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;
48 char *fio_inst_prefix = _INST_PREFIX;
50 static volatile int startup_sem;
52 #define TERMINATE_ALL (-1)
53 #define JOB_START_TIMEOUT (5 * 1000)
55 static void terminate_threads(int group_id)
57 struct thread_data *td;
61 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
68 static void sig_handler(int sig)
73 disk_util_timer_arm();
74 print_thread_status();
77 printf("\nfio: terminating on signal\n");
79 terminate_threads(TERMINATE_ALL);
85 * Check if we are above the minimum rate given.
87 static int check_min_rate(struct thread_data *td, struct timeval *now)
94 * allow a 2 second settle period in the beginning
96 if (mtime_since(&td->start, now) < 2000)
100 * if rate blocks is set, sample is running
102 if (td->rate_bytes) {
103 spent = mtime_since(&td->lastrate, now);
104 if (spent < td->ratecycle)
107 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
108 if (rate < td->ratemin) {
109 fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
114 td->rate_bytes = td->this_io_bytes[ddir];
115 memcpy(&td->lastrate, now, sizeof(*now));
119 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
123 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
129 static inline void td_set_runstate(struct thread_data *td, int runstate)
131 td->runstate = runstate;
134 static struct fio_file *get_next_file(struct thread_data *td)
136 unsigned int old_next_file = td->next_file;
140 f = &td->files[td->next_file];
143 if (td->next_file >= td->nr_files)
150 } while (td->next_file != old_next_file);
156 * When job exits, we can cancel the in-flight IO if we are using async
157 * io. Attempt to do so.
159 static void cleanup_pending_aio(struct thread_data *td)
161 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
162 struct list_head *entry, *n;
163 struct io_completion_data icd;
168 * get immediately available events, if any
170 r = td_io_getevents(td, 0, td->cur_depth, &ts);
173 ios_completed(td, &icd);
177 * now cancel remaining active events
179 if (td->io_ops->cancel) {
180 list_for_each_safe(entry, n, &td->io_u_busylist) {
181 io_u = list_entry(entry, struct io_u, list);
183 r = td->io_ops->cancel(td, io_u);
190 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
193 ios_completed(td, &icd);
199 * Helper to handle the final sync of a file. Works just like the normal
200 * io path, just does everything sync.
202 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
204 struct io_u *io_u = __get_io_u(td);
205 struct io_completion_data icd;
211 io_u->ddir = DDIR_SYNC;
214 if (td_io_prep(td, io_u)) {
219 ret = td_io_queue(td, io_u);
221 td_verror(td, io_u->error);
226 ret = td_io_getevents(td, 1, td->cur_depth, NULL);
233 ios_completed(td, &icd);
235 td_verror(td, icd.error);
243 * The main verify engine. Runs over the writes we previusly submitted,
244 * reads the blocks back in, and checks the crc/md5 of the data.
246 void do_verify(struct thread_data *td)
249 struct io_u *io_u, *v_io_u = NULL;
250 struct io_completion_data icd;
255 * sync io first and invalidate cache, to make sure we really
258 for_each_file(td, f, i) {
260 file_invalidate_cache(td, f);
263 td_set_runstate(td, TD_VERIFYING);
269 gettimeofday(&t, NULL);
270 if (runtime_exceeded(td, &t))
273 io_u = __get_io_u(td);
277 if (get_next_verify(td, io_u)) {
282 f = get_next_file(td);
288 if (td_io_prep(td, io_u)) {
293 ret = td_io_queue(td, io_u);
295 td_verror(td, io_u->error);
301 * we have one pending to verify, do that while
302 * we are doing io on the next one
304 if (do_io_u_verify(td, &v_io_u))
307 ret = td_io_getevents(td, 1, 1, NULL);
314 v_io_u = td->io_ops->event(td, 0);
317 io_completed(td, v_io_u, &icd);
320 td_verror(td, icd.error);
321 put_io_u(td, v_io_u);
327 * if we can't submit more io, we need to verify now
329 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
334 do_io_u_verify(td, &v_io_u);
337 cleanup_pending_aio(td);
339 td_set_runstate(td, TD_RUNNING);
343 * Not really an io thread, all it does is burn CPU cycles in the specified
346 static void do_cpuio(struct thread_data *td)
349 int split = 100 / td->cpuload;
352 while (!td->terminate) {
353 gettimeofday(&e, NULL);
355 if (runtime_exceeded(td, &e))
361 usec_sleep(td, 10000);
368 * Main IO worker function. It retrieves io_u's to process and queues
369 * and reaps them, checking for rate and errors along the way.
371 static void do_io(struct thread_data *td)
373 struct io_completion_data icd;
379 td_set_runstate(td, TD_RUNNING);
381 while (td->this_io_bytes[td->ddir] < td->io_size) {
382 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
383 struct timespec *timeout;
390 f = get_next_file(td);
394 io_u = get_io_u(td, f);
398 memcpy(&s, &io_u->start_time, sizeof(s));
400 ret = td_io_queue(td, io_u);
402 td_verror(td, io_u->error);
407 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
409 if (td->cur_depth < td->iodepth) {
417 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
425 ios_completed(td, &icd);
427 td_verror(td, icd.error);
432 * the rate is batched for now, it should work for batches
433 * of completions except the very first one which may look
436 gettimeofday(&e, NULL);
437 usec = utime_since(&s, &e);
439 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
441 if (check_min_rate(td, &e)) {
443 terminate_threads(td->groupid);
444 td_verror(td, ENOMEM);
448 if (runtime_exceeded(td, &e))
452 usec_sleep(td, td->thinktime);
457 cleanup_pending_aio(td);
459 if (should_fsync(td) && td->end_fsync) {
460 td_set_runstate(td, TD_FSYNCING);
461 for_each_file(td, f, i)
467 static void cleanup_io_u(struct thread_data *td)
469 struct list_head *entry, *n;
472 list_for_each_safe(entry, n, &td->io_u_freelist) {
473 io_u = list_entry(entry, struct io_u, list);
475 list_del(&io_u->list);
483 * "randomly" fill the buffer contents
485 static void fill_rand_buf(struct io_u *io_u, int max_bs)
487 int *ptr = io_u->buf;
489 while ((void *) ptr - io_u->buf < max_bs) {
490 *ptr = rand() * 0x9e370001;
495 static int init_io_u(struct thread_data *td)
502 if (td->io_ops->flags & FIO_CPUIO)
505 if (td->io_ops->flags & FIO_SYNCIO)
508 max_units = td->iodepth;
510 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
511 td->orig_buffer_size = max_bs * max_units + MASK;
513 if (allocate_io_mem(td))
516 p = ALIGN(td->orig_buffer);
517 for (i = 0; i < max_units; i++) {
518 io_u = malloc(sizeof(*io_u));
519 memset(io_u, 0, sizeof(*io_u));
520 INIT_LIST_HEAD(&io_u->list);
522 io_u->buf = p + max_bs * i;
523 if (td_write(td) || td_rw(td))
524 fill_rand_buf(io_u, max_bs);
527 list_add(&io_u->list, &td->io_u_freelist);
533 static int switch_ioscheduler(struct thread_data *td)
535 char tmp[256], tmp2[128];
539 if (td->io_ops->flags & FIO_CPUIO)
542 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
544 f = fopen(tmp, "r+");
546 td_verror(td, errno);
553 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
554 if (ferror(f) || ret != 1) {
555 td_verror(td, errno);
563 * Read back and check that the selected scheduler is now the default.
565 ret = fread(tmp, 1, sizeof(tmp), f);
566 if (ferror(f) || ret < 0) {
567 td_verror(td, errno);
572 sprintf(tmp2, "[%s]", td->ioscheduler);
573 if (!strstr(tmp, tmp2)) {
574 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
575 td_verror(td, EINVAL);
584 static void clear_io_state(struct thread_data *td)
589 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
590 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
593 for_each_file(td, f, i) {
595 if (td->io_ops->flags & FIO_SYNCIO)
596 lseek(f->fd, SEEK_SET, 0);
599 memset(f->file_map, 0, f->num_maps * sizeof(long));
604 * Entry point for the thread based jobs. The process based jobs end up
605 * here as well, after a little setup.
607 static void *thread_main(void *data)
609 struct thread_data *td = data;
616 INIT_LIST_HEAD(&td->io_u_freelist);
617 INIT_LIST_HEAD(&td->io_u_busylist);
618 INIT_LIST_HEAD(&td->io_hist_list);
619 INIT_LIST_HEAD(&td->io_log_list);
624 if (fio_setaffinity(td) == -1) {
625 td_verror(td, errno);
636 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
637 td_verror(td, errno);
642 if (nice(td->nice) == -1) {
643 td_verror(td, errno);
647 if (init_random_state(td))
650 if (td->ioscheduler && switch_ioscheduler(td))
653 td_set_runstate(td, TD_INITIALIZED);
654 fio_sem_up(&startup_sem);
655 fio_sem_down(&td->mutex);
657 if (!td->create_serialize && setup_files(td))
660 gettimeofday(&td->epoch, NULL);
663 system(td->exec_prerun);
665 while (td->loops--) {
666 getrusage(RUSAGE_SELF, &td->ru_start);
667 gettimeofday(&td->start, NULL);
668 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
671 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
674 prune_io_piece_log(td);
676 if (td->io_ops->flags & FIO_CPUIO)
681 td->runtime[td->ddir] += mtime_since_now(&td->start);
682 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
683 td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
685 update_rusage_stat(td);
687 if (td->error || td->terminate)
690 if (td->verify == VERIFY_NONE)
694 gettimeofday(&td->start, NULL);
698 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
700 if (td->error || td->terminate)
705 finish_log(td, td->bw_log, "bw");
707 finish_log(td, td->slat_log, "slat");
709 finish_log(td, td->clat_log, "clat");
710 if (td->write_iolog_file)
711 write_iolog_close(td);
712 if (td->exec_postrun)
713 system(td->exec_postrun);
715 if (exitall_on_terminate)
716 terminate_threads(td->groupid);
722 td_set_runstate(td, TD_EXITED);
728 * We cannot pass the td data into a forked process, so attach the td and
729 * pass it to the thread worker.
731 static void *fork_main(int shmid, int offset)
733 struct thread_data *td;
736 data = shmat(shmid, NULL, 0);
737 if (data == (void *) -1) {
742 td = data + offset * sizeof(struct thread_data);
749 * Run over the job map and reap the threads that have exited, if any.
751 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
753 struct thread_data *td;
757 * reap exited threads (TD_EXITED -> TD_REAPED)
762 * ->io_ops is NULL for a thread that has closed its
765 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
768 if (td->runstate != TD_EXITED)
771 td_set_runstate(td, TD_REAPED);
773 if (td->use_thread) {
776 if (pthread_join(td->thread, (void *) &ret))
777 perror("thread_join");
779 waitpid(td->pid, NULL, 0);
782 (*m_rate) -= td->ratemin;
783 (*t_rate) -= td->rate;
786 if (*nr_running == cputhreads)
787 terminate_threads(TERMINATE_ALL);
791 * Main function for kicking off and reaping jobs, as needed.
793 static void run_threads(void)
795 struct thread_data *td;
797 int i, todo, nr_running, m_rate, t_rate, nr_started;
799 if (fio_pin_memory())
803 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
807 signal(SIGINT, sig_handler);
808 signal(SIGALRM, sig_handler);
810 todo = thread_number;
816 print_status_init(td->thread_number - 1);
820 if (!td->create_serialize)
824 * do file setup here so it happens sequentially,
825 * we don't want X number of threads getting their
826 * client data interspersed on disk
828 if (setup_files(td)) {
829 td_set_runstate(td, TD_REAPED);
837 struct thread_data *map[MAX_JOBS];
838 struct timeval this_start;
839 int this_jobs = 0, left;
842 * create threads (TD_NOT_CREATED -> TD_CREATED)
845 if (td->runstate != TD_NOT_CREATED)
849 * never got a chance to start, killed by other
850 * thread for some reason
857 if (td->start_delay) {
858 spent = mtime_since_genesis();
860 if (td->start_delay * 1000 > spent)
864 if (td->stonewall && (nr_started || nr_running))
868 * Set state to created. Thread will transition
869 * to TD_INITIALIZED when it's done setting up.
871 td_set_runstate(td, TD_CREATED);
872 map[this_jobs++] = td;
873 fio_sem_init(&startup_sem, 1);
876 if (td->use_thread) {
877 if (pthread_create(&td->thread, NULL, thread_main, td)) {
878 perror("thread_create");
883 fio_sem_down(&startup_sem);
885 fork_main(shm_id, i);
892 * Wait for the started threads to transition to
895 gettimeofday(&this_start, NULL);
898 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
903 for (i = 0; i < this_jobs; i++) {
907 if (td->runstate == TD_INITIALIZED) {
910 } else if (td->runstate >= TD_EXITED) {
914 nr_running++; /* work-around... */
920 log_err("fio: %d jobs failed to start\n", left);
921 for (i = 0; i < this_jobs; i++) {
925 kill(td->pid, SIGTERM);
931 * start created threads (TD_INITIALIZED -> TD_RUNNING).
934 if (td->runstate != TD_INITIALIZED)
937 td_set_runstate(td, TD_RUNNING);
940 m_rate += td->ratemin;
943 fio_sem_up(&td->mutex);
946 reap_threads(&nr_running, &t_rate, &m_rate);
953 reap_threads(&nr_running, &t_rate, &m_rate);
961 int main(int argc, char *argv[])
963 if (parse_options(argc, argv))
966 if (!thread_number) {
967 log_err("Nothing to do\n");
971 disk_util_timer_arm();