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)
346 if (in_ramp_time(td))
350 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
357 * When job exits, we can cancel the in-flight IO if we are using async
358 * io. Attempt to do so.
360 static void cleanup_pending_aio(struct thread_data *td)
362 struct flist_head *entry, *n;
367 * get immediately available events, if any
369 r = io_u_queued_complete(td, 0, NULL);
374 * now cancel remaining active events
376 if (td->io_ops->cancel) {
377 flist_for_each_safe(entry, n, &td->io_u_busylist) {
378 io_u = flist_entry(entry, struct io_u, list);
381 * if the io_u isn't in flight, then that generally
382 * means someone leaked an io_u. complain but fix
383 * it up, so we don't stall here.
385 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
386 log_err("fio: non-busy IO on busy list\n");
389 r = td->io_ops->cancel(td, io_u);
397 r = io_u_queued_complete(td, td->cur_depth, NULL);
401 * Helper to handle the final sync of a file. Works just like the normal
402 * io path, just does everything sync.
404 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
406 struct io_u *io_u = __get_io_u(td);
412 io_u->ddir = DDIR_SYNC;
415 if (td_io_prep(td, io_u)) {
421 ret = td_io_queue(td, io_u);
423 td_verror(td, io_u->error, "td_io_queue");
426 } else if (ret == FIO_Q_QUEUED) {
427 if (io_u_queued_complete(td, 1, NULL) < 0)
429 } else if (ret == FIO_Q_COMPLETED) {
431 td_verror(td, io_u->error, "td_io_queue");
435 if (io_u_sync_complete(td, io_u, NULL) < 0)
437 } else if (ret == FIO_Q_BUSY) {
438 if (td_io_commit(td))
446 static inline void __update_tv_cache(struct thread_data *td)
448 fio_gettime(&td->tv_cache, NULL);
451 static inline void update_tv_cache(struct thread_data *td)
453 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
454 __update_tv_cache(td);
457 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
462 if (ret < 0 || td->error) {
470 if (!(td->o.continue_on_error & td_error_type(ddir, err)))
473 if (td_non_fatal_error(err)) {
475 * Continue with the I/Os in case of
478 update_error_count(td, err);
482 } else if (td->o.fill_device && err == ENOSPC) {
484 * We expect to hit this error if
485 * fill_device option is set.
492 * Stop the I/O in case of a fatal
495 update_error_count(td, err);
504 * The main verify engine. Runs over the writes we previously submitted,
505 * reads the blocks back in, and checks the crc/md5 of the data.
507 static void do_verify(struct thread_data *td)
514 dprint(FD_VERIFY, "starting loop\n");
517 * sync io first and invalidate cache, to make sure we really
520 for_each_file(td, f, i) {
521 if (!fio_file_open(f))
523 if (fio_io_sync(td, f))
525 if (file_invalidate_cache(td, f))
532 td_set_runstate(td, TD_VERIFYING);
535 while (!td->terminate) {
540 if (runtime_exceeded(td, &td->tv_cache)) {
541 __update_tv_cache(td);
542 if (runtime_exceeded(td, &td->tv_cache)) {
548 io_u = __get_io_u(td);
552 if (get_next_verify(td, io_u)) {
557 if (td_io_prep(td, io_u)) {
562 if (td->o.verify_async)
563 io_u->end_io = verify_io_u_async;
565 io_u->end_io = verify_io_u;
567 ret = td_io_queue(td, io_u);
569 case FIO_Q_COMPLETED:
572 clear_io_u(td, io_u);
573 } else if (io_u->resid) {
574 int bytes = io_u->xfer_buflen - io_u->resid;
580 td_verror(td, EIO, "full resid");
585 io_u->xfer_buflen = io_u->resid;
586 io_u->xfer_buf += bytes;
587 io_u->offset += bytes;
589 if (ddir_rw(io_u->ddir))
590 td->ts.short_io_u[io_u->ddir]++;
593 if (io_u->offset == f->real_file_size)
596 requeue_io_u(td, &io_u);
599 ret = io_u_sync_complete(td, io_u, NULL);
607 requeue_io_u(td, &io_u);
608 ret2 = td_io_commit(td);
614 td_verror(td, -ret, "td_io_queue");
618 if (break_on_this_error(td, io_u->ddir, &ret))
622 * if we can queue more, do so. but check if there are
623 * completed io_u's first. Note that we can get BUSY even
624 * without IO queued, if the system is resource starved.
626 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
627 if (full || !td->o.iodepth_batch_complete) {
628 min_events = min(td->o.iodepth_batch_complete,
630 if (full && !min_events && td->o.iodepth_batch_complete != 0)
635 * Reap required number of io units, if any,
636 * and do the verification on them through
637 * the callback handler
639 if (io_u_queued_complete(td, min_events, NULL) < 0) {
643 } while (full && (td->cur_depth > td->o.iodepth_low));
650 min_events = td->cur_depth;
653 ret = io_u_queued_complete(td, min_events, NULL);
655 cleanup_pending_aio(td);
657 td_set_runstate(td, TD_RUNNING);
659 dprint(FD_VERIFY, "exiting loop\n");
663 * Main IO worker function. It retrieves io_u's to process and queues
664 * and reaps them, checking for rate and errors along the way.
666 static void do_io(struct thread_data *td)
671 if (in_ramp_time(td))
672 td_set_runstate(td, TD_RAMP);
674 td_set_runstate(td, TD_RUNNING);
676 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
677 (!flist_empty(&td->trim_list)) ||
678 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
679 struct timeval comp_time;
680 unsigned long bytes_done[2] = { 0, 0 };
691 if (runtime_exceeded(td, &td->tv_cache)) {
692 __update_tv_cache(td);
693 if (runtime_exceeded(td, &td->tv_cache)) {
706 * Add verification end_io handler, if asked to verify
707 * a previously written file.
709 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
711 if (td->o.verify_async)
712 io_u->end_io = verify_io_u_async;
714 io_u->end_io = verify_io_u;
715 td_set_runstate(td, TD_VERIFYING);
716 } else if (in_ramp_time(td))
717 td_set_runstate(td, TD_RAMP);
719 td_set_runstate(td, TD_RUNNING);
721 ret = td_io_queue(td, io_u);
723 case FIO_Q_COMPLETED:
726 clear_io_u(td, io_u);
727 } else if (io_u->resid) {
728 int bytes = io_u->xfer_buflen - io_u->resid;
729 struct fio_file *f = io_u->file;
735 td_verror(td, EIO, "full resid");
740 io_u->xfer_buflen = io_u->resid;
741 io_u->xfer_buf += bytes;
742 io_u->offset += bytes;
744 if (ddir_rw(io_u->ddir))
745 td->ts.short_io_u[io_u->ddir]++;
747 if (io_u->offset == f->real_file_size)
750 requeue_io_u(td, &io_u);
753 if (__should_check_rate(td, 0) ||
754 __should_check_rate(td, 1))
755 fio_gettime(&comp_time, NULL);
757 ret = io_u_sync_complete(td, io_u, bytes_done);
764 * if the engine doesn't have a commit hook,
765 * the io_u is really queued. if it does have such
766 * a hook, it has to call io_u_queued() itself.
768 if (td->io_ops->commit == NULL)
769 io_u_queued(td, io_u);
772 requeue_io_u(td, &io_u);
773 ret2 = td_io_commit(td);
783 if (break_on_this_error(td, ddir, &ret))
787 * See if we need to complete some commands. Note that we
788 * can get BUSY even without IO queued, if the system is
791 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
792 if (full || !td->o.iodepth_batch_complete) {
793 min_evts = min(td->o.iodepth_batch_complete,
795 if (full && !min_evts && td->o.iodepth_batch_complete != 0)
798 if (__should_check_rate(td, 0) ||
799 __should_check_rate(td, 1))
800 fio_gettime(&comp_time, NULL);
803 ret = io_u_queued_complete(td, min_evts, bytes_done);
807 } while (full && (td->cur_depth > td->o.iodepth_low));
812 if (!(bytes_done[0] + bytes_done[1]))
815 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
816 if (check_min_rate(td, &comp_time, bytes_done)) {
817 if (exitall_on_terminate)
818 fio_terminate_threads(td->groupid);
819 td_verror(td, EIO, "check_min_rate");
824 if (td->o.thinktime) {
825 unsigned long long b;
827 b = td->io_blocks[0] + td->io_blocks[1];
828 if (!(b % td->o.thinktime_blocks)) {
831 if (td->o.thinktime_spin)
832 usec_spin(td->o.thinktime_spin);
834 left = td->o.thinktime - td->o.thinktime_spin;
836 usec_sleep(td, left);
841 if (td->trim_entries)
842 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
844 if (td->o.fill_device && td->error == ENOSPC) {
853 ret = io_u_queued_complete(td, i, NULL);
854 if (td->o.fill_device && td->error == ENOSPC)
858 if (should_fsync(td) && td->o.end_fsync) {
859 td_set_runstate(td, TD_FSYNCING);
861 for_each_file(td, f, i) {
862 if (!fio_file_open(f))
868 cleanup_pending_aio(td);
871 * stop job if we failed doing any IO
873 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
877 static void cleanup_io_u(struct thread_data *td)
879 struct flist_head *entry, *n;
882 flist_for_each_safe(entry, n, &td->io_u_freelist) {
883 io_u = flist_entry(entry, struct io_u, list);
885 flist_del(&io_u->list);
886 fio_memfree(io_u, sizeof(*io_u));
892 static int init_io_u(struct thread_data *td)
896 int cl_align, i, max_units;
899 max_units = td->o.iodepth;
900 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
901 td->orig_buffer_size = (unsigned long long) max_bs
902 * (unsigned long long) max_units;
904 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
907 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
908 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
911 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
912 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
916 if (allocate_io_mem(td))
919 if (td->o.odirect || td->o.mem_align ||
920 (td->io_ops->flags & FIO_RAWIO))
921 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
925 cl_align = os_cache_line_size();
927 for (i = 0; i < max_units; i++) {
933 ptr = fio_memalign(cl_align, sizeof(*io_u));
935 log_err("fio: unable to allocate aligned memory\n");
940 memset(io_u, 0, sizeof(*io_u));
941 INIT_FLIST_HEAD(&io_u->list);
942 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
944 if (!(td->io_ops->flags & FIO_NOIO)) {
946 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
949 io_u_fill_buffer(td, io_u, max_bs);
950 if (td_write(td) && td->o.verify_pattern_bytes) {
952 * Fill the buffer with the pattern if we are
953 * going to be doing writes.
955 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
960 io_u->flags = IO_U_F_FREE;
961 flist_add(&io_u->list, &td->io_u_freelist);
968 static int switch_ioscheduler(struct thread_data *td)
970 char tmp[256], tmp2[128];
974 if (td->io_ops->flags & FIO_DISKLESSIO)
977 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
979 f = fopen(tmp, "r+");
981 if (errno == ENOENT) {
982 log_err("fio: os or kernel doesn't support IO scheduler"
986 td_verror(td, errno, "fopen iosched");
993 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
994 if (ferror(f) || ret != 1) {
995 td_verror(td, errno, "fwrite");
1003 * Read back and check that the selected scheduler is now the default.
1005 ret = fread(tmp, 1, sizeof(tmp), f);
1006 if (ferror(f) || ret < 0) {
1007 td_verror(td, errno, "fread");
1012 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1013 if (!strstr(tmp, tmp2)) {
1014 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1015 td_verror(td, EINVAL, "iosched_switch");
1024 static int keep_running(struct thread_data *td)
1026 unsigned long long io_done;
1030 if (td->o.time_based)
1037 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
1038 + td->io_skip_bytes;
1039 if (io_done < td->o.size)
1045 static void reset_io_counters(struct thread_data *td)
1047 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1048 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
1049 td->stat_io_blocks[0] = td->stat_io_blocks[1] = 0;
1050 td->this_io_blocks[0] = td->this_io_blocks[1] = 0;
1052 td->rate_bytes[0] = td->rate_bytes[1] = 0;
1053 td->rate_blocks[0] = td->rate_blocks[1] = 0;
1055 td->last_was_sync = 0;
1058 * reset file done count if we are to start over
1060 if (td->o.time_based || td->o.loops)
1061 td->nr_done_files = 0;
1064 void reset_all_stats(struct thread_data *td)
1069 reset_io_counters(td);
1071 for (i = 0; i < 2; i++) {
1072 td->io_bytes[i] = 0;
1073 td->io_blocks[i] = 0;
1074 td->io_issues[i] = 0;
1075 td->ts.total_io_u[i] = 0;
1078 fio_gettime(&tv, NULL);
1079 td->ts.runtime[0] = 0;
1080 td->ts.runtime[1] = 0;
1081 memcpy(&td->epoch, &tv, sizeof(tv));
1082 memcpy(&td->start, &tv, sizeof(tv));
1085 static void clear_io_state(struct thread_data *td)
1090 reset_io_counters(td);
1093 for_each_file(td, f, i)
1094 fio_file_clear_done(f);
1097 * Set the same seed to get repeatable runs
1099 td_fill_rand_seeds(td);
1102 static int exec_string(const char *string)
1104 int ret, newlen = strlen(string) + 1 + 8;
1107 str = malloc(newlen);
1108 sprintf(str, "sh -c %s", string);
1112 log_err("fio: exec of cmd <%s> failed\n", str);
1119 * Entry point for the thread based jobs. The process based jobs end up
1120 * here as well, after a little setup.
1122 static void *thread_main(void *data)
1124 unsigned long long elapsed;
1125 struct thread_data *td = data;
1126 pthread_condattr_t attr;
1129 if (!td->o.use_thread) {
1135 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1137 INIT_FLIST_HEAD(&td->io_u_freelist);
1138 INIT_FLIST_HEAD(&td->io_u_busylist);
1139 INIT_FLIST_HEAD(&td->io_u_requeues);
1140 INIT_FLIST_HEAD(&td->io_log_list);
1141 INIT_FLIST_HEAD(&td->io_hist_list);
1142 INIT_FLIST_HEAD(&td->verify_list);
1143 INIT_FLIST_HEAD(&td->trim_list);
1144 pthread_mutex_init(&td->io_u_lock, NULL);
1145 td->io_hist_tree = RB_ROOT;
1147 pthread_condattr_init(&attr);
1148 pthread_cond_init(&td->verify_cond, &attr);
1149 pthread_cond_init(&td->free_cond, &attr);
1151 td_set_runstate(td, TD_INITIALIZED);
1152 dprint(FD_MUTEX, "up startup_mutex\n");
1153 fio_mutex_up(startup_mutex);
1154 dprint(FD_MUTEX, "wait on td->mutex\n");
1155 fio_mutex_down(td->mutex);
1156 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1159 * the ->mutex mutex is now no longer used, close it to avoid
1160 * eating a file descriptor
1162 fio_mutex_remove(td->mutex);
1165 * A new gid requires privilege, so we need to do this before setting
1168 if (td->o.gid != -1U && setgid(td->o.gid)) {
1169 td_verror(td, errno, "setgid");
1172 if (td->o.uid != -1U && setuid(td->o.uid)) {
1173 td_verror(td, errno, "setuid");
1178 * If we have a gettimeofday() thread, make sure we exclude that
1179 * thread from this job
1182 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
1185 * Set affinity first, in case it has an impact on the memory
1188 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1189 td_verror(td, errno, "cpu_set_affinity");
1194 * May alter parameters that init_io_u() will use, so we need to
1203 if (td->o.verify_async && verify_async_init(td))
1206 if (td->ioprio_set) {
1207 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1208 td_verror(td, errno, "ioprio_set");
1213 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1216 if (nice(td->o.nice) == -1) {
1217 td_verror(td, errno, "nice");
1221 if (td->o.ioscheduler && switch_ioscheduler(td))
1224 if (!td->o.create_serialize && setup_files(td))
1230 if (init_random_map(td))
1233 if (td->o.exec_prerun) {
1234 if (exec_string(td->o.exec_prerun))
1238 if (td->o.pre_read) {
1239 if (pre_read_files(td) < 0)
1243 fio_gettime(&td->epoch, NULL);
1244 getrusage(RUSAGE_SELF, &td->ru_start);
1247 while (keep_running(td)) {
1248 fio_gettime(&td->start, NULL);
1249 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1250 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1251 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1253 if (td->o.ratemin[0] || td->o.ratemin[1]) {
1254 memcpy(&td->lastrate[0], &td->bw_sample_time,
1255 sizeof(td->bw_sample_time));
1256 memcpy(&td->lastrate[1], &td->bw_sample_time,
1257 sizeof(td->bw_sample_time));
1263 prune_io_piece_log(td);
1269 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1270 elapsed = utime_since_now(&td->start);
1271 td->ts.runtime[DDIR_READ] += elapsed;
1273 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1274 elapsed = utime_since_now(&td->start);
1275 td->ts.runtime[DDIR_WRITE] += elapsed;
1278 if (td->error || td->terminate)
1281 if (!td->o.do_verify ||
1282 td->o.verify == VERIFY_NONE ||
1283 (td->io_ops->flags & FIO_UNIDIR))
1288 fio_gettime(&td->start, NULL);
1292 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1294 if (td->error || td->terminate)
1298 update_rusage_stat(td);
1299 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1300 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1301 td->ts.total_run_time = mtime_since_now(&td->epoch);
1302 td->ts.io_bytes[0] = td->io_bytes[0];
1303 td->ts.io_bytes[1] = td->io_bytes[1];
1305 fio_mutex_down(writeout_mutex);
1307 if (td->o.bw_log_file) {
1308 finish_log_named(td, td->bw_log,
1309 td->o.bw_log_file, "bw");
1311 finish_log(td, td->bw_log, "bw");
1314 if (td->o.lat_log_file) {
1315 finish_log_named(td, td->lat_log,
1316 td->o.lat_log_file, "lat");
1318 finish_log(td, td->lat_log, "lat");
1321 if (td->o.lat_log_file) {
1322 finish_log_named(td, td->slat_log,
1323 td->o.lat_log_file, "slat");
1325 finish_log(td, td->slat_log, "slat");
1328 if (td->o.lat_log_file) {
1329 finish_log_named(td, td->clat_log,
1330 td->o.lat_log_file, "clat");
1332 finish_log(td, td->clat_log, "clat");
1335 if (td->o.iops_log_file) {
1336 finish_log_named(td, td->iops_log,
1337 td->o.iops_log_file, "iops");
1339 finish_log(td, td->iops_log, "iops");
1342 fio_mutex_up(writeout_mutex);
1343 if (td->o.exec_postrun)
1344 exec_string(td->o.exec_postrun);
1346 if (exitall_on_terminate)
1347 fio_terminate_threads(td->groupid);
1351 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1354 if (td->o.verify_async)
1355 verify_async_exit(td);
1357 close_and_free_files(td);
1360 cgroup_shutdown(td, &cgroup_mnt);
1362 if (td->o.cpumask_set) {
1363 int ret = fio_cpuset_exit(&td->o.cpumask);
1365 td_verror(td, ret, "fio_cpuset_exit");
1369 * do this very late, it will log file closing as well
1371 if (td->o.write_iolog_file)
1372 write_iolog_close(td);
1374 td_set_runstate(td, TD_EXITED);
1375 return (void *) (unsigned long) td->error;
1379 * We cannot pass the td data into a forked process, so attach the td and
1380 * pass it to the thread worker.
1382 static int fork_main(int shmid, int offset)
1384 struct thread_data *td;
1388 data = shmat(shmid, NULL, 0);
1389 if (data == (void *) -1) {
1397 * HP-UX inherits shm mappings?
1402 td = data + offset * sizeof(struct thread_data);
1403 ret = thread_main(td);
1405 return (int) (unsigned long) ret;
1409 * Run over the job map and reap the threads that have exited, if any.
1411 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1412 unsigned int *m_rate)
1414 struct thread_data *td;
1415 unsigned int cputhreads, realthreads, pending;
1419 * reap exited threads (TD_EXITED -> TD_REAPED)
1421 realthreads = pending = cputhreads = 0;
1422 for_each_td(td, i) {
1426 * ->io_ops is NULL for a thread that has closed its
1429 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1438 if (td->runstate == TD_REAPED)
1440 if (td->o.use_thread) {
1441 if (td->runstate == TD_EXITED) {
1442 td_set_runstate(td, TD_REAPED);
1449 if (td->runstate == TD_EXITED)
1453 * check if someone quit or got killed in an unusual way
1455 ret = waitpid(td->pid, &status, flags);
1457 if (errno == ECHILD) {
1458 log_err("fio: pid=%d disappeared %d\n",
1459 (int) td->pid, td->runstate);
1460 td_set_runstate(td, TD_REAPED);
1464 } else if (ret == td->pid) {
1465 if (WIFSIGNALED(status)) {
1466 int sig = WTERMSIG(status);
1469 log_err("fio: pid=%d, got signal=%d\n",
1470 (int) td->pid, sig);
1471 td_set_runstate(td, TD_REAPED);
1474 if (WIFEXITED(status)) {
1475 if (WEXITSTATUS(status) && !td->error)
1476 td->error = WEXITSTATUS(status);
1478 td_set_runstate(td, TD_REAPED);
1484 * thread is not dead, continue
1490 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1491 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1498 done_secs += mtime_since_now(&td->epoch) / 1000;
1501 if (*nr_running == cputhreads && !pending && realthreads)
1502 fio_terminate_threads(TERMINATE_ALL);
1505 static void *gtod_thread_main(void *data)
1507 fio_mutex_up(startup_mutex);
1510 * As long as we have jobs around, update the clock. It would be nice
1511 * to have some way of NOT hammering that CPU with gettimeofday(),
1512 * but I'm not sure what to use outside of a simple CPU nop to relax
1513 * it - we don't want to lose precision.
1523 static int fio_start_gtod_thread(void)
1525 pthread_attr_t attr;
1528 pthread_attr_init(&attr);
1529 pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
1530 ret = pthread_create(>od_thread, &attr, gtod_thread_main, NULL);
1531 pthread_attr_destroy(&attr);
1533 log_err("Can't create gtod thread: %s\n", strerror(ret));
1537 ret = pthread_detach(gtod_thread);
1539 log_err("Can't detatch gtod thread: %s\n", strerror(ret));
1543 dprint(FD_MUTEX, "wait on startup_mutex\n");
1544 fio_mutex_down(startup_mutex);
1545 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1550 * Main function for kicking off and reaping jobs, as needed.
1552 static void run_threads(void)
1554 struct thread_data *td;
1555 unsigned long spent;
1556 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1558 if (fio_pin_memory())
1561 if (fio_gtod_offload && fio_start_gtod_thread())
1566 if (!terse_output) {
1567 log_info("Starting ");
1569 log_info("%d thread%s", nr_thread,
1570 nr_thread > 1 ? "s" : "");
1574 log_info("%d process%s", nr_process,
1575 nr_process > 1 ? "es" : "");
1581 todo = thread_number;
1584 m_rate = t_rate = 0;
1586 for_each_td(td, i) {
1587 print_status_init(td->thread_number - 1);
1589 if (!td->o.create_serialize)
1593 * do file setup here so it happens sequentially,
1594 * we don't want X number of threads getting their
1595 * client data interspersed on disk
1597 if (setup_files(td)) {
1600 log_err("fio: pid=%d, err=%d/%s\n",
1601 (int) td->pid, td->error, td->verror);
1602 td_set_runstate(td, TD_REAPED);
1609 * for sharing to work, each job must always open
1610 * its own files. so close them, if we opened them
1613 for_each_file(td, f, j) {
1614 if (fio_file_open(f))
1615 td_io_close_file(td, f);
1623 struct thread_data *map[REAL_MAX_JOBS];
1624 struct timeval this_start;
1625 int this_jobs = 0, left;
1628 * create threads (TD_NOT_CREATED -> TD_CREATED)
1630 for_each_td(td, i) {
1631 if (td->runstate != TD_NOT_CREATED)
1635 * never got a chance to start, killed by other
1636 * thread for some reason
1638 if (td->terminate) {
1643 if (td->o.start_delay) {
1644 spent = mtime_since_genesis();
1646 if (td->o.start_delay * 1000 > spent)
1650 if (td->o.stonewall && (nr_started || nr_running)) {
1651 dprint(FD_PROCESS, "%s: stonewall wait\n",
1659 * Set state to created. Thread will transition
1660 * to TD_INITIALIZED when it's done setting up.
1662 td_set_runstate(td, TD_CREATED);
1663 map[this_jobs++] = td;
1666 if (td->o.use_thread) {
1669 dprint(FD_PROCESS, "will pthread_create\n");
1670 ret = pthread_create(&td->thread, NULL,
1673 log_err("pthread_create: %s\n",
1678 ret = pthread_detach(td->thread);
1680 log_err("pthread_detach: %s",
1684 dprint(FD_PROCESS, "will fork\n");
1687 int ret = fork_main(shm_id, i);
1690 } else if (i == fio_debug_jobno)
1691 *fio_debug_jobp = pid;
1693 dprint(FD_MUTEX, "wait on startup_mutex\n");
1694 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1695 log_err("fio: job startup hung? exiting.\n");
1696 fio_terminate_threads(TERMINATE_ALL);
1701 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1705 * Wait for the started threads to transition to
1708 fio_gettime(&this_start, NULL);
1710 while (left && !fio_abort) {
1711 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1716 for (i = 0; i < this_jobs; i++) {
1720 if (td->runstate == TD_INITIALIZED) {
1723 } else if (td->runstate >= TD_EXITED) {
1727 nr_running++; /* work-around... */
1733 log_err("fio: %d jobs failed to start\n", left);
1734 for (i = 0; i < this_jobs; i++) {
1738 kill(td->pid, SIGTERM);
1744 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1746 for_each_td(td, i) {
1747 if (td->runstate != TD_INITIALIZED)
1750 if (in_ramp_time(td))
1751 td_set_runstate(td, TD_RAMP);
1753 td_set_runstate(td, TD_RUNNING);
1756 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1757 t_rate += td->o.rate[0] + td->o.rate[1];
1759 fio_mutex_up(td->mutex);
1762 reap_threads(&nr_running, &t_rate, &m_rate);
1766 fio_server_idle_loop();
1772 while (nr_running) {
1773 reap_threads(&nr_running, &t_rate, &m_rate);
1776 fio_server_idle_loop();
1787 struct thread_data *td;
1791 return fio_handle_clients();
1793 if (load_profile(exec_profile))
1796 exec_profile = NULL;
1802 setup_log(&agg_io_log[DDIR_READ], 0);
1803 setup_log(&agg_io_log[DDIR_WRITE], 0);
1806 startup_mutex = fio_mutex_init(0);
1807 if (startup_mutex == NULL)
1809 writeout_mutex = fio_mutex_init(1);
1810 if (writeout_mutex == NULL)
1814 create_disk_util_thread();
1816 cgroup_list = smalloc(sizeof(*cgroup_list));
1817 INIT_FLIST_HEAD(cgroup_list);
1824 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1825 __finish_log(agg_io_log[DDIR_WRITE],
1826 "agg-write_bw.log");
1831 fio_options_free(td);
1833 cgroup_kill(cgroup_list);
1837 fio_mutex_remove(startup_mutex);
1838 fio_mutex_remove(writeout_mutex);
1842 void reset_fio_state(void)
1851 static int endian_check(void)
1862 else if (u.c[0] == 0x12)
1865 #if defined(FIO_LITTLE_ENDIAN)
1868 #elif defined(FIO_BIG_ENDIAN)
1881 int main(int argc, char *argv[], char *envp[])
1885 if (endian_check()) {
1886 log_err("fio: endianness settings appear wrong.\n");
1887 log_err("fio: please report this to fio@vger.kernel.org\n");
1896 * We need locale for number printing, if it isn't set then just
1897 * go with the US format.
1899 if (!getenv("LC_NUMERIC"))
1900 setlocale(LC_NUMERIC, "en_US");
1902 ps = sysconf(_SC_PAGESIZE);
1904 log_err("Failed to get page size\n");
1911 fio_keywords_init();
1913 if (parse_options(argc, argv))