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)
65 struct thread_data *td;
69 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
70 kill(td->pid, SIGQUIT);
77 static void sig_handler(int sig)
82 disk_util_timer_arm();
83 print_thread_status();
86 printf("\nfio: terminating on signal %d\n", sig);
88 terminate_threads(TERMINATE_ALL);
94 * Check if we are above the minimum rate given.
96 static int check_min_rate(struct thread_data *td, struct timeval *now)
98 unsigned long long bytes = 0;
103 * No minimum rate set, always ok
109 * allow a 2 second settle period in the beginning
111 if (mtime_since(&td->start, now) < 2000)
115 bytes += td->this_io_bytes[DDIR_READ];
117 bytes += td->this_io_bytes[DDIR_WRITE];
120 * if rate blocks is set, sample is running
122 if (td->rate_bytes) {
123 spent = mtime_since(&td->lastrate, now);
124 if (spent < td->ratecycle)
127 if (bytes < td->rate_bytes) {
128 fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
131 rate = (bytes - td->rate_bytes) / spent;
132 if (rate < td->ratemin || bytes < td->rate_bytes) {
133 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
139 td->rate_bytes = bytes;
140 memcpy(&td->lastrate, now, sizeof(*now));
144 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
148 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
155 * When job exits, we can cancel the in-flight IO if we are using async
156 * io. Attempt to do so.
158 static void cleanup_pending_aio(struct thread_data *td)
160 struct list_head *entry, *n;
165 * get immediately available events, if any
167 r = io_u_queued_complete(td, 0);
172 * now cancel remaining active events
174 if (td->io_ops->cancel) {
175 list_for_each_safe(entry, n, &td->io_u_busylist) {
176 io_u = list_entry(entry, struct io_u, list);
179 * if the io_u isn't in flight, then that generally
180 * means someone leaked an io_u. complain but fix
181 * it up, so we don't stall here.
183 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
184 log_err("fio: non-busy IO on busy list\n");
187 r = td->io_ops->cancel(td, io_u);
195 r = io_u_queued_complete(td, td->cur_depth);
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);
210 io_u->ddir = DDIR_SYNC;
213 if (td_io_prep(td, io_u)) {
219 ret = td_io_queue(td, io_u);
221 td_verror(td, io_u->error, "td_io_queue");
224 } else if (ret == FIO_Q_QUEUED) {
225 if (io_u_queued_complete(td, 1) < 0)
227 } else if (ret == FIO_Q_COMPLETED) {
229 td_verror(td, io_u->error, "td_io_queue");
233 if (io_u_sync_complete(td, io_u) < 0)
235 } else if (ret == FIO_Q_BUSY) {
236 if (td_io_commit(td))
245 * The main verify engine. Runs over the writes we previously submitted,
246 * reads the blocks back in, and checks the crc/md5 of the data.
248 static void do_verify(struct thread_data *td)
252 int ret, i, min_events;
255 * sync io first and invalidate cache, to make sure we really
258 for_each_file(td, f, i) {
259 if (fio_io_sync(td, f))
261 if (file_invalidate_cache(td, f))
268 td_set_runstate(td, TD_VERIFYING);
271 while (!td->terminate) {
274 io_u = __get_io_u(td);
278 if (runtime_exceeded(td, &io_u->start_time)) {
283 if (get_next_verify(td, io_u)) {
288 if (td_io_prep(td, io_u)) {
293 io_u->end_io = verify_io_u;
295 ret = td_io_queue(td, io_u);
298 case FIO_Q_COMPLETED:
301 else if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
302 int bytes = io_u->xfer_buflen - io_u->resid;
304 io_u->xfer_buflen = io_u->resid;
305 io_u->xfer_buf += bytes;
308 ret = io_u_sync_complete(td, io_u);
315 requeue_io_u(td, &io_u);
316 ret2 = td_io_commit(td);
322 td_verror(td, -ret, "td_io_queue");
326 if (ret < 0 || td->error)
330 * if we can queue more, do so. but check if there are
331 * completed io_u's first.
334 if (queue_full(td) || ret == FIO_Q_BUSY) {
337 if (td->cur_depth > td->iodepth_low)
338 min_events = td->cur_depth - td->iodepth_low;
342 * Reap required number of io units, if any, and do the
343 * verification on them through the callback handler
345 if (io_u_queued_complete(td, min_events) < 0)
350 min_events = td->cur_depth;
353 ret = io_u_queued_complete(td, min_events);
355 cleanup_pending_aio(td);
357 td_set_runstate(td, TD_RUNNING);
361 * Not really an io thread, all it does is burn CPU cycles in the specified
364 static void do_cpuio(struct thread_data *td)
367 int split = 100 / td->cpuload;
370 while (!td->terminate) {
371 fio_gettime(&e, NULL);
373 if (runtime_exceeded(td, &e))
379 usec_sleep(td, 10000);
386 * Main IO worker function. It retrieves io_u's to process and queues
387 * and reaps them, checking for rate and errors along the way.
389 static void do_io(struct thread_data *td)
395 td_set_runstate(td, TD_RUNNING);
397 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
398 struct timeval comp_time;
411 memcpy(&s, &io_u->start_time, sizeof(s));
413 if (runtime_exceeded(td, &s)) {
418 ret = td_io_queue(td, io_u);
421 case FIO_Q_COMPLETED:
424 else if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
425 int bytes = io_u->xfer_buflen - io_u->resid;
427 io_u->xfer_buflen = io_u->resid;
428 io_u->xfer_buf += bytes;
431 fio_gettime(&comp_time, NULL);
432 bytes_done = io_u_sync_complete(td, io_u);
438 * if the engine doesn't have a commit hook,
439 * the io_u is really queued. if it does have such
440 * a hook, it has to call io_u_queued() itself.
442 if (td->io_ops->commit == NULL)
443 io_u_queued(td, io_u);
446 requeue_io_u(td, &io_u);
447 ret2 = td_io_commit(td);
457 if (ret < 0 || td->error)
461 * See if we need to complete some commands
463 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
465 if (queue_full(td) || ret == FIO_Q_BUSY) {
468 if (td->cur_depth > td->iodepth_low)
469 min_evts = td->cur_depth - td->iodepth_low;
472 fio_gettime(&comp_time, NULL);
473 bytes_done = io_u_queued_complete(td, min_evts);
482 * the rate is batched for now, it should work for batches
483 * of completions except the very first one which may look
486 usec = utime_since(&s, &comp_time);
488 rate_throttle(td, usec, bytes_done);
490 if (check_min_rate(td, &comp_time)) {
491 if (exitall_on_terminate)
492 terminate_threads(td->groupid);
493 td_verror(td, ENODATA, "check_min_rate");
498 unsigned long long b;
500 b = td->io_blocks[0] + td->io_blocks[1];
501 if (!(b % td->thinktime_blocks)) {
504 if (td->thinktime_spin)
505 __usec_sleep(td->thinktime_spin);
507 left = td->thinktime - td->thinktime_spin;
509 usec_sleep(td, left);
519 ret = io_u_queued_complete(td, i);
521 if (should_fsync(td) && td->end_fsync) {
522 td_set_runstate(td, TD_FSYNCING);
523 for_each_file(td, f, i)
527 cleanup_pending_aio(td);
530 static void cleanup_io_u(struct thread_data *td)
532 struct list_head *entry, *n;
535 list_for_each_safe(entry, n, &td->io_u_freelist) {
536 io_u = list_entry(entry, struct io_u, list);
538 list_del(&io_u->list);
546 * "randomly" fill the buffer contents
548 static void fill_rand_buf(struct io_u *io_u, int max_bs)
550 int *ptr = io_u->buf;
552 while ((void *) ptr - io_u->buf < max_bs) {
553 *ptr = rand() * 0x9e370001;
558 static int init_io_u(struct thread_data *td)
565 if (td->io_ops->flags & FIO_CPUIO)
568 if (td->io_ops->flags & FIO_SYNCIO)
571 max_units = td->iodepth;
573 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
574 td->orig_buffer_size = max_bs * max_units;
576 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
577 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
579 td->orig_buffer_size += page_mask;
581 if (allocate_io_mem(td))
584 p = ALIGN(td->orig_buffer);
585 for (i = 0; i < max_units; i++) {
586 io_u = malloc(sizeof(*io_u));
587 memset(io_u, 0, sizeof(*io_u));
588 INIT_LIST_HEAD(&io_u->list);
590 io_u->buf = p + max_bs * i;
591 if (td_write(td) || td_rw(td))
592 fill_rand_buf(io_u, max_bs);
595 io_u->flags = IO_U_F_FREE;
596 list_add(&io_u->list, &td->io_u_freelist);
604 static int switch_ioscheduler(struct thread_data *td)
606 char tmp[256], tmp2[128];
610 if (td->io_ops->flags & FIO_CPUIO)
613 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
615 f = fopen(tmp, "r+");
617 td_verror(td, errno, "fopen");
624 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
625 if (ferror(f) || ret != 1) {
626 td_verror(td, errno, "fwrite");
634 * Read back and check that the selected scheduler is now the default.
636 ret = fread(tmp, 1, sizeof(tmp), f);
637 if (ferror(f) || ret < 0) {
638 td_verror(td, errno, "fread");
643 sprintf(tmp2, "[%s]", td->ioscheduler);
644 if (!strstr(tmp, tmp2)) {
645 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
646 td_verror(td, EINVAL, "iosched_switch");
655 static void clear_io_state(struct thread_data *td)
660 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
661 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
664 td->last_was_sync = 0;
666 for_each_file(td, f, i) {
667 f->last_completed_pos = 0;
670 if (td->io_ops->flags & FIO_SYNCIO)
671 lseek(f->fd, SEEK_SET, 0);
674 memset(f->file_map, 0, f->num_maps * sizeof(long));
679 * Entry point for the thread based jobs. The process based jobs end up
680 * here as well, after a little setup.
682 static void *thread_main(void *data)
684 unsigned long long runtime[2];
685 struct thread_data *td = data;
692 INIT_LIST_HEAD(&td->io_u_freelist);
693 INIT_LIST_HEAD(&td->io_u_busylist);
694 INIT_LIST_HEAD(&td->io_u_requeues);
695 INIT_LIST_HEAD(&td->io_hist_list);
696 INIT_LIST_HEAD(&td->io_log_list);
701 if (fio_setaffinity(td) == -1) {
702 td_verror(td, errno, "cpu_set_affinity");
710 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
711 td_verror(td, errno, "ioprio_set");
716 if (nice(td->nice) == -1) {
717 td_verror(td, errno, "nice");
721 if (init_random_state(td))
724 if (td->ioscheduler && switch_ioscheduler(td))
727 td_set_runstate(td, TD_INITIALIZED);
728 fio_sem_up(&startup_sem);
729 fio_sem_down(&td->mutex);
731 if (!td->create_serialize && setup_files(td))
737 * Do this late, as some IO engines would like to have the
738 * files setup prior to initializing structures.
743 if (td->exec_prerun) {
744 if (system(td->exec_prerun) < 0)
748 fio_gettime(&td->epoch, NULL);
749 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
750 getrusage(RUSAGE_SELF, &td->ts.ru_start);
752 runtime[0] = runtime[1] = 0;
753 while (td->loops--) {
754 fio_gettime(&td->start, NULL);
755 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
758 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
761 prune_io_piece_log(td);
763 if (td->io_ops->flags & FIO_CPUIO)
768 if (td_read(td) && td->io_bytes[DDIR_READ])
769 runtime[DDIR_READ] += utime_since_now(&td->start);
770 if (td_write(td) && td->io_bytes[DDIR_WRITE])
771 runtime[DDIR_WRITE] += utime_since_now(&td->start);
773 if (td->error || td->terminate)
776 if (td->verify == VERIFY_NONE)
780 fio_gettime(&td->start, NULL);
784 runtime[DDIR_READ] += utime_since_now(&td->start);
786 if (td->error || td->terminate)
790 update_rusage_stat(td);
791 fio_gettime(&td->end_time, NULL);
792 td->runtime[0] = runtime[0] / 1000;
793 td->runtime[1] = runtime[1] / 1000;
796 finish_log(td, td->ts.bw_log, "bw");
798 finish_log(td, td->ts.slat_log, "slat");
800 finish_log(td, td->ts.clat_log, "clat");
801 if (td->write_iolog_file)
802 write_iolog_close(td);
803 if (td->exec_postrun) {
804 if (system(td->exec_postrun) < 0)
805 log_err("fio: postrun %s failed\n", td->exec_postrun);
808 if (exitall_on_terminate)
809 terminate_threads(td->groupid);
813 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
817 td_set_runstate(td, TD_EXITED);
818 return (void *) (unsigned long) td->error;
822 * We cannot pass the td data into a forked process, so attach the td and
823 * pass it to the thread worker.
825 static int fork_main(int shmid, int offset)
827 struct thread_data *td;
830 data = shmat(shmid, NULL, 0);
831 if (data == (void *) -1) {
838 td = data + offset * sizeof(struct thread_data);
839 ret = thread_main(td);
841 return (int) (unsigned long) ret;
845 * Run over the job map and reap the threads that have exited, if any.
847 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
849 struct thread_data *td;
850 int i, cputhreads, pending, status, ret;
853 * reap exited threads (TD_EXITED -> TD_REAPED)
855 pending = cputhreads = 0;
860 * ->io_ops is NULL for a thread that has closed its
863 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
866 if (!td->pid || td->runstate == TD_REAPED)
868 if (td->use_thread) {
869 if (td->runstate == TD_EXITED) {
870 td_set_runstate(td, TD_REAPED);
877 if (td->runstate == TD_EXITED)
881 * check if someone quit or got killed in an unusual way
883 ret = waitpid(td->pid, &status, flags);
885 if (errno == ECHILD) {
886 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
887 td_set_runstate(td, TD_REAPED);
891 } else if (ret == td->pid) {
892 if (WIFSIGNALED(status)) {
893 int sig = WTERMSIG(status);
896 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
897 td_set_runstate(td, TD_REAPED);
900 if (WIFEXITED(status)) {
901 if (WEXITSTATUS(status) && !td->error)
902 td->error = WEXITSTATUS(status);
904 td_set_runstate(td, TD_REAPED);
910 * thread is not dead, continue
914 if (td->use_thread) {
917 if (pthread_join(td->thread, (void *) &ret))
918 perror("pthread_join");
922 (*m_rate) -= td->ratemin;
923 (*t_rate) -= td->rate;
929 if (*nr_running == cputhreads && !pending)
930 terminate_threads(TERMINATE_ALL);
934 * Main function for kicking off and reaping jobs, as needed.
936 static void run_threads(void)
938 struct thread_data *td;
940 int i, todo, nr_running, m_rate, t_rate, nr_started;
942 if (fio_pin_memory())
946 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
950 signal(SIGINT, sig_handler);
951 signal(SIGALRM, sig_handler);
953 todo = thread_number;
959 print_status_init(td->thread_number - 1);
961 if (!td->create_serialize) {
967 * do file setup here so it happens sequentially,
968 * we don't want X number of threads getting their
969 * client data interspersed on disk
971 if (setup_files(td)) {
974 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
975 td_set_runstate(td, TD_REAPED);
985 struct thread_data *map[MAX_JOBS];
986 struct timeval this_start;
987 int this_jobs = 0, left;
990 * create threads (TD_NOT_CREATED -> TD_CREATED)
993 if (td->runstate != TD_NOT_CREATED)
997 * never got a chance to start, killed by other
998 * thread for some reason
1000 if (td->terminate) {
1005 if (td->start_delay) {
1006 spent = mtime_since_genesis();
1008 if (td->start_delay * 1000 > spent)
1012 if (td->stonewall && (nr_started || nr_running))
1016 * Set state to created. Thread will transition
1017 * to TD_INITIALIZED when it's done setting up.
1019 td_set_runstate(td, TD_CREATED);
1020 map[this_jobs++] = td;
1021 fio_sem_init(&startup_sem, 1);
1024 if (td->use_thread) {
1025 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1026 perror("thread_create");
1031 fio_sem_down(&startup_sem);
1033 int ret = fork_main(shm_id, i);
1041 * Wait for the started threads to transition to
1044 fio_gettime(&this_start, NULL);
1046 while (left && !fio_abort) {
1047 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1052 for (i = 0; i < this_jobs; i++) {
1056 if (td->runstate == TD_INITIALIZED) {
1059 } else if (td->runstate >= TD_EXITED) {
1063 nr_running++; /* work-around... */
1069 log_err("fio: %d jobs failed to start\n", left);
1070 for (i = 0; i < this_jobs; i++) {
1074 kill(td->pid, SIGTERM);
1080 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1082 for_each_td(td, i) {
1083 if (td->runstate != TD_INITIALIZED)
1086 td_set_runstate(td, TD_RUNNING);
1089 m_rate += td->ratemin;
1092 fio_sem_up(&td->mutex);
1095 reap_threads(&nr_running, &t_rate, &m_rate);
1101 while (nr_running) {
1102 reap_threads(&nr_running, &t_rate, &m_rate);
1110 int main(int argc, char *argv[])
1115 * We need locale for number printing, if it isn't set then just
1116 * go with the US format.
1118 if (!getenv("LC_NUMERIC"))
1119 setlocale(LC_NUMERIC, "en_US");
1121 if (parse_options(argc, argv))
1124 if (!thread_number) {
1125 log_err("Nothing to do\n");
1129 ps = sysconf(_SC_PAGESIZE);
1131 log_err("Failed to get page size\n");
1138 setup_log(&agg_io_log[DDIR_READ]);
1139 setup_log(&agg_io_log[DDIR_WRITE]);
1144 disk_util_timer_arm();
1151 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1152 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");