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)
139 static struct fio_file *get_next_file(struct thread_data *td)
141 unsigned int old_next_file = td->next_file;
145 f = &td->files[td->next_file];
148 if (td->next_file >= td->nr_files)
155 } while (td->next_file != old_next_file);
161 * When job exits, we can cancel the in-flight IO if we are using async
162 * io. Attempt to do so.
164 static void cleanup_pending_aio(struct thread_data *td)
166 struct list_head *entry, *n;
171 * get immediately available events, if any
173 io_u_queued_complete(td, 0, NULL);
176 * now cancel remaining active events
178 if (td->io_ops->cancel) {
179 list_for_each_safe(entry, n, &td->io_u_busylist) {
180 io_u = list_entry(entry, struct io_u, list);
182 r = td->io_ops->cancel(td, io_u);
189 io_u_queued_complete(td, td->cur_depth, NULL);
193 * Helper to handle the final sync of a file. Works just like the normal
194 * io path, just does everything sync.
196 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
198 struct io_u *io_u = __get_io_u(td);
204 io_u->ddir = DDIR_SYNC;
207 if (td_io_prep(td, io_u)) {
212 ret = td_io_queue(td, io_u);
214 td_verror(td, io_u->error);
217 } else if (ret == FIO_Q_QUEUED) {
218 if (io_u_queued_complete(td, 1, NULL))
220 } else if (ret == FIO_Q_COMPLETED) {
222 td_verror(td, io_u->error);
226 io_u_sync_complete(td, io_u, NULL);
233 * The main verify engine. Runs over the writes we previusly submitted,
234 * reads the blocks back in, and checks the crc/md5 of the data.
236 static void do_verify(struct thread_data *td)
240 int ret, i, min_events;
243 * sync io first and invalidate cache, to make sure we really
246 for_each_file(td, f, i) {
248 file_invalidate_cache(td, f);
251 td_set_runstate(td, TD_VERIFYING);
254 while (!td->terminate) {
255 io_u = __get_io_u(td);
259 if (runtime_exceeded(td, &io_u->start_time))
262 if (get_next_verify(td, io_u))
265 if (td_io_prep(td, io_u))
269 ret = td_io_queue(td, io_u);
272 case FIO_Q_COMPLETED:
275 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
276 int bytes = io_u->xfer_buflen - io_u->resid;
278 io_u->xfer_buflen = io_u->resid;
279 io_u->xfer_buf += bytes;
282 if (io_u_sync_complete(td, io_u, verify_io_u))
297 * if we can queue more, do so. but check if there are
298 * completed io_u's first.
305 * Reap required number of io units, if any, and do the
306 * verification on them through the callback handler
308 if (io_u_queued_complete(td, min_events, verify_io_u))
316 cleanup_pending_aio(td);
318 td_set_runstate(td, TD_RUNNING);
322 * Not really an io thread, all it does is burn CPU cycles in the specified
325 static void do_cpuio(struct thread_data *td)
328 int split = 100 / td->cpuload;
331 while (!td->terminate) {
332 fio_gettime(&e, NULL);
334 if (runtime_exceeded(td, &e))
340 usec_sleep(td, 10000);
347 * Main IO worker function. It retrieves io_u's to process and queues
348 * and reaps them, checking for rate and errors along the way.
350 static void do_io(struct thread_data *td)
357 td_set_runstate(td, TD_RUNNING);
359 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
360 struct timeval comp_time;
368 f = get_next_file(td);
372 io_u = get_io_u(td, f);
376 memcpy(&s, &io_u->start_time, sizeof(s));
378 if (runtime_exceeded(td, &s)) {
383 ret = td_io_queue(td, io_u);
386 case FIO_Q_COMPLETED:
391 if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
392 int bytes = io_u->xfer_buflen - io_u->resid;
394 io_u->xfer_buflen = io_u->resid;
395 io_u->xfer_buf += bytes;
398 fio_gettime(&comp_time, NULL);
399 bytes_done = io_u_sync_complete(td, io_u, NULL);
412 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
415 * See if we need to complete some commands
417 if (ret == FIO_Q_QUEUED) {
422 fio_gettime(&comp_time, NULL);
423 bytes_done = io_u_queued_complete(td, min_evts, NULL);
432 * the rate is batched for now, it should work for batches
433 * of completions except the very first one which may look
436 usec = utime_since(&s, &comp_time);
438 rate_throttle(td, usec, bytes_done, td->ddir);
440 if (check_min_rate(td, &comp_time)) {
441 if (exitall_on_terminate)
442 terminate_threads(td->groupid, 0);
443 td_verror(td, ENODATA);
448 unsigned long long b;
450 b = td->io_blocks[0] + td->io_blocks[1];
451 if (!(b % td->thinktime_blocks)) {
454 if (td->thinktime_spin)
455 __usec_sleep(td->thinktime_spin);
457 left = td->thinktime - td->thinktime_spin;
459 usec_sleep(td, left);
466 cleanup_pending_aio(td);
468 if (should_fsync(td) && td->end_fsync) {
469 td_set_runstate(td, TD_FSYNCING);
470 for_each_file(td, f, i)
476 static void cleanup_io_u(struct thread_data *td)
478 struct list_head *entry, *n;
481 list_for_each_safe(entry, n, &td->io_u_freelist) {
482 io_u = list_entry(entry, struct io_u, list);
484 list_del(&io_u->list);
492 * "randomly" fill the buffer contents
494 static void fill_rand_buf(struct io_u *io_u, int max_bs)
496 int *ptr = io_u->buf;
498 while ((void *) ptr - io_u->buf < max_bs) {
499 *ptr = rand() * 0x9e370001;
504 static int init_io_u(struct thread_data *td)
511 if (td->io_ops->flags & FIO_CPUIO)
514 if (td->io_ops->flags & FIO_SYNCIO)
517 max_units = td->iodepth;
519 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
520 td->orig_buffer_size = max_bs * max_units;
522 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
523 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
525 td->orig_buffer_size += page_mask;
527 if (allocate_io_mem(td))
530 p = ALIGN(td->orig_buffer);
531 for (i = 0; i < max_units; i++) {
532 io_u = malloc(sizeof(*io_u));
533 memset(io_u, 0, sizeof(*io_u));
534 INIT_LIST_HEAD(&io_u->list);
536 io_u->buf = p + max_bs * i;
537 if (td_write(td) || td_rw(td))
538 fill_rand_buf(io_u, max_bs);
541 list_add(&io_u->list, &td->io_u_freelist);
547 static int switch_ioscheduler(struct thread_data *td)
549 char tmp[256], tmp2[128];
553 if (td->io_ops->flags & FIO_CPUIO)
556 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
558 f = fopen(tmp, "r+");
560 td_verror(td, errno);
567 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
568 if (ferror(f) || ret != 1) {
569 td_verror(td, errno);
577 * Read back and check that the selected scheduler is now the default.
579 ret = fread(tmp, 1, sizeof(tmp), f);
580 if (ferror(f) || ret < 0) {
581 td_verror(td, errno);
586 sprintf(tmp2, "[%s]", td->ioscheduler);
587 if (!strstr(tmp, tmp2)) {
588 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
589 td_verror(td, EINVAL);
598 static void clear_io_state(struct thread_data *td)
603 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
604 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
607 for_each_file(td, f, i) {
609 if (td->io_ops->flags & FIO_SYNCIO)
610 lseek(f->fd, SEEK_SET, 0);
613 memset(f->file_map, 0, f->num_maps * sizeof(long));
618 * Entry point for the thread based jobs. The process based jobs end up
619 * here as well, after a little setup.
621 static void *thread_main(void *data)
623 unsigned long long runtime[2];
624 struct thread_data *td = data;
631 INIT_LIST_HEAD(&td->io_u_freelist);
632 INIT_LIST_HEAD(&td->io_u_busylist);
633 INIT_LIST_HEAD(&td->io_hist_list);
634 INIT_LIST_HEAD(&td->io_log_list);
639 if (fio_setaffinity(td) == -1) {
640 td_verror(td, errno);
648 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
649 td_verror(td, errno);
654 if (nice(td->nice) == -1) {
655 td_verror(td, errno);
659 if (init_random_state(td))
662 if (td->ioscheduler && switch_ioscheduler(td))
665 td_set_runstate(td, TD_INITIALIZED);
666 fio_sem_up(&startup_sem);
667 fio_sem_down(&td->mutex);
669 if (!td->create_serialize && setup_files(td))
675 * Do this late, as some IO engines would like to have the
676 * files setup prior to initializing structures.
681 if (td->exec_prerun) {
682 if (system(td->exec_prerun) < 0)
686 fio_gettime(&td->epoch, NULL);
687 getrusage(RUSAGE_SELF, &td->ru_start);
689 runtime[0] = runtime[1] = 0;
690 while (td->loops--) {
691 fio_gettime(&td->start, NULL);
692 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
695 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
698 prune_io_piece_log(td);
700 if (td->io_ops->flags & FIO_CPUIO)
705 runtime[td->ddir] += utime_since_now(&td->start);
706 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
707 runtime[td->ddir ^ 1] = runtime[td->ddir];
709 if (td->error || td->terminate)
712 if (td->verify == VERIFY_NONE)
716 fio_gettime(&td->start, NULL);
720 runtime[DDIR_READ] += utime_since_now(&td->start);
722 if (td->error || td->terminate)
726 update_rusage_stat(td);
727 fio_gettime(&td->end_time, NULL);
728 td->runtime[0] = runtime[0] / 1000;
729 td->runtime[1] = runtime[1] / 1000;
732 finish_log(td, td->bw_log, "bw");
734 finish_log(td, td->slat_log, "slat");
736 finish_log(td, td->clat_log, "clat");
737 if (td->write_iolog_file)
738 write_iolog_close(td);
739 if (td->exec_postrun) {
740 if (system(td->exec_postrun) < 0)
741 log_err("fio: postrun %s failed\n", td->exec_postrun);
744 if (exitall_on_terminate)
745 terminate_threads(td->groupid, 0);
749 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
753 td_set_runstate(td, TD_EXITED);
754 return (void *) td->error;
758 * We cannot pass the td data into a forked process, so attach the td and
759 * pass it to the thread worker.
761 static int fork_main(int shmid, int offset)
763 struct thread_data *td;
766 data = shmat(shmid, NULL, 0);
767 if (data == (void *) -1) {
774 td = data + offset * sizeof(struct thread_data);
775 ret = thread_main(td);
781 * Run over the job map and reap the threads that have exited, if any.
783 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
785 struct thread_data *td;
786 int i, cputhreads, pending, status, ret;
789 * reap exited threads (TD_EXITED -> TD_REAPED)
791 pending = cputhreads = 0;
794 * ->io_ops is NULL for a thread that has closed its
797 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
800 if (td->runstate < TD_EXITED) {
802 * check if someone quit or got killed in an unusual way
804 ret = waitpid(td->pid, &status, WNOHANG);
807 else if ((ret == td->pid) && WIFSIGNALED(status)) {
808 int sig = WTERMSIG(status);
810 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
811 td_set_runstate(td, TD_REAPED);
816 if (td->runstate != TD_EXITED) {
817 if (td->runstate < TD_RUNNING)
826 td_set_runstate(td, TD_REAPED);
828 if (td->use_thread) {
831 if (pthread_join(td->thread, (void *) &ret))
832 perror("thread_join");
836 ret = waitpid(td->pid, &status, 0);
839 else if (WIFEXITED(status) && WEXITSTATUS(status)) {
847 (*m_rate) -= td->ratemin;
848 (*t_rate) -= td->rate;
851 if (*nr_running == cputhreads && !pending)
852 terminate_threads(TERMINATE_ALL, 0);
856 * Main function for kicking off and reaping jobs, as needed.
858 static void run_threads(void)
860 struct thread_data *td;
862 int i, todo, nr_running, m_rate, t_rate, nr_started;
864 if (fio_pin_memory())
868 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
872 signal(SIGINT, sig_handler);
873 signal(SIGALRM, sig_handler);
875 todo = thread_number;
881 print_status_init(td->thread_number - 1);
883 if (!td->create_serialize) {
889 * do file setup here so it happens sequentially,
890 * we don't want X number of threads getting their
891 * client data interspersed on disk
893 if (setup_files(td)) {
896 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
897 td_set_runstate(td, TD_REAPED);
905 struct thread_data *map[MAX_JOBS];
906 struct timeval this_start;
907 int this_jobs = 0, left;
910 * create threads (TD_NOT_CREATED -> TD_CREATED)
913 if (td->runstate != TD_NOT_CREATED)
917 * never got a chance to start, killed by other
918 * thread for some reason
925 if (td->start_delay) {
926 spent = mtime_since_genesis();
928 if (td->start_delay * 1000 > spent)
932 if (td->stonewall && (nr_started || nr_running))
936 * Set state to created. Thread will transition
937 * to TD_INITIALIZED when it's done setting up.
939 td_set_runstate(td, TD_CREATED);
940 map[this_jobs++] = td;
941 fio_sem_init(&startup_sem, 1);
944 if (td->use_thread) {
945 if (pthread_create(&td->thread, NULL, thread_main, td)) {
946 perror("thread_create");
951 fio_sem_down(&startup_sem);
953 int ret = fork_main(shm_id, i);
961 * Wait for the started threads to transition to
964 fio_gettime(&this_start, NULL);
966 while (left && !fio_abort) {
967 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
972 for (i = 0; i < this_jobs; i++) {
976 if (td->runstate == TD_INITIALIZED) {
979 } else if (td->runstate >= TD_EXITED) {
983 nr_running++; /* work-around... */
989 log_err("fio: %d jobs failed to start\n", left);
990 for (i = 0; i < this_jobs; i++) {
994 kill(td->pid, SIGTERM);
1000 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1002 for_each_td(td, i) {
1003 if (td->runstate != TD_INITIALIZED)
1006 td_set_runstate(td, TD_RUNNING);
1009 m_rate += td->ratemin;
1012 fio_sem_up(&td->mutex);
1015 reap_threads(&nr_running, &t_rate, &m_rate);
1021 while (nr_running) {
1022 reap_threads(&nr_running, &t_rate, &m_rate);
1030 int main(int argc, char *argv[])
1035 * We need locale for number printing, if it isn't set then just
1036 * go with the US format.
1038 if (!getenv("LC_NUMERIC"))
1039 setlocale(LC_NUMERIC, "en_US");
1041 if (parse_options(argc, argv))
1044 if (!thread_number) {
1045 log_err("Nothing to do\n");
1049 ps = sysconf(_SC_PAGESIZE);
1051 log_err("Failed to get page size\n");
1058 setup_log(&agg_io_log[DDIR_READ]);
1059 setup_log(&agg_io_log[DDIR_WRITE]);
1062 disk_util_timer_arm();
1069 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1070 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");