// 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 * distribute 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 <perf/cpumap.h> #include "../util/stat.h" #include "../util/mutex.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 struct mutex thread_lock; static unsigned int threads_starting; static struct stats throughput_stats; static struct cond 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; mutex_lock(&thread_lock); threads_starting--; if (!threads_starting) cond_signal(&thread_parent); cond_wait(&thread_worker, &thread_lock); mutex_unlock(&thread_lock); do { /* * Block indefinitely 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; int nrcpus; size_t size; 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); nrcpus = perf_cpu_map__nr(cpu); cpuset = CPU_ALLOC(nrcpus); BUG_ON(!cpuset); size = CPU_ALLOC_SIZE(nrcpus); 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_S(size, cpuset); CPU_SET_S(perf_cpu_map__cpu(cpu, i % perf_cpu_map__nr(cpu)).cpu, size, cpuset); ret = pthread_attr_setaffinity_np(&thread_attr, size, cpuset); if (ret) { CPU_FREE(cpuset); err(EXIT_FAILURE, "pthread_attr_setaffinity_np"); } attrp = &thread_attr; } ret = pthread_create(&w->thread, attrp, workerfn, (void *)(struct worker *) w); if (ret) { CPU_FREE(cpuset); err(EXIT_FAILURE, "pthread_create"); } } CPU_FREE(cpuset); 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 = perf_cpu_map__nr(cpu) - 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); mutex_init(&thread_lock); cond_init(&thread_parent); cond_init(&thread_worker); threads_starting = nthreads; gettimeofday(&bench__start, NULL); do_threads(worker, cpu); mutex_lock(&thread_lock); while (threads_starting) cond_wait(&thread_parent, &thread_lock); cond_broadcast(&thread_worker); 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 */ cond_destroy(&thread_parent); cond_destroy(&thread_worker); 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); perf_cpu_map__put(cpu); for (i = 0; i < nthreads; i++) free(worker[i].fdmap); free(worker); return ret; errmem: err(EXIT_FAILURE, "calloc"); } #endif // HAVE_EVENTFD_SUPPORT