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;
52 unsigned long done_secs = 0;
54 static struct fio_mutex *startup_mutex;
55 static volatile int fio_abort;
56 static int exit_value;
58 struct io_log *agg_io_log[2];
60 #define TERMINATE_ALL (-1)
61 #define JOB_START_TIMEOUT (5 * 1000)
63 static inline void td_set_runstate(struct thread_data *td, int runstate)
65 if (td->runstate == runstate)
68 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", td->pid, td->runstate,
70 td->runstate = runstate;
73 static void terminate_threads(int group_id)
75 struct thread_data *td;
78 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
81 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
82 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
85 td->o.start_delay = 0;
88 * if the thread is running, just let it exit
90 if (td->runstate < TD_RUNNING)
91 kill(td->pid, SIGQUIT);
93 struct ioengine_ops *ops = td->io_ops;
95 if (ops && (ops->flags & FIO_SIGQUIT))
96 kill(td->pid, SIGQUIT);
102 static void sig_handler(int sig)
107 disk_util_timer_arm();
108 print_thread_status();
111 printf("\nfio: terminating on signal %d\n", sig);
113 terminate_threads(TERMINATE_ALL);
119 * Check if we are above the minimum rate given.
121 static int check_min_rate(struct thread_data *td, struct timeval *now)
123 unsigned long long bytes = 0;
124 unsigned long iops = 0;
129 * No minimum rate set, always ok
131 if (!td->o.ratemin && !td->o.rate_iops_min)
135 * allow a 2 second settle period in the beginning
137 if (mtime_since(&td->start, now) < 2000)
141 iops += td->io_blocks[DDIR_READ];
142 bytes += td->this_io_bytes[DDIR_READ];
145 iops += td->io_blocks[DDIR_WRITE];
146 bytes += td->this_io_bytes[DDIR_WRITE];
150 * if rate blocks is set, sample is running
152 if (td->rate_bytes || td->rate_blocks) {
153 spent = mtime_since(&td->lastrate, now);
154 if (spent < td->o.ratecycle)
159 * check bandwidth specified rate
161 if (bytes < td->rate_bytes) {
162 log_err("%s: min rate %u not met\n", td->o.name,
166 rate = (bytes - td->rate_bytes) / spent;
167 if (rate < td->o.ratemin ||
168 bytes < td->rate_bytes) {
169 log_err("%s: min rate %u not met, got"
170 " %luKiB/sec\n", td->o.name,
171 td->o.ratemin, rate);
177 * checks iops specified rate
179 if (iops < td->o.rate_iops) {
180 log_err("%s: min iops rate %u not met\n",
181 td->o.name, td->o.rate_iops);
184 rate = (iops - td->rate_blocks) / spent;
185 if (rate < td->o.rate_iops_min ||
186 iops < td->rate_blocks) {
187 log_err("%s: min iops rate %u not met,"
188 " got %lu\n", td->o.name,
196 td->rate_bytes = bytes;
197 td->rate_blocks = iops;
198 memcpy(&td->lastrate, now, sizeof(*now));
202 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
206 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
213 * When job exits, we can cancel the in-flight IO if we are using async
214 * io. Attempt to do so.
216 static void cleanup_pending_aio(struct thread_data *td)
218 struct list_head *entry, *n;
223 * get immediately available events, if any
225 r = io_u_queued_complete(td, 0);
230 * now cancel remaining active events
232 if (td->io_ops->cancel) {
233 list_for_each_safe(entry, n, &td->io_u_busylist) {
234 io_u = list_entry(entry, struct io_u, list);
237 * if the io_u isn't in flight, then that generally
238 * means someone leaked an io_u. complain but fix
239 * it up, so we don't stall here.
241 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
242 log_err("fio: non-busy IO on busy list\n");
245 r = td->io_ops->cancel(td, io_u);
253 r = io_u_queued_complete(td, td->cur_depth);
257 * Helper to handle the final sync of a file. Works just like the normal
258 * io path, just does everything sync.
260 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
262 struct io_u *io_u = __get_io_u(td);
268 io_u->ddir = DDIR_SYNC;
271 if (td_io_prep(td, io_u)) {
277 ret = td_io_queue(td, io_u);
279 td_verror(td, io_u->error, "td_io_queue");
282 } else if (ret == FIO_Q_QUEUED) {
283 if (io_u_queued_complete(td, 1) < 0)
285 } else if (ret == FIO_Q_COMPLETED) {
287 td_verror(td, io_u->error, "td_io_queue");
291 if (io_u_sync_complete(td, io_u) < 0)
293 } else if (ret == FIO_Q_BUSY) {
294 if (td_io_commit(td))
303 * The main verify engine. Runs over the writes we previously submitted,
304 * reads the blocks back in, and checks the crc/md5 of the data.
306 static void do_verify(struct thread_data *td)
314 * sync io first and invalidate cache, to make sure we really
317 for_each_file(td, f, i) {
318 if (!(f->flags & FIO_FILE_OPEN))
320 if (fio_io_sync(td, f))
322 if (file_invalidate_cache(td, f))
329 td_set_runstate(td, TD_VERIFYING);
332 while (!td->terminate) {
335 io_u = __get_io_u(td);
339 if (runtime_exceeded(td, &io_u->start_time)) {
345 if (get_next_verify(td, io_u)) {
350 if (td_io_prep(td, io_u)) {
355 io_u->end_io = verify_io_u;
357 ret = td_io_queue(td, io_u);
359 case FIO_Q_COMPLETED:
362 else if (io_u->resid) {
363 int bytes = io_u->xfer_buflen - io_u->resid;
364 struct fio_file *f = io_u->file;
370 td_verror(td, ENODATA, "full resid");
375 io_u->xfer_buflen = io_u->resid;
376 io_u->xfer_buf += bytes;
377 io_u->offset += bytes;
379 td->ts.short_io_u[io_u->ddir]++;
381 if (io_u->offset == f->real_file_size)
384 requeue_io_u(td, &io_u);
387 ret = io_u_sync_complete(td, io_u);
395 requeue_io_u(td, &io_u);
396 ret2 = td_io_commit(td);
402 td_verror(td, -ret, "td_io_queue");
406 if (ret < 0 || td->error)
410 * if we can queue more, do so. but check if there are
411 * completed io_u's first.
414 if (queue_full(td) || ret == FIO_Q_BUSY) {
417 if (td->cur_depth > td->o.iodepth_low)
418 min_events = td->cur_depth - td->o.iodepth_low;
422 * Reap required number of io units, if any, and do the
423 * verification on them through the callback handler
425 if (io_u_queued_complete(td, min_events) < 0)
430 min_events = td->cur_depth;
433 ret = io_u_queued_complete(td, min_events);
435 cleanup_pending_aio(td);
437 td_set_runstate(td, TD_RUNNING);
441 * Main IO worker function. It retrieves io_u's to process and queues
442 * and reaps them, checking for rate and errors along the way.
444 static void do_io(struct thread_data *td)
451 td_set_runstate(td, TD_RUNNING);
453 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
454 struct timeval comp_time;
467 memcpy(&s, &io_u->start_time, sizeof(s));
469 if (runtime_exceeded(td, &s)) {
476 * Add verification end_io handler, if asked to verify
477 * a previously written file.
479 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
480 io_u->end_io = verify_io_u;
481 td_set_runstate(td, TD_VERIFYING);
483 td_set_runstate(td, TD_RUNNING);
485 ret = td_io_queue(td, io_u);
487 case FIO_Q_COMPLETED:
490 else if (io_u->resid) {
491 int bytes = io_u->xfer_buflen - io_u->resid;
492 struct fio_file *f = io_u->file;
498 td_verror(td, ENODATA, "full resid");
503 io_u->xfer_buflen = io_u->resid;
504 io_u->xfer_buf += bytes;
505 io_u->offset += bytes;
507 td->ts.short_io_u[io_u->ddir]++;
509 if (io_u->offset == f->real_file_size)
512 requeue_io_u(td, &io_u);
515 fio_gettime(&comp_time, NULL);
516 bytes_done = io_u_sync_complete(td, io_u);
523 * if the engine doesn't have a commit hook,
524 * the io_u is really queued. if it does have such
525 * a hook, it has to call io_u_queued() itself.
527 if (td->io_ops->commit == NULL)
528 io_u_queued(td, io_u);
531 requeue_io_u(td, &io_u);
532 ret2 = td_io_commit(td);
542 if (ret < 0 || td->error)
546 * See if we need to complete some commands
548 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
550 if (queue_full(td) || ret == FIO_Q_BUSY) {
553 if (td->cur_depth > td->o.iodepth_low)
554 min_evts = td->cur_depth
558 fio_gettime(&comp_time, NULL);
559 bytes_done = io_u_queued_complete(td, min_evts);
568 * the rate is batched for now, it should work for batches
569 * of completions except the very first one which may look
572 usec = utime_since(&s, &comp_time);
574 rate_throttle(td, usec, bytes_done);
576 if (check_min_rate(td, &comp_time)) {
577 if (exitall_on_terminate)
578 terminate_threads(td->groupid);
579 td_verror(td, ENODATA, "check_min_rate");
583 if (td->o.thinktime) {
584 unsigned long long b;
586 b = td->io_blocks[0] + td->io_blocks[1];
587 if (!(b % td->o.thinktime_blocks)) {
590 if (td->o.thinktime_spin)
591 __usec_sleep(td->o.thinktime_spin);
593 left = td->o.thinktime - td->o.thinktime_spin;
595 usec_sleep(td, left);
600 if (td->o.fill_device && td->error == ENOSPC) {
609 ret = io_u_queued_complete(td, i);
611 if (should_fsync(td) && td->o.end_fsync) {
612 td_set_runstate(td, TD_FSYNCING);
614 for_each_file(td, f, i) {
615 if (!(f->flags & FIO_FILE_OPEN))
621 cleanup_pending_aio(td);
624 * stop job if we failed doing any IO
626 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
630 static void cleanup_io_u(struct thread_data *td)
632 struct list_head *entry, *n;
635 list_for_each_safe(entry, n, &td->io_u_freelist) {
636 io_u = list_entry(entry, struct io_u, list);
638 list_del(&io_u->list);
646 * "randomly" fill the buffer contents
648 static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
650 long *ptr = io_u->buf;
652 if (!td->o.zero_buffers) {
653 while ((void *) ptr - io_u->buf < max_bs) {
654 *ptr = rand() * GOLDEN_RATIO_PRIME;
658 memset(ptr, 0, max_bs);
661 static int init_io_u(struct thread_data *td)
668 max_units = td->o.iodepth;
669 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
670 td->orig_buffer_size = (unsigned long long) max_bs
671 * (unsigned long long) max_units;
673 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
676 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
677 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
680 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
681 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
685 if (allocate_io_mem(td))
689 p = ALIGN(td->orig_buffer);
693 for (i = 0; i < max_units; i++) {
696 io_u = malloc(sizeof(*io_u));
697 memset(io_u, 0, sizeof(*io_u));
698 INIT_LIST_HEAD(&io_u->list);
700 if (!(td->io_ops->flags & FIO_NOIO)) {
701 io_u->buf = p + max_bs * i;
704 fill_io_buf(td, io_u, max_bs);
708 io_u->flags = IO_U_F_FREE;
709 list_add(&io_u->list, &td->io_u_freelist);
717 static int switch_ioscheduler(struct thread_data *td)
719 char tmp[256], tmp2[128];
723 if (td->io_ops->flags & FIO_DISKLESSIO)
726 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
728 f = fopen(tmp, "r+");
730 if (errno == ENOENT) {
731 log_err("fio: os or kernel doesn't support IO scheduler"
735 td_verror(td, errno, "fopen iosched");
742 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
743 if (ferror(f) || ret != 1) {
744 td_verror(td, errno, "fwrite");
752 * Read back and check that the selected scheduler is now the default.
754 ret = fread(tmp, 1, sizeof(tmp), f);
755 if (ferror(f) || ret < 0) {
756 td_verror(td, errno, "fread");
761 sprintf(tmp2, "[%s]", td->o.ioscheduler);
762 if (!strstr(tmp, tmp2)) {
763 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
764 td_verror(td, EINVAL, "iosched_switch");
773 static int keep_running(struct thread_data *td)
775 unsigned long long io_done;
779 if (td->o.time_based)
786 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
788 if (io_done < td->o.size)
794 static int clear_io_state(struct thread_data *td)
800 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
801 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
805 td->rw_end_set[0] = td->rw_end_set[1] = 0;
807 td->last_was_sync = 0;
810 * reset file done count if we are to start over
812 if (td->o.time_based || td->o.loops)
813 td->nr_done_files = 0;
818 for_each_file(td, f, i) {
819 f->flags &= ~FIO_FILE_DONE;
820 ret = td_io_open_file(td, f);
829 * Entry point for the thread based jobs. The process based jobs end up
830 * here as well, after a little setup.
832 static void *thread_main(void *data)
834 unsigned long long runtime[2], elapsed;
835 struct thread_data *td = data;
838 if (!td->o.use_thread)
843 dprint(FD_PROCESS, "jobs pid=%d started\n", td->pid);
845 INIT_LIST_HEAD(&td->io_u_freelist);
846 INIT_LIST_HEAD(&td->io_u_busylist);
847 INIT_LIST_HEAD(&td->io_u_requeues);
848 INIT_LIST_HEAD(&td->io_log_list);
849 INIT_LIST_HEAD(&td->io_hist_list);
850 td->io_hist_tree = RB_ROOT;
852 td_set_runstate(td, TD_INITIALIZED);
853 fio_mutex_up(startup_mutex);
854 fio_mutex_down(td->mutex);
857 * the ->mutex mutex is now no longer used, close it to avoid
858 * eating a file descriptor
860 fio_mutex_remove(td->mutex);
863 * May alter parameters that init_io_u() will use, so we need to
872 if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
873 td_verror(td, errno, "cpu_set_affinity");
877 if (td->ioprio_set) {
878 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
879 td_verror(td, errno, "ioprio_set");
884 if (nice(td->o.nice) == -1) {
885 td_verror(td, errno, "nice");
889 if (td->o.ioscheduler && switch_ioscheduler(td))
892 if (!td->o.create_serialize && setup_files(td))
901 if (init_random_map(td))
904 if (td->o.exec_prerun) {
905 if (system(td->o.exec_prerun) < 0)
909 fio_gettime(&td->epoch, NULL);
910 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
911 getrusage(RUSAGE_SELF, &td->ts.ru_start);
913 runtime[0] = runtime[1] = 0;
915 while (keep_running(td)) {
916 fio_gettime(&td->start, NULL);
917 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
920 memcpy(&td->lastrate, &td->ts.stat_sample_time,
921 sizeof(td->lastrate));
923 if (clear_state && clear_io_state(td))
926 prune_io_piece_log(td);
932 if (td_read(td) && td->io_bytes[DDIR_READ]) {
933 if (td->rw_end_set[DDIR_READ])
934 elapsed = utime_since(&td->start,
935 &td->rw_end[DDIR_READ]);
937 elapsed = utime_since_now(&td->start);
939 runtime[DDIR_READ] += elapsed;
941 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
942 if (td->rw_end_set[DDIR_WRITE])
943 elapsed = utime_since(&td->start,
944 &td->rw_end[DDIR_WRITE]);
946 elapsed = utime_since_now(&td->start);
948 runtime[DDIR_WRITE] += elapsed;
951 if (td->error || td->terminate)
954 if (!td->o.do_verify ||
955 td->o.verify == VERIFY_NONE ||
956 (td->io_ops->flags & FIO_UNIDIR))
959 if (clear_io_state(td))
962 fio_gettime(&td->start, NULL);
966 runtime[DDIR_READ] += utime_since_now(&td->start);
968 if (td->error || td->terminate)
972 update_rusage_stat(td);
973 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
974 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
975 td->ts.total_run_time = mtime_since_now(&td->epoch);
976 td->ts.io_bytes[0] = td->io_bytes[0];
977 td->ts.io_bytes[1] = td->io_bytes[1];
980 finish_log(td, td->ts.bw_log, "bw");
982 finish_log(td, td->ts.slat_log, "slat");
984 finish_log(td, td->ts.clat_log, "clat");
985 if (td->o.exec_postrun) {
986 if (system(td->o.exec_postrun) < 0)
987 log_err("fio: postrun %s failed\n", td->o.exec_postrun);
990 if (exitall_on_terminate)
991 terminate_threads(td->groupid);
995 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error,
997 close_and_free_files(td);
1002 * do this very late, it will log file closing as well
1004 if (td->o.write_iolog_file)
1005 write_iolog_close(td);
1007 options_mem_free(td);
1008 td_set_runstate(td, TD_EXITED);
1009 return (void *) (unsigned long) td->error;
1013 * We cannot pass the td data into a forked process, so attach the td and
1014 * pass it to the thread worker.
1016 static int fork_main(int shmid, int offset)
1018 struct thread_data *td;
1021 data = shmat(shmid, NULL, 0);
1022 if (data == (void *) -1) {
1029 td = data + offset * sizeof(struct thread_data);
1030 ret = thread_main(td);
1032 return (int) (unsigned long) ret;
1036 * Run over the job map and reap the threads that have exited, if any.
1038 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1040 struct thread_data *td;
1041 int i, cputhreads, realthreads, pending, status, ret;
1044 * reap exited threads (TD_EXITED -> TD_REAPED)
1046 realthreads = pending = cputhreads = 0;
1047 for_each_td(td, i) {
1051 * ->io_ops is NULL for a thread that has closed its
1054 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1063 if (td->runstate == TD_REAPED)
1065 if (td->o.use_thread) {
1066 if (td->runstate == TD_EXITED) {
1067 td_set_runstate(td, TD_REAPED);
1074 if (td->runstate == TD_EXITED)
1078 * check if someone quit or got killed in an unusual way
1080 ret = waitpid(td->pid, &status, flags);
1082 if (errno == ECHILD) {
1083 log_err("fio: pid=%d disappeared %d\n", td->pid,
1085 td_set_runstate(td, TD_REAPED);
1089 } else if (ret == td->pid) {
1090 if (WIFSIGNALED(status)) {
1091 int sig = WTERMSIG(status);
1094 log_err("fio: pid=%d, got signal=%d\n",
1096 td_set_runstate(td, TD_REAPED);
1099 if (WIFEXITED(status)) {
1100 if (WEXITSTATUS(status) && !td->error)
1101 td->error = WEXITSTATUS(status);
1103 td_set_runstate(td, TD_REAPED);
1109 * thread is not dead, continue
1115 (*m_rate) -= td->o.ratemin;
1116 (*t_rate) -= td->o.rate;
1123 done_secs += mtime_since_now(&td->epoch) / 1000;
1126 if (*nr_running == cputhreads && !pending && realthreads)
1127 terminate_threads(TERMINATE_ALL);
1131 * Main function for kicking off and reaping jobs, as needed.
1133 static void run_threads(void)
1135 struct thread_data *td;
1136 unsigned long spent;
1137 int i, todo, nr_running, m_rate, t_rate, nr_started;
1139 if (fio_pin_memory())
1142 if (!terse_output) {
1143 printf("Starting ");
1145 printf("%d thread%s", nr_thread,
1146 nr_thread > 1 ? "s" : "");
1150 printf("%d process%s", nr_process,
1151 nr_process > 1 ? "es" : "");
1157 signal(SIGINT, sig_handler);
1158 signal(SIGALRM, sig_handler);
1160 todo = thread_number;
1163 m_rate = t_rate = 0;
1165 for_each_td(td, i) {
1166 print_status_init(td->thread_number - 1);
1168 if (!td->o.create_serialize) {
1174 * do file setup here so it happens sequentially,
1175 * we don't want X number of threads getting their
1176 * client data interspersed on disk
1178 if (setup_files(td)) {
1181 log_err("fio: pid=%d, err=%d/%s\n", td->pid,
1182 td->error, td->verror);
1183 td_set_runstate(td, TD_REAPED);
1190 * for sharing to work, each job must always open
1191 * its own files. so close them, if we opened them
1194 for_each_file(td, f, i)
1195 td_io_close_file(td, f);
1204 struct thread_data *map[MAX_JOBS];
1205 struct timeval this_start;
1206 int this_jobs = 0, left;
1209 * create threads (TD_NOT_CREATED -> TD_CREATED)
1211 for_each_td(td, i) {
1212 if (td->runstate != TD_NOT_CREATED)
1216 * never got a chance to start, killed by other
1217 * thread for some reason
1219 if (td->terminate) {
1224 if (td->o.start_delay) {
1225 spent = mtime_since_genesis();
1227 if (td->o.start_delay * 1000 > spent)
1231 if (td->o.stonewall && (nr_started || nr_running)) {
1232 dprint(FD_PROCESS, "%s: stonewall wait\n",
1238 * Set state to created. Thread will transition
1239 * to TD_INITIALIZED when it's done setting up.
1241 td_set_runstate(td, TD_CREATED);
1242 map[this_jobs++] = td;
1245 if (td->o.use_thread) {
1246 dprint(FD_PROCESS, "will pthread_create\n");
1247 if (pthread_create(&td->thread, NULL,
1249 perror("pthread_create");
1253 if (pthread_detach(td->thread) < 0)
1254 perror("pthread_detach");
1256 dprint(FD_PROCESS, "will fork\n");
1258 int ret = fork_main(shm_id, i);
1263 fio_mutex_down(startup_mutex);
1267 * Wait for the started threads to transition to
1270 fio_gettime(&this_start, NULL);
1272 while (left && !fio_abort) {
1273 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1278 for (i = 0; i < this_jobs; i++) {
1282 if (td->runstate == TD_INITIALIZED) {
1285 } else if (td->runstate >= TD_EXITED) {
1289 nr_running++; /* work-around... */
1295 log_err("fio: %d jobs failed to start\n", left);
1296 for (i = 0; i < this_jobs; i++) {
1300 kill(td->pid, SIGTERM);
1306 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1308 for_each_td(td, i) {
1309 if (td->runstate != TD_INITIALIZED)
1312 td_set_runstate(td, TD_RUNNING);
1315 m_rate += td->o.ratemin;
1316 t_rate += td->o.rate;
1318 fio_mutex_up(td->mutex);
1321 reap_threads(&nr_running, &t_rate, &m_rate);
1327 while (nr_running) {
1328 reap_threads(&nr_running, &t_rate, &m_rate);
1336 int main(int argc, char *argv[])
1343 * We need locale for number printing, if it isn't set then just
1344 * go with the US format.
1346 if (!getenv("LC_NUMERIC"))
1347 setlocale(LC_NUMERIC, "en_US");
1349 if (parse_options(argc, argv))
1355 ps = sysconf(_SC_PAGESIZE);
1357 log_err("Failed to get page size\n");
1365 setup_log(&agg_io_log[DDIR_READ]);
1366 setup_log(&agg_io_log[DDIR_WRITE]);
1369 startup_mutex = fio_mutex_init(0);
1373 disk_util_timer_arm();
1380 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1381 __finish_log(agg_io_log[DDIR_WRITE],
1382 "agg-write_bw.log");
1386 fio_mutex_remove(startup_mutex);