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
40 static unsigned long page_mask;
42 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
45 int thread_number = 0;
49 static volatile int startup_sem;
50 static volatile int fio_abort;
51 static int exit_value;
53 struct io_log *agg_io_log[2];
55 #define TERMINATE_ALL (-1)
56 #define JOB_START_TIMEOUT (5 * 1000)
58 static inline void td_set_runstate(struct thread_data *td, int runstate)
60 td->runstate = runstate;
63 static void terminate_threads(int group_id, int forced_kill)
65 struct thread_data *td;
69 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
73 td_set_runstate(td, TD_EXITED);
78 static void sig_handler(int sig)
83 disk_util_timer_arm();
84 print_thread_status();
87 printf("\nfio: terminating on signal %d\n", sig);
89 terminate_threads(TERMINATE_ALL, 0);
95 * Check if we are above the minimum rate given.
97 static int check_min_rate(struct thread_data *td, struct timeval *now)
104 * allow a 2 second settle period in the beginning
106 if (mtime_since(&td->start, now) < 2000)
110 * if rate blocks is set, sample is running
112 if (td->rate_bytes) {
113 spent = mtime_since(&td->lastrate, now);
114 if (spent < td->ratecycle)
117 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
118 if (rate < td->ratemin) {
119 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
124 td->rate_bytes = td->this_io_bytes[ddir];
125 memcpy(&td->lastrate, now, sizeof(*now));
129 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
133 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
140 * When job exits, we can cancel the in-flight IO if we are using async
141 * io. Attempt to do so.
143 static void cleanup_pending_aio(struct thread_data *td)
145 struct list_head *entry, *n;
150 * get immediately available events, if any
152 io_u_queued_complete(td, 0, NULL);
155 * now cancel remaining active events
157 if (td->io_ops->cancel) {
158 list_for_each_safe(entry, n, &td->io_u_busylist) {
159 io_u = list_entry(entry, struct io_u, list);
161 r = td->io_ops->cancel(td, io_u);
168 io_u_queued_complete(td, td->cur_depth, NULL);
172 * Helper to handle the final sync of a file. Works just like the normal
173 * io path, just does everything sync.
175 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
177 struct io_u *io_u = __get_io_u(td);
183 io_u->ddir = DDIR_SYNC;
186 if (td_io_prep(td, io_u)) {
192 ret = td_io_queue(td, io_u);
194 td_verror(td, io_u->error);
197 } else if (ret == FIO_Q_QUEUED) {
198 if (io_u_queued_complete(td, 1, NULL))
200 } else if (ret == FIO_Q_COMPLETED) {
202 td_verror(td, io_u->error);
206 io_u_sync_complete(td, io_u, NULL);
207 } else if (ret == FIO_Q_BUSY) {
208 if (td_io_commit(td))
217 * The main verify engine. Runs over the writes we previusly submitted,
218 * reads the blocks back in, and checks the crc/md5 of the data.
220 static void do_verify(struct thread_data *td)
224 int ret, i, min_events;
227 * sync io first and invalidate cache, to make sure we really
230 for_each_file(td, f, i) {
232 file_invalidate_cache(td, f);
235 td_set_runstate(td, TD_VERIFYING);
238 while (!td->terminate) {
239 io_u = __get_io_u(td);
243 if (runtime_exceeded(td, &io_u->start_time))
246 if (get_next_verify(td, io_u))
249 if (td_io_prep(td, io_u))
253 ret = td_io_queue(td, io_u);
256 case FIO_Q_COMPLETED:
259 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
260 int bytes = io_u->xfer_buflen - io_u->resid;
262 io_u->xfer_buflen = io_u->resid;
263 io_u->xfer_buf += bytes;
266 ret = io_u_sync_complete(td, io_u, verify_io_u);
273 requeue_io_u(td, &io_u);
274 ret = td_io_commit(td);
282 if (ret < 0 || td->error)
286 * if we can queue more, do so. but check if there are
287 * completed io_u's first.
290 if (queue_full(td) || ret == FIO_Q_BUSY) {
293 if (td->cur_depth > td->iodepth_low)
294 min_events = td->cur_depth - td->iodepth_low;
298 * Reap required number of io units, if any, and do the
299 * verification on them through the callback handler
301 if (io_u_queued_complete(td, min_events, verify_io_u))
309 cleanup_pending_aio(td);
311 td_set_runstate(td, TD_RUNNING);
315 * Not really an io thread, all it does is burn CPU cycles in the specified
318 static void do_cpuio(struct thread_data *td)
321 int split = 100 / td->cpuload;
324 while (!td->terminate) {
325 fio_gettime(&e, NULL);
327 if (runtime_exceeded(td, &e))
333 usec_sleep(td, 10000);
340 * Main IO worker function. It retrieves io_u's to process and queues
341 * and reaps them, checking for rate and errors along the way.
343 static void do_io(struct thread_data *td)
349 td_set_runstate(td, TD_RUNNING);
351 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
352 struct timeval comp_time;
364 memcpy(&s, &io_u->start_time, sizeof(s));
366 if (runtime_exceeded(td, &s)) {
371 ret = td_io_queue(td, io_u);
374 case FIO_Q_COMPLETED:
379 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
380 int bytes = io_u->xfer_buflen - io_u->resid;
382 io_u->xfer_buflen = io_u->resid;
383 io_u->xfer_buf += bytes;
386 fio_gettime(&comp_time, NULL);
387 bytes_done = io_u_sync_complete(td, io_u, NULL);
393 * if the engine doesn't have a commit hook,
394 * the io_u is really queued. if it does have such
395 * a hook, it has to call io_u_queued() itself.
397 if (td->io_ops->commit == NULL)
398 io_u_queued(td, io_u);
401 requeue_io_u(td, &io_u);
402 ret = td_io_commit(td);
410 if (ret < 0 || td->error)
414 * See if we need to complete some commands
416 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
418 if (queue_full(td) || ret == FIO_Q_BUSY) {
421 if (td->cur_depth > td->iodepth_low)
422 min_evts = td->cur_depth - td->iodepth_low;
425 fio_gettime(&comp_time, NULL);
426 bytes_done = io_u_queued_complete(td, min_evts, NULL);
435 * the rate is batched for now, it should work for batches
436 * of completions except the very first one which may look
439 usec = utime_since(&s, &comp_time);
441 rate_throttle(td, usec, bytes_done, td->ddir);
443 if (check_min_rate(td, &comp_time)) {
444 if (exitall_on_terminate)
445 terminate_threads(td->groupid, 0);
446 td_verror(td, ENODATA);
451 unsigned long long b;
453 b = td->io_blocks[0] + td->io_blocks[1];
454 if (!(b % td->thinktime_blocks)) {
457 if (td->thinktime_spin)
458 __usec_sleep(td->thinktime_spin);
460 left = td->thinktime - td->thinktime_spin;
462 usec_sleep(td, left);
471 cleanup_pending_aio(td);
473 if (should_fsync(td) && td->end_fsync) {
474 td_set_runstate(td, TD_FSYNCING);
475 for_each_file(td, f, i)
481 static void cleanup_io_u(struct thread_data *td)
483 struct list_head *entry, *n;
486 list_for_each_safe(entry, n, &td->io_u_freelist) {
487 io_u = list_entry(entry, struct io_u, list);
489 list_del(&io_u->list);
497 * "randomly" fill the buffer contents
499 static void fill_rand_buf(struct io_u *io_u, int max_bs)
501 int *ptr = io_u->buf;
503 while ((void *) ptr - io_u->buf < max_bs) {
504 *ptr = rand() * 0x9e370001;
509 static int init_io_u(struct thread_data *td)
516 if (td->io_ops->flags & FIO_CPUIO)
519 if (td->io_ops->flags & FIO_SYNCIO)
522 max_units = td->iodepth;
524 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
525 td->orig_buffer_size = max_bs * max_units;
527 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
528 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
530 td->orig_buffer_size += page_mask;
532 if (allocate_io_mem(td))
535 p = ALIGN(td->orig_buffer);
536 for (i = 0; i < max_units; i++) {
537 io_u = malloc(sizeof(*io_u));
538 memset(io_u, 0, sizeof(*io_u));
539 INIT_LIST_HEAD(&io_u->list);
541 io_u->buf = p + max_bs * i;
542 if (td_write(td) || td_rw(td))
543 fill_rand_buf(io_u, max_bs);
546 list_add(&io_u->list, &td->io_u_freelist);
552 static int switch_ioscheduler(struct thread_data *td)
554 char tmp[256], tmp2[128];
558 if (td->io_ops->flags & FIO_CPUIO)
561 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
563 f = fopen(tmp, "r+");
565 td_verror(td, errno);
572 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
573 if (ferror(f) || ret != 1) {
574 td_verror(td, errno);
582 * Read back and check that the selected scheduler is now the default.
584 ret = fread(tmp, 1, sizeof(tmp), f);
585 if (ferror(f) || ret < 0) {
586 td_verror(td, errno);
591 sprintf(tmp2, "[%s]", td->ioscheduler);
592 if (!strstr(tmp, tmp2)) {
593 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
594 td_verror(td, EINVAL);
603 static void clear_io_state(struct thread_data *td)
608 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
609 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
612 td->last_was_sync = 0;
614 for_each_file(td, f, i) {
615 f->last_completed_pos = 0;
618 if (td->io_ops->flags & FIO_SYNCIO)
619 lseek(f->fd, SEEK_SET, 0);
622 memset(f->file_map, 0, f->num_maps * sizeof(long));
627 * Entry point for the thread based jobs. The process based jobs end up
628 * here as well, after a little setup.
630 static void *thread_main(void *data)
632 unsigned long long runtime[2];
633 struct thread_data *td = data;
640 INIT_LIST_HEAD(&td->io_u_freelist);
641 INIT_LIST_HEAD(&td->io_u_busylist);
642 INIT_LIST_HEAD(&td->io_u_requeues);
643 INIT_LIST_HEAD(&td->io_hist_list);
644 INIT_LIST_HEAD(&td->io_log_list);
649 if (fio_setaffinity(td) == -1) {
650 td_verror(td, errno);
658 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
659 td_verror(td, errno);
664 if (nice(td->nice) == -1) {
665 td_verror(td, errno);
669 if (init_random_state(td))
672 if (td->ioscheduler && switch_ioscheduler(td))
675 td_set_runstate(td, TD_INITIALIZED);
676 fio_sem_up(&startup_sem);
677 fio_sem_down(&td->mutex);
679 if (!td->create_serialize && setup_files(td))
685 * Do this late, as some IO engines would like to have the
686 * files setup prior to initializing structures.
691 if (td->exec_prerun) {
692 if (system(td->exec_prerun) < 0)
696 fio_gettime(&td->epoch, NULL);
697 getrusage(RUSAGE_SELF, &td->ts.ru_start);
699 runtime[0] = runtime[1] = 0;
700 while (td->loops--) {
701 fio_gettime(&td->start, NULL);
702 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
705 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
708 prune_io_piece_log(td);
710 if (td->io_ops->flags & FIO_CPUIO)
715 runtime[td->ddir] += utime_since_now(&td->start);
716 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
717 runtime[td->ddir ^ 1] = runtime[td->ddir];
719 if (td->error || td->terminate)
722 if (td->verify == VERIFY_NONE)
726 fio_gettime(&td->start, NULL);
730 runtime[DDIR_READ] += utime_since_now(&td->start);
732 if (td->error || td->terminate)
736 update_rusage_stat(td);
737 fio_gettime(&td->end_time, NULL);
738 td->runtime[0] = runtime[0] / 1000;
739 td->runtime[1] = runtime[1] / 1000;
742 finish_log(td, td->ts.bw_log, "bw");
744 finish_log(td, td->ts.slat_log, "slat");
746 finish_log(td, td->ts.clat_log, "clat");
747 if (td->write_iolog_file)
748 write_iolog_close(td);
749 if (td->exec_postrun) {
750 if (system(td->exec_postrun) < 0)
751 log_err("fio: postrun %s failed\n", td->exec_postrun);
754 if (exitall_on_terminate)
755 terminate_threads(td->groupid, 0);
759 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
763 td_set_runstate(td, TD_EXITED);
764 return (void *) td->error;
768 * We cannot pass the td data into a forked process, so attach the td and
769 * pass it to the thread worker.
771 static int fork_main(int shmid, int offset)
773 struct thread_data *td;
776 data = shmat(shmid, NULL, 0);
777 if (data == (void *) -1) {
784 td = data + offset * sizeof(struct thread_data);
785 ret = thread_main(td);
791 * Run over the job map and reap the threads that have exited, if any.
793 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
795 struct thread_data *td;
796 int i, cputhreads, pending, status, ret;
799 * reap exited threads (TD_EXITED -> TD_REAPED)
801 pending = cputhreads = 0;
804 * ->io_ops is NULL for a thread that has closed its
807 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
810 if (td->runstate < TD_EXITED) {
812 * check if someone quit or got killed in an unusual way
814 ret = waitpid(td->pid, &status, WNOHANG);
817 else if ((ret == td->pid) && WIFSIGNALED(status)) {
818 int sig = WTERMSIG(status);
820 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
821 td_set_runstate(td, TD_REAPED);
826 if (td->runstate != TD_EXITED) {
827 if (td->runstate < TD_RUNNING)
836 td_set_runstate(td, TD_REAPED);
838 if (td->use_thread) {
841 if (pthread_join(td->thread, (void *) &ret))
842 perror("thread_join");
846 ret = waitpid(td->pid, &status, 0);
849 else if (WIFEXITED(status) && WEXITSTATUS(status)) {
857 (*m_rate) -= td->ratemin;
858 (*t_rate) -= td->rate;
861 if (*nr_running == cputhreads && !pending)
862 terminate_threads(TERMINATE_ALL, 0);
866 * Main function for kicking off and reaping jobs, as needed.
868 static void run_threads(void)
870 struct thread_data *td;
872 int i, todo, nr_running, m_rate, t_rate, nr_started;
874 if (fio_pin_memory())
878 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
882 signal(SIGINT, sig_handler);
883 signal(SIGALRM, sig_handler);
885 todo = thread_number;
891 print_status_init(td->thread_number - 1);
893 if (!td->create_serialize) {
899 * do file setup here so it happens sequentially,
900 * we don't want X number of threads getting their
901 * client data interspersed on disk
903 if (setup_files(td)) {
906 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
907 td_set_runstate(td, TD_REAPED);
915 struct thread_data *map[MAX_JOBS];
916 struct timeval this_start;
917 int this_jobs = 0, left;
920 * create threads (TD_NOT_CREATED -> TD_CREATED)
923 if (td->runstate != TD_NOT_CREATED)
927 * never got a chance to start, killed by other
928 * thread for some reason
935 if (td->start_delay) {
936 spent = mtime_since_genesis();
938 if (td->start_delay * 1000 > spent)
942 if (td->stonewall && (nr_started || nr_running))
946 * Set state to created. Thread will transition
947 * to TD_INITIALIZED when it's done setting up.
949 td_set_runstate(td, TD_CREATED);
950 map[this_jobs++] = td;
951 fio_sem_init(&startup_sem, 1);
954 if (td->use_thread) {
955 if (pthread_create(&td->thread, NULL, thread_main, td)) {
956 perror("thread_create");
961 fio_sem_down(&startup_sem);
963 int ret = fork_main(shm_id, i);
971 * Wait for the started threads to transition to
974 fio_gettime(&this_start, NULL);
976 while (left && !fio_abort) {
977 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
982 for (i = 0; i < this_jobs; i++) {
986 if (td->runstate == TD_INITIALIZED) {
989 } else if (td->runstate >= TD_EXITED) {
993 nr_running++; /* work-around... */
999 log_err("fio: %d jobs failed to start\n", left);
1000 for (i = 0; i < this_jobs; i++) {
1004 kill(td->pid, SIGTERM);
1010 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1012 for_each_td(td, i) {
1013 if (td->runstate != TD_INITIALIZED)
1016 td_set_runstate(td, TD_RUNNING);
1019 m_rate += td->ratemin;
1022 fio_sem_up(&td->mutex);
1025 reap_threads(&nr_running, &t_rate, &m_rate);
1031 while (nr_running) {
1032 reap_threads(&nr_running, &t_rate, &m_rate);
1040 int main(int argc, char *argv[])
1045 * We need locale for number printing, if it isn't set then just
1046 * go with the US format.
1048 if (!getenv("LC_NUMERIC"))
1049 setlocale(LC_NUMERIC, "en_US");
1051 if (parse_options(argc, argv))
1054 if (!thread_number) {
1055 log_err("Nothing to do\n");
1059 ps = sysconf(_SC_PAGESIZE);
1061 log_err("Failed to get page size\n");
1068 setup_log(&agg_io_log[DDIR_READ]);
1069 setup_log(&agg_io_log[DDIR_WRITE]);
1072 disk_util_timer_arm();
1079 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1080 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");