1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2018 Davidlohr Bueso.
5 * Benchmark the various operations allowed for epoll_ctl(2).
6 * The idea is to concurrently stress a single epoll instance
9 /* For the CLR_() macros */
18 #include <linux/compiler.h>
19 #include <linux/kernel.h>
21 #include <sys/resource.h>
22 #include <sys/epoll.h>
23 #include <sys/eventfd.h>
24 #include <internal/cpumap.h>
25 #include <perf/cpumap.h>
27 #include "../util/stat.h"
28 #include <subcmd/parse-options.h>
33 #define printinfo(fmt, arg...) \
34 do { if (__verbose) printf(fmt, ## arg); } while (0)
36 static unsigned int nthreads = 0;
37 static unsigned int nsecs = 8;
38 static bool done, __verbose, randomize;
41 * epoll related shared variables.
44 /* Maximum number of nesting allowed inside epoll sets */
45 #define EPOLL_MAXNESTS 4
56 static bool noaffinity;
57 static unsigned int nested = 0;
59 /* amount of fds to monitor, per thread */
60 static unsigned int nfds = 64;
62 static pthread_mutex_t thread_lock;
63 static unsigned int threads_starting;
64 static struct stats all_stats[EPOLL_NR_OPS];
65 static pthread_cond_t thread_parent, thread_worker;
70 unsigned long ops[EPOLL_NR_OPS];
74 static const struct option options[] = {
75 OPT_UINTEGER('t', "threads", &nthreads, "Specify amount of threads"),
76 OPT_UINTEGER('r', "runtime", &nsecs, "Specify runtime (in seconds)"),
77 OPT_UINTEGER('f', "nfds", &nfds, "Specify amount of file descriptors to monitor for each thread"),
78 OPT_BOOLEAN( 'n', "noaffinity", &noaffinity, "Disables CPU affinity"),
79 OPT_UINTEGER( 'N', "nested", &nested, "Nesting level epoll hierarchy (default is 0, no nesting)"),
80 OPT_BOOLEAN( 'R', "randomize", &randomize, "Perform random operations on random fds"),
81 OPT_BOOLEAN( 'v', "verbose", &__verbose, "Verbose mode"),
85 static const char * const bench_epoll_ctl_usage[] = {
86 "perf bench epoll ctl <options>",
90 static void toggle_done(int sig __maybe_unused,
91 siginfo_t *info __maybe_unused,
92 void *uc __maybe_unused)
94 /* inform all threads that we're done for the day */
96 gettimeofday(&bench__end, NULL);
97 timersub(&bench__end, &bench__start, &bench__runtime);
100 static void nest_epollfd(void)
103 struct epoll_event ev;
105 if (nested > EPOLL_MAXNESTS)
106 nested = EPOLL_MAXNESTS;
107 printinfo("Nesting level(s): %d\n", nested);
109 epollfdp = calloc(nested, sizeof(int));
111 err(EXIT_FAILURE, "calloc");
113 for (i = 0; i < nested; i++) {
114 epollfdp[i] = epoll_create(1);
116 err(EXIT_FAILURE, "epoll_create");
119 ev.events = EPOLLHUP; /* anything */
120 ev.data.u64 = i; /* any number */
122 for (i = nested - 1; i; i--) {
123 if (epoll_ctl(epollfdp[i - 1], EPOLL_CTL_ADD,
124 epollfdp[i], &ev) < 0)
125 err(EXIT_FAILURE, "epoll_ctl");
128 if (epoll_ctl(epollfd, EPOLL_CTL_ADD, *epollfdp, &ev) < 0)
129 err(EXIT_FAILURE, "epoll_ctl");
132 static inline void do_epoll_op(struct worker *w, int op, int fd)
135 struct epoll_event ev;
142 error = epoll_ctl(epollfd, EPOLL_CTL_ADD, fd, &ev);
145 ev.events = EPOLLOUT;
146 error = epoll_ctl(epollfd, EPOLL_CTL_MOD, fd, &ev);
149 error = epoll_ctl(epollfd, EPOLL_CTL_DEL, fd, NULL);
160 static inline void do_random_epoll_op(struct worker *w)
162 unsigned long rnd1 = random(), rnd2 = random();
165 fd = w->fdmap[rnd1 % nfds];
166 op = rnd2 % EPOLL_NR_OPS;
168 do_epoll_op(w, op, fd);
171 static void *workerfn(void *arg)
174 struct worker *w = (struct worker *) arg;
175 struct timespec ts = { .tv_sec = 0,
178 pthread_mutex_lock(&thread_lock);
180 if (!threads_starting)
181 pthread_cond_signal(&thread_parent);
182 pthread_cond_wait(&thread_worker, &thread_lock);
183 pthread_mutex_unlock(&thread_lock);
189 do_random_epoll_op(w);
191 for (i = 0; i < nfds; i++) {
192 do_epoll_op(w, OP_EPOLL_ADD, w->fdmap[i]);
193 do_epoll_op(w, OP_EPOLL_MOD, w->fdmap[i]);
194 do_epoll_op(w, OP_EPOLL_DEL, w->fdmap[i]);
198 nanosleep(&ts, NULL);
204 static void init_fdmaps(struct worker *w, int pct)
208 struct epoll_event ev;
214 for (i = 0; i < nfds; i+=inc) {
215 ev.data.fd = w->fdmap[i];
218 if (epoll_ctl(epollfd, EPOLL_CTL_ADD, w->fdmap[i], &ev) < 0)
219 err(EXIT_FAILURE, "epoll_ct");
223 static int do_threads(struct worker *worker, struct perf_cpu_map *cpu)
225 pthread_attr_t thread_attr, *attrp = NULL;
231 pthread_attr_init(&thread_attr);
233 for (i = 0; i < nthreads; i++) {
234 struct worker *w = &worker[i];
237 w->fdmap = calloc(nfds, sizeof(int));
241 for (j = 0; j < nfds; j++) {
242 w->fdmap[j] = eventfd(0, EFD_NONBLOCK);
244 err(EXIT_FAILURE, "eventfd");
248 * Lets add 50% of the fdmap to the epoll instance, and
249 * do it before any threads are started; otherwise there is
250 * an initial bias of the call failing (mod and del ops).
257 CPU_SET(cpu->map[i % cpu->nr], &cpuset);
259 ret = pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpuset);
261 err(EXIT_FAILURE, "pthread_attr_setaffinity_np");
263 attrp = &thread_attr;
266 ret = pthread_create(&w->thread, attrp, workerfn,
267 (void *)(struct worker *) w);
269 err(EXIT_FAILURE, "pthread_create");
273 pthread_attr_destroy(&thread_attr);
278 static void print_summary(void)
281 unsigned long avg[EPOLL_NR_OPS];
282 double stddev[EPOLL_NR_OPS];
284 for (i = 0; i < EPOLL_NR_OPS; i++) {
285 avg[i] = avg_stats(&all_stats[i]);
286 stddev[i] = stddev_stats(&all_stats[i]);
289 printf("\nAveraged %ld ADD operations (+- %.2f%%)\n",
290 avg[OP_EPOLL_ADD], rel_stddev_stats(stddev[OP_EPOLL_ADD],
292 printf("Averaged %ld MOD operations (+- %.2f%%)\n",
293 avg[OP_EPOLL_MOD], rel_stddev_stats(stddev[OP_EPOLL_MOD],
295 printf("Averaged %ld DEL operations (+- %.2f%%)\n",
296 avg[OP_EPOLL_DEL], rel_stddev_stats(stddev[OP_EPOLL_DEL],
300 int bench_epoll_ctl(int argc, const char **argv)
303 struct sigaction act;
304 struct worker *worker = NULL;
305 struct perf_cpu_map *cpu;
306 struct rlimit rl, prevrl;
309 argc = parse_options(argc, argv, options, bench_epoll_ctl_usage, 0);
311 usage_with_options(bench_epoll_ctl_usage, options);
315 memset(&act, 0, sizeof(act));
316 sigfillset(&act.sa_mask);
317 act.sa_sigaction = toggle_done;
318 sigaction(SIGINT, &act, NULL);
320 cpu = perf_cpu_map__new(NULL);
324 /* a single, main epoll instance */
325 epollfd = epoll_create(1);
327 err(EXIT_FAILURE, "epoll_create");
330 * Deal with nested epolls, if any.
335 /* default to the number of CPUs */
339 worker = calloc(nthreads, sizeof(*worker));
343 if (getrlimit(RLIMIT_NOFILE, &prevrl))
344 err(EXIT_FAILURE, "getrlimit");
345 rl.rlim_cur = rl.rlim_max = nfds * nthreads * 2 + 50;
346 printinfo("Setting RLIMIT_NOFILE rlimit from %" PRIu64 " to: %" PRIu64 "\n",
347 (uint64_t)prevrl.rlim_max, (uint64_t)rl.rlim_max);
348 if (setrlimit(RLIMIT_NOFILE, &rl) < 0)
349 err(EXIT_FAILURE, "setrlimit");
351 printf("Run summary [PID %d]: %d threads doing epoll_ctl ops "
352 "%d file-descriptors for %d secs.\n\n",
353 getpid(), nthreads, nfds, nsecs);
355 for (i = 0; i < EPOLL_NR_OPS; i++)
356 init_stats(&all_stats[i]);
358 pthread_mutex_init(&thread_lock, NULL);
359 pthread_cond_init(&thread_parent, NULL);
360 pthread_cond_init(&thread_worker, NULL);
362 threads_starting = nthreads;
364 gettimeofday(&bench__start, NULL);
366 do_threads(worker, cpu);
368 pthread_mutex_lock(&thread_lock);
369 while (threads_starting)
370 pthread_cond_wait(&thread_parent, &thread_lock);
371 pthread_cond_broadcast(&thread_worker);
372 pthread_mutex_unlock(&thread_lock);
375 toggle_done(0, NULL, NULL);
376 printinfo("main thread: toggling done\n");
378 for (i = 0; i < nthreads; i++) {
379 ret = pthread_join(worker[i].thread, NULL);
381 err(EXIT_FAILURE, "pthread_join");
384 /* cleanup & report results */
385 pthread_cond_destroy(&thread_parent);
386 pthread_cond_destroy(&thread_worker);
387 pthread_mutex_destroy(&thread_lock);
389 for (i = 0; i < nthreads; i++) {
390 unsigned long t[EPOLL_NR_OPS];
392 for (j = 0; j < EPOLL_NR_OPS; j++) {
393 t[j] = worker[i].ops[j];
394 update_stats(&all_stats[j], t[j]);
398 printf("[thread %2d] fdmap: %p [ add: %04ld; mod: %04ld; del: %04lds ops ]\n",
399 worker[i].tid, &worker[i].fdmap[0],
400 t[OP_EPOLL_ADD], t[OP_EPOLL_MOD], t[OP_EPOLL_DEL]);
402 printf("[thread %2d] fdmap: %p ... %p [ add: %04ld ops; mod: %04ld ops; del: %04ld ops ]\n",
403 worker[i].tid, &worker[i].fdmap[0],
404 &worker[i].fdmap[nfds-1],
405 t[OP_EPOLL_ADD], t[OP_EPOLL_MOD], t[OP_EPOLL_DEL]);
413 err(EXIT_FAILURE, "calloc");
415 #endif // HAVE_EVENTFD