root/tools/perf/bench/epoll-wait.c

DEFINITIONS

1. shuffle
2. workerfn
3. nest_epollfd
4. toggle_done
5. print_summary
6. do_threads
7. writerfn
8. cmpworker
9. bench_epoll_wait

   1 // SPDX-License-Identifier: GPL-2.0
   2 #ifdef HAVE_EVENTFD
   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 
  91 static unsigned int nthreads = 0;
  92 static unsigned int nsecs    = 8;
  93 struct timeval start, end, runtime;
  94 static bool wdone, done, __verbose, randomize, nonblocking;
  95 
  96 /*
  97  * epoll related shared variables.
  98  */
  99 
 100 /* Maximum number of nesting allowed inside epoll sets */
 101 #define EPOLL_MAXNESTS 4
 102 
 103 static int epollfd;
 104 static int *epollfdp;
 105 static bool noaffinity;
 106 static unsigned int nested = 0;
 107 static bool et; /* edge-trigger */
 108 static bool oneshot;
 109 static bool multiq; /* use an epoll instance per thread */
 110 
 111 /* amount of fds to monitor, per thread */
 112 static unsigned int nfds = 64;
 113 
 114 static pthread_mutex_t thread_lock;
 115 static unsigned int threads_starting;
 116 static struct stats throughput_stats;
 117 static pthread_cond_t thread_parent, thread_worker;
 118 
 119 struct worker {
 120         int tid;
 121         int epollfd; /* for --multiq */
 122         pthread_t thread;
 123         unsigned long ops;
 124         int *fdmap;
 125 };
 126 
 127 static const struct option options[] = {
 128         /* general benchmark options */
 129         OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
 130         OPT_UINTEGER('r', "runtime", &nsecs, "Specify runtime (in seconds)"),
 131         OPT_UINTEGER('f', "nfds",    &nfds,  "Specify amount of file descriptors to monitor for each thread"),
 132         OPT_BOOLEAN( 'n', "noaffinity",  &noaffinity,   "Disables CPU affinity"),
 133         OPT_BOOLEAN('R', "randomize", &randomize,   "Enable random write behaviour (default is lineal)"),
 134         OPT_BOOLEAN( 'v', "verbose", &__verbose, "Verbose mode"),
 135 
 136         /* epoll specific options */
 137         OPT_BOOLEAN( 'm', "multiq",  &multiq,   "Use multiple epoll instances (one per thread)"),
 138         OPT_BOOLEAN( 'B', "nonblocking", &nonblocking, "Nonblocking epoll_wait(2) behaviour"),
 139         OPT_UINTEGER( 'N', "nested",  &nested,   "Nesting level epoll hierarchy (default is 0, no nesting)"),
 140         OPT_BOOLEAN( 'S', "oneshot",  &oneshot,   "Use EPOLLONESHOT semantics"),
 141         OPT_BOOLEAN( 'E', "edge",  &et,   "Use Edge-triggered interface (default is LT)"),
 142 
 143         OPT_END()
 144 };
 145 
 146 static const char * const bench_epoll_wait_usage[] = {
 147         "perf bench epoll wait <options>",
 148         NULL
 149 };
 150 
 151 
 152 /*
 153  * Arrange the N elements of ARRAY in random order.
 154  * Only effective if N is much smaller than RAND_MAX;
 155  * if this may not be the case, use a better random
 156  * number generator. -- Ben Pfaff.
 157  */
 158 static void shuffle(void *array, size_t n, size_t size)
 159 {
 160         char *carray = array;
 161         void *aux;
 162         size_t i;
 163 
 164         if (n <= 1)
 165                 return;
 166 
 167         aux = calloc(1, size);
 168         if (!aux)
 169                 err(EXIT_FAILURE, "calloc");
 170 
 171         for (i = 1; i < n; ++i) {
 172                 size_t j =   i + rand() / (RAND_MAX / (n - i) + 1);
 173                 j *= size;
 174 
 175                 memcpy(aux, &carray[j], size);
 176                 memcpy(&carray[j], &carray[i*size], size);
 177                 memcpy(&carray[i*size], aux, size);
 178         }
 179 
 180         free(aux);
 181 }
 182 
 183 
 184 static void *workerfn(void *arg)
 185 {
 186         int fd, ret, r;
 187         struct worker *w = (struct worker *) arg;
 188         unsigned long ops = w->ops;
 189         struct epoll_event ev;
 190         uint64_t val;
 191         int to = nonblocking? 0 : -1;
 192         int efd = multiq ? w->epollfd : epollfd;
 193 
 194         pthread_mutex_lock(&thread_lock);
 195         threads_starting--;
 196         if (!threads_starting)
 197                 pthread_cond_signal(&thread_parent);
 198         pthread_cond_wait(&thread_worker, &thread_lock);
 199         pthread_mutex_unlock(&thread_lock);
 200 
 201         do {
 202                 /*
 203                  * Block undefinitely waiting for the IN event.
 204                  * In order to stress the epoll_wait(2) syscall,
 205                  * call it event per event, instead of a larger
 206                  * batch (max)limit.
 207                  */
 208                 do {
 209                         ret = epoll_wait(efd, &ev, 1, to);
 210                 } while (ret < 0 && errno == EINTR);
 211                 if (ret < 0)
 212                         err(EXIT_FAILURE, "epoll_wait");
 213 
 214                 fd = ev.data.fd;
 215 
 216                 do {
 217                         r = read(fd, &val, sizeof(val));
 218                 } while (!done && (r < 0 && errno == EAGAIN));
 219 
 220                 if (et) {
 221                         ev.events = EPOLLIN | EPOLLET;
 222                         ret = epoll_ctl(efd, EPOLL_CTL_ADD, fd, &ev);
 223                 }
 224 
 225                 if (oneshot) {
 226                         /* rearm the file descriptor with a new event mask */
 227                         ev.events |= EPOLLIN | EPOLLONESHOT;
 228                         ret = epoll_ctl(efd, EPOLL_CTL_MOD, fd, &ev);
 229                 }
 230 
 231                 ops++;
 232         }  while (!done);
 233 
 234         if (multiq)
 235                 close(w->epollfd);
 236 
 237         w->ops = ops;
 238         return NULL;
 239 }
 240 
 241 static void nest_epollfd(struct worker *w)
 242 {
 243         unsigned int i;
 244         struct epoll_event ev;
 245         int efd = multiq ? w->epollfd : epollfd;
 246 
 247         if (nested > EPOLL_MAXNESTS)
 248                 nested = EPOLL_MAXNESTS;
 249 
 250         epollfdp = calloc(nested, sizeof(*epollfdp));
 251         if (!epollfdp)
 252                 err(EXIT_FAILURE, "calloc");
 253 
 254         for (i = 0; i < nested; i++) {
 255                 epollfdp[i] = epoll_create(1);
 256                 if (epollfdp[i] < 0)
 257                         err(EXIT_FAILURE, "epoll_create");
 258         }
 259 
 260         ev.events = EPOLLHUP; /* anything */
 261         ev.data.u64 = i; /* any number */
 262 
 263         for (i = nested - 1; i; i--) {
 264                 if (epoll_ctl(epollfdp[i - 1], EPOLL_CTL_ADD,
 265                               epollfdp[i], &ev) < 0)
 266                         err(EXIT_FAILURE, "epoll_ctl");
 267         }
 268 
 269         if (epoll_ctl(efd, EPOLL_CTL_ADD, *epollfdp, &ev) < 0)
 270                 err(EXIT_FAILURE, "epoll_ctl");
 271 }
 272 
 273 static void toggle_done(int sig __maybe_unused,
 274                         siginfo_t *info __maybe_unused,
 275                         void *uc __maybe_unused)
 276 {
 277         /* inform all threads that we're done for the day */
 278         done = true;
 279         gettimeofday(&end, NULL);
 280         timersub(&end, &start, &runtime);
 281 }
 282 
 283 static void print_summary(void)
 284 {
 285         unsigned long avg = avg_stats(&throughput_stats);
 286         double stddev = stddev_stats(&throughput_stats);
 287 
 288         printf("\nAveraged %ld operations/sec (+- %.2f%%), total secs = %d\n",
 289                avg, rel_stddev_stats(stddev, avg),
 290                (int) runtime.tv_sec);
 291 }
 292 
 293 static int do_threads(struct worker *worker, struct perf_cpu_map *cpu)
 294 {
 295         pthread_attr_t thread_attr, *attrp = NULL;
 296         cpu_set_t cpuset;
 297         unsigned int i, j;
 298         int ret = 0, events = EPOLLIN;
 299 
 300         if (oneshot)
 301                 events |= EPOLLONESHOT;
 302         if (et)
 303                 events |= EPOLLET;
 304 
 305         printinfo("starting worker/consumer %sthreads%s\n",
 306                   noaffinity ?  "":"CPU affinity ",
 307                   nonblocking ? " (nonblocking)":"");
 308         if (!noaffinity)
 309                 pthread_attr_init(&thread_attr);
 310 
 311         for (i = 0; i < nthreads; i++) {
 312                 struct worker *w = &worker[i];
 313 
 314                 if (multiq) {
 315                         w->epollfd = epoll_create(1);
 316                         if (w->epollfd < 0)
 317                                 err(EXIT_FAILURE, "epoll_create");
 318 
 319                         if (nested)
 320                                 nest_epollfd(w);
 321                 }
 322 
 323                 w->tid = i;
 324                 w->fdmap = calloc(nfds, sizeof(int));
 325                 if (!w->fdmap)
 326                         return 1;
 327 
 328                 for (j = 0; j < nfds; j++) {
 329                         int efd = multiq ? w->epollfd : epollfd;
 330                         struct epoll_event ev;
 331 
 332                         w->fdmap[j] = eventfd(0, EFD_NONBLOCK);
 333                         if (w->fdmap[j] < 0)
 334                                 err(EXIT_FAILURE, "eventfd");
 335 
 336                         ev.data.fd = w->fdmap[j];
 337                         ev.events = events;
 338 
 339                         ret = epoll_ctl(efd, EPOLL_CTL_ADD,
 340                                         w->fdmap[j], &ev);
 341                         if (ret < 0)
 342                                 err(EXIT_FAILURE, "epoll_ctl");
 343                 }
 344 
 345                 if (!noaffinity) {
 346                         CPU_ZERO(&cpuset);
 347                         CPU_SET(cpu->map[i % cpu->nr], &cpuset);
 348 
 349                         ret = pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset);
 350                         if (ret)
 351                                 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
 352 
 353                         attrp = &thread_attr;
 354                 }
 355 
 356                 ret = pthread_create(&w->thread, attrp, workerfn,
 357                                      (void *)(struct worker *) w);
 358                 if (ret)
 359                         err(EXIT_FAILURE, "pthread_create");
 360         }
 361 
 362         if (!noaffinity)
 363                 pthread_attr_destroy(&thread_attr);
 364 
 365         return ret;
 366 }
 367 
 368 static void *writerfn(void *p)
 369 {
 370         struct worker *worker = p;
 371         size_t i, j, iter;
 372         const uint64_t val = 1;
 373         ssize_t sz;
 374         struct timespec ts = { .tv_sec = 0,
 375                                .tv_nsec = 500 };
 376 
 377         printinfo("starting writer-thread: doing %s writes ...\n",
 378                   randomize? "random":"lineal");
 379 
 380         for (iter = 0; !wdone; iter++) {
 381                 if (randomize) {
 382                         shuffle((void *)worker, nthreads, sizeof(*worker));
 383                 }
 384 
 385                 for (i = 0; i < nthreads; i++) {
 386                         struct worker *w = &worker[i];
 387 
 388                         if (randomize) {
 389                                 shuffle((void *)w->fdmap, nfds, sizeof(int));
 390                         }
 391 
 392                         for (j = 0; j < nfds; j++) {
 393                                 do {
 394                                         sz = write(w->fdmap[j], &val, sizeof(val));
 395                                 } while (!wdone && (sz < 0 && errno == EAGAIN));
 396                         }
 397                 }
 398 
 399                 nanosleep(&ts, NULL);
 400         }
 401 
 402         printinfo("exiting writer-thread (total full-loops: %zd)\n", iter);
 403         return NULL;
 404 }
 405 
 406 static int cmpworker(const void *p1, const void *p2)
 407 {
 408 
 409         struct worker *w1 = (struct worker *) p1;
 410         struct worker *w2 = (struct worker *) p2;
 411         return w1->tid > w2->tid;
 412 }
 413 
 414 int bench_epoll_wait(int argc, const char **argv)
 415 {
 416         int ret = 0;
 417         struct sigaction act;
 418         unsigned int i;
 419         struct worker *worker = NULL;
 420         struct perf_cpu_map *cpu;
 421         pthread_t wthread;
 422         struct rlimit rl, prevrl;
 423 
 424         argc = parse_options(argc, argv, options, bench_epoll_wait_usage, 0);
 425         if (argc) {
 426                 usage_with_options(bench_epoll_wait_usage, options);
 427                 exit(EXIT_FAILURE);
 428         }
 429 
 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(&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 = worker[i].ops/runtime.tv_sec;
 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