2 * fio - the flexible io tester
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
7 * The license below covers all files distributed with fio unless otherwise
8 * noted in the file itself.
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License version 2 as
12 * published by the Free Software Foundation.
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
34 #include <sys/ioctl.h>
42 #define ALIGN(buf) (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
45 int thread_number = 0;
49 static volatile int startup_sem;
50 static volatile int fio_abort;
52 #define TERMINATE_ALL (-1)
53 #define JOB_START_TIMEOUT (5 * 1000)
55 static inline void td_set_runstate(struct thread_data *td, int runstate)
57 td->runstate = runstate;
60 static void terminate_threads(int group_id, int forced_kill)
62 struct thread_data *td;
66 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
70 td_set_runstate(td, TD_EXITED);
75 static void sig_handler(int sig)
80 disk_util_timer_arm();
81 print_thread_status();
84 fprintf(stderr, "fio: got segfault, aborting\n");
85 terminate_threads(TERMINATE_ALL, 1);
89 printf("\nfio: terminating on signal %d\n", sig);
91 terminate_threads(TERMINATE_ALL, 0);
97 * Check if we are above the minimum rate given.
99 static int check_min_rate(struct thread_data *td, struct timeval *now)
106 * allow a 2 second settle period in the beginning
108 if (mtime_since(&td->start, now) < 2000)
112 * if rate blocks is set, sample is running
114 if (td->rate_bytes) {
115 spent = mtime_since(&td->lastrate, now);
116 if (spent < td->ratecycle)
119 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
120 if (rate < td->ratemin) {
121 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
126 td->rate_bytes = td->this_io_bytes[ddir];
127 memcpy(&td->lastrate, now, sizeof(*now));
131 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
135 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
141 static struct fio_file *get_next_file(struct thread_data *td)
143 unsigned int old_next_file = td->next_file;
147 f = &td->files[td->next_file];
150 if (td->next_file >= td->nr_files)
157 } while (td->next_file != old_next_file);
163 * When job exits, we can cancel the in-flight IO if we are using async
164 * io. Attempt to do so.
166 static void cleanup_pending_aio(struct thread_data *td)
168 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
169 struct list_head *entry, *n;
170 struct io_completion_data icd;
175 * get immediately available events, if any
177 r = td_io_getevents(td, 0, td->cur_depth, &ts);
180 ios_completed(td, &icd);
184 * now cancel remaining active events
186 if (td->io_ops->cancel) {
187 list_for_each_safe(entry, n, &td->io_u_busylist) {
188 io_u = list_entry(entry, struct io_u, list);
190 r = td->io_ops->cancel(td, io_u);
197 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
200 ios_completed(td, &icd);
206 * Helper to handle the final sync of a file. Works just like the normal
207 * io path, just does everything sync.
209 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
211 struct io_u *io_u = __get_io_u(td);
212 struct io_completion_data icd;
218 io_u->ddir = DDIR_SYNC;
221 if (td_io_prep(td, io_u)) {
226 ret = td_io_queue(td, io_u);
228 td_verror(td, io_u->error);
233 ret = td_io_getevents(td, 1, td->cur_depth, NULL);
240 ios_completed(td, &icd);
242 td_verror(td, icd.error);
250 * The main verify engine. Runs over the writes we previusly submitted,
251 * reads the blocks back in, and checks the crc/md5 of the data.
253 static void do_verify(struct thread_data *td)
255 struct io_u *io_u, *v_io_u = NULL;
256 struct io_completion_data icd;
261 * sync io first and invalidate cache, to make sure we really
264 for_each_file(td, f, i) {
266 file_invalidate_cache(td, f);
269 td_set_runstate(td, TD_VERIFYING);
275 io_u = __get_io_u(td);
279 if (runtime_exceeded(td, &io_u->start_time)) {
284 if (get_next_verify(td, io_u)) {
289 f = get_next_file(td);
295 if (td_io_prep(td, io_u)) {
300 ret = td_io_queue(td, io_u);
302 td_verror(td, io_u->error);
308 * we have one pending to verify, do that while
309 * we are doing io on the next one
311 if (do_io_u_verify(td, &v_io_u))
314 ret = td_io_getevents(td, 1, 1, NULL);
321 v_io_u = td->io_ops->event(td, 0);
324 fio_gettime(&icd.time, NULL);
325 io_completed(td, v_io_u, &icd);
328 td_verror(td, icd.error);
329 put_io_u(td, v_io_u);
335 * if we can't submit more io, we need to verify now
337 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
342 do_io_u_verify(td, &v_io_u);
345 cleanup_pending_aio(td);
347 td_set_runstate(td, TD_RUNNING);
351 * Not really an io thread, all it does is burn CPU cycles in the specified
354 static void do_cpuio(struct thread_data *td)
357 int split = 100 / td->cpuload;
360 while (!td->terminate) {
361 fio_gettime(&e, NULL);
363 if (runtime_exceeded(td, &e))
369 usec_sleep(td, 10000);
376 * Main IO worker function. It retrieves io_u's to process and queues
377 * and reaps them, checking for rate and errors along the way.
379 static void do_io(struct thread_data *td)
381 struct io_completion_data icd;
387 td_set_runstate(td, TD_RUNNING);
389 while (td->this_io_bytes[td->ddir] < td->io_size) {
390 struct timespec *timeout;
397 f = get_next_file(td);
401 io_u = get_io_u(td, f);
405 memcpy(&s, &io_u->start_time, sizeof(s));
407 ret = td_io_queue(td, io_u);
409 td_verror(td, io_u->error);
414 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
416 if (td->cur_depth < td->iodepth) {
417 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
426 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
434 ios_completed(td, &icd);
436 td_verror(td, icd.error);
441 * the rate is batched for now, it should work for batches
442 * of completions except the very first one which may look
445 usec = utime_since(&s, &icd.time);
447 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
449 if (check_min_rate(td, &icd.time)) {
450 if (exitall_on_terminate)
451 terminate_threads(td->groupid, 0);
452 td_verror(td, ENOMEM);
456 if (runtime_exceeded(td, &icd.time))
460 unsigned long long b;
462 b = td->io_blocks[0] + td->io_blocks[1];
463 if (!(b % td->thinktime_blocks))
464 usec_sleep(td, td->thinktime);
470 cleanup_pending_aio(td);
472 if (should_fsync(td) && td->end_fsync) {
473 td_set_runstate(td, TD_FSYNCING);
474 for_each_file(td, f, i)
480 static void cleanup_io_u(struct thread_data *td)
482 struct list_head *entry, *n;
485 list_for_each_safe(entry, n, &td->io_u_freelist) {
486 io_u = list_entry(entry, struct io_u, list);
488 list_del(&io_u->list);
496 * "randomly" fill the buffer contents
498 static void fill_rand_buf(struct io_u *io_u, int max_bs)
500 int *ptr = io_u->buf;
502 while ((void *) ptr - io_u->buf < max_bs) {
503 *ptr = rand() * 0x9e370001;
508 static int init_io_u(struct thread_data *td)
515 if (td->io_ops->flags & FIO_CPUIO)
518 if (td->io_ops->flags & FIO_SYNCIO)
521 max_units = td->iodepth;
523 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
524 td->orig_buffer_size = max_bs * max_units;
526 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
527 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
529 td->orig_buffer_size += MASK;
531 if (allocate_io_mem(td))
534 p = ALIGN(td->orig_buffer);
535 for (i = 0; i < max_units; i++) {
536 io_u = malloc(sizeof(*io_u));
537 memset(io_u, 0, sizeof(*io_u));
538 INIT_LIST_HEAD(&io_u->list);
540 io_u->buf = p + max_bs * i;
541 if (td_write(td) || td_rw(td))
542 fill_rand_buf(io_u, max_bs);
545 list_add(&io_u->list, &td->io_u_freelist);
551 static int switch_ioscheduler(struct thread_data *td)
553 char tmp[256], tmp2[128];
557 if (td->io_ops->flags & FIO_CPUIO)
560 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
562 f = fopen(tmp, "r+");
564 td_verror(td, errno);
571 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
572 if (ferror(f) || ret != 1) {
573 td_verror(td, errno);
581 * Read back and check that the selected scheduler is now the default.
583 ret = fread(tmp, 1, sizeof(tmp), f);
584 if (ferror(f) || ret < 0) {
585 td_verror(td, errno);
590 sprintf(tmp2, "[%s]", td->ioscheduler);
591 if (!strstr(tmp, tmp2)) {
592 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
593 td_verror(td, EINVAL);
602 static void clear_io_state(struct thread_data *td)
607 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
608 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
611 for_each_file(td, f, i) {
613 if (td->io_ops->flags & FIO_SYNCIO)
614 lseek(f->fd, SEEK_SET, 0);
617 memset(f->file_map, 0, f->num_maps * sizeof(long));
622 * Entry point for the thread based jobs. The process based jobs end up
623 * here as well, after a little setup.
625 static void *thread_main(void *data)
627 unsigned long long runtime[2];
628 struct thread_data *td = data;
635 INIT_LIST_HEAD(&td->io_u_freelist);
636 INIT_LIST_HEAD(&td->io_u_busylist);
637 INIT_LIST_HEAD(&td->io_hist_list);
638 INIT_LIST_HEAD(&td->io_log_list);
643 if (fio_setaffinity(td) == -1) {
644 td_verror(td, errno);
655 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
656 td_verror(td, errno);
661 if (nice(td->nice) == -1) {
662 td_verror(td, errno);
666 if (init_random_state(td))
669 if (td->ioscheduler && switch_ioscheduler(td))
672 td_set_runstate(td, TD_INITIALIZED);
673 fio_sem_up(&startup_sem);
674 fio_sem_down(&td->mutex);
676 if (!td->create_serialize && setup_files(td))
682 system(td->exec_prerun);
684 fio_gettime(&td->epoch, NULL);
685 getrusage(RUSAGE_SELF, &td->ru_start);
687 runtime[0] = runtime[1] = 0;
688 while (td->loops--) {
689 fio_gettime(&td->start, NULL);
690 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
693 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
696 prune_io_piece_log(td);
698 if (td->io_ops->flags & FIO_CPUIO)
703 runtime[td->ddir] += utime_since_now(&td->start);
704 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
705 runtime[td->ddir ^ 1] = runtime[td->ddir];
707 if (td->error || td->terminate)
710 if (td->verify == VERIFY_NONE)
714 fio_gettime(&td->start, NULL);
718 runtime[DDIR_READ] += utime_since_now(&td->start);
720 if (td->error || td->terminate)
724 update_rusage_stat(td);
725 fio_gettime(&td->end_time, NULL);
726 td->runtime[0] = runtime[0] / 1000;
727 td->runtime[1] = runtime[1] / 1000;
730 finish_log(td, td->bw_log, "bw");
732 finish_log(td, td->slat_log, "slat");
734 finish_log(td, td->clat_log, "clat");
735 if (td->write_iolog_file)
736 write_iolog_close(td);
737 if (td->exec_postrun)
738 system(td->exec_postrun);
740 if (exitall_on_terminate)
741 terminate_threads(td->groupid, 0);
747 td_set_runstate(td, TD_EXITED);
753 * We cannot pass the td data into a forked process, so attach the td and
754 * pass it to the thread worker.
756 static void *fork_main(int shmid, int offset)
758 struct thread_data *td;
761 data = shmat(shmid, NULL, 0);
762 if (data == (void *) -1) {
767 td = data + offset * sizeof(struct thread_data);
774 * Run over the job map and reap the threads that have exited, if any.
776 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
778 struct thread_data *td;
779 int i, cputhreads, pending;
782 * reap exited threads (TD_EXITED -> TD_REAPED)
784 pending = cputhreads = 0;
787 * ->io_ops is NULL for a thread that has closed its
790 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
793 if (td->runstate != TD_EXITED) {
794 if (td->runstate < TD_RUNNING)
800 td_set_runstate(td, TD_REAPED);
802 if (td->use_thread) {
805 if (pthread_join(td->thread, (void *) &ret))
806 perror("thread_join");
808 waitpid(td->pid, NULL, 0);
811 (*m_rate) -= td->ratemin;
812 (*t_rate) -= td->rate;
815 if (*nr_running == cputhreads && !pending)
816 terminate_threads(TERMINATE_ALL, 0);
820 * Main function for kicking off and reaping jobs, as needed.
822 static void run_threads(void)
824 struct thread_data *td;
826 int i, todo, nr_running, m_rate, t_rate, nr_started;
828 if (fio_pin_memory())
832 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
836 signal(SIGINT, sig_handler);
837 signal(SIGALRM, sig_handler);
838 signal(SIGSEGV, sig_handler);
840 todo = thread_number;
846 print_status_init(td->thread_number - 1);
848 if (!td->create_serialize) {
854 * do file setup here so it happens sequentially,
855 * we don't want X number of threads getting their
856 * client data interspersed on disk
858 if (setup_files(td)) {
859 td_set_runstate(td, TD_REAPED);
867 struct thread_data *map[MAX_JOBS];
868 struct timeval this_start;
869 int this_jobs = 0, left;
872 * create threads (TD_NOT_CREATED -> TD_CREATED)
875 if (td->runstate != TD_NOT_CREATED)
879 * never got a chance to start, killed by other
880 * thread for some reason
887 if (td->start_delay) {
888 spent = mtime_since_genesis();
890 if (td->start_delay * 1000 > spent)
894 if (td->stonewall && (nr_started || nr_running))
898 * Set state to created. Thread will transition
899 * to TD_INITIALIZED when it's done setting up.
901 td_set_runstate(td, TD_CREATED);
902 map[this_jobs++] = td;
903 fio_sem_init(&startup_sem, 1);
906 if (td->use_thread) {
907 if (pthread_create(&td->thread, NULL, thread_main, td)) {
908 perror("thread_create");
913 fio_sem_down(&startup_sem);
915 fork_main(shm_id, i);
922 * Wait for the started threads to transition to
925 fio_gettime(&this_start, NULL);
927 while (left && !fio_abort) {
928 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
933 for (i = 0; i < this_jobs; i++) {
937 if (td->runstate == TD_INITIALIZED) {
940 } else if (td->runstate >= TD_EXITED) {
944 nr_running++; /* work-around... */
950 log_err("fio: %d jobs failed to start\n", left);
951 for (i = 0; i < this_jobs; i++) {
955 kill(td->pid, SIGTERM);
961 * start created threads (TD_INITIALIZED -> TD_RUNNING).
964 if (td->runstate != TD_INITIALIZED)
967 td_set_runstate(td, TD_RUNNING);
970 m_rate += td->ratemin;
973 fio_sem_up(&td->mutex);
976 reap_threads(&nr_running, &t_rate, &m_rate);
983 reap_threads(&nr_running, &t_rate, &m_rate);
991 int main(int argc, char *argv[])
993 if (parse_options(argc, argv))
996 if (!thread_number) {
997 log_err("Nothing to do\n");
1001 disk_util_timer_arm();