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 r = io_u_queued_complete(td, 0, NULL);
157 * now cancel remaining active events
159 if (td->io_ops->cancel) {
160 list_for_each_safe(entry, n, &td->io_u_busylist) {
161 io_u = list_entry(entry, struct io_u, list);
163 r = td->io_ops->cancel(td, io_u);
170 r = io_u_queued_complete(td, td->cur_depth, NULL);
174 * Helper to handle the final sync of a file. Works just like the normal
175 * io path, just does everything sync.
177 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
179 struct io_u *io_u = __get_io_u(td);
185 io_u->ddir = DDIR_SYNC;
188 if (td_io_prep(td, io_u)) {
194 ret = td_io_queue(td, io_u);
196 td_verror(td, io_u->error);
199 } else if (ret == FIO_Q_QUEUED) {
200 if (io_u_queued_complete(td, 1, NULL) < 0)
202 } else if (ret == FIO_Q_COMPLETED) {
204 td_verror(td, io_u->error);
208 if (io_u_sync_complete(td, io_u, NULL) < 0)
210 } else if (ret == FIO_Q_BUSY) {
211 if (td_io_commit(td))
220 * The main verify engine. Runs over the writes we previusly submitted,
221 * reads the blocks back in, and checks the crc/md5 of the data.
223 static void do_verify(struct thread_data *td)
227 int ret, i, min_events;
230 * sync io first and invalidate cache, to make sure we really
233 for_each_file(td, f, i) {
234 if (fio_io_sync(td, f))
236 if (file_invalidate_cache(td, f))
243 td_set_runstate(td, TD_VERIFYING);
246 while (!td->terminate) {
247 io_u = __get_io_u(td);
251 if (runtime_exceeded(td, &io_u->start_time))
254 if (get_next_verify(td, io_u))
257 if (td_io_prep(td, io_u))
261 ret = td_io_queue(td, io_u);
264 case FIO_Q_COMPLETED:
267 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
268 int bytes = io_u->xfer_buflen - io_u->resid;
270 io_u->xfer_buflen = io_u->resid;
271 io_u->xfer_buf += bytes;
274 ret = io_u_sync_complete(td, io_u, verify_io_u);
281 requeue_io_u(td, &io_u);
282 ret = td_io_commit(td);
290 if (ret < 0 || td->error)
294 * if we can queue more, do so. but check if there are
295 * completed io_u's first.
298 if (queue_full(td) || ret == FIO_Q_BUSY) {
301 if (td->cur_depth > td->iodepth_low)
302 min_events = td->cur_depth - td->iodepth_low;
306 * Reap required number of io units, if any, and do the
307 * verification on them through the callback handler
309 if (io_u_queued_complete(td, min_events, verify_io_u) < 0)
314 cleanup_pending_aio(td);
316 td_set_runstate(td, TD_RUNNING);
320 * Not really an io thread, all it does is burn CPU cycles in the specified
323 static void do_cpuio(struct thread_data *td)
326 int split = 100 / td->cpuload;
329 while (!td->terminate) {
330 fio_gettime(&e, NULL);
332 if (runtime_exceeded(td, &e))
338 usec_sleep(td, 10000);
345 * Main IO worker function. It retrieves io_u's to process and queues
346 * and reaps them, checking for rate and errors along the way.
348 static void do_io(struct thread_data *td)
354 td_set_runstate(td, TD_RUNNING);
356 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
357 struct timeval comp_time;
369 memcpy(&s, &io_u->start_time, sizeof(s));
371 if (runtime_exceeded(td, &s)) {
376 ret = td_io_queue(td, io_u);
379 case FIO_Q_COMPLETED:
384 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
385 int bytes = io_u->xfer_buflen - io_u->resid;
387 io_u->xfer_buflen = io_u->resid;
388 io_u->xfer_buf += bytes;
391 fio_gettime(&comp_time, NULL);
392 bytes_done = io_u_sync_complete(td, io_u, NULL);
398 * if the engine doesn't have a commit hook,
399 * the io_u is really queued. if it does have such
400 * a hook, it has to call io_u_queued() itself.
402 if (td->io_ops->commit == NULL)
403 io_u_queued(td, io_u);
406 requeue_io_u(td, &io_u);
407 ret = td_io_commit(td);
415 if (ret < 0 || td->error)
419 * See if we need to complete some commands
421 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
423 if (queue_full(td) || ret == FIO_Q_BUSY) {
426 if (td->cur_depth > td->iodepth_low)
427 min_evts = td->cur_depth - td->iodepth_low;
430 fio_gettime(&comp_time, NULL);
431 bytes_done = io_u_queued_complete(td, min_evts, NULL);
440 * the rate is batched for now, it should work for batches
441 * of completions except the very first one which may look
444 usec = utime_since(&s, &comp_time);
446 rate_throttle(td, usec, bytes_done, td->ddir);
448 if (check_min_rate(td, &comp_time)) {
449 if (exitall_on_terminate)
450 terminate_threads(td->groupid, 0);
451 td_verror(td, ENODATA);
456 unsigned long long b;
458 b = td->io_blocks[0] + td->io_blocks[1];
459 if (!(b % td->thinktime_blocks)) {
462 if (td->thinktime_spin)
463 __usec_sleep(td->thinktime_spin);
465 left = td->thinktime - td->thinktime_spin;
467 usec_sleep(td, left);
476 cleanup_pending_aio(td);
478 if (should_fsync(td) && td->end_fsync) {
479 td_set_runstate(td, TD_FSYNCING);
480 for_each_file(td, f, i)
486 static void cleanup_io_u(struct thread_data *td)
488 struct list_head *entry, *n;
491 list_for_each_safe(entry, n, &td->io_u_freelist) {
492 io_u = list_entry(entry, struct io_u, list);
494 list_del(&io_u->list);
502 * "randomly" fill the buffer contents
504 static void fill_rand_buf(struct io_u *io_u, int max_bs)
506 int *ptr = io_u->buf;
508 while ((void *) ptr - io_u->buf < max_bs) {
509 *ptr = rand() * 0x9e370001;
514 static int init_io_u(struct thread_data *td)
521 if (td->io_ops->flags & FIO_CPUIO)
524 if (td->io_ops->flags & FIO_SYNCIO)
527 max_units = td->iodepth;
529 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
530 td->orig_buffer_size = max_bs * max_units;
532 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
533 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
535 td->orig_buffer_size += page_mask;
537 if (allocate_io_mem(td))
540 p = ALIGN(td->orig_buffer);
541 for (i = 0; i < max_units; i++) {
542 io_u = malloc(sizeof(*io_u));
543 memset(io_u, 0, sizeof(*io_u));
544 INIT_LIST_HEAD(&io_u->list);
546 io_u->buf = p + max_bs * i;
547 if (td_write(td) || td_rw(td))
548 fill_rand_buf(io_u, max_bs);
551 list_add(&io_u->list, &td->io_u_freelist);
557 static int switch_ioscheduler(struct thread_data *td)
559 char tmp[256], tmp2[128];
563 if (td->io_ops->flags & FIO_CPUIO)
566 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
568 f = fopen(tmp, "r+");
570 td_verror(td, errno);
577 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
578 if (ferror(f) || ret != 1) {
579 td_verror(td, errno);
587 * Read back and check that the selected scheduler is now the default.
589 ret = fread(tmp, 1, sizeof(tmp), f);
590 if (ferror(f) || ret < 0) {
591 td_verror(td, errno);
596 sprintf(tmp2, "[%s]", td->ioscheduler);
597 if (!strstr(tmp, tmp2)) {
598 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
599 td_verror(td, EINVAL);
608 static void clear_io_state(struct thread_data *td)
613 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
614 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
617 td->last_was_sync = 0;
619 for_each_file(td, f, i) {
620 f->last_completed_pos = 0;
623 if (td->io_ops->flags & FIO_SYNCIO)
624 lseek(f->fd, SEEK_SET, 0);
627 memset(f->file_map, 0, f->num_maps * sizeof(long));
632 * Entry point for the thread based jobs. The process based jobs end up
633 * here as well, after a little setup.
635 static void *thread_main(void *data)
637 unsigned long long runtime[2];
638 struct thread_data *td = data;
645 INIT_LIST_HEAD(&td->io_u_freelist);
646 INIT_LIST_HEAD(&td->io_u_busylist);
647 INIT_LIST_HEAD(&td->io_u_requeues);
648 INIT_LIST_HEAD(&td->io_hist_list);
649 INIT_LIST_HEAD(&td->io_log_list);
654 if (fio_setaffinity(td) == -1) {
655 td_verror(td, errno);
663 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
664 td_verror(td, errno);
669 if (nice(td->nice) == -1) {
670 td_verror(td, errno);
674 if (init_random_state(td))
677 if (td->ioscheduler && switch_ioscheduler(td))
680 td_set_runstate(td, TD_INITIALIZED);
681 fio_sem_up(&startup_sem);
682 fio_sem_down(&td->mutex);
684 if (!td->create_serialize && setup_files(td))
690 * Do this late, as some IO engines would like to have the
691 * files setup prior to initializing structures.
696 if (td->exec_prerun) {
697 if (system(td->exec_prerun) < 0)
701 fio_gettime(&td->epoch, NULL);
702 getrusage(RUSAGE_SELF, &td->ts.ru_start);
704 runtime[0] = runtime[1] = 0;
705 while (td->loops--) {
706 fio_gettime(&td->start, NULL);
707 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
710 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
713 prune_io_piece_log(td);
715 if (td->io_ops->flags & FIO_CPUIO)
720 runtime[td->ddir] += utime_since_now(&td->start);
721 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
722 runtime[td->ddir ^ 1] = runtime[td->ddir];
724 if (td->error || td->terminate)
727 if (td->verify == VERIFY_NONE)
731 fio_gettime(&td->start, NULL);
735 runtime[DDIR_READ] += utime_since_now(&td->start);
737 if (td->error || td->terminate)
741 update_rusage_stat(td);
742 fio_gettime(&td->end_time, NULL);
743 td->runtime[0] = runtime[0] / 1000;
744 td->runtime[1] = runtime[1] / 1000;
747 finish_log(td, td->ts.bw_log, "bw");
749 finish_log(td, td->ts.slat_log, "slat");
751 finish_log(td, td->ts.clat_log, "clat");
752 if (td->write_iolog_file)
753 write_iolog_close(td);
754 if (td->exec_postrun) {
755 if (system(td->exec_postrun) < 0)
756 log_err("fio: postrun %s failed\n", td->exec_postrun);
759 if (exitall_on_terminate)
760 terminate_threads(td->groupid, 0);
764 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
768 td_set_runstate(td, TD_EXITED);
769 return (void *) (unsigned long) td->error;
773 * We cannot pass the td data into a forked process, so attach the td and
774 * pass it to the thread worker.
776 static int fork_main(int shmid, int offset)
778 struct thread_data *td;
781 data = shmat(shmid, NULL, 0);
782 if (data == (void *) -1) {
789 td = data + offset * sizeof(struct thread_data);
790 ret = thread_main(td);
792 return (int) (unsigned long) ret;
796 * Run over the job map and reap the threads that have exited, if any.
798 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
800 struct thread_data *td;
801 int i, cputhreads, pending, status, ret;
804 * reap exited threads (TD_EXITED -> TD_REAPED)
806 pending = cputhreads = 0;
809 * ->io_ops is NULL for a thread that has closed its
812 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
815 if (td->runstate < TD_EXITED) {
817 * check if someone quit or got killed in an unusual way
819 ret = waitpid(td->pid, &status, WNOHANG);
822 else if ((ret == td->pid) && WIFSIGNALED(status)) {
823 int sig = WTERMSIG(status);
825 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
826 td_set_runstate(td, TD_REAPED);
831 if (td->runstate != TD_EXITED) {
832 if (td->runstate < TD_RUNNING)
841 td_set_runstate(td, TD_REAPED);
843 if (td->use_thread) {
846 if (pthread_join(td->thread, (void *) &ret))
847 perror("thread_join");
851 ret = waitpid(td->pid, &status, 0);
854 else if (WIFEXITED(status) && WEXITSTATUS(status)) {
862 (*m_rate) -= td->ratemin;
863 (*t_rate) -= td->rate;
866 if (*nr_running == cputhreads && !pending)
867 terminate_threads(TERMINATE_ALL, 0);
871 * Main function for kicking off and reaping jobs, as needed.
873 static void run_threads(void)
875 struct thread_data *td;
877 int i, todo, nr_running, m_rate, t_rate, nr_started;
879 if (fio_pin_memory())
883 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
887 signal(SIGINT, sig_handler);
888 signal(SIGALRM, sig_handler);
890 todo = thread_number;
896 print_status_init(td->thread_number - 1);
898 if (!td->create_serialize) {
904 * do file setup here so it happens sequentially,
905 * we don't want X number of threads getting their
906 * client data interspersed on disk
908 if (setup_files(td)) {
911 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
912 td_set_runstate(td, TD_REAPED);
920 struct thread_data *map[MAX_JOBS];
921 struct timeval this_start;
922 int this_jobs = 0, left;
925 * create threads (TD_NOT_CREATED -> TD_CREATED)
928 if (td->runstate != TD_NOT_CREATED)
932 * never got a chance to start, killed by other
933 * thread for some reason
940 if (td->start_delay) {
941 spent = mtime_since_genesis();
943 if (td->start_delay * 1000 > spent)
947 if (td->stonewall && (nr_started || nr_running))
951 * Set state to created. Thread will transition
952 * to TD_INITIALIZED when it's done setting up.
954 td_set_runstate(td, TD_CREATED);
955 map[this_jobs++] = td;
956 fio_sem_init(&startup_sem, 1);
959 if (td->use_thread) {
960 if (pthread_create(&td->thread, NULL, thread_main, td)) {
961 perror("thread_create");
966 fio_sem_down(&startup_sem);
968 int ret = fork_main(shm_id, i);
976 * Wait for the started threads to transition to
979 fio_gettime(&this_start, NULL);
981 while (left && !fio_abort) {
982 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
987 for (i = 0; i < this_jobs; i++) {
991 if (td->runstate == TD_INITIALIZED) {
994 } else if (td->runstate >= TD_EXITED) {
998 nr_running++; /* work-around... */
1004 log_err("fio: %d jobs failed to start\n", left);
1005 for (i = 0; i < this_jobs; i++) {
1009 kill(td->pid, SIGTERM);
1015 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1017 for_each_td(td, i) {
1018 if (td->runstate != TD_INITIALIZED)
1021 td_set_runstate(td, TD_RUNNING);
1024 m_rate += td->ratemin;
1027 fio_sem_up(&td->mutex);
1030 reap_threads(&nr_running, &t_rate, &m_rate);
1036 while (nr_running) {
1037 reap_threads(&nr_running, &t_rate, &m_rate);
1045 int main(int argc, char *argv[])
1050 * We need locale for number printing, if it isn't set then just
1051 * go with the US format.
1053 if (!getenv("LC_NUMERIC"))
1054 setlocale(LC_NUMERIC, "en_US");
1056 if (parse_options(argc, argv))
1059 if (!thread_number) {
1060 log_err("Nothing to do\n");
1064 ps = sysconf(_SC_PAGESIZE);
1066 log_err("Failed to get page size\n");
1073 setup_log(&agg_io_log[DDIR_READ]);
1074 setup_log(&agg_io_log[DDIR_WRITE]);
1077 disk_util_timer_arm();
1084 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1085 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");