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
41 unsigned long page_mask;
42 unsigned long page_size;
44 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
47 int thread_number = 0;
53 static struct fio_mutex *startup_mutex;
54 static volatile int fio_abort;
55 static int exit_value;
57 struct io_log *agg_io_log[2];
59 #define TERMINATE_ALL (-1)
60 #define JOB_START_TIMEOUT (5 * 1000)
62 static inline void td_set_runstate(struct thread_data *td, int runstate)
64 if (td->runstate == runstate)
67 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", td->pid, td->runstate,
69 td->runstate = runstate;
72 static void terminate_threads(int group_id)
74 struct thread_data *td;
78 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
79 dprint(FD_PROCESS, "setting terminate on %d\n",td->pid);
82 td->o.start_delay = 0;
85 * if the thread is running, just let it exit
87 if (td->runstate < TD_RUNNING)
88 kill(td->pid, SIGQUIT);
90 struct ioengine_ops *ops = td->io_ops;
92 if (ops && (ops->flags & FIO_SIGQUIT))
93 kill(td->pid, SIGQUIT);
99 static void sig_handler(int sig)
104 disk_util_timer_arm();
105 print_thread_status();
108 printf("\nfio: terminating on signal %d\n", sig);
110 terminate_threads(TERMINATE_ALL);
116 * Check if we are above the minimum rate given.
118 static int check_min_rate(struct thread_data *td, struct timeval *now)
120 unsigned long long bytes = 0;
121 unsigned long iops = 0;
126 * No minimum rate set, always ok
128 if (!td->o.ratemin && !td->o.rate_iops_min)
132 * allow a 2 second settle period in the beginning
134 if (mtime_since(&td->start, now) < 2000)
138 iops += td->io_blocks[DDIR_READ];
139 bytes += td->this_io_bytes[DDIR_READ];
142 iops += td->io_blocks[DDIR_WRITE];
143 bytes += td->this_io_bytes[DDIR_WRITE];
147 * if rate blocks is set, sample is running
149 if (td->rate_bytes || td->rate_blocks) {
150 spent = mtime_since(&td->lastrate, now);
151 if (spent < td->o.ratecycle)
156 * check bandwidth specified rate
158 if (bytes < td->rate_bytes) {
159 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
162 rate = (bytes - td->rate_bytes) / spent;
163 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
164 log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
170 * checks iops specified rate
172 if (iops < td->o.rate_iops) {
173 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
176 rate = (iops - td->rate_blocks) / spent;
177 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
178 log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
184 td->rate_bytes = bytes;
185 td->rate_blocks = iops;
186 memcpy(&td->lastrate, now, sizeof(*now));
190 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
194 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
201 * When job exits, we can cancel the in-flight IO if we are using async
202 * io. Attempt to do so.
204 static void cleanup_pending_aio(struct thread_data *td)
206 struct list_head *entry, *n;
211 * get immediately available events, if any
213 r = io_u_queued_complete(td, 0);
218 * now cancel remaining active events
220 if (td->io_ops->cancel) {
221 list_for_each_safe(entry, n, &td->io_u_busylist) {
222 io_u = list_entry(entry, struct io_u, list);
225 * if the io_u isn't in flight, then that generally
226 * means someone leaked an io_u. complain but fix
227 * it up, so we don't stall here.
229 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
230 log_err("fio: non-busy IO on busy list\n");
233 r = td->io_ops->cancel(td, io_u);
241 r = io_u_queued_complete(td, td->cur_depth);
245 * Helper to handle the final sync of a file. Works just like the normal
246 * io path, just does everything sync.
248 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
250 struct io_u *io_u = __get_io_u(td);
256 io_u->ddir = DDIR_SYNC;
259 if (td_io_prep(td, io_u)) {
265 ret = td_io_queue(td, io_u);
267 td_verror(td, io_u->error, "td_io_queue");
270 } else if (ret == FIO_Q_QUEUED) {
271 if (io_u_queued_complete(td, 1) < 0)
273 } else if (ret == FIO_Q_COMPLETED) {
275 td_verror(td, io_u->error, "td_io_queue");
279 if (io_u_sync_complete(td, io_u) < 0)
281 } else if (ret == FIO_Q_BUSY) {
282 if (td_io_commit(td))
291 * The main verify engine. Runs over the writes we previously submitted,
292 * reads the blocks back in, and checks the crc/md5 of the data.
294 static void do_verify(struct thread_data *td)
302 * sync io first and invalidate cache, to make sure we really
305 for_each_file(td, f, i) {
306 if (!(f->flags & FIO_FILE_OPEN))
308 if (fio_io_sync(td, f))
310 if (file_invalidate_cache(td, f))
317 td_set_runstate(td, TD_VERIFYING);
320 while (!td->terminate) {
323 io_u = __get_io_u(td);
327 if (runtime_exceeded(td, &io_u->start_time)) {
333 if (get_next_verify(td, io_u)) {
338 if (td_io_prep(td, io_u)) {
343 io_u->end_io = verify_io_u;
345 ret = td_io_queue(td, io_u);
347 case FIO_Q_COMPLETED:
350 else if (io_u->resid) {
351 int bytes = io_u->xfer_buflen - io_u->resid;
352 struct fio_file *f = io_u->file;
358 td_verror(td, ENODATA, "full resid");
363 io_u->xfer_buflen = io_u->resid;
364 io_u->xfer_buf += bytes;
365 io_u->offset += bytes;
367 td->ts.short_io_u[io_u->ddir]++;
369 if (io_u->offset == f->real_file_size)
372 requeue_io_u(td, &io_u);
375 ret = io_u_sync_complete(td, io_u);
383 requeue_io_u(td, &io_u);
384 ret2 = td_io_commit(td);
390 td_verror(td, -ret, "td_io_queue");
394 if (ret < 0 || td->error)
398 * if we can queue more, do so. but check if there are
399 * completed io_u's first.
402 if (queue_full(td) || ret == FIO_Q_BUSY) {
405 if (td->cur_depth > td->o.iodepth_low)
406 min_events = td->cur_depth - td->o.iodepth_low;
410 * Reap required number of io units, if any, and do the
411 * verification on them through the callback handler
413 if (io_u_queued_complete(td, min_events) < 0)
418 min_events = td->cur_depth;
421 ret = io_u_queued_complete(td, min_events);
423 cleanup_pending_aio(td);
425 td_set_runstate(td, TD_RUNNING);
429 * Main IO worker function. It retrieves io_u's to process and queues
430 * and reaps them, checking for rate and errors along the way.
432 static void do_io(struct thread_data *td)
439 td_set_runstate(td, TD_RUNNING);
441 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
442 struct timeval comp_time;
455 memcpy(&s, &io_u->start_time, sizeof(s));
457 if (runtime_exceeded(td, &s)) {
464 * Add verification end_io handler, if asked to verify
465 * a previously written file.
467 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
468 io_u->end_io = verify_io_u;
469 td_set_runstate(td, TD_VERIFYING);
471 td_set_runstate(td, TD_RUNNING);
473 ret = td_io_queue(td, io_u);
475 case FIO_Q_COMPLETED:
478 else if (io_u->resid) {
479 int bytes = io_u->xfer_buflen - io_u->resid;
480 struct fio_file *f = io_u->file;
486 td_verror(td, ENODATA, "full resid");
491 io_u->xfer_buflen = io_u->resid;
492 io_u->xfer_buf += bytes;
493 io_u->offset += bytes;
495 td->ts.short_io_u[io_u->ddir]++;
497 if (io_u->offset == f->real_file_size)
500 requeue_io_u(td, &io_u);
503 fio_gettime(&comp_time, NULL);
504 bytes_done = io_u_sync_complete(td, io_u);
511 * if the engine doesn't have a commit hook,
512 * the io_u is really queued. if it does have such
513 * a hook, it has to call io_u_queued() itself.
515 if (td->io_ops->commit == NULL)
516 io_u_queued(td, io_u);
519 requeue_io_u(td, &io_u);
520 ret2 = td_io_commit(td);
530 if (ret < 0 || td->error)
534 * See if we need to complete some commands
536 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
538 if (queue_full(td) || ret == FIO_Q_BUSY) {
541 if (td->cur_depth > td->o.iodepth_low)
542 min_evts = td->cur_depth - td->o.iodepth_low;
545 fio_gettime(&comp_time, NULL);
546 bytes_done = io_u_queued_complete(td, min_evts);
555 * the rate is batched for now, it should work for batches
556 * of completions except the very first one which may look
559 usec = utime_since(&s, &comp_time);
561 rate_throttle(td, usec, bytes_done);
563 if (check_min_rate(td, &comp_time)) {
564 if (exitall_on_terminate)
565 terminate_threads(td->groupid);
566 td_verror(td, ENODATA, "check_min_rate");
570 if (td->o.thinktime) {
571 unsigned long long b;
573 b = td->io_blocks[0] + td->io_blocks[1];
574 if (!(b % td->o.thinktime_blocks)) {
577 if (td->o.thinktime_spin)
578 __usec_sleep(td->o.thinktime_spin);
580 left = td->o.thinktime - td->o.thinktime_spin;
582 usec_sleep(td, left);
587 if (td->o.fill_device && td->error == ENOSPC) {
596 ret = io_u_queued_complete(td, i);
598 if (should_fsync(td) && td->o.end_fsync) {
599 td_set_runstate(td, TD_FSYNCING);
601 for_each_file(td, f, i) {
602 if (!(f->flags & FIO_FILE_OPEN))
608 cleanup_pending_aio(td);
611 * stop job if we failed doing any IO
613 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
617 static void cleanup_io_u(struct thread_data *td)
619 struct list_head *entry, *n;
622 list_for_each_safe(entry, n, &td->io_u_freelist) {
623 io_u = list_entry(entry, struct io_u, list);
625 list_del(&io_u->list);
633 * "randomly" fill the buffer contents
635 static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
637 long *ptr = io_u->buf;
639 if (!td->o.zero_buffers) {
640 while ((void *) ptr - io_u->buf < max_bs) {
641 *ptr = rand() * GOLDEN_RATIO_PRIME;
645 memset(ptr, 0, max_bs);
648 static int init_io_u(struct thread_data *td)
655 max_units = td->o.iodepth;
656 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
657 td->orig_buffer_size = (unsigned long long) max_bs * (unsigned long long) max_units;
659 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
660 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
662 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
663 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
667 if (allocate_io_mem(td))
671 p = ALIGN(td->orig_buffer);
675 for (i = 0; i < max_units; i++) {
678 io_u = malloc(sizeof(*io_u));
679 memset(io_u, 0, sizeof(*io_u));
680 INIT_LIST_HEAD(&io_u->list);
682 if (!(td->io_ops->flags & FIO_NOIO)) {
683 io_u->buf = p + max_bs * i;
686 fill_io_buf(td, io_u, max_bs);
690 io_u->flags = IO_U_F_FREE;
691 list_add(&io_u->list, &td->io_u_freelist);
699 static int switch_ioscheduler(struct thread_data *td)
701 char tmp[256], tmp2[128];
705 if (td->io_ops->flags & FIO_DISKLESSIO)
708 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
710 f = fopen(tmp, "r+");
712 if (errno == ENOENT) {
713 log_err("fio: os or kernel doesn't support IO scheduler switching\n");
716 td_verror(td, errno, "fopen iosched");
723 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
724 if (ferror(f) || ret != 1) {
725 td_verror(td, errno, "fwrite");
733 * Read back and check that the selected scheduler is now the default.
735 ret = fread(tmp, 1, sizeof(tmp), f);
736 if (ferror(f) || ret < 0) {
737 td_verror(td, errno, "fread");
742 sprintf(tmp2, "[%s]", td->o.ioscheduler);
743 if (!strstr(tmp, tmp2)) {
744 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
745 td_verror(td, EINVAL, "iosched_switch");
754 static int keep_running(struct thread_data *td)
756 unsigned long long io_done;
760 if (td->o.time_based)
767 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE] + td->io_skip_bytes;
768 if (io_done < td->o.size)
774 static int clear_io_state(struct thread_data *td)
780 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
781 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
785 td->rw_end_set[0] = td->rw_end_set[1] = 0;
787 td->last_was_sync = 0;
790 * reset file done count if we are to start over
792 if (td->o.time_based || td->o.loops)
793 td->nr_done_files = 0;
798 for_each_file(td, f, i) {
799 f->flags &= ~FIO_FILE_DONE;
800 ret = td_io_open_file(td, f);
809 * Entry point for the thread based jobs. The process based jobs end up
810 * here as well, after a little setup.
812 static void *thread_main(void *data)
814 unsigned long long runtime[2], elapsed;
815 struct thread_data *td = data;
818 if (!td->o.use_thread)
823 dprint(FD_PROCESS, "jobs pid=%d started\n", td->pid);
825 INIT_LIST_HEAD(&td->io_u_freelist);
826 INIT_LIST_HEAD(&td->io_u_busylist);
827 INIT_LIST_HEAD(&td->io_u_requeues);
828 INIT_LIST_HEAD(&td->io_log_list);
829 INIT_LIST_HEAD(&td->io_hist_list);
830 td->io_hist_tree = RB_ROOT;
832 td_set_runstate(td, TD_INITIALIZED);
833 fio_mutex_up(startup_mutex);
834 fio_mutex_down(td->mutex);
837 * the ->mutex mutex is now no longer used, close it to avoid
838 * eating a file descriptor
840 fio_mutex_remove(td->mutex);
843 * May alter parameters that init_io_u() will use, so we need to
852 if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
853 td_verror(td, errno, "cpu_set_affinity");
857 if (td->ioprio_set) {
858 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
859 td_verror(td, errno, "ioprio_set");
864 if (nice(td->o.nice) == -1) {
865 td_verror(td, errno, "nice");
869 if (td->o.ioscheduler && switch_ioscheduler(td))
872 if (!td->o.create_serialize && setup_files(td))
881 if (init_random_map(td))
884 if (td->o.exec_prerun) {
885 if (system(td->o.exec_prerun) < 0)
889 fio_gettime(&td->epoch, NULL);
890 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
891 getrusage(RUSAGE_SELF, &td->ts.ru_start);
893 runtime[0] = runtime[1] = 0;
895 while (keep_running(td)) {
896 fio_gettime(&td->start, NULL);
897 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
900 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
902 if (clear_state && clear_io_state(td))
905 prune_io_piece_log(td);
911 if (td_read(td) && td->io_bytes[DDIR_READ]) {
912 if (td->rw_end_set[DDIR_READ])
913 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
915 elapsed = utime_since_now(&td->start);
917 runtime[DDIR_READ] += elapsed;
919 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
920 if (td->rw_end_set[DDIR_WRITE])
921 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
923 elapsed = utime_since_now(&td->start);
925 runtime[DDIR_WRITE] += elapsed;
928 if (td->error || td->terminate)
931 if (!td->o.do_verify ||
932 td->o.verify == VERIFY_NONE ||
933 (td->io_ops->flags & FIO_UNIDIR))
936 if (clear_io_state(td))
939 fio_gettime(&td->start, NULL);
943 runtime[DDIR_READ] += utime_since_now(&td->start);
945 if (td->error || td->terminate)
949 update_rusage_stat(td);
950 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
951 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
952 td->ts.total_run_time = mtime_since_now(&td->epoch);
953 td->ts.io_bytes[0] = td->io_bytes[0];
954 td->ts.io_bytes[1] = td->io_bytes[1];
957 finish_log(td, td->ts.bw_log, "bw");
959 finish_log(td, td->ts.slat_log, "slat");
961 finish_log(td, td->ts.clat_log, "clat");
962 if (td->o.exec_postrun) {
963 if (system(td->o.exec_postrun) < 0)
964 log_err("fio: postrun %s failed\n", td->o.exec_postrun);
967 if (exitall_on_terminate)
968 terminate_threads(td->groupid);
972 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
973 close_and_free_files(td);
978 * do this very late, it will log file closing as well
980 if (td->o.write_iolog_file)
981 write_iolog_close(td);
983 options_mem_free(td);
984 td_set_runstate(td, TD_EXITED);
985 return (void *) (unsigned long) td->error;
989 * We cannot pass the td data into a forked process, so attach the td and
990 * pass it to the thread worker.
992 static int fork_main(int shmid, int offset)
994 struct thread_data *td;
997 data = shmat(shmid, NULL, 0);
998 if (data == (void *) -1) {
1005 td = data + offset * sizeof(struct thread_data);
1006 ret = thread_main(td);
1008 return (int) (unsigned long) ret;
1012 * Run over the job map and reap the threads that have exited, if any.
1014 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1016 struct thread_data *td;
1017 int i, cputhreads, realthreads, pending, status, ret;
1020 * reap exited threads (TD_EXITED -> TD_REAPED)
1022 realthreads = pending = cputhreads = 0;
1023 for_each_td(td, i) {
1027 * ->io_ops is NULL for a thread that has closed its
1030 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1039 if (td->runstate == TD_REAPED)
1041 if (td->o.use_thread) {
1042 if (td->runstate == TD_EXITED) {
1043 td_set_runstate(td, TD_REAPED);
1050 if (td->runstate == TD_EXITED)
1054 * check if someone quit or got killed in an unusual way
1056 ret = waitpid(td->pid, &status, flags);
1058 if (errno == ECHILD) {
1059 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
1060 td_set_runstate(td, TD_REAPED);
1064 } else if (ret == td->pid) {
1065 if (WIFSIGNALED(status)) {
1066 int sig = WTERMSIG(status);
1069 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
1070 td_set_runstate(td, TD_REAPED);
1073 if (WIFEXITED(status)) {
1074 if (WEXITSTATUS(status) && !td->error)
1075 td->error = WEXITSTATUS(status);
1077 td_set_runstate(td, TD_REAPED);
1083 * thread is not dead, continue
1088 if (td->o.use_thread) {
1091 dprint(FD_PROCESS, "joining tread %d\n", td->pid);
1092 if (pthread_join(td->thread, (void *) &ret)) {
1093 dprint(FD_PROCESS, "join failed %ld\n", ret);
1094 perror("pthread_join");
1099 (*m_rate) -= td->o.ratemin;
1100 (*t_rate) -= td->o.rate;
1107 if (*nr_running == cputhreads && !pending && realthreads)
1108 terminate_threads(TERMINATE_ALL);
1112 * Main function for kicking off and reaping jobs, as needed.
1114 static void run_threads(void)
1116 struct thread_data *td;
1117 unsigned long spent;
1118 int i, todo, nr_running, m_rate, t_rate, nr_started;
1120 if (fio_pin_memory())
1123 if (!terse_output) {
1124 printf("Starting ");
1126 printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1130 printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1136 signal(SIGINT, sig_handler);
1137 signal(SIGALRM, sig_handler);
1139 todo = thread_number;
1142 m_rate = t_rate = 0;
1144 for_each_td(td, i) {
1145 print_status_init(td->thread_number - 1);
1147 if (!td->o.create_serialize) {
1153 * do file setup here so it happens sequentially,
1154 * we don't want X number of threads getting their
1155 * client data interspersed on disk
1157 if (setup_files(td)) {
1160 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1161 td_set_runstate(td, TD_REAPED);
1168 * for sharing to work, each job must always open
1169 * its own files. so close them, if we opened them
1172 for_each_file(td, f, i)
1173 td_io_close_file(td, f);
1182 struct thread_data *map[MAX_JOBS];
1183 struct timeval this_start;
1184 int this_jobs = 0, left;
1187 * create threads (TD_NOT_CREATED -> TD_CREATED)
1189 for_each_td(td, i) {
1190 if (td->runstate != TD_NOT_CREATED)
1194 * never got a chance to start, killed by other
1195 * thread for some reason
1197 if (td->terminate) {
1202 if (td->o.start_delay) {
1203 spent = mtime_since_genesis();
1205 if (td->o.start_delay * 1000 > spent)
1209 if (td->o.stonewall && (nr_started || nr_running))
1213 * Set state to created. Thread will transition
1214 * to TD_INITIALIZED when it's done setting up.
1216 td_set_runstate(td, TD_CREATED);
1217 map[this_jobs++] = td;
1220 if (td->o.use_thread) {
1221 dprint(FD_PROCESS, "will pthread_create\n");
1222 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1223 perror("thread_create");
1228 dprint(FD_PROCESS, "will fork\n");
1230 int ret = fork_main(shm_id, i);
1235 fio_mutex_down(startup_mutex);
1239 * Wait for the started threads to transition to
1242 fio_gettime(&this_start, NULL);
1244 while (left && !fio_abort) {
1245 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1250 for (i = 0; i < this_jobs; i++) {
1254 if (td->runstate == TD_INITIALIZED) {
1257 } else if (td->runstate >= TD_EXITED) {
1261 nr_running++; /* work-around... */
1267 log_err("fio: %d jobs failed to start\n", left);
1268 for (i = 0; i < this_jobs; i++) {
1272 kill(td->pid, SIGTERM);
1278 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1280 for_each_td(td, i) {
1281 if (td->runstate != TD_INITIALIZED)
1284 td_set_runstate(td, TD_RUNNING);
1287 m_rate += td->o.ratemin;
1288 t_rate += td->o.rate;
1290 fio_mutex_up(td->mutex);
1293 reap_threads(&nr_running, &t_rate, &m_rate);
1299 while (nr_running) {
1300 reap_threads(&nr_running, &t_rate, &m_rate);
1308 int main(int argc, char *argv[])
1315 * We need locale for number printing, if it isn't set then just
1316 * go with the US format.
1318 if (!getenv("LC_NUMERIC"))
1319 setlocale(LC_NUMERIC, "en_US");
1321 if (parse_options(argc, argv))
1327 ps = sysconf(_SC_PAGESIZE);
1329 log_err("Failed to get page size\n");
1337 setup_log(&agg_io_log[DDIR_READ]);
1338 setup_log(&agg_io_log[DDIR_WRITE]);
1341 startup_mutex = fio_mutex_init(0);
1345 disk_util_timer_arm();
1352 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1353 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1357 fio_mutex_remove(startup_mutex);