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 unsigned long page_mask;
41 unsigned long page_size;
43 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
46 int thread_number = 0;
52 static struct fio_sem *startup_sem;
53 static volatile int fio_abort;
54 static int exit_value;
56 struct io_log *agg_io_log[2];
58 #define TERMINATE_ALL (-1)
59 #define JOB_START_TIMEOUT (5 * 1000)
61 static inline void td_set_runstate(struct thread_data *td, int runstate)
63 td->runstate = runstate;
66 static void terminate_threads(int group_id)
68 struct thread_data *td;
72 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
73 if (td->runstate <= TD_RUNNING)
74 kill(td->pid, SIGQUIT);
76 td->o.start_delay = 0;
81 static void sig_handler(int sig)
86 disk_util_timer_arm();
87 print_thread_status();
90 printf("\nfio: terminating on signal %d\n", sig);
92 terminate_threads(TERMINATE_ALL);
98 * Check if we are above the minimum rate given.
100 static int check_min_rate(struct thread_data *td, struct timeval *now)
102 unsigned long long bytes = 0;
103 unsigned long iops = 0;
108 * No minimum rate set, always ok
110 if (!td->o.ratemin && !td->o.rate_iops_min)
114 * allow a 2 second settle period in the beginning
116 if (mtime_since(&td->start, now) < 2000)
120 iops += td->io_blocks[DDIR_READ];
121 bytes += td->this_io_bytes[DDIR_READ];
124 iops += td->io_blocks[DDIR_WRITE];
125 bytes += td->this_io_bytes[DDIR_WRITE];
129 * if rate blocks is set, sample is running
131 if (td->rate_bytes || td->rate_blocks) {
132 spent = mtime_since(&td->lastrate, now);
133 if (spent < td->o.ratecycle)
138 * check bandwidth specified rate
140 if (bytes < td->rate_bytes) {
141 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
144 rate = (bytes - td->rate_bytes) / spent;
145 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
146 log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
152 * checks iops specified rate
154 if (iops < td->o.rate_iops) {
155 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
158 rate = (iops - td->rate_blocks) / spent;
159 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
160 log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
166 td->rate_bytes = bytes;
167 td->rate_blocks = iops;
168 memcpy(&td->lastrate, now, sizeof(*now));
172 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
176 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
183 * When job exits, we can cancel the in-flight IO if we are using async
184 * io. Attempt to do so.
186 static void cleanup_pending_aio(struct thread_data *td)
188 struct list_head *entry, *n;
193 * get immediately available events, if any
195 r = io_u_queued_complete(td, 0);
200 * now cancel remaining active events
202 if (td->io_ops->cancel) {
203 list_for_each_safe(entry, n, &td->io_u_busylist) {
204 io_u = list_entry(entry, struct io_u, list);
207 * if the io_u isn't in flight, then that generally
208 * means someone leaked an io_u. complain but fix
209 * it up, so we don't stall here.
211 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
212 log_err("fio: non-busy IO on busy list\n");
215 r = td->io_ops->cancel(td, io_u);
223 r = io_u_queued_complete(td, td->cur_depth);
227 * Helper to handle the final sync of a file. Works just like the normal
228 * io path, just does everything sync.
230 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
232 struct io_u *io_u = __get_io_u(td);
238 io_u->ddir = DDIR_SYNC;
241 if (td_io_prep(td, io_u)) {
247 ret = td_io_queue(td, io_u);
249 td_verror(td, io_u->error, "td_io_queue");
252 } else if (ret == FIO_Q_QUEUED) {
253 if (io_u_queued_complete(td, 1) < 0)
255 } else if (ret == FIO_Q_COMPLETED) {
257 td_verror(td, io_u->error, "td_io_queue");
261 if (io_u_sync_complete(td, io_u) < 0)
263 } else if (ret == FIO_Q_BUSY) {
264 if (td_io_commit(td))
273 * The main verify engine. Runs over the writes we previously submitted,
274 * reads the blocks back in, and checks the crc/md5 of the data.
276 static void do_verify(struct thread_data *td)
284 * sync io first and invalidate cache, to make sure we really
287 for_each_file(td, f, i) {
288 if (!(f->flags & FIO_FILE_OPEN))
290 if (fio_io_sync(td, f))
292 if (file_invalidate_cache(td, f))
299 td_set_runstate(td, TD_VERIFYING);
302 while (!td->terminate) {
305 io_u = __get_io_u(td);
309 if (runtime_exceeded(td, &io_u->start_time)) {
314 if (get_next_verify(td, io_u)) {
319 if (td_io_prep(td, io_u)) {
324 io_u->end_io = verify_io_u;
326 ret = td_io_queue(td, io_u);
328 case FIO_Q_COMPLETED:
331 else if (io_u->resid) {
332 int bytes = io_u->xfer_buflen - io_u->resid;
333 struct fio_file *f = io_u->file;
339 td_verror(td, ENODATA, "full resid");
344 io_u->xfer_buflen = io_u->resid;
345 io_u->xfer_buf += bytes;
346 io_u->offset += bytes;
347 f->last_completed_pos = io_u->offset;
349 td->ts.short_io_u[io_u->ddir]++;
351 if (io_u->offset == f->real_file_size)
354 requeue_io_u(td, &io_u);
357 ret = io_u_sync_complete(td, io_u);
365 requeue_io_u(td, &io_u);
366 ret2 = td_io_commit(td);
372 td_verror(td, -ret, "td_io_queue");
376 if (ret < 0 || td->error)
380 * if we can queue more, do so. but check if there are
381 * completed io_u's first.
384 if (queue_full(td) || ret == FIO_Q_BUSY) {
387 if (td->cur_depth > td->o.iodepth_low)
388 min_events = td->cur_depth - td->o.iodepth_low;
392 * Reap required number of io units, if any, and do the
393 * verification on them through the callback handler
395 if (io_u_queued_complete(td, min_events) < 0)
400 min_events = td->cur_depth;
403 ret = io_u_queued_complete(td, min_events);
405 cleanup_pending_aio(td);
407 td_set_runstate(td, TD_RUNNING);
411 * Main IO worker function. It retrieves io_u's to process and queues
412 * and reaps them, checking for rate and errors along the way.
414 static void do_io(struct thread_data *td)
421 td_set_runstate(td, TD_RUNNING);
423 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
424 struct timeval comp_time;
437 memcpy(&s, &io_u->start_time, sizeof(s));
439 if (runtime_exceeded(td, &s)) {
444 ret = td_io_queue(td, io_u);
446 case FIO_Q_COMPLETED:
449 else if (io_u->resid) {
450 int bytes = io_u->xfer_buflen - io_u->resid;
451 struct fio_file *f = io_u->file;
457 td_verror(td, ENODATA, "full resid");
462 io_u->xfer_buflen = io_u->resid;
463 io_u->xfer_buf += bytes;
464 io_u->offset += bytes;
465 f->last_completed_pos = io_u->offset;
467 td->ts.short_io_u[io_u->ddir]++;
469 if (io_u->offset == f->real_file_size)
472 requeue_io_u(td, &io_u);
475 fio_gettime(&comp_time, NULL);
476 bytes_done = io_u_sync_complete(td, io_u);
483 * if the engine doesn't have a commit hook,
484 * the io_u is really queued. if it does have such
485 * a hook, it has to call io_u_queued() itself.
487 if (td->io_ops->commit == NULL)
488 io_u_queued(td, io_u);
491 requeue_io_u(td, &io_u);
492 ret2 = td_io_commit(td);
502 if (ret < 0 || td->error)
506 * See if we need to complete some commands
508 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
510 if (queue_full(td) || ret == FIO_Q_BUSY) {
513 if (td->cur_depth > td->o.iodepth_low)
514 min_evts = td->cur_depth - td->o.iodepth_low;
517 fio_gettime(&comp_time, NULL);
518 bytes_done = io_u_queued_complete(td, min_evts);
527 * the rate is batched for now, it should work for batches
528 * of completions except the very first one which may look
531 usec = utime_since(&s, &comp_time);
533 rate_throttle(td, usec, bytes_done);
535 if (check_min_rate(td, &comp_time)) {
536 if (exitall_on_terminate)
537 terminate_threads(td->groupid);
538 td_verror(td, ENODATA, "check_min_rate");
542 if (td->o.thinktime) {
543 unsigned long long b;
545 b = td->io_blocks[0] + td->io_blocks[1];
546 if (!(b % td->o.thinktime_blocks)) {
549 if (td->o.thinktime_spin)
550 __usec_sleep(td->o.thinktime_spin);
552 left = td->o.thinktime - td->o.thinktime_spin;
554 usec_sleep(td, left);
564 ret = io_u_queued_complete(td, i);
566 if (should_fsync(td) && td->o.end_fsync) {
567 td_set_runstate(td, TD_FSYNCING);
569 for_each_file(td, f, i) {
570 if (!(f->flags & FIO_FILE_OPEN))
576 cleanup_pending_aio(td);
579 static void cleanup_io_u(struct thread_data *td)
581 struct list_head *entry, *n;
584 list_for_each_safe(entry, n, &td->io_u_freelist) {
585 io_u = list_entry(entry, struct io_u, list);
587 list_del(&io_u->list);
595 * "randomly" fill the buffer contents
597 static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
599 long *ptr = io_u->buf;
601 if (!td->o.zero_buffers) {
602 while ((void *) ptr - io_u->buf < max_bs) {
603 *ptr = rand() * GOLDEN_RATIO_PRIME;
607 memset(ptr, 0, max_bs);
610 static int init_io_u(struct thread_data *td)
612 unsigned long long buf_size;
618 if (td->io_ops->flags & FIO_SYNCIO)
621 max_units = td->o.iodepth;
623 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
624 buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
625 buf_size += page_mask;
626 if (buf_size != (size_t) buf_size) {
627 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
631 td->orig_buffer_size = buf_size;
633 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
634 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
635 else if (td->orig_buffer_size & page_mask)
636 td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
638 if (allocate_io_mem(td))
641 p = ALIGN(td->orig_buffer);
642 for (i = 0; i < max_units; i++) {
643 io_u = malloc(sizeof(*io_u));
644 memset(io_u, 0, sizeof(*io_u));
645 INIT_LIST_HEAD(&io_u->list);
647 io_u->buf = p + max_bs * i;
650 fill_io_buf(td, io_u, max_bs);
653 io_u->flags = IO_U_F_FREE;
654 list_add(&io_u->list, &td->io_u_freelist);
662 static int switch_ioscheduler(struct thread_data *td)
664 char tmp[256], tmp2[128];
668 if (td->io_ops->flags & FIO_DISKLESSIO)
671 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
673 f = fopen(tmp, "r+");
675 if (errno == ENOENT) {
676 log_err("fio: os or kernel doesn't support IO scheduler switching\n");
679 td_verror(td, errno, "fopen iosched");
686 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
687 if (ferror(f) || ret != 1) {
688 td_verror(td, errno, "fwrite");
696 * Read back and check that the selected scheduler is now the default.
698 ret = fread(tmp, 1, sizeof(tmp), f);
699 if (ferror(f) || ret < 0) {
700 td_verror(td, errno, "fread");
705 sprintf(tmp2, "[%s]", td->o.ioscheduler);
706 if (!strstr(tmp, tmp2)) {
707 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
708 td_verror(td, EINVAL, "iosched_switch");
717 static int clear_io_state(struct thread_data *td)
723 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
724 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
728 td->rw_end_set[0] = td->rw_end_set[1] = 0;
730 td->last_was_sync = 0;
732 for_each_file(td, f, i)
733 td_io_close_file(td, f);
736 for_each_file(td, f, i) {
737 ret = td_io_open_file(td, f);
746 * Entry point for the thread based jobs. The process based jobs end up
747 * here as well, after a little setup.
749 static void *thread_main(void *data)
751 unsigned long long runtime[2];
752 struct thread_data *td = data;
753 unsigned long elapsed;
757 if (!td->o.use_thread)
762 INIT_LIST_HEAD(&td->io_u_freelist);
763 INIT_LIST_HEAD(&td->io_u_busylist);
764 INIT_LIST_HEAD(&td->io_u_requeues);
765 INIT_LIST_HEAD(&td->io_log_list);
766 INIT_LIST_HEAD(&td->io_hist_list);
767 td->io_hist_tree = RB_ROOT;
772 if (fio_setaffinity(td) == -1) {
773 td_verror(td, errno, "cpu_set_affinity");
781 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
782 td_verror(td, errno, "ioprio_set");
787 if (nice(td->o.nice) == -1) {
788 td_verror(td, errno, "nice");
792 if (td->o.ioscheduler && switch_ioscheduler(td))
795 td_set_runstate(td, TD_INITIALIZED);
796 fio_sem_up(startup_sem);
797 fio_sem_down(td->mutex);
800 * the ->mutex semaphore is now no longer used, close it to avoid
801 * eating a file descriptor
803 fio_sem_remove(td->mutex);
805 if (!td->o.create_serialize && setup_files(td))
814 if (init_random_map(td))
817 if (td->o.exec_prerun) {
818 if (system(td->o.exec_prerun) < 0)
822 fio_gettime(&td->epoch, NULL);
823 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
824 getrusage(RUSAGE_SELF, &td->ts.ru_start);
826 runtime[0] = runtime[1] = 0;
828 while (td->o.time_based || td->o.loops--) {
829 fio_gettime(&td->start, NULL);
830 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
833 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
835 if (clear_state && clear_io_state(td))
838 prune_io_piece_log(td);
844 if (td_read(td) && td->io_bytes[DDIR_READ]) {
845 if (td->rw_end_set[DDIR_READ])
846 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
848 elapsed = utime_since_now(&td->start);
850 runtime[DDIR_READ] += elapsed;
852 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
853 if (td->rw_end_set[DDIR_WRITE])
854 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
856 elapsed = utime_since_now(&td->start);
858 runtime[DDIR_WRITE] += elapsed;
861 if (td->error || td->terminate)
864 fio_gettime(&t, NULL);
865 if (runtime_exceeded(td, &t))
868 if (td->o.verify == VERIFY_NONE)
871 if (clear_io_state(td))
874 fio_gettime(&td->start, NULL);
878 runtime[DDIR_READ] += utime_since_now(&td->start);
880 if (td->error || td->terminate)
884 update_rusage_stat(td);
885 td->ts.runtime[0] = runtime[0] / 1000;
886 td->ts.runtime[1] = runtime[1] / 1000;
887 td->ts.total_run_time = mtime_since_now(&td->epoch);
888 td->ts.io_bytes[0] = td->io_bytes[0];
889 td->ts.io_bytes[1] = td->io_bytes[1];
892 finish_log(td, td->ts.bw_log, "bw");
894 finish_log(td, td->ts.slat_log, "slat");
896 finish_log(td, td->ts.clat_log, "clat");
897 if (td->o.write_iolog_file)
898 write_iolog_close(td);
899 if (td->o.exec_postrun) {
900 if (system(td->o.exec_postrun) < 0)
901 log_err("fio: postrun %s failed\n", td->o.exec_postrun);
904 if (exitall_on_terminate)
905 terminate_threads(td->groupid);
909 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
913 options_mem_free(td);
914 td_set_runstate(td, TD_EXITED);
915 return (void *) (unsigned long) td->error;
917 fio_sem_up(startup_sem);
922 * We cannot pass the td data into a forked process, so attach the td and
923 * pass it to the thread worker.
925 static int fork_main(int shmid, int offset)
927 struct thread_data *td;
930 data = shmat(shmid, NULL, 0);
931 if (data == (void *) -1) {
938 td = data + offset * sizeof(struct thread_data);
939 ret = thread_main(td);
941 return (int) (unsigned long) ret;
945 * Run over the job map and reap the threads that have exited, if any.
947 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
949 struct thread_data *td;
950 int i, cputhreads, pending, status, ret;
953 * reap exited threads (TD_EXITED -> TD_REAPED)
955 pending = cputhreads = 0;
960 * ->io_ops is NULL for a thread that has closed its
963 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
966 if (!td->pid || td->runstate == TD_REAPED)
968 if (td->o.use_thread) {
969 if (td->runstate == TD_EXITED) {
970 td_set_runstate(td, TD_REAPED);
977 if (td->runstate == TD_EXITED)
981 * check if someone quit or got killed in an unusual way
983 ret = waitpid(td->pid, &status, flags);
985 if (errno == ECHILD) {
986 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
987 td_set_runstate(td, TD_REAPED);
991 } else if (ret == td->pid) {
992 if (WIFSIGNALED(status)) {
993 int sig = WTERMSIG(status);
996 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
997 td_set_runstate(td, TD_REAPED);
1000 if (WIFEXITED(status)) {
1001 if (WEXITSTATUS(status) && !td->error)
1002 td->error = WEXITSTATUS(status);
1004 td_set_runstate(td, TD_REAPED);
1010 * thread is not dead, continue
1015 if (td->o.use_thread) {
1018 if (pthread_join(td->thread, (void *) &ret))
1019 perror("pthread_join");
1023 (*m_rate) -= td->o.ratemin;
1024 (*t_rate) -= td->o.rate;
1031 if (*nr_running == cputhreads && !pending)
1032 terminate_threads(TERMINATE_ALL);
1036 * Main function for kicking off and reaping jobs, as needed.
1038 static void run_threads(void)
1040 struct thread_data *td;
1041 unsigned long spent;
1042 int i, todo, nr_running, m_rate, t_rate, nr_started;
1044 if (fio_pin_memory())
1047 if (!terse_output) {
1048 printf("Starting ");
1050 printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1054 printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1060 signal(SIGINT, sig_handler);
1061 signal(SIGALRM, sig_handler);
1063 todo = thread_number;
1066 m_rate = t_rate = 0;
1068 for_each_td(td, i) {
1069 print_status_init(td->thread_number - 1);
1071 if (!td->o.create_serialize) {
1077 * do file setup here so it happens sequentially,
1078 * we don't want X number of threads getting their
1079 * client data interspersed on disk
1081 if (setup_files(td)) {
1084 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1085 td_set_runstate(td, TD_REAPED);
1095 struct thread_data *map[MAX_JOBS];
1096 struct timeval this_start;
1097 int this_jobs = 0, left;
1100 * create threads (TD_NOT_CREATED -> TD_CREATED)
1102 for_each_td(td, i) {
1103 if (td->runstate != TD_NOT_CREATED)
1107 * never got a chance to start, killed by other
1108 * thread for some reason
1110 if (td->terminate) {
1115 if (td->o.start_delay) {
1116 spent = mtime_since_genesis();
1118 if (td->o.start_delay * 1000 > spent)
1122 if (td->o.stonewall && (nr_started || nr_running))
1126 * Set state to created. Thread will transition
1127 * to TD_INITIALIZED when it's done setting up.
1129 td_set_runstate(td, TD_CREATED);
1130 map[this_jobs++] = td;
1133 if (td->o.use_thread) {
1134 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1135 perror("thread_create");
1141 int ret = fork_main(shm_id, i);
1146 fio_sem_down(startup_sem);
1150 * Wait for the started threads to transition to
1153 fio_gettime(&this_start, NULL);
1155 while (left && !fio_abort) {
1156 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1161 for (i = 0; i < this_jobs; i++) {
1165 if (td->runstate == TD_INITIALIZED) {
1168 } else if (td->runstate >= TD_EXITED) {
1172 nr_running++; /* work-around... */
1178 log_err("fio: %d jobs failed to start\n", left);
1179 for (i = 0; i < this_jobs; i++) {
1183 kill(td->pid, SIGTERM);
1189 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1191 for_each_td(td, i) {
1192 if (td->runstate != TD_INITIALIZED)
1195 td_set_runstate(td, TD_RUNNING);
1198 m_rate += td->o.ratemin;
1199 t_rate += td->o.rate;
1201 fio_sem_up(td->mutex);
1204 reap_threads(&nr_running, &t_rate, &m_rate);
1210 while (nr_running) {
1211 reap_threads(&nr_running, &t_rate, &m_rate);
1219 int main(int argc, char *argv[])
1224 * We need locale for number printing, if it isn't set then just
1225 * go with the US format.
1227 if (!getenv("LC_NUMERIC"))
1228 setlocale(LC_NUMERIC, "en_US");
1230 if (parse_options(argc, argv))
1236 ps = sysconf(_SC_PAGESIZE);
1238 log_err("Failed to get page size\n");
1246 setup_log(&agg_io_log[DDIR_READ]);
1247 setup_log(&agg_io_log[DDIR_WRITE]);
1250 startup_sem = fio_sem_init(0);
1254 disk_util_timer_arm();
1261 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1262 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1266 fio_sem_remove(startup_sem);