summaryrefslogtreecommitdiffstats
path: root/tools/perf/bench/epoll-wait.c
blob: 75dca9773186245b3d79ce02a33e613a5b4c71f1 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
// SPDX-License-Identifier: GPL-2.0
#ifdef HAVE_EVENTFD_SUPPORT
/*
 * Copyright (C) 2018 Davidlohr Bueso.
 *
 * This program benchmarks concurrent epoll_wait(2) monitoring multiple
 * file descriptors under one or two load balancing models. The first,
 * and default, is the single/combined queueing (which refers to a single
 * epoll instance for N worker threads):
 *
 *                          |---> [worker A]
 *                          |---> [worker B]
 *        [combined queue]  .---> [worker C]
 *                          |---> [worker D]
 *                          |---> [worker E]
 *
 * While the second model, enabled via --multiq option, uses multiple
 * queueing (which refers to one epoll instance per worker). For example,
 * short lived tcp connections in a high throughput httpd server will
 * ditribute the accept()'ing  connections across CPUs. In this case each
 * worker does a limited  amount of processing.
 *
 *             [queue A]  ---> [worker]
 *             [queue B]  ---> [worker]
 *             [queue C]  ---> [worker]
 *             [queue D]  ---> [worker]
 *             [queue E]  ---> [worker]
 *
 * Naturally, the single queue will enforce more concurrency on the epoll
 * instance, and can therefore scale poorly compared to multiple queues.
 * However, this is a benchmark raw data and must be taken with a grain of
 * salt when choosing how to make use of sys_epoll.

 * Each thread has a number of private, nonblocking file descriptors,
 * referred to as fdmap. A writer thread will constantly be writing to
 * the fdmaps of all threads, minimizing each threads's chances of
 * epoll_wait not finding any ready read events and blocking as this
 * is not what we want to stress. The size of the fdmap can be adjusted
 * by the user; enlarging the value will increase the chances of
 * epoll_wait(2) blocking as the lineal writer thread will take "longer",
 * at least at a high level.
 *
 * Note that because fds are private to each thread, this workload does
 * not stress scenarios where multiple tasks are awoken per ready IO; ie:
 * EPOLLEXCLUSIVE semantics.
 *
 * The end result/metric is throughput: number of ops/second where an
 * operation consists of:
 *
 *   epoll_wait(2) + [others]
 *
 *        ... where [others] is the cost of re-adding the fd (EPOLLET),
 *            or rearming it (EPOLLONESHOT).
 *
 *
 * The purpose of this is program is that it be useful for measuring
 * kernel related changes to the sys_epoll, and not comparing different
 * IO polling methods, for example. Hence everything is very adhoc and
 * outputs raw microbenchmark numbers. Also this uses eventfd, similar
 * tools tend to use pipes or sockets, but the result is the same.
 */

/* For the CLR_() macros */
#include <string.h>
#include <pthread.h>
#include <unistd.h>

#include <errno.h>
#include <inttypes.h>
#include <signal.h>
#include <stdlib.h>
#include <linux/compiler.h>
#include <linux/kernel.h>
#include <sys/time.h>
#include <sys/resource.h>
#include <sys/epoll.h>
#include <sys/eventfd.h>
#include <sys/types.h>
#include <internal/cpumap.h>
#include <perf/cpumap.h>

#include "../util/stat.h"
#include <subcmd/parse-options.h>
#include "bench.h"

#include <err.h>

#define printinfo(fmt, arg...) \
	do { if (__verbose) { printf(fmt, ## arg); fflush(stdout); } } while (0)

static unsigned int nthreads = 0;
static unsigned int nsecs    = 8;
static bool wdone, done, __verbose, randomize, nonblocking;

/*
 * epoll related shared variables.
 */

/* Maximum number of nesting allowed inside epoll sets */
#define EPOLL_MAXNESTS 4

static int epollfd;
static int *epollfdp;
static bool noaffinity;
static unsigned int nested = 0;
static bool et; /* edge-trigger */
static bool oneshot;
static bool multiq; /* use an epoll instance per thread */

/* amount of fds to monitor, per thread */
static unsigned int nfds = 64;

static pthread_mutex_t thread_lock;
static unsigned int threads_starting;
static struct stats throughput_stats;
static pthread_cond_t thread_parent, thread_worker;

struct worker {
	int tid;
	int epollfd; /* for --multiq */
	pthread_t thread;
	unsigned long ops;
	int *fdmap;
};

static const struct option options[] = {
	/* general benchmark options */
	OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
	OPT_UINTEGER('r', "runtime", &nsecs, "Specify runtime (in seconds)"),
	OPT_UINTEGER('f', "nfds",    &nfds,  "Specify amount of file descriptors to monitor for each thread"),
	OPT_BOOLEAN( 'n', "noaffinity",  &noaffinity,   "Disables CPU affinity"),
	OPT_BOOLEAN('R', "randomize", &randomize,   "Enable random write behaviour (default is lineal)"),
	OPT_BOOLEAN( 'v', "verbose", &__verbose, "Verbose mode"),

	/* epoll specific options */
	OPT_BOOLEAN( 'm', "multiq",  &multiq,   "Use multiple epoll instances (one per thread)"),
	OPT_BOOLEAN( 'B', "nonblocking", &nonblocking, "Nonblocking epoll_wait(2) behaviour"),
	OPT_UINTEGER( 'N', "nested",  &nested,   "Nesting level epoll hierarchy (default is 0, no nesting)"),
	OPT_BOOLEAN( 'S', "oneshot",  &oneshot,   "Use EPOLLONESHOT semantics"),
	OPT_BOOLEAN( 'E', "edge",  &et,   "Use Edge-triggered interface (default is LT)"),

	OPT_END()
};

static const char * const bench_epoll_wait_usage[] = {
	"perf bench epoll wait <options>",
	NULL
};


/*
 * Arrange the N elements of ARRAY in random order.
 * Only effective if N is much smaller than RAND_MAX;
 * if this may not be the case, use a better random
 * number generator. -- Ben Pfaff.
 */
static void shuffle(void *array, size_t n, size_t size)
{
	char *carray = array;
	void *aux;
	size_t i;

	if (n <= 1)
		return;

	aux = calloc(1, size);
	if (!aux)
		err(EXIT_FAILURE, "calloc");

	for (i = 1; i < n; ++i) {
		size_t j =   i + rand() / (RAND_MAX / (n - i) + 1);
		j *= size;

		memcpy(aux, &carray[j], size);
		memcpy(&carray[j], &carray[i*size], size);
		memcpy(&carray[i*size], aux, size);
	}

	free(aux);
}


static void *workerfn(void *arg)
{
	int fd, ret, r;
	struct worker *w = (struct worker *) arg;
	unsigned long ops = w->ops;
	struct epoll_event ev;
	uint64_t val;
	int to = nonblocking? 0 : -1;
	int efd = multiq ? w->epollfd : epollfd;

	pthread_mutex_lock(&thread_lock);
	threads_starting--;
	if (!threads_starting)
		pthread_cond_signal(&thread_parent);
	pthread_cond_wait(&thread_worker, &thread_lock);
	pthread_mutex_unlock(&thread_lock);

	do {
		/*
		 * Block undefinitely waiting for the IN event.
		 * In order to stress the epoll_wait(2) syscall,
		 * call it event per event, instead of a larger
		 * batch (max)limit.
		 */
		do {
			ret = epoll_wait(efd, &ev, 1, to);
		} while (ret < 0 && errno == EINTR);
		if (ret < 0)
			err(EXIT_FAILURE, "epoll_wait");

		fd = ev.data.fd;

		do {
			r = read(fd, &val, sizeof(val));
		} while (!done && (r < 0 && errno == EAGAIN));

		if (et) {
			ev.events = EPOLLIN | EPOLLET;
			ret = epoll_ctl(efd, EPOLL_CTL_ADD, fd, &ev);
		}

		if (oneshot) {
			/* rearm the file descriptor with a new event mask */
			ev.events |= EPOLLIN | EPOLLONESHOT;
			ret = epoll_ctl(efd, EPOLL_CTL_MOD, fd, &ev);
		}

		ops++;
	}  while (!done);

	if (multiq)
		close(w->epollfd);

	w->ops = ops;
	return NULL;
}

static void nest_epollfd(struct worker *w)
{
	unsigned int i;
	struct epoll_event ev;
	int efd = multiq ? w->epollfd : epollfd;

	if (nested > EPOLL_MAXNESTS)
		nested = EPOLL_MAXNESTS;

	epollfdp = calloc(nested, sizeof(*epollfdp));
	if (!epollfdp)
		err(EXIT_FAILURE, "calloc");

	for (i = 0; i < nested; i++) {
		epollfdp[i] = epoll_create(1);
		if (epollfdp[i] < 0)
			err(EXIT_FAILURE, "epoll_create");
	}

	ev.events = EPOLLHUP; /* anything */
	ev.data.u64 = i; /* any number */

	for (i = nested - 1; i; i--) {
		if (epoll_ctl(epollfdp[i - 1], EPOLL_CTL_ADD,
			      epollfdp[i], &ev) < 0)
			err(EXIT_FAILURE, "epoll_ctl");
	}

	if (epoll_ctl(efd, EPOLL_CTL_ADD, *epollfdp, &ev) < 0)
		err(EXIT_FAILURE, "epoll_ctl");
}

static void toggle_done(int sig __maybe_unused,
			siginfo_t *info __maybe_unused,
			void *uc __maybe_unused)
{
	/* inform all threads that we're done for the day */
	done = true;
	gettimeofday(&bench__end, NULL);
	timersub(&bench__end, &bench__start, &bench__runtime);
}

static void print_summary(void)
{
	unsigned long avg = avg_stats(&throughput_stats);
	double stddev = stddev_stats(&throughput_stats);

	printf("\nAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
	       avg, rel_stddev_stats(stddev, avg),
	       (int)bench__runtime.tv_sec);
}

static int do_threads(struct worker *worker, struct perf_cpu_map *cpu)
{
	pthread_attr_t thread_attr, *attrp = NULL;
	cpu_set_t cpuset;
	unsigned int i, j;
	int ret = 0, events = EPOLLIN;

	if (oneshot)
		events |= EPOLLONESHOT;
	if (et)
		events |= EPOLLET;

	printinfo("starting worker/consumer %sthreads%s\n",
		  noaffinity ?  "":"CPU affinity ",
		  nonblocking ? " (nonblocking)":"");
	if (!noaffinity)
		pthread_attr_init(&thread_attr);

	for (i = 0; i < nthreads; i++) {
		struct worker *w = &worker[i];

		if (multiq) {
			w->epollfd = epoll_create(1);
			if (w->epollfd < 0)
				err(EXIT_FAILURE, "epoll_create");

			if (nested)
				nest_epollfd(w);
		}

		w->tid = i;
		w->fdmap = calloc(nfds, sizeof(int));
		if (!w->fdmap)
			return 1;

		for (j = 0; j < nfds; j++) {
			int efd = multiq ? w->epollfd : epollfd;
			struct epoll_event ev;

			w->fdmap[j] = eventfd(0, EFD_NONBLOCK);
			if (w->fdmap[j] < 0)
				err(EXIT_FAILURE, "eventfd");

			ev.data.fd = w->fdmap[j];
			ev.events = events;

			ret = epoll_ctl(efd, EPOLL_CTL_ADD,
					w->fdmap[j], &ev);
			if (ret < 0)
				err(EXIT_FAILURE, "epoll_ctl");
		}

		if (!noaffinity) {
			CPU_ZERO(&cpuset);
			CPU_SET(cpu->map[i % cpu->nr], &cpuset);

			ret = pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset);
			if (ret)
				err(EXIT_FAILURE, "pthread_attr_setaffinity_np");

			attrp = &thread_attr;
		}

		ret = pthread_create(&w->thread, attrp, workerfn,
				     (void *)(struct worker *) w);
		if (ret)
			err(EXIT_FAILURE, "pthread_create");
	}

	if (!noaffinity)
		pthread_attr_destroy(&thread_attr);

	return ret;
}

static void *writerfn(void *p)
{
	struct worker *worker = p;
	size_t i, j, iter;
	const uint64_t val = 1;
	ssize_t sz;
	struct timespec ts = { .tv_sec = 0,
			       .tv_nsec = 500 };

	printinfo("starting writer-thread: doing %s writes ...\n",
		  randomize? "random":"lineal");

	for (iter = 0; !wdone; iter++) {
		if (randomize) {
			shuffle((void *)worker, nthreads, sizeof(*worker));
		}

		for (i = 0; i < nthreads; i++) {
			struct worker *w = &worker[i];

			if (randomize) {
				shuffle((void *)w->fdmap, nfds, sizeof(int));
			}

			for (j = 0; j < nfds; j++) {
				do {
					sz = write(w->fdmap[j], &val, sizeof(val));
				} while (!wdone && (sz < 0 && errno == EAGAIN));
			}
		}

		nanosleep(&ts, NULL);
	}

	printinfo("exiting writer-thread (total full-loops: %zd)\n", iter);
	return NULL;
}

static int cmpworker(const void *p1, const void *p2)
{

	struct worker *w1 = (struct worker *) p1;
	struct worker *w2 = (struct worker *) p2;
	return w1->tid > w2->tid;
}

int bench_epoll_wait(int argc, const char **argv)
{
	int ret = 0;
	struct sigaction act;
	unsigned int i;
	struct worker *worker = NULL;
	struct perf_cpu_map *cpu;
	pthread_t wthread;
	struct rlimit rl, prevrl;

	argc = parse_options(argc, argv, options, bench_epoll_wait_usage, 0);
	if (argc) {
		usage_with_options(bench_epoll_wait_usage, options);
		exit(EXIT_FAILURE);
	}

	memset(&act, 0, sizeof(act));
	sigfillset(&act.sa_mask);
	act.sa_sigaction = toggle_done;
	sigaction(SIGINT, &act, NULL);

	cpu = perf_cpu_map__new(NULL);
	if (!cpu)
		goto errmem;

	/* a single, main epoll instance */
	if (!multiq) {
		epollfd = epoll_create(1);
		if (epollfd < 0)
			err(EXIT_FAILURE, "epoll_create");

		/*
		 * Deal with nested epolls, if any.
		 */
		if (nested)
			nest_epollfd(NULL);
	}

	printinfo("Using %s queue model\n", multiq ? "multi" : "single");
	printinfo("Nesting level(s): %d\n", nested);

	/* default to the number of CPUs and leave one for the writer pthread */
	if (!nthreads)
		nthreads = cpu->nr - 1;

	worker = calloc(nthreads, sizeof(*worker));
	if (!worker) {
		goto errmem;
	}

	if (getrlimit(RLIMIT_NOFILE, &prevrl))
		err(EXIT_FAILURE, "getrlimit");
	rl.rlim_cur = rl.rlim_max = nfds * nthreads * 2 + 50;
	printinfo("Setting RLIMIT_NOFILE rlimit from %" PRIu64 " to: %" PRIu64 "\n",
		  (uint64_t)prevrl.rlim_max, (uint64_t)rl.rlim_max);
	if (setrlimit(RLIMIT_NOFILE, &rl) < 0)
		err(EXIT_FAILURE, "setrlimit");

	printf("Run summary [PID %d]: %d threads monitoring%s on "
	       "%d file-descriptors for %d secs.\n\n",
	       getpid(), nthreads, oneshot ? " (EPOLLONESHOT semantics)": "", nfds, nsecs);

	init_stats(&throughput_stats);
	pthread_mutex_init(&thread_lock, NULL);
	pthread_cond_init(&thread_parent, NULL);
	pthread_cond_init(&thread_worker, NULL);

	threads_starting = nthreads;

	gettimeofday(&bench__start, NULL);

	do_threads(worker, cpu);

	pthread_mutex_lock(&thread_lock);
	while (threads_starting)
		pthread_cond_wait(&thread_parent, &thread_lock);
	pthread_cond_broadcast(&thread_worker);
	pthread_mutex_unlock(&thread_lock);

	/*
	 * At this point the workers should be blocked waiting for read events
	 * to become ready. Launch the writer which will constantly be writing
	 * to each thread's fdmap.
	 */
	ret = pthread_create(&wthread, NULL, writerfn,
			     (void *)(struct worker *) worker);
	if (ret)
		err(EXIT_FAILURE, "pthread_create");

	sleep(nsecs);
	toggle_done(0, NULL, NULL);
	printinfo("main thread: toggling done\n");

	sleep(1); /* meh */
	wdone = true;
	ret = pthread_join(wthread, NULL);
	if (ret)
		err(EXIT_FAILURE, "pthread_join");

	/* cleanup & report results */
	pthread_cond_destroy(&thread_parent);
	pthread_cond_destroy(&thread_worker);
	pthread_mutex_destroy(&thread_lock);

	/* sort the array back before reporting */
	if (randomize)
		qsort(worker, nthreads, sizeof(struct worker), cmpworker);

	for (i = 0; i < nthreads; i++) {
		unsigned long t = bench__runtime.tv_sec > 0 ?
			worker[i].ops / bench__runtime.tv_sec : 0;

		update_stats(&throughput_stats, t);

		if (nfds == 1)
			printf("[thread %2d] fdmap: %p [ %04ld ops/sec ]\n",
			       worker[i].tid, &worker[i].fdmap[0], t);
		else
			printf("[thread %2d] fdmap: %p ... %p [ %04ld ops/sec ]\n",
			       worker[i].tid, &worker[i].fdmap[0],
			       &worker[i].fdmap[nfds-1], t);
	}

	print_summary();

	close(epollfd);
	return ret;
errmem:
	err(EXIT_FAILURE, "calloc");
}
#endif // HAVE_EVENTFD_SUPPORT