n", iter);
401 return NULL;
402 }
403
cmpworker(const void * p1,const void * p2)404 static int cmpworker(const void *p1, const void *p2)
405 {
406
407 struct worker *w1 = (struct worker *) p1;
408 struct worker *w2 = (struct worker *) p2;
409 return w1->tid > w2->tid;
410 }
411
bench_epoll_wait(int argc,const char ** argv)412 int bench_epoll_wait(int argc, const char **argv)
413 {
414 int ret = 0;
415 struct sigaction act;
416 unsigned int i;
417 struct worker *worker = NULL;
418 struct perf_cpu_map *cpu;
419 pthread_t wthread;
420 struct rlimit rl, prevrl;
421
422 argc = parse_options(argc, argv, options, bench_epoll_wait_usage, 0);
423 if (argc) {
424 usage_with_options(bench_epoll_wait_usage, options);
425 exit(EXIT_FAILURE);
426 }
427
428 memset(&act, 0, sizeof(act));
429 sigfillset(&act.sa_mask);
430 act.sa_sigaction = toggle_done;
431 sigaction(SIGINT, &act, NULL);
432
433 cpu = perf_cpu_map__new(NULL);
434 if (!cpu)
435 goto errmem;
436
437 /* a single, main epoll instance */
438 if (!multiq) {
439 epollfd = epoll_create(1);
440 if (epollfd < 0)
441 err(EXIT_FAILURE, "epoll_create");
442
443 /*
444 * Deal with nested epolls, if any.
445 */
446 if (nested)
447 nest_epollfd(NULL);
448 }
449
450 printinfo("Using %s queue model\n", multiq ? "multi" : "single");
451 printinfo("Nesting level(s): %d\n", nested);
452
453 /* default to the number of CPUs and leave one for the writer pthread */
454 if (!nthreads)
455 nthreads = cpu->nr - 1;
456
457 worker = calloc(nthreads, sizeof(*worker));
458 if (!worker) {
459 goto errmem;
460 }
461
462 if (getrlimit(RLIMIT_NOFILE, &prevrl))
463 err(EXIT_FAILURE, "getrlimit");
464 rl.rlim_cur = rl.rlim_max = nfds * nthreads * 2 + 50;
465 printinfo("Setting RLIMIT_NOFILE rlimit from %" PRIu64 " to: %" PRIu64 "\n",
466 (uint64_t)prevrl.rlim_max, (uint64_t)rl.rlim_max);
467 if (setrlimit(RLIMIT_NOFILE, &rl) < 0)
468 err(EXIT_FAILURE, "setrlimit");
469
470 printf("Run summary [PID %d]: %d threads monitoring%s on "
471 "%d file-descriptors for %d secs.\n\n",
472 getpid(), nthreads, oneshot ? " (EPOLLONESHOT semantics)": "", nfds, nsecs);
473
474 init_stats(&throughput_stats);
475 pthread_mutex_init(&thread_lock, NULL);
476 pthread_cond_init(&thread_parent, NULL);
477 pthread_cond_init(&thread_worker, NULL);
478
479 threads_starting = nthreads;
480
481 gettimeofday(&bench__start, NULL);
482
483 do_threads(worker, cpu);
484
485 pthread_mutex_lock(&thread_lock);
486 while (threads_starting)
487 pthread_cond_wait(&thread_parent, &thread_lock);
488 pthread_cond_broadcast(&thread_worker);
489 pthread_mutex_unlock(&thread_lock);
490
491 /*
492 * At this point the workers should be blocked waiting for read events
493 * to become ready. Launch the writer which will constantly be writing
494 * to each thread's fdmap.
495 */
496 ret = pthread_create(&wthread, NULL, writerfn,
497 (void *)(struct worker *) worker);
498 if (ret)
499 err(EXIT_FAILURE, "pthread_create");
500
501 sleep(nsecs);
502 toggle_done(0, NULL, NULL);
503 printinfo("main thread: toggling done\n");
504
505 sleep(1); /* meh */
506 wdone = true;
507 ret = pthread_join(wthread, NULL);
508 if (ret)
509 err(EXIT_FAILURE, "pthread_join");
510
511 /* cleanup & report results */
512 pthread_cond_destroy(&thread_parent);
513 pthread_cond_destroy(&thread_worker);
514 pthread_mutex_destroy(&thread_lock);
515
516 /* sort the array back before reporting */
517 if (randomize)
518 qsort(worker, nthreads, sizeof(struct worker), cmpworker);
519
520 for (i = 0; i < nthreads; i++) {
521 unsigned long t = bench__runtime.tv_sec > 0 ?
522 worker[i].ops / bench__runtime.tv_sec : 0;
523
524 update_stats(&throughput_stats, t);
525
526 if (nfds == 1)
527 printf("[thread %2d] fdmap: %p [ %04ld ops/sec ]\n",
528 worker[i].tid, &worker[i].fdmap[0], t);
529 else
530 printf("[thread %2d] fdmap: %p ... %p [ %04ld ops/sec ]\n",
531 worker[i].tid, &worker[i].fdmap[0],
532 &worker[i].fdmap[nfds-1], t);
533 }
534
535 print_summary();
536
537 close(epollfd);
538 return ret;
539 errmem:
540 err(EXIT_FAILURE, "calloc");
541 }
542 #endif // HAVE_EVENTFD_SUPPORT
543