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
51 unsigned long page_mask;
52 unsigned long page_size;
54 #define PAGE_ALIGN(buf) \
55 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
58 unsigned int thread_number = 0;
59 unsigned int nr_process = 0;
60 unsigned int nr_thread = 0;
63 unsigned long done_secs = 0;
66 * Just expose an empty list, if the OS does not support disk util stats
68 #ifndef FIO_HAVE_DISK_UTIL
69 FLIST_HEAD(disk_list);
72 static struct fio_mutex *startup_mutex;
73 static struct fio_mutex *writeout_mutex;
74 static volatile int fio_abort;
75 static int exit_value;
76 static pthread_t gtod_thread;
77 static pthread_t disk_util_thread;
78 static struct flist_head *cgroup_list;
79 static char *cgroup_mnt;
81 unsigned long arch_flags = 0;
83 struct io_log *agg_io_log[2];
85 #define JOB_START_TIMEOUT (5 * 1000)
87 static const char *fio_os_strings[os_nr] = {
99 static const char *fio_arch_strings[arch_nr] = {
115 const char *fio_get_os_string(int nr)
118 return fio_os_strings[nr];
123 const char *fio_get_arch_string(int nr)
126 return fio_arch_strings[nr];
131 void td_set_runstate(struct thread_data *td, int runstate)
133 if (td->runstate == runstate)
136 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
137 td->runstate, runstate);
138 td->runstate = runstate;
141 void fio_terminate_threads(int group_id)
143 struct thread_data *td;
146 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
149 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
150 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
151 td->o.name, (int) td->pid);
153 td->o.start_delay = 0;
156 * if the thread is running, just let it exit
160 else if (td->runstate < TD_RAMP)
161 kill(td->pid, SIGTERM);
163 struct ioengine_ops *ops = td->io_ops;
165 if (ops && (ops->flags & FIO_SIGTERM))
166 kill(td->pid, SIGTERM);
172 static void sig_int(int sig)
176 fio_server_got_signal(sig);
178 log_info("\nfio: terminating on signal %d\n", sig);
183 fio_terminate_threads(TERMINATE_ALL);
187 static void *disk_thread_main(void *data)
189 fio_mutex_up(startup_mutex);
192 usleep(DISK_UTIL_MSEC * 1000);
198 print_thread_status();
204 static int create_disk_util_thread(void)
208 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
210 log_err("Can't create disk util thread: %s\n", strerror(ret));
214 ret = pthread_detach(disk_util_thread);
216 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
220 dprint(FD_MUTEX, "wait on startup_mutex\n");
221 fio_mutex_down(startup_mutex);
222 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
226 static void set_sig_handlers(void)
228 struct sigaction act;
230 memset(&act, 0, sizeof(act));
231 act.sa_handler = sig_int;
232 act.sa_flags = SA_RESTART;
233 sigaction(SIGINT, &act, NULL);
235 memset(&act, 0, sizeof(act));
236 act.sa_handler = sig_int;
237 act.sa_flags = SA_RESTART;
238 sigaction(SIGTERM, &act, NULL);
241 memset(&act, 0, sizeof(act));
242 act.sa_handler = sig_int;
243 act.sa_flags = SA_RESTART;
244 sigaction(SIGPIPE, &act, NULL);
249 * Check if we are above the minimum rate given.
251 static int __check_min_rate(struct thread_data *td, struct timeval *now,
254 unsigned long long bytes = 0;
255 unsigned long iops = 0;
258 unsigned int ratemin = 0;
259 unsigned int rate_iops = 0;
260 unsigned int rate_iops_min = 0;
262 assert(ddir_rw(ddir));
264 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
268 * allow a 2 second settle period in the beginning
270 if (mtime_since(&td->start, now) < 2000)
273 iops += td->this_io_blocks[ddir];
274 bytes += td->this_io_bytes[ddir];
275 ratemin += td->o.ratemin[ddir];
276 rate_iops += td->o.rate_iops[ddir];
277 rate_iops_min += td->o.rate_iops_min[ddir];
280 * if rate blocks is set, sample is running
282 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
283 spent = mtime_since(&td->lastrate[ddir], now);
284 if (spent < td->o.ratecycle)
287 if (td->o.rate[ddir]) {
289 * check bandwidth specified rate
291 if (bytes < td->rate_bytes[ddir]) {
292 log_err("%s: min rate %u not met\n", td->o.name,
296 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
297 if (rate < ratemin ||
298 bytes < td->rate_bytes[ddir]) {
299 log_err("%s: min rate %u not met, got"
300 " %luKB/sec\n", td->o.name,
307 * checks iops specified rate
309 if (iops < rate_iops) {
310 log_err("%s: min iops rate %u not met\n",
311 td->o.name, rate_iops);
314 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
315 if (rate < rate_iops_min ||
316 iops < td->rate_blocks[ddir]) {
317 log_err("%s: min iops rate %u not met,"
318 " got %lu\n", td->o.name,
319 rate_iops_min, rate);
325 td->rate_bytes[ddir] = bytes;
326 td->rate_blocks[ddir] = iops;
327 memcpy(&td->lastrate[ddir], now, sizeof(*now));
331 static int check_min_rate(struct thread_data *td, struct timeval *now,
332 unsigned long *bytes_done)
337 ret |= __check_min_rate(td, now, 0);
339 ret |= __check_min_rate(td, now, 1);
344 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
348 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
355 * When job exits, we can cancel the in-flight IO if we are using async
356 * io. Attempt to do so.
358 static void cleanup_pending_aio(struct thread_data *td)
360 struct flist_head *entry, *n;
365 * get immediately available events, if any
367 r = io_u_queued_complete(td, 0, NULL);
372 * now cancel remaining active events
374 if (td->io_ops->cancel) {
375 flist_for_each_safe(entry, n, &td->io_u_busylist) {
376 io_u = flist_entry(entry, struct io_u, list);
379 * if the io_u isn't in flight, then that generally
380 * means someone leaked an io_u. complain but fix
381 * it up, so we don't stall here.
383 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
384 log_err("fio: non-busy IO on busy list\n");
387 r = td->io_ops->cancel(td, io_u);
395 r = io_u_queued_complete(td, td->cur_depth, NULL);
399 * Helper to handle the final sync of a file. Works just like the normal
400 * io path, just does everything sync.
402 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
404 struct io_u *io_u = __get_io_u(td);
410 io_u->ddir = DDIR_SYNC;
413 if (td_io_prep(td, io_u)) {
419 ret = td_io_queue(td, io_u);
421 td_verror(td, io_u->error, "td_io_queue");
424 } else if (ret == FIO_Q_QUEUED) {
425 if (io_u_queued_complete(td, 1, NULL) < 0)
427 } else if (ret == FIO_Q_COMPLETED) {
429 td_verror(td, io_u->error, "td_io_queue");
433 if (io_u_sync_complete(td, io_u, NULL) < 0)
435 } else if (ret == FIO_Q_BUSY) {
436 if (td_io_commit(td))
444 static inline void __update_tv_cache(struct thread_data *td)
446 fio_gettime(&td->tv_cache, NULL);
449 static inline void update_tv_cache(struct thread_data *td)
451 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
452 __update_tv_cache(td);
455 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
460 if (ret < 0 || td->error) {
468 if (!(td->o.continue_on_error & td_error_type(ddir, err)))
471 if (td_non_fatal_error(err)) {
473 * Continue with the I/Os in case of
476 update_error_count(td, err);
480 } else if (td->o.fill_device && err == ENOSPC) {
482 * We expect to hit this error if
483 * fill_device option is set.
490 * Stop the I/O in case of a fatal
493 update_error_count(td, err);
502 * The main verify engine. Runs over the writes we previously submitted,
503 * reads the blocks back in, and checks the crc/md5 of the data.
505 static void do_verify(struct thread_data *td)
512 dprint(FD_VERIFY, "starting loop\n");
515 * sync io first and invalidate cache, to make sure we really
518 for_each_file(td, f, i) {
519 if (!fio_file_open(f))
521 if (fio_io_sync(td, f))
523 if (file_invalidate_cache(td, f))
530 td_set_runstate(td, TD_VERIFYING);
533 while (!td->terminate) {
538 if (runtime_exceeded(td, &td->tv_cache)) {
539 __update_tv_cache(td);
540 if (runtime_exceeded(td, &td->tv_cache)) {
546 io_u = __get_io_u(td);
550 if (get_next_verify(td, io_u)) {
555 if (td_io_prep(td, io_u)) {
560 if (td->o.verify_async)
561 io_u->end_io = verify_io_u_async;
563 io_u->end_io = verify_io_u;
565 ret = td_io_queue(td, io_u);
567 case FIO_Q_COMPLETED:
570 clear_io_u(td, io_u);
571 } else if (io_u->resid) {
572 int bytes = io_u->xfer_buflen - io_u->resid;
578 td_verror(td, EIO, "full resid");
583 io_u->xfer_buflen = io_u->resid;
584 io_u->xfer_buf += bytes;
585 io_u->offset += bytes;
587 if (ddir_rw(io_u->ddir))
588 td->ts.short_io_u[io_u->ddir]++;
591 if (io_u->offset == f->real_file_size)
594 requeue_io_u(td, &io_u);
597 ret = io_u_sync_complete(td, io_u, NULL);
605 requeue_io_u(td, &io_u);
606 ret2 = td_io_commit(td);
612 td_verror(td, -ret, "td_io_queue");
616 if (break_on_this_error(td, io_u->ddir, &ret))
620 * if we can queue more, do so. but check if there are
621 * completed io_u's first. Note that we can get BUSY even
622 * without IO queued, if the system is resource starved.
624 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
625 if (full || !td->o.iodepth_batch_complete) {
626 min_events = min(td->o.iodepth_batch_complete,
628 if (full && !min_events && td->o.iodepth_batch_complete != 0)
633 * Reap required number of io units, if any,
634 * and do the verification on them through
635 * the callback handler
637 if (io_u_queued_complete(td, min_events, NULL) < 0) {
641 } while (full && (td->cur_depth > td->o.iodepth_low));
648 min_events = td->cur_depth;
651 ret = io_u_queued_complete(td, min_events, NULL);
653 cleanup_pending_aio(td);
655 td_set_runstate(td, TD_RUNNING);
657 dprint(FD_VERIFY, "exiting loop\n");
661 * Main IO worker function. It retrieves io_u's to process and queues
662 * and reaps them, checking for rate and errors along the way.
664 static void do_io(struct thread_data *td)
669 if (in_ramp_time(td))
670 td_set_runstate(td, TD_RAMP);
672 td_set_runstate(td, TD_RUNNING);
674 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
675 (!flist_empty(&td->trim_list)) ||
676 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
677 struct timeval comp_time;
678 unsigned long bytes_done[2] = { 0, 0 };
689 if (runtime_exceeded(td, &td->tv_cache)) {
690 __update_tv_cache(td);
691 if (runtime_exceeded(td, &td->tv_cache)) {
704 * Add verification end_io handler, if asked to verify
705 * a previously written file.
707 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
709 if (td->o.verify_async)
710 io_u->end_io = verify_io_u_async;
712 io_u->end_io = verify_io_u;
713 td_set_runstate(td, TD_VERIFYING);
714 } else if (in_ramp_time(td))
715 td_set_runstate(td, TD_RAMP);
717 td_set_runstate(td, TD_RUNNING);
719 ret = td_io_queue(td, io_u);
721 case FIO_Q_COMPLETED:
724 clear_io_u(td, io_u);
725 } else if (io_u->resid) {
726 int bytes = io_u->xfer_buflen - io_u->resid;
727 struct fio_file *f = io_u->file;
733 td_verror(td, EIO, "full resid");
738 io_u->xfer_buflen = io_u->resid;
739 io_u->xfer_buf += bytes;
740 io_u->offset += bytes;
742 if (ddir_rw(io_u->ddir))
743 td->ts.short_io_u[io_u->ddir]++;
745 if (io_u->offset == f->real_file_size)
748 requeue_io_u(td, &io_u);
751 if (__should_check_rate(td, 0) ||
752 __should_check_rate(td, 1))
753 fio_gettime(&comp_time, NULL);
755 ret = io_u_sync_complete(td, io_u, bytes_done);
762 * if the engine doesn't have a commit hook,
763 * the io_u is really queued. if it does have such
764 * a hook, it has to call io_u_queued() itself.
766 if (td->io_ops->commit == NULL)
767 io_u_queued(td, io_u);
770 requeue_io_u(td, &io_u);
771 ret2 = td_io_commit(td);
781 if (break_on_this_error(td, ddir, &ret))
785 * See if we need to complete some commands. Note that we
786 * can get BUSY even without IO queued, if the system is
789 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
790 if (full || !td->o.iodepth_batch_complete) {
791 min_evts = min(td->o.iodepth_batch_complete,
793 if (full && !min_evts && td->o.iodepth_batch_complete != 0)
796 if (__should_check_rate(td, 0) ||
797 __should_check_rate(td, 1))
798 fio_gettime(&comp_time, NULL);
801 ret = io_u_queued_complete(td, min_evts, bytes_done);
805 } while (full && (td->cur_depth > td->o.iodepth_low));
810 if (!(bytes_done[0] + bytes_done[1]))
813 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
814 if (check_min_rate(td, &comp_time, bytes_done)) {
815 if (exitall_on_terminate)
816 fio_terminate_threads(td->groupid);
817 td_verror(td, EIO, "check_min_rate");
822 if (td->o.thinktime) {
823 unsigned long long b;
825 b = td->io_blocks[0] + td->io_blocks[1];
826 if (!(b % td->o.thinktime_blocks)) {
829 if (td->o.thinktime_spin)
830 usec_spin(td->o.thinktime_spin);
832 left = td->o.thinktime - td->o.thinktime_spin;
834 usec_sleep(td, left);
839 if (td->trim_entries)
840 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
842 if (td->o.fill_device && td->error == ENOSPC) {
851 ret = io_u_queued_complete(td, i, NULL);
852 if (td->o.fill_device && td->error == ENOSPC)
856 if (should_fsync(td) && td->o.end_fsync) {
857 td_set_runstate(td, TD_FSYNCING);
859 for_each_file(td, f, i) {
860 if (!fio_file_open(f))
866 cleanup_pending_aio(td);
869 * stop job if we failed doing any IO
871 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
875 static void cleanup_io_u(struct thread_data *td)
877 struct flist_head *entry, *n;
880 flist_for_each_safe(entry, n, &td->io_u_freelist) {
881 io_u = flist_entry(entry, struct io_u, list);
883 flist_del(&io_u->list);
884 fio_memfree(io_u, sizeof(*io_u));
890 static int init_io_u(struct thread_data *td)
894 int cl_align, i, max_units;
897 max_units = td->o.iodepth;
898 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
899 td->orig_buffer_size = (unsigned long long) max_bs
900 * (unsigned long long) max_units;
902 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
905 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
906 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
909 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
910 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
914 if (allocate_io_mem(td))
917 if (td->o.odirect || td->o.mem_align ||
918 (td->io_ops->flags & FIO_RAWIO))
919 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
923 cl_align = os_cache_line_size();
925 for (i = 0; i < max_units; i++) {
931 ptr = fio_memalign(cl_align, sizeof(*io_u));
933 log_err("fio: unable to allocate aligned memory\n");
938 memset(io_u, 0, sizeof(*io_u));
939 INIT_FLIST_HEAD(&io_u->list);
940 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
942 if (!(td->io_ops->flags & FIO_NOIO)) {
944 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
947 io_u_fill_buffer(td, io_u, max_bs);
948 if (td_write(td) && td->o.verify_pattern_bytes) {
950 * Fill the buffer with the pattern if we are
951 * going to be doing writes.
953 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
958 io_u->flags = IO_U_F_FREE;
959 flist_add(&io_u->list, &td->io_u_freelist);
966 static int switch_ioscheduler(struct thread_data *td)
968 char tmp[256], tmp2[128];
972 if (td->io_ops->flags & FIO_DISKLESSIO)
975 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
977 f = fopen(tmp, "r+");
979 if (errno == ENOENT) {
980 log_err("fio: os or kernel doesn't support IO scheduler"
984 td_verror(td, errno, "fopen iosched");
991 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
992 if (ferror(f) || ret != 1) {
993 td_verror(td, errno, "fwrite");
1001 * Read back and check that the selected scheduler is now the default.
1003 ret = fread(tmp, 1, sizeof(tmp), f);
1004 if (ferror(f) || ret < 0) {
1005 td_verror(td, errno, "fread");
1010 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1011 if (!strstr(tmp, tmp2)) {
1012 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1013 td_verror(td, EINVAL, "iosched_switch");
1022 static int keep_running(struct thread_data *td)
1024 unsigned long long io_done;
1028 if (td->o.time_based)
1035 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
1036 + td->io_skip_bytes;
1037 if (io_done < td->o.size)
1043 static void reset_io_counters(struct thread_data *td)
1045 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1046 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1047 td->stat_io_blocks[0] = td->stat_io_blocks[1] = 0;
1048 td->this_io_blocks[0] = td->this_io_blocks[1] = 0;
1050 td->rate_bytes[0] = td->rate_bytes[1] = 0;
1051 td->rate_blocks[0] = td->rate_blocks[1] = 0;
1053 td->last_was_sync = 0;
1056 * reset file done count if we are to start over
1058 if (td->o.time_based || td->o.loops)
1059 td->nr_done_files = 0;
1062 void reset_all_stats(struct thread_data *td)
1067 reset_io_counters(td);
1069 for (i = 0; i < 2; i++) {
1070 td->io_bytes[i] = 0;
1071 td->io_blocks[i] = 0;
1072 td->io_issues[i] = 0;
1073 td->ts.total_io_u[i] = 0;
1076 fio_gettime(&tv, NULL);
1077 td->ts.runtime[0] = 0;
1078 td->ts.runtime[1] = 0;
1079 memcpy(&td->epoch, &tv, sizeof(tv));
1080 memcpy(&td->start, &tv, sizeof(tv));
1083 static void clear_io_state(struct thread_data *td)
1088 reset_io_counters(td);
1091 for_each_file(td, f, i)
1092 fio_file_clear_done(f);
1095 * Set the same seed to get repeatable runs
1097 td_fill_rand_seeds(td);
1100 static int exec_string(const char *string)
1102 int ret, newlen = strlen(string) + 1 + 8;
1105 str = malloc(newlen);
1106 sprintf(str, "sh -c %s", string);
1110 log_err("fio: exec of cmd <%s> failed\n", str);
1117 * Entry point for the thread based jobs. The process based jobs end up
1118 * here as well, after a little setup.
1120 static void *thread_main(void *data)
1122 unsigned long long elapsed;
1123 struct thread_data *td = data;
1124 pthread_condattr_t attr;
1127 if (!td->o.use_thread) {
1133 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1135 INIT_FLIST_HEAD(&td->io_u_freelist);
1136 INIT_FLIST_HEAD(&td->io_u_busylist);
1137 INIT_FLIST_HEAD(&td->io_u_requeues);
1138 INIT_FLIST_HEAD(&td->io_log_list);
1139 INIT_FLIST_HEAD(&td->io_hist_list);
1140 INIT_FLIST_HEAD(&td->verify_list);
1141 INIT_FLIST_HEAD(&td->trim_list);
1142 pthread_mutex_init(&td->io_u_lock, NULL);
1143 td->io_hist_tree = RB_ROOT;
1145 pthread_condattr_init(&attr);
1146 pthread_cond_init(&td->verify_cond, &attr);
1147 pthread_cond_init(&td->free_cond, &attr);
1149 td_set_runstate(td, TD_INITIALIZED);
1150 dprint(FD_MUTEX, "up startup_mutex\n");
1151 fio_mutex_up(startup_mutex);
1152 dprint(FD_MUTEX, "wait on td->mutex\n");
1153 fio_mutex_down(td->mutex);
1154 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1157 * the ->mutex mutex is now no longer used, close it to avoid
1158 * eating a file descriptor
1160 fio_mutex_remove(td->mutex);
1163 * A new gid requires privilege, so we need to do this before setting
1166 if (td->o.gid != -1U && setgid(td->o.gid)) {
1167 td_verror(td, errno, "setgid");
1170 if (td->o.uid != -1U && setuid(td->o.uid)) {
1171 td_verror(td, errno, "setuid");
1176 * If we have a gettimeofday() thread, make sure we exclude that
1177 * thread from this job
1180 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1183 * Set affinity first, in case it has an impact on the memory
1186 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1187 td_verror(td, errno, "cpu_set_affinity");
1192 * May alter parameters that init_io_u() will use, so we need to
1201 if (td->o.verify_async && verify_async_init(td))
1204 if (td->ioprio_set) {
1205 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1206 td_verror(td, errno, "ioprio_set");
1211 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1214 if (nice(td->o.nice) == -1) {
1215 td_verror(td, errno, "nice");
1219 if (td->o.ioscheduler && switch_ioscheduler(td))
1222 if (!td->o.create_serialize && setup_files(td))
1228 if (init_random_map(td))
1231 if (td->o.exec_prerun) {
1232 if (exec_string(td->o.exec_prerun))
1236 if (td->o.pre_read) {
1237 if (pre_read_files(td) < 0)
1241 fio_gettime(&td->epoch, NULL);
1242 getrusage(RUSAGE_SELF, &td->ru_start);
1245 while (keep_running(td)) {
1246 fio_gettime(&td->start, NULL);
1247 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1248 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1249 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1251 if (td->o.ratemin[0] || td->o.ratemin[1]) {
1252 memcpy(&td->lastrate[0], &td->bw_sample_time,
1253 sizeof(td->bw_sample_time));
1254 memcpy(&td->lastrate[1], &td->bw_sample_time,
1255 sizeof(td->bw_sample_time));
1261 prune_io_piece_log(td);
1267 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1268 elapsed = utime_since_now(&td->start);
1269 td->ts.runtime[DDIR_READ] += elapsed;
1271 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1272 elapsed = utime_since_now(&td->start);
1273 td->ts.runtime[DDIR_WRITE] += elapsed;
1276 if (td->error || td->terminate)
1279 if (!td->o.do_verify ||
1280 td->o.verify == VERIFY_NONE ||
1281 (td->io_ops->flags & FIO_UNIDIR))
1286 fio_gettime(&td->start, NULL);
1290 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1292 if (td->error || td->terminate)
1296 update_rusage_stat(td);
1297 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1298 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1299 td->ts.total_run_time = mtime_since_now(&td->epoch);
1300 td->ts.io_bytes[0] = td->io_bytes[0];
1301 td->ts.io_bytes[1] = td->io_bytes[1];
1303 fio_mutex_down(writeout_mutex);
1305 if (td->o.bw_log_file) {
1306 finish_log_named(td, td->bw_log,
1307 td->o.bw_log_file, "bw");
1309 finish_log(td, td->bw_log, "bw");
1312 if (td->o.lat_log_file) {
1313 finish_log_named(td, td->lat_log,
1314 td->o.lat_log_file, "lat");
1316 finish_log(td, td->lat_log, "lat");
1319 if (td->o.lat_log_file) {
1320 finish_log_named(td, td->slat_log,
1321 td->o.lat_log_file, "slat");
1323 finish_log(td, td->slat_log, "slat");
1326 if (td->o.lat_log_file) {
1327 finish_log_named(td, td->clat_log,
1328 td->o.lat_log_file, "clat");
1330 finish_log(td, td->clat_log, "clat");
1333 if (td->o.iops_log_file) {
1334 finish_log_named(td, td->iops_log,
1335 td->o.iops_log_file, "iops");
1337 finish_log(td, td->iops_log, "iops");
1340 fio_mutex_up(writeout_mutex);
1341 if (td->o.exec_postrun)
1342 exec_string(td->o.exec_postrun);
1344 if (exitall_on_terminate)
1345 fio_terminate_threads(td->groupid);
1349 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1352 if (td->o.verify_async)
1353 verify_async_exit(td);
1355 close_and_free_files(td);
1358 cgroup_shutdown(td, &cgroup_mnt);
1360 if (td->o.cpumask_set) {
1361 int ret = fio_cpuset_exit(&td->o.cpumask);
1363 td_verror(td, ret, "fio_cpuset_exit");
1367 * do this very late, it will log file closing as well
1369 if (td->o.write_iolog_file)
1370 write_iolog_close(td);
1372 td_set_runstate(td, TD_EXITED);
1373 return (void *) (unsigned long) td->error;
1377 * We cannot pass the td data into a forked process, so attach the td and
1378 * pass it to the thread worker.
1380 static int fork_main(int shmid, int offset)
1382 struct thread_data *td;
1386 data = shmat(shmid, NULL, 0);
1387 if (data == (void *) -1) {
1395 * HP-UX inherits shm mappings?
1400 td = data + offset * sizeof(struct thread_data);
1401 ret = thread_main(td);
1403 return (int) (unsigned long) ret;
1407 * Run over the job map and reap the threads that have exited, if any.
1409 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1410 unsigned int *m_rate)
1412 struct thread_data *td;
1413 unsigned int cputhreads, realthreads, pending;
1417 * reap exited threads (TD_EXITED -> TD_REAPED)
1419 realthreads = pending = cputhreads = 0;
1420 for_each_td(td, i) {
1424 * ->io_ops is NULL for a thread that has closed its
1427 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1436 if (td->runstate == TD_REAPED)
1438 if (td->o.use_thread) {
1439 if (td->runstate == TD_EXITED) {
1440 td_set_runstate(td, TD_REAPED);
1447 if (td->runstate == TD_EXITED)
1451 * check if someone quit or got killed in an unusual way
1453 ret = waitpid(td->pid, &status, flags);
1455 if (errno == ECHILD) {
1456 log_err("fio: pid=%d disappeared %d\n",
1457 (int) td->pid, td->runstate);
1458 td_set_runstate(td, TD_REAPED);
1462 } else if (ret == td->pid) {
1463 if (WIFSIGNALED(status)) {
1464 int sig = WTERMSIG(status);
1467 log_err("fio: pid=%d, got signal=%d\n",
1468 (int) td->pid, sig);
1469 td_set_runstate(td, TD_REAPED);
1472 if (WIFEXITED(status)) {
1473 if (WEXITSTATUS(status) && !td->error)
1474 td->error = WEXITSTATUS(status);
1476 td_set_runstate(td, TD_REAPED);
1482 * thread is not dead, continue
1488 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1489 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1496 done_secs += mtime_since_now(&td->epoch) / 1000;
1499 if (*nr_running == cputhreads && !pending && realthreads)
1500 fio_terminate_threads(TERMINATE_ALL);
1503 static void *gtod_thread_main(void *data)
1505 fio_mutex_up(startup_mutex);
1508 * As long as we have jobs around, update the clock. It would be nice
1509 * to have some way of NOT hammering that CPU with gettimeofday(),
1510 * but I'm not sure what to use outside of a simple CPU nop to relax
1511 * it - we don't want to lose precision.
1521 static int fio_start_gtod_thread(void)
1523 pthread_attr_t attr;
1526 pthread_attr_init(&attr);
1527 pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1528 ret = pthread_create(>od_thread, &attr, gtod_thread_main, NULL);
1529 pthread_attr_destroy(&attr);
1531 log_err("Can't create gtod thread: %s\n", strerror(ret));
1535 ret = pthread_detach(gtod_thread);
1537 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1541 dprint(FD_MUTEX, "wait on startup_mutex\n");
1542 fio_mutex_down(startup_mutex);
1543 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1548 * Main function for kicking off and reaping jobs, as needed.
1550 static void run_threads(void)
1552 struct thread_data *td;
1553 unsigned long spent;
1554 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1556 if (fio_pin_memory())
1559 if (fio_gtod_offload && fio_start_gtod_thread())
1564 if (!terse_output) {
1565 log_info("Starting ");
1567 log_info("%d thread%s", nr_thread,
1568 nr_thread > 1 ? "s" : "");
1572 log_info("%d process%s", nr_process,
1573 nr_process > 1 ? "es" : "");
1579 todo = thread_number;
1582 m_rate = t_rate = 0;
1584 for_each_td(td, i) {
1585 print_status_init(td->thread_number - 1);
1587 if (!td->o.create_serialize)
1591 * do file setup here so it happens sequentially,
1592 * we don't want X number of threads getting their
1593 * client data interspersed on disk
1595 if (setup_files(td)) {
1598 log_err("fio: pid=%d, err=%d/%s\n",
1599 (int) td->pid, td->error, td->verror);
1600 td_set_runstate(td, TD_REAPED);
1607 * for sharing to work, each job must always open
1608 * its own files. so close them, if we opened them
1611 for_each_file(td, f, j) {
1612 if (fio_file_open(f))
1613 td_io_close_file(td, f);
1621 struct thread_data *map[REAL_MAX_JOBS];
1622 struct timeval this_start;
1623 int this_jobs = 0, left;
1626 * create threads (TD_NOT_CREATED -> TD_CREATED)
1628 for_each_td(td, i) {
1629 if (td->runstate != TD_NOT_CREATED)
1633 * never got a chance to start, killed by other
1634 * thread for some reason
1636 if (td->terminate) {
1641 if (td->o.start_delay) {
1642 spent = mtime_since_genesis();
1644 if (td->o.start_delay * 1000 > spent)
1648 if (td->o.stonewall && (nr_started || nr_running)) {
1649 dprint(FD_PROCESS, "%s: stonewall wait\n",
1657 * Set state to created. Thread will transition
1658 * to TD_INITIALIZED when it's done setting up.
1660 td_set_runstate(td, TD_CREATED);
1661 map[this_jobs++] = td;
1664 if (td->o.use_thread) {
1667 dprint(FD_PROCESS, "will pthread_create\n");
1668 ret = pthread_create(&td->thread, NULL,
1671 log_err("pthread_create: %s\n",
1676 ret = pthread_detach(td->thread);
1678 log_err("pthread_detach: %s",
1682 dprint(FD_PROCESS, "will fork\n");
1685 int ret = fork_main(shm_id, i);
1688 } else if (i == fio_debug_jobno)
1689 *fio_debug_jobp = pid;
1691 dprint(FD_MUTEX, "wait on startup_mutex\n");
1692 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1693 log_err("fio: job startup hung? exiting.\n");
1694 fio_terminate_threads(TERMINATE_ALL);
1699 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1703 * Wait for the started threads to transition to
1706 fio_gettime(&this_start, NULL);
1708 while (left && !fio_abort) {
1709 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1714 for (i = 0; i < this_jobs; i++) {
1718 if (td->runstate == TD_INITIALIZED) {
1721 } else if (td->runstate >= TD_EXITED) {
1725 nr_running++; /* work-around... */
1731 log_err("fio: %d jobs failed to start\n", left);
1732 for (i = 0; i < this_jobs; i++) {
1736 kill(td->pid, SIGTERM);
1742 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1744 for_each_td(td, i) {
1745 if (td->runstate != TD_INITIALIZED)
1748 if (in_ramp_time(td))
1749 td_set_runstate(td, TD_RAMP);
1751 td_set_runstate(td, TD_RUNNING);
1754 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1755 t_rate += td->o.rate[0] + td->o.rate[1];
1757 fio_mutex_up(td->mutex);
1760 reap_threads(&nr_running, &t_rate, &m_rate);
1764 fio_server_idle_loop();
1770 while (nr_running) {
1771 reap_threads(&nr_running, &t_rate, &m_rate);
1774 fio_server_idle_loop();
1785 struct thread_data *td;
1789 return fio_handle_clients();
1791 if (load_profile(exec_profile))
1794 exec_profile = NULL;
1800 setup_log(&agg_io_log[DDIR_READ], 0);
1801 setup_log(&agg_io_log[DDIR_WRITE], 0);
1804 startup_mutex = fio_mutex_init(0);
1805 if (startup_mutex == NULL)
1807 writeout_mutex = fio_mutex_init(1);
1808 if (writeout_mutex == NULL)
1812 create_disk_util_thread();
1814 cgroup_list = smalloc(sizeof(*cgroup_list));
1815 INIT_FLIST_HEAD(cgroup_list);
1822 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1823 __finish_log(agg_io_log[DDIR_WRITE],
1824 "agg-write_bw.log");
1829 fio_options_free(td);
1831 cgroup_kill(cgroup_list);
1835 fio_mutex_remove(startup_mutex);
1836 fio_mutex_remove(writeout_mutex);
1840 void reset_fio_state(void)
1849 static int endian_check(void)
1860 else if (u.c[0] == 0x12)
1863 #if defined(FIO_LITTLE_ENDIAN)
1866 #elif defined(FIO_BIG_ENDIAN)
1879 int main(int argc, char *argv[], char *envp[])
1883 if (endian_check()) {
1884 log_err("fio: endianness settings appear wrong.\n");
1885 log_err("fio: please report this to fio@vger.kernel.org\n");
1894 * We need locale for number printing, if it isn't set then just
1895 * go with the US format.
1897 if (!getenv("LC_NUMERIC"))
1898 setlocale(LC_NUMERIC, "en_US");
1900 ps = sysconf(_SC_PAGESIZE);
1902 log_err("Failed to get page size\n");
1909 fio_keywords_init();
1911 if (parse_options(argc, argv))