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 int thread_number = 0;
63 unsigned long done_secs = 0;
65 static struct fio_mutex *startup_mutex;
66 static struct fio_mutex *writeout_mutex;
67 static volatile int fio_abort;
68 static int exit_value;
69 static pthread_t gtod_thread;
70 static pthread_t disk_util_thread;
71 static struct flist_head *cgroup_list;
72 static char *cgroup_mnt;
74 unsigned long arch_flags = 0;
76 struct io_log *agg_io_log[2];
78 #define JOB_START_TIMEOUT (5 * 1000)
80 static const char *fio_os_strings[os_nr] = {
92 static const char *fio_arch_strings[arch_nr] = {
108 const char *fio_get_os_string(int nr)
111 return fio_os_strings[nr];
116 const char *fio_get_arch_string(int nr)
119 return fio_arch_strings[nr];
124 void td_set_runstate(struct thread_data *td, int runstate)
126 if (td->runstate == runstate)
129 dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
130 td->runstate, runstate);
131 td->runstate = runstate;
134 void fio_terminate_threads(int group_id)
136 struct thread_data *td;
139 dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
142 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
143 dprint(FD_PROCESS, "setting terminate on %s/%d\n",
144 td->o.name, (int) td->pid);
146 td->o.start_delay = 0;
149 * if the thread is running, just let it exit
153 else if (td->runstate < TD_RAMP)
154 kill(td->pid, SIGTERM);
156 struct ioengine_ops *ops = td->io_ops;
158 if (ops && (ops->flags & FIO_SIGTERM))
159 kill(td->pid, SIGTERM);
165 static void sig_int(int sig)
169 fio_server_got_signal(sig);
171 log_info("\nfio: terminating on signal %d\n", sig);
176 fio_terminate_threads(TERMINATE_ALL);
180 static void *disk_thread_main(void *data)
182 fio_mutex_up(startup_mutex);
185 usleep(DISK_UTIL_MSEC * 1000);
191 print_thread_status();
197 static int create_disk_util_thread(void)
201 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
203 log_err("Can't create disk util thread: %s\n", strerror(ret));
207 ret = pthread_detach(disk_util_thread);
209 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
213 dprint(FD_MUTEX, "wait on startup_mutex\n");
214 fio_mutex_down(startup_mutex);
215 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
219 static void set_sig_handlers(void)
221 struct sigaction act;
223 memset(&act, 0, sizeof(act));
224 act.sa_handler = sig_int;
225 act.sa_flags = SA_RESTART;
226 sigaction(SIGINT, &act, NULL);
228 memset(&act, 0, sizeof(act));
229 act.sa_handler = sig_int;
230 act.sa_flags = SA_RESTART;
231 sigaction(SIGTERM, &act, NULL);
234 memset(&act, 0, sizeof(act));
235 act.sa_handler = sig_int;
236 act.sa_flags = SA_RESTART;
237 sigaction(SIGPIPE, &act, NULL);
242 * Check if we are above the minimum rate given.
244 static int __check_min_rate(struct thread_data *td, struct timeval *now,
247 unsigned long long bytes = 0;
248 unsigned long iops = 0;
251 unsigned int ratemin = 0;
252 unsigned int rate_iops = 0;
253 unsigned int rate_iops_min = 0;
255 assert(ddir_rw(ddir));
257 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
261 * allow a 2 second settle period in the beginning
263 if (mtime_since(&td->start, now) < 2000)
266 iops += td->this_io_blocks[ddir];
267 bytes += td->this_io_bytes[ddir];
268 ratemin += td->o.ratemin[ddir];
269 rate_iops += td->o.rate_iops[ddir];
270 rate_iops_min += td->o.rate_iops_min[ddir];
273 * if rate blocks is set, sample is running
275 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
276 spent = mtime_since(&td->lastrate[ddir], now);
277 if (spent < td->o.ratecycle)
280 if (td->o.rate[ddir]) {
282 * check bandwidth specified rate
284 if (bytes < td->rate_bytes[ddir]) {
285 log_err("%s: min rate %u not met\n", td->o.name,
289 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
290 if (rate < ratemin ||
291 bytes < td->rate_bytes[ddir]) {
292 log_err("%s: min rate %u not met, got"
293 " %luKB/sec\n", td->o.name,
300 * checks iops specified rate
302 if (iops < rate_iops) {
303 log_err("%s: min iops rate %u not met\n",
304 td->o.name, rate_iops);
307 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
308 if (rate < rate_iops_min ||
309 iops < td->rate_blocks[ddir]) {
310 log_err("%s: min iops rate %u not met,"
311 " got %lu\n", td->o.name,
312 rate_iops_min, rate);
318 td->rate_bytes[ddir] = bytes;
319 td->rate_blocks[ddir] = iops;
320 memcpy(&td->lastrate[ddir], now, sizeof(*now));
324 static int check_min_rate(struct thread_data *td, struct timeval *now,
325 unsigned long *bytes_done)
330 ret |= __check_min_rate(td, now, 0);
332 ret |= __check_min_rate(td, now, 1);
337 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
341 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
348 * When job exits, we can cancel the in-flight IO if we are using async
349 * io. Attempt to do so.
351 static void cleanup_pending_aio(struct thread_data *td)
353 struct flist_head *entry, *n;
358 * get immediately available events, if any
360 r = io_u_queued_complete(td, 0, NULL);
365 * now cancel remaining active events
367 if (td->io_ops->cancel) {
368 flist_for_each_safe(entry, n, &td->io_u_busylist) {
369 io_u = flist_entry(entry, struct io_u, list);
372 * if the io_u isn't in flight, then that generally
373 * means someone leaked an io_u. complain but fix
374 * it up, so we don't stall here.
376 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
377 log_err("fio: non-busy IO on busy list\n");
380 r = td->io_ops->cancel(td, io_u);
388 r = io_u_queued_complete(td, td->cur_depth, NULL);
392 * Helper to handle the final sync of a file. Works just like the normal
393 * io path, just does everything sync.
395 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
397 struct io_u *io_u = __get_io_u(td);
403 io_u->ddir = DDIR_SYNC;
406 if (td_io_prep(td, io_u)) {
412 ret = td_io_queue(td, io_u);
414 td_verror(td, io_u->error, "td_io_queue");
417 } else if (ret == FIO_Q_QUEUED) {
418 if (io_u_queued_complete(td, 1, NULL) < 0)
420 } else if (ret == FIO_Q_COMPLETED) {
422 td_verror(td, io_u->error, "td_io_queue");
426 if (io_u_sync_complete(td, io_u, NULL) < 0)
428 } else if (ret == FIO_Q_BUSY) {
429 if (td_io_commit(td))
437 static inline void __update_tv_cache(struct thread_data *td)
439 fio_gettime(&td->tv_cache, NULL);
442 static inline void update_tv_cache(struct thread_data *td)
444 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
445 __update_tv_cache(td);
448 static int break_on_this_error(struct thread_data *td, int *retptr)
452 if (ret < 0 || td->error) {
455 if (!td->o.continue_on_error)
463 if (td_non_fatal_error(err)) {
465 * Continue with the I/Os in case of
468 update_error_count(td, err);
472 } else if (td->o.fill_device && err == ENOSPC) {
474 * We expect to hit this error if
475 * fill_device option is set.
482 * Stop the I/O in case of a fatal
485 update_error_count(td, err);
494 * The main verify engine. Runs over the writes we previously submitted,
495 * reads the blocks back in, and checks the crc/md5 of the data.
497 static void do_verify(struct thread_data *td)
504 dprint(FD_VERIFY, "starting loop\n");
507 * sync io first and invalidate cache, to make sure we really
510 for_each_file(td, f, i) {
511 if (!fio_file_open(f))
513 if (fio_io_sync(td, f))
515 if (file_invalidate_cache(td, f))
522 td_set_runstate(td, TD_VERIFYING);
525 while (!td->terminate) {
530 if (runtime_exceeded(td, &td->tv_cache)) {
531 __update_tv_cache(td);
532 if (runtime_exceeded(td, &td->tv_cache)) {
538 io_u = __get_io_u(td);
542 if (get_next_verify(td, io_u)) {
547 if (td_io_prep(td, io_u)) {
552 if (td->o.verify_async)
553 io_u->end_io = verify_io_u_async;
555 io_u->end_io = verify_io_u;
557 ret = td_io_queue(td, io_u);
559 case FIO_Q_COMPLETED:
562 clear_io_u(td, io_u);
563 } else if (io_u->resid) {
564 int bytes = io_u->xfer_buflen - io_u->resid;
570 td_verror(td, EIO, "full resid");
575 io_u->xfer_buflen = io_u->resid;
576 io_u->xfer_buf += bytes;
577 io_u->offset += bytes;
579 if (ddir_rw(io_u->ddir))
580 td->ts.short_io_u[io_u->ddir]++;
583 if (io_u->offset == f->real_file_size)
586 requeue_io_u(td, &io_u);
589 ret = io_u_sync_complete(td, io_u, NULL);
597 requeue_io_u(td, &io_u);
598 ret2 = td_io_commit(td);
604 td_verror(td, -ret, "td_io_queue");
608 if (break_on_this_error(td, &ret))
612 * if we can queue more, do so. but check if there are
613 * completed io_u's first. Note that we can get BUSY even
614 * without IO queued, if the system is resource starved.
616 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
617 if (full || !td->o.iodepth_batch_complete) {
618 min_events = min(td->o.iodepth_batch_complete,
620 if (full && !min_events && td->o.iodepth_batch_complete != 0)
625 * Reap required number of io units, if any,
626 * and do the verification on them through
627 * the callback handler
629 if (io_u_queued_complete(td, min_events, NULL) < 0) {
633 } while (full && (td->cur_depth > td->o.iodepth_low));
640 min_events = td->cur_depth;
643 ret = io_u_queued_complete(td, min_events, NULL);
645 cleanup_pending_aio(td);
647 td_set_runstate(td, TD_RUNNING);
649 dprint(FD_VERIFY, "exiting loop\n");
653 * Main IO worker function. It retrieves io_u's to process and queues
654 * and reaps them, checking for rate and errors along the way.
656 static void do_io(struct thread_data *td)
661 if (in_ramp_time(td))
662 td_set_runstate(td, TD_RAMP);
664 td_set_runstate(td, TD_RUNNING);
666 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
667 (!flist_empty(&td->trim_list)) ||
668 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
669 struct timeval comp_time;
670 unsigned long bytes_done[2] = { 0, 0 };
680 if (runtime_exceeded(td, &td->tv_cache)) {
681 __update_tv_cache(td);
682 if (runtime_exceeded(td, &td->tv_cache)) {
693 * Add verification end_io handler, if asked to verify
694 * a previously written file.
696 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
698 if (td->o.verify_async)
699 io_u->end_io = verify_io_u_async;
701 io_u->end_io = verify_io_u;
702 td_set_runstate(td, TD_VERIFYING);
703 } else if (in_ramp_time(td))
704 td_set_runstate(td, TD_RAMP);
706 td_set_runstate(td, TD_RUNNING);
708 ret = td_io_queue(td, io_u);
710 case FIO_Q_COMPLETED:
713 clear_io_u(td, io_u);
714 } else if (io_u->resid) {
715 int bytes = io_u->xfer_buflen - io_u->resid;
716 struct fio_file *f = io_u->file;
722 td_verror(td, EIO, "full resid");
727 io_u->xfer_buflen = io_u->resid;
728 io_u->xfer_buf += bytes;
729 io_u->offset += bytes;
731 if (ddir_rw(io_u->ddir))
732 td->ts.short_io_u[io_u->ddir]++;
734 if (io_u->offset == f->real_file_size)
737 requeue_io_u(td, &io_u);
740 if (__should_check_rate(td, 0) ||
741 __should_check_rate(td, 1))
742 fio_gettime(&comp_time, NULL);
744 ret = io_u_sync_complete(td, io_u, bytes_done);
751 * if the engine doesn't have a commit hook,
752 * the io_u is really queued. if it does have such
753 * a hook, it has to call io_u_queued() itself.
755 if (td->io_ops->commit == NULL)
756 io_u_queued(td, io_u);
759 requeue_io_u(td, &io_u);
760 ret2 = td_io_commit(td);
770 if (break_on_this_error(td, &ret))
774 * See if we need to complete some commands. Note that we
775 * can get BUSY even without IO queued, if the system is
778 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
779 if (full || !td->o.iodepth_batch_complete) {
780 min_evts = min(td->o.iodepth_batch_complete,
782 if (full && !min_evts && td->o.iodepth_batch_complete != 0)
785 if (__should_check_rate(td, 0) ||
786 __should_check_rate(td, 1))
787 fio_gettime(&comp_time, NULL);
790 ret = io_u_queued_complete(td, min_evts, bytes_done);
794 } while (full && (td->cur_depth > td->o.iodepth_low));
799 if (!(bytes_done[0] + bytes_done[1]))
802 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
803 if (check_min_rate(td, &comp_time, bytes_done)) {
804 if (exitall_on_terminate)
805 fio_terminate_threads(td->groupid);
806 td_verror(td, EIO, "check_min_rate");
811 if (td->o.thinktime) {
812 unsigned long long b;
814 b = td->io_blocks[0] + td->io_blocks[1];
815 if (!(b % td->o.thinktime_blocks)) {
818 if (td->o.thinktime_spin)
819 usec_spin(td->o.thinktime_spin);
821 left = td->o.thinktime - td->o.thinktime_spin;
823 usec_sleep(td, left);
828 if (td->trim_entries)
829 printf("trim entries %ld\n", td->trim_entries);
831 if (td->o.fill_device && td->error == ENOSPC) {
840 ret = io_u_queued_complete(td, i, NULL);
841 if (td->o.fill_device && td->error == ENOSPC)
845 if (should_fsync(td) && td->o.end_fsync) {
846 td_set_runstate(td, TD_FSYNCING);
848 for_each_file(td, f, i) {
849 if (!fio_file_open(f))
855 cleanup_pending_aio(td);
858 * stop job if we failed doing any IO
860 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
864 static void cleanup_io_u(struct thread_data *td)
866 struct flist_head *entry, *n;
869 flist_for_each_safe(entry, n, &td->io_u_freelist) {
870 io_u = flist_entry(entry, struct io_u, list);
872 flist_del(&io_u->list);
873 fio_memfree(io_u, sizeof(*io_u));
879 static int init_io_u(struct thread_data *td)
883 int cl_align, i, max_units;
886 max_units = td->o.iodepth;
887 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
888 td->orig_buffer_size = (unsigned long long) max_bs
889 * (unsigned long long) max_units;
891 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
894 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
895 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
898 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
899 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
903 if (allocate_io_mem(td))
906 if (td->o.odirect || td->o.mem_align ||
907 (td->io_ops->flags & FIO_RAWIO))
908 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
912 cl_align = os_cache_line_size();
914 for (i = 0; i < max_units; i++) {
920 ptr = fio_memalign(cl_align, sizeof(*io_u));
922 log_err("fio: unable to allocate aligned memory\n");
927 memset(io_u, 0, sizeof(*io_u));
928 INIT_FLIST_HEAD(&io_u->list);
929 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
931 if (!(td->io_ops->flags & FIO_NOIO)) {
932 io_u->buf = p + max_bs * i;
933 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
936 io_u_fill_buffer(td, io_u, max_bs);
937 if (td_write(td) && td->o.verify_pattern_bytes) {
939 * Fill the buffer with the pattern if we are
940 * going to be doing writes.
942 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
947 io_u->flags = IO_U_F_FREE;
948 flist_add(&io_u->list, &td->io_u_freelist);
954 static int switch_ioscheduler(struct thread_data *td)
956 char tmp[256], tmp2[128];
960 if (td->io_ops->flags & FIO_DISKLESSIO)
963 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
965 f = fopen(tmp, "r+");
967 if (errno == ENOENT) {
968 log_err("fio: os or kernel doesn't support IO scheduler"
972 td_verror(td, errno, "fopen iosched");
979 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
980 if (ferror(f) || ret != 1) {
981 td_verror(td, errno, "fwrite");
989 * Read back and check that the selected scheduler is now the default.
991 ret = fread(tmp, 1, sizeof(tmp), f);
992 if (ferror(f) || ret < 0) {
993 td_verror(td, errno, "fread");
998 sprintf(tmp2, "[%s]", td->o.ioscheduler);
999 if (!strstr(tmp, tmp2)) {
1000 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1001 td_verror(td, EINVAL, "iosched_switch");
1010 static int keep_running(struct thread_data *td)
1012 unsigned long long io_done;
1016 if (td->o.time_based)
1023 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
1024 + td->io_skip_bytes;
1025 if (io_done < td->o.size)
1031 static void reset_io_counters(struct thread_data *td)
1033 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1034 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1035 td->stat_io_blocks[0] = td->stat_io_blocks[1] = 0;
1036 td->this_io_blocks[0] = td->this_io_blocks[1] = 0;
1038 td->rate_bytes[0] = td->rate_bytes[1] = 0;
1039 td->rate_blocks[0] = td->rate_blocks[1] = 0;
1041 td->last_was_sync = 0;
1044 * reset file done count if we are to start over
1046 if (td->o.time_based || td->o.loops)
1047 td->nr_done_files = 0;
1050 void reset_all_stats(struct thread_data *td)
1055 reset_io_counters(td);
1057 for (i = 0; i < 2; i++) {
1058 td->io_bytes[i] = 0;
1059 td->io_blocks[i] = 0;
1060 td->io_issues[i] = 0;
1061 td->ts.total_io_u[i] = 0;
1064 fio_gettime(&tv, NULL);
1065 td->ts.runtime[0] = 0;
1066 td->ts.runtime[1] = 0;
1067 memcpy(&td->epoch, &tv, sizeof(tv));
1068 memcpy(&td->start, &tv, sizeof(tv));
1071 static void clear_io_state(struct thread_data *td)
1076 reset_io_counters(td);
1079 for_each_file(td, f, i)
1080 fio_file_clear_done(f);
1083 * Set the same seed to get repeatable runs
1085 td_fill_rand_seeds(td);
1088 static int exec_string(const char *string)
1090 int ret, newlen = strlen(string) + 1 + 8;
1093 str = malloc(newlen);
1094 sprintf(str, "sh -c %s", string);
1098 log_err("fio: exec of cmd <%s> failed\n", str);
1105 * Entry point for the thread based jobs. The process based jobs end up
1106 * here as well, after a little setup.
1108 static void *thread_main(void *data)
1110 unsigned long long elapsed;
1111 struct thread_data *td = data;
1112 pthread_condattr_t attr;
1115 if (!td->o.use_thread) {
1121 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1123 INIT_FLIST_HEAD(&td->io_u_freelist);
1124 INIT_FLIST_HEAD(&td->io_u_busylist);
1125 INIT_FLIST_HEAD(&td->io_u_requeues);
1126 INIT_FLIST_HEAD(&td->io_log_list);
1127 INIT_FLIST_HEAD(&td->io_hist_list);
1128 INIT_FLIST_HEAD(&td->verify_list);
1129 INIT_FLIST_HEAD(&td->trim_list);
1130 pthread_mutex_init(&td->io_u_lock, NULL);
1131 td->io_hist_tree = RB_ROOT;
1133 pthread_condattr_init(&attr);
1134 pthread_cond_init(&td->verify_cond, &attr);
1135 pthread_cond_init(&td->free_cond, &attr);
1137 td_set_runstate(td, TD_INITIALIZED);
1138 dprint(FD_MUTEX, "up startup_mutex\n");
1139 fio_mutex_up(startup_mutex);
1140 dprint(FD_MUTEX, "wait on td->mutex\n");
1141 fio_mutex_down(td->mutex);
1142 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1145 * the ->mutex mutex is now no longer used, close it to avoid
1146 * eating a file descriptor
1148 fio_mutex_remove(td->mutex);
1151 * A new gid requires privilege, so we need to do this before setting
1154 if (td->o.gid != -1U && setgid(td->o.gid)) {
1155 td_verror(td, errno, "setgid");
1158 if (td->o.uid != -1U && setuid(td->o.uid)) {
1159 td_verror(td, errno, "setuid");
1164 * If we have a gettimeofday() thread, make sure we exclude that
1165 * thread from this job
1168 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1171 * Set affinity first, in case it has an impact on the memory
1174 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1175 td_verror(td, errno, "cpu_set_affinity");
1180 * May alter parameters that init_io_u() will use, so we need to
1189 if (td->o.verify_async && verify_async_init(td))
1192 if (td->ioprio_set) {
1193 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1194 td_verror(td, errno, "ioprio_set");
1199 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1202 if (nice(td->o.nice) == -1) {
1203 td_verror(td, errno, "nice");
1207 if (td->o.ioscheduler && switch_ioscheduler(td))
1210 if (!td->o.create_serialize && setup_files(td))
1216 if (init_random_map(td))
1219 if (td->o.exec_prerun) {
1220 if (exec_string(td->o.exec_prerun))
1224 if (td->o.pre_read) {
1225 if (pre_read_files(td) < 0)
1229 fio_gettime(&td->epoch, NULL);
1230 getrusage(RUSAGE_SELF, &td->ru_start);
1233 while (keep_running(td)) {
1234 fio_gettime(&td->start, NULL);
1235 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1236 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1237 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1239 if (td->o.ratemin[0] || td->o.ratemin[1])
1240 memcpy(&td->lastrate, &td->bw_sample_time,
1241 sizeof(td->lastrate));
1246 prune_io_piece_log(td);
1252 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1253 elapsed = utime_since_now(&td->start);
1254 td->ts.runtime[DDIR_READ] += elapsed;
1256 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1257 elapsed = utime_since_now(&td->start);
1258 td->ts.runtime[DDIR_WRITE] += elapsed;
1261 if (td->error || td->terminate)
1264 if (!td->o.do_verify ||
1265 td->o.verify == VERIFY_NONE ||
1266 (td->io_ops->flags & FIO_UNIDIR))
1271 fio_gettime(&td->start, NULL);
1275 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1277 if (td->error || td->terminate)
1281 update_rusage_stat(td);
1282 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1283 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1284 td->ts.total_run_time = mtime_since_now(&td->epoch);
1285 td->ts.io_bytes[0] = td->io_bytes[0];
1286 td->ts.io_bytes[1] = td->io_bytes[1];
1288 fio_mutex_down(writeout_mutex);
1290 if (td->o.bw_log_file) {
1291 finish_log_named(td, td->bw_log,
1292 td->o.bw_log_file, "bw");
1294 finish_log(td, td->bw_log, "bw");
1297 if (td->o.lat_log_file) {
1298 finish_log_named(td, td->lat_log,
1299 td->o.lat_log_file, "lat");
1301 finish_log(td, td->lat_log, "lat");
1304 if (td->o.lat_log_file) {
1305 finish_log_named(td, td->slat_log,
1306 td->o.lat_log_file, "slat");
1308 finish_log(td, td->slat_log, "slat");
1311 if (td->o.lat_log_file) {
1312 finish_log_named(td, td->clat_log,
1313 td->o.lat_log_file, "clat");
1315 finish_log(td, td->clat_log, "clat");
1318 if (td->o.iops_log_file) {
1319 finish_log_named(td, td->iops_log,
1320 td->o.iops_log_file, "iops");
1322 finish_log(td, td->iops_log, "iops");
1325 fio_mutex_up(writeout_mutex);
1326 if (td->o.exec_postrun)
1327 exec_string(td->o.exec_postrun);
1329 if (exitall_on_terminate)
1330 fio_terminate_threads(td->groupid);
1334 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1337 if (td->o.verify_async)
1338 verify_async_exit(td);
1340 close_and_free_files(td);
1343 cgroup_shutdown(td, &cgroup_mnt);
1345 if (td->o.cpumask_set) {
1346 int ret = fio_cpuset_exit(&td->o.cpumask);
1348 td_verror(td, ret, "fio_cpuset_exit");
1352 * do this very late, it will log file closing as well
1354 if (td->o.write_iolog_file)
1355 write_iolog_close(td);
1357 options_mem_free(td);
1358 td_set_runstate(td, TD_EXITED);
1359 return (void *) (unsigned long) td->error;
1363 * We cannot pass the td data into a forked process, so attach the td and
1364 * pass it to the thread worker.
1366 static int fork_main(int shmid, int offset)
1368 struct thread_data *td;
1372 data = shmat(shmid, NULL, 0);
1373 if (data == (void *) -1) {
1381 * HP-UX inherits shm mappings?
1386 td = data + offset * sizeof(struct thread_data);
1387 ret = thread_main(td);
1389 return (int) (unsigned long) ret;
1393 * Run over the job map and reap the threads that have exited, if any.
1395 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1397 struct thread_data *td;
1398 int i, cputhreads, realthreads, pending, status, ret;
1401 * reap exited threads (TD_EXITED -> TD_REAPED)
1403 realthreads = pending = cputhreads = 0;
1404 for_each_td(td, i) {
1408 * ->io_ops is NULL for a thread that has closed its
1411 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1420 if (td->runstate == TD_REAPED)
1422 if (td->o.use_thread) {
1423 if (td->runstate == TD_EXITED) {
1424 td_set_runstate(td, TD_REAPED);
1431 if (td->runstate == TD_EXITED)
1435 * check if someone quit or got killed in an unusual way
1437 ret = waitpid(td->pid, &status, flags);
1439 if (errno == ECHILD) {
1440 log_err("fio: pid=%d disappeared %d\n",
1441 (int) td->pid, td->runstate);
1442 td_set_runstate(td, TD_REAPED);
1446 } else if (ret == td->pid) {
1447 if (WIFSIGNALED(status)) {
1448 int sig = WTERMSIG(status);
1451 log_err("fio: pid=%d, got signal=%d\n",
1452 (int) td->pid, sig);
1453 td_set_runstate(td, TD_REAPED);
1456 if (WIFEXITED(status)) {
1457 if (WEXITSTATUS(status) && !td->error)
1458 td->error = WEXITSTATUS(status);
1460 td_set_runstate(td, TD_REAPED);
1466 * thread is not dead, continue
1472 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1473 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1480 done_secs += mtime_since_now(&td->epoch) / 1000;
1483 if (*nr_running == cputhreads && !pending && realthreads)
1484 fio_terminate_threads(TERMINATE_ALL);
1487 static void *gtod_thread_main(void *data)
1489 fio_mutex_up(startup_mutex);
1492 * As long as we have jobs around, update the clock. It would be nice
1493 * to have some way of NOT hammering that CPU with gettimeofday(),
1494 * but I'm not sure what to use outside of a simple CPU nop to relax
1495 * it - we don't want to lose precision.
1505 static int fio_start_gtod_thread(void)
1507 pthread_attr_t attr;
1510 pthread_attr_init(&attr);
1511 pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1512 ret = pthread_create(>od_thread, &attr, gtod_thread_main, NULL);
1513 pthread_attr_destroy(&attr);
1515 log_err("Can't create gtod thread: %s\n", strerror(ret));
1519 ret = pthread_detach(gtod_thread);
1521 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1525 dprint(FD_MUTEX, "wait on startup_mutex\n");
1526 fio_mutex_down(startup_mutex);
1527 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1532 * Main function for kicking off and reaping jobs, as needed.
1534 static void run_threads(void)
1536 struct thread_data *td;
1537 unsigned long spent;
1538 int i, todo, nr_running, m_rate, t_rate, nr_started;
1540 if (fio_pin_memory())
1543 if (fio_gtod_offload && fio_start_gtod_thread())
1548 if (!terse_output) {
1549 log_info("Starting ");
1551 log_info("%d thread%s", nr_thread,
1552 nr_thread > 1 ? "s" : "");
1556 log_info("%d process%s", nr_process,
1557 nr_process > 1 ? "es" : "");
1563 todo = thread_number;
1566 m_rate = t_rate = 0;
1568 for_each_td(td, i) {
1569 print_status_init(td->thread_number - 1);
1571 if (!td->o.create_serialize)
1575 * do file setup here so it happens sequentially,
1576 * we don't want X number of threads getting their
1577 * client data interspersed on disk
1579 if (setup_files(td)) {
1582 log_err("fio: pid=%d, err=%d/%s\n",
1583 (int) td->pid, td->error, td->verror);
1584 td_set_runstate(td, TD_REAPED);
1591 * for sharing to work, each job must always open
1592 * its own files. so close them, if we opened them
1595 for_each_file(td, f, j) {
1596 if (fio_file_open(f))
1597 td_io_close_file(td, f);
1605 struct thread_data *map[REAL_MAX_JOBS];
1606 struct timeval this_start;
1607 int this_jobs = 0, left;
1610 * create threads (TD_NOT_CREATED -> TD_CREATED)
1612 for_each_td(td, i) {
1613 if (td->runstate != TD_NOT_CREATED)
1617 * never got a chance to start, killed by other
1618 * thread for some reason
1620 if (td->terminate) {
1625 if (td->o.start_delay) {
1626 spent = mtime_since_genesis();
1628 if (td->o.start_delay * 1000 > spent)
1632 if (td->o.stonewall && (nr_started || nr_running)) {
1633 dprint(FD_PROCESS, "%s: stonewall wait\n",
1641 * Set state to created. Thread will transition
1642 * to TD_INITIALIZED when it's done setting up.
1644 td_set_runstate(td, TD_CREATED);
1645 map[this_jobs++] = td;
1648 if (td->o.use_thread) {
1651 dprint(FD_PROCESS, "will pthread_create\n");
1652 ret = pthread_create(&td->thread, NULL,
1655 log_err("pthread_create: %s\n",
1660 ret = pthread_detach(td->thread);
1662 log_err("pthread_detach: %s",
1666 dprint(FD_PROCESS, "will fork\n");
1669 int ret = fork_main(shm_id, i);
1672 } else if (i == fio_debug_jobno)
1673 *fio_debug_jobp = pid;
1675 dprint(FD_MUTEX, "wait on startup_mutex\n");
1676 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1677 log_err("fio: job startup hung? exiting.\n");
1678 fio_terminate_threads(TERMINATE_ALL);
1683 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1687 * Wait for the started threads to transition to
1690 fio_gettime(&this_start, NULL);
1692 while (left && !fio_abort) {
1693 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1698 for (i = 0; i < this_jobs; i++) {
1702 if (td->runstate == TD_INITIALIZED) {
1705 } else if (td->runstate >= TD_EXITED) {
1709 nr_running++; /* work-around... */
1715 log_err("fio: %d jobs failed to start\n", left);
1716 for (i = 0; i < this_jobs; i++) {
1720 kill(td->pid, SIGTERM);
1726 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1728 for_each_td(td, i) {
1729 if (td->runstate != TD_INITIALIZED)
1732 if (in_ramp_time(td))
1733 td_set_runstate(td, TD_RAMP);
1735 td_set_runstate(td, TD_RUNNING);
1738 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1739 t_rate += td->o.rate[0] + td->o.rate[1];
1741 fio_mutex_up(td->mutex);
1744 reap_threads(&nr_running, &t_rate, &m_rate);
1748 fio_server_idle_loop();
1754 while (nr_running) {
1755 reap_threads(&nr_running, &t_rate, &m_rate);
1758 fio_server_idle_loop();
1770 return fio_handle_clients();
1771 if (exec_profile && load_profile(exec_profile))
1778 setup_log(&agg_io_log[DDIR_READ]);
1779 setup_log(&agg_io_log[DDIR_WRITE]);
1782 startup_mutex = fio_mutex_init(0);
1783 if (startup_mutex == NULL)
1785 writeout_mutex = fio_mutex_init(1);
1786 if (writeout_mutex == NULL)
1790 create_disk_util_thread();
1792 cgroup_list = smalloc(sizeof(*cgroup_list));
1793 INIT_FLIST_HEAD(cgroup_list);
1800 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1801 __finish_log(agg_io_log[DDIR_WRITE],
1802 "agg-write_bw.log");
1806 cgroup_kill(cgroup_list);
1810 fio_mutex_remove(startup_mutex);
1811 fio_mutex_remove(writeout_mutex);
1815 void reset_fio_state(void)
1824 static int endian_check(void)
1835 else if (u.c[0] == 0x12)
1838 #if defined(FIO_LITTLE_ENDIAN)
1841 #elif defined(FIO_BIG_ENDIAN)
1854 int main(int argc, char *argv[], char *envp[])
1858 if (endian_check()) {
1859 log_err("fio: endianness settings appear wrong.\n");
1860 log_err("fio: please report this to fio@vger.kernel.org\n");
1869 * We need locale for number printing, if it isn't set then just
1870 * go with the US format.
1872 if (!getenv("LC_NUMERIC"))
1873 setlocale(LC_NUMERIC, "en_US");
1875 ps = sysconf(_SC_PAGESIZE);
1877 log_err("Failed to get page size\n");
1884 fio_keywords_init();
1886 if (parse_options(argc, argv))