2 * fio - the flexible io tester
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
7 * The license below covers all files distributed with fio unless otherwise
8 * noted in the file itself.
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License version 2 as
12 * published by the Free Software Foundation.
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
41 #ifndef FIO_NO_HAVE_SHM_H
52 #include "lib/memalign.h"
54 #include "lib/getrusage.h"
57 #include "workqueue.h"
58 #include "lib/mountcheck.h"
59 #include "rate-submit.h"
60 #include "helper_thread.h"
62 static struct fio_mutex *startup_mutex;
63 static struct flist_head *cgroup_list;
64 static char *cgroup_mnt;
65 static int exit_value;
66 static volatile int fio_abort;
67 static unsigned int nr_process = 0;
68 static unsigned int nr_thread = 0;
70 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
73 unsigned int thread_number = 0;
74 unsigned int stat_number = 0;
77 unsigned long done_secs = 0;
79 #define JOB_START_TIMEOUT (5 * 1000)
81 static void sig_int(int sig)
85 fio_server_got_signal(sig);
87 log_info("\nfio: terminating on signal %d\n", sig);
92 fio_terminate_threads(TERMINATE_ALL);
96 void sig_show_status(int sig)
98 show_running_run_stats();
101 static void set_sig_handlers(void)
103 struct sigaction act;
105 memset(&act, 0, sizeof(act));
106 act.sa_handler = sig_int;
107 act.sa_flags = SA_RESTART;
108 sigaction(SIGINT, &act, NULL);
110 memset(&act, 0, sizeof(act));
111 act.sa_handler = sig_int;
112 act.sa_flags = SA_RESTART;
113 sigaction(SIGTERM, &act, NULL);
115 /* Windows uses SIGBREAK as a quit signal from other applications */
117 memset(&act, 0, sizeof(act));
118 act.sa_handler = sig_int;
119 act.sa_flags = SA_RESTART;
120 sigaction(SIGBREAK, &act, NULL);
123 memset(&act, 0, sizeof(act));
124 act.sa_handler = sig_show_status;
125 act.sa_flags = SA_RESTART;
126 sigaction(SIGUSR1, &act, NULL);
129 memset(&act, 0, sizeof(act));
130 act.sa_handler = sig_int;
131 act.sa_flags = SA_RESTART;
132 sigaction(SIGPIPE, &act, NULL);
137 * Check if we are above the minimum rate given.
139 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
142 unsigned long long bytes = 0;
143 unsigned long iops = 0;
146 unsigned int ratemin = 0;
147 unsigned int rate_iops = 0;
148 unsigned int rate_iops_min = 0;
150 assert(ddir_rw(ddir));
152 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
156 * allow a 2 second settle period in the beginning
158 if (mtime_since(&td->start, now) < 2000)
161 iops += td->this_io_blocks[ddir];
162 bytes += td->this_io_bytes[ddir];
163 ratemin += td->o.ratemin[ddir];
164 rate_iops += td->o.rate_iops[ddir];
165 rate_iops_min += td->o.rate_iops_min[ddir];
168 * if rate blocks is set, sample is running
170 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
171 spent = mtime_since(&td->lastrate[ddir], now);
172 if (spent < td->o.ratecycle)
175 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
177 * check bandwidth specified rate
179 if (bytes < td->rate_bytes[ddir]) {
180 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
181 td->o.name, ratemin, bytes);
185 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
189 if (rate < ratemin ||
190 bytes < td->rate_bytes[ddir]) {
191 log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
192 td->o.name, ratemin, rate);
198 * checks iops specified rate
200 if (iops < rate_iops) {
201 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
202 td->o.name, rate_iops, iops);
206 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
210 if (rate < rate_iops_min ||
211 iops < td->rate_blocks[ddir]) {
212 log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
213 td->o.name, rate_iops_min, rate);
220 td->rate_bytes[ddir] = bytes;
221 td->rate_blocks[ddir] = iops;
222 memcpy(&td->lastrate[ddir], now, sizeof(*now));
226 static bool check_min_rate(struct thread_data *td, struct timespec *now)
230 if (td->bytes_done[DDIR_READ])
231 ret |= __check_min_rate(td, now, DDIR_READ);
232 if (td->bytes_done[DDIR_WRITE])
233 ret |= __check_min_rate(td, now, DDIR_WRITE);
234 if (td->bytes_done[DDIR_TRIM])
235 ret |= __check_min_rate(td, now, DDIR_TRIM);
241 * When job exits, we can cancel the in-flight IO if we are using async
242 * io. Attempt to do so.
244 static void cleanup_pending_aio(struct thread_data *td)
249 * get immediately available events, if any
251 r = io_u_queued_complete(td, 0);
256 * now cancel remaining active events
258 if (td->io_ops->cancel) {
262 io_u_qiter(&td->io_u_all, io_u, i) {
263 if (io_u->flags & IO_U_F_FLIGHT) {
264 r = td->io_ops->cancel(td, io_u);
272 r = io_u_queued_complete(td, td->cur_depth);
276 * Helper to handle the final sync of a file. Works just like the normal
277 * io path, just does everything sync.
279 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
281 struct io_u *io_u = __get_io_u(td);
287 io_u->ddir = DDIR_SYNC;
290 if (td_io_prep(td, io_u)) {
296 ret = td_io_queue(td, io_u);
298 td_verror(td, io_u->error, "td_io_queue");
301 } else if (ret == FIO_Q_QUEUED) {
302 if (td_io_commit(td))
304 if (io_u_queued_complete(td, 1) < 0)
306 } else if (ret == FIO_Q_COMPLETED) {
308 td_verror(td, io_u->error, "td_io_queue");
312 if (io_u_sync_complete(td, io_u) < 0)
314 } else if (ret == FIO_Q_BUSY) {
315 if (td_io_commit(td))
323 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
327 if (fio_file_open(f))
328 return fio_io_sync(td, f);
330 if (td_io_open_file(td, f))
333 ret = fio_io_sync(td, f);
334 td_io_close_file(td, f);
338 static inline void __update_ts_cache(struct thread_data *td)
340 fio_gettime(&td->ts_cache, NULL);
343 static inline void update_ts_cache(struct thread_data *td)
345 if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
346 __update_ts_cache(td);
349 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
351 if (in_ramp_time(td))
355 if (utime_since(&td->epoch, t) >= td->o.timeout)
362 * We need to update the runtime consistently in ms, but keep a running
363 * tally of the current elapsed time in microseconds for sub millisecond
366 static inline void update_runtime(struct thread_data *td,
367 unsigned long long *elapsed_us,
368 const enum fio_ddir ddir)
370 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
373 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
374 elapsed_us[ddir] += utime_since_now(&td->start);
375 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
378 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
383 if (ret < 0 || td->error) {
385 enum error_type_bit eb;
390 eb = td_error_type(ddir, err);
391 if (!(td->o.continue_on_error & (1 << eb)))
394 if (td_non_fatal_error(td, eb, err)) {
396 * Continue with the I/Os in case of
399 update_error_count(td, err);
403 } else if (td->o.fill_device && err == ENOSPC) {
405 * We expect to hit this error if
406 * fill_device option is set.
409 fio_mark_td_terminate(td);
413 * Stop the I/O in case of a fatal
416 update_error_count(td, err);
424 static void check_update_rusage(struct thread_data *td)
426 if (td->update_rusage) {
427 td->update_rusage = 0;
428 update_rusage_stat(td);
429 fio_mutex_up(td->rusage_sem);
433 static int wait_for_completions(struct thread_data *td, struct timespec *time)
435 const int full = queue_full(td);
439 if (td->flags & TD_F_REGROW_LOGS)
440 return io_u_quiesce(td);
443 * if the queue is full, we MUST reap at least 1 event
445 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
446 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
449 if (time && (__should_check_rate(td, DDIR_READ) ||
450 __should_check_rate(td, DDIR_WRITE) ||
451 __should_check_rate(td, DDIR_TRIM)))
452 fio_gettime(time, NULL);
455 ret = io_u_queued_complete(td, min_evts);
458 } while (full && (td->cur_depth > td->o.iodepth_low));
463 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
464 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
465 struct timespec *comp_time)
470 case FIO_Q_COMPLETED:
473 clear_io_u(td, io_u);
474 } else if (io_u->resid) {
475 int bytes = io_u->xfer_buflen - io_u->resid;
476 struct fio_file *f = io_u->file;
479 *bytes_issued += bytes;
482 trim_io_piece(td, io_u);
489 unlog_io_piece(td, io_u);
490 td_verror(td, EIO, "full resid");
495 io_u->xfer_buflen = io_u->resid;
496 io_u->xfer_buf += bytes;
497 io_u->offset += bytes;
499 if (ddir_rw(io_u->ddir))
500 td->ts.short_io_u[io_u->ddir]++;
503 if (io_u->offset == f->real_file_size)
506 requeue_io_u(td, &io_u);
509 if (comp_time && (__should_check_rate(td, DDIR_READ) ||
510 __should_check_rate(td, DDIR_WRITE) ||
511 __should_check_rate(td, DDIR_TRIM)))
512 fio_gettime(comp_time, NULL);
514 *ret = io_u_sync_complete(td, io_u);
519 if (td->flags & TD_F_REGROW_LOGS)
523 * when doing I/O (not when verifying),
524 * check for any errors that are to be ignored
532 * if the engine doesn't have a commit hook,
533 * the io_u is really queued. if it does have such
534 * a hook, it has to call io_u_queued() itself.
536 if (td->io_ops->commit == NULL)
537 io_u_queued(td, io_u);
539 *bytes_issued += io_u->xfer_buflen;
543 unlog_io_piece(td, io_u);
544 requeue_io_u(td, &io_u);
545 ret2 = td_io_commit(td);
551 td_verror(td, -(*ret), "td_io_queue");
555 if (break_on_this_error(td, ddir, ret))
561 static inline bool io_in_polling(struct thread_data *td)
563 return !td->o.iodepth_batch_complete_min &&
564 !td->o.iodepth_batch_complete_max;
567 * Unlinks files from thread data fio_file structure
569 static int unlink_all_files(struct thread_data *td)
575 for_each_file(td, f, i) {
576 if (f->filetype != FIO_TYPE_FILE)
578 ret = td_io_unlink_file(td, f);
584 td_verror(td, ret, "unlink_all_files");
590 * The main verify engine. Runs over the writes we previously submitted,
591 * reads the blocks back in, and checks the crc/md5 of the data.
593 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
600 dprint(FD_VERIFY, "starting loop\n");
603 * sync io first and invalidate cache, to make sure we really
606 for_each_file(td, f, i) {
607 if (!fio_file_open(f))
609 if (fio_io_sync(td, f))
611 if (file_invalidate_cache(td, f))
615 check_update_rusage(td);
621 * verify_state needs to be reset before verification
622 * proceeds so that expected random seeds match actual
623 * random seeds in headers. The main loop will reset
624 * all random number generators if randrepeat is set.
626 if (!td->o.rand_repeatable)
627 td_fill_verify_state_seed(td);
629 td_set_runstate(td, TD_VERIFYING);
632 while (!td->terminate) {
637 check_update_rusage(td);
639 if (runtime_exceeded(td, &td->ts_cache)) {
640 __update_ts_cache(td);
641 if (runtime_exceeded(td, &td->ts_cache)) {
642 fio_mark_td_terminate(td);
647 if (flow_threshold_exceeded(td))
650 if (!td->o.experimental_verify) {
651 io_u = __get_io_u(td);
655 if (get_next_verify(td, io_u)) {
660 if (td_io_prep(td, io_u)) {
665 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
668 while ((io_u = get_io_u(td)) != NULL) {
669 if (IS_ERR_OR_NULL(io_u)) {
676 * We are only interested in the places where
677 * we wrote or trimmed IOs. Turn those into
678 * reads for verification purposes.
680 if (io_u->ddir == DDIR_READ) {
682 * Pretend we issued it for rwmix
685 td->io_issues[DDIR_READ]++;
688 } else if (io_u->ddir == DDIR_TRIM) {
689 io_u->ddir = DDIR_READ;
690 io_u_set(td, io_u, IO_U_F_TRIMMED);
692 } else if (io_u->ddir == DDIR_WRITE) {
693 io_u->ddir = DDIR_READ;
705 if (verify_state_should_stop(td, io_u)) {
710 if (td->o.verify_async)
711 io_u->end_io = verify_io_u_async;
713 io_u->end_io = verify_io_u;
716 if (!td->o.disable_slat)
717 fio_gettime(&io_u->start_time, NULL);
719 ret = td_io_queue(td, io_u);
721 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
725 * if we can queue more, do so. but check if there are
726 * completed io_u's first. Note that we can get BUSY even
727 * without IO queued, if the system is resource starved.
730 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
731 if (full || io_in_polling(td))
732 ret = wait_for_completions(td, NULL);
738 check_update_rusage(td);
741 min_events = td->cur_depth;
744 ret = io_u_queued_complete(td, min_events);
746 cleanup_pending_aio(td);
748 td_set_runstate(td, TD_RUNNING);
750 dprint(FD_VERIFY, "exiting loop\n");
753 static bool exceeds_number_ios(struct thread_data *td)
755 unsigned long long number_ios;
757 if (!td->o.number_ios)
760 number_ios = ddir_rw_sum(td->io_blocks);
761 number_ios += td->io_u_queued + td->io_u_in_flight;
763 return number_ios >= (td->o.number_ios * td->loops);
766 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
768 unsigned long long bytes, limit;
771 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
772 else if (td_write(td))
773 bytes = this_bytes[DDIR_WRITE];
774 else if (td_read(td))
775 bytes = this_bytes[DDIR_READ];
777 bytes = this_bytes[DDIR_TRIM];
780 limit = td->o.io_size;
785 return bytes >= limit || exceeds_number_ios(td);
788 static bool io_issue_bytes_exceeded(struct thread_data *td)
790 return io_bytes_exceeded(td, td->io_issue_bytes);
793 static bool io_complete_bytes_exceeded(struct thread_data *td)
795 return io_bytes_exceeded(td, td->this_io_bytes);
799 * used to calculate the next io time for rate control
802 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
804 uint64_t secs, remainder, bps, bytes, iops;
806 assert(!(td->flags & TD_F_CHILD));
807 bytes = td->rate_io_issue_bytes[ddir];
808 bps = td->rate_bps[ddir];
810 if (td->o.rate_process == RATE_PROCESS_POISSON) {
812 iops = bps / td->o.bs[ddir];
813 val = (int64_t) (1000000 / iops) *
814 -logf(__rand_0_1(&td->poisson_state[ddir]));
816 dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
817 (unsigned long long) 1000000 / val,
820 td->last_usec[ddir] += val;
821 return td->last_usec[ddir];
824 remainder = bytes % bps;
825 return remainder * 1000000 / bps + secs * 1000000;
832 * Main IO worker function. It retrieves io_u's to process and queues
833 * and reaps them, checking for rate and errors along the way.
835 * Returns number of bytes written and trimmed.
837 static void do_io(struct thread_data *td, uint64_t *bytes_done)
841 uint64_t total_bytes, bytes_issued = 0;
843 for (i = 0; i < DDIR_RWDIR_CNT; i++)
844 bytes_done[i] = td->bytes_done[i];
846 if (in_ramp_time(td))
847 td_set_runstate(td, TD_RAMP);
849 td_set_runstate(td, TD_RUNNING);
853 total_bytes = td->o.size;
855 * Allow random overwrite workloads to write up to io_size
856 * before starting verification phase as 'size' doesn't apply.
858 if (td_write(td) && td_random(td) && td->o.norandommap)
859 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
861 * If verify_backlog is enabled, we'll run the verify in this
862 * handler as well. For that case, we may need up to twice the
865 if (td->o.verify != VERIFY_NONE &&
866 (td_write(td) && td->o.verify_backlog))
867 total_bytes += td->o.size;
869 /* In trimwrite mode, each byte is trimmed and then written, so
870 * allow total_bytes to be twice as big */
871 if (td_trimwrite(td))
872 total_bytes += td->total_io_size;
874 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
875 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
877 struct timespec comp_time;
882 check_update_rusage(td);
884 if (td->terminate || td->done)
889 if (runtime_exceeded(td, &td->ts_cache)) {
890 __update_ts_cache(td);
891 if (runtime_exceeded(td, &td->ts_cache)) {
892 fio_mark_td_terminate(td);
897 if (flow_threshold_exceeded(td))
901 * Break if we exceeded the bytes. The exception is time
902 * based runs, but we still need to break out of the loop
903 * for those to run verification, if enabled.
905 if (bytes_issued >= total_bytes &&
906 (!td->o.time_based ||
907 (td->o.time_based && td->o.verify != VERIFY_NONE)))
911 if (IS_ERR_OR_NULL(io_u)) {
912 int err = PTR_ERR(io_u);
919 if (td->o.latency_target)
927 * Add verification end_io handler if:
928 * - Asked to verify (!td_rw(td))
929 * - Or the io_u is from our verify list (mixed write/ver)
931 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
932 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
934 if (!td->o.verify_pattern_bytes) {
935 io_u->rand_seed = __rand(&td->verify_state);
936 if (sizeof(int) != sizeof(long *))
937 io_u->rand_seed *= __rand(&td->verify_state);
940 if (verify_state_should_stop(td, io_u)) {
945 if (td->o.verify_async)
946 io_u->end_io = verify_io_u_async;
948 io_u->end_io = verify_io_u;
949 td_set_runstate(td, TD_VERIFYING);
950 } else if (in_ramp_time(td))
951 td_set_runstate(td, TD_RAMP);
953 td_set_runstate(td, TD_RUNNING);
956 * Always log IO before it's issued, so we know the specific
957 * order of it. The logged unit will track when the IO has
960 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
962 td->o.verify != VERIFY_NONE &&
963 !td->o.experimental_verify)
964 log_io_piece(td, io_u);
966 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
967 const unsigned long blen = io_u->xfer_buflen;
968 const enum fio_ddir ddir = acct_ddir(io_u);
973 workqueue_enqueue(&td->io_wq, &io_u->work);
977 td->io_issues[ddir]++;
978 td->io_issue_bytes[ddir] += blen;
979 td->rate_io_issue_bytes[ddir] += blen;
982 if (should_check_rate(td))
983 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
986 ret = td_io_queue(td, io_u);
988 if (should_check_rate(td))
989 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
991 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
995 * See if we need to complete some commands. Note that
996 * we can get BUSY even without IO queued, if the
997 * system is resource starved.
1000 full = queue_full(td) ||
1001 (ret == FIO_Q_BUSY && td->cur_depth);
1002 if (full || io_in_polling(td))
1003 ret = wait_for_completions(td, &comp_time);
1007 if (!ddir_rw_sum(td->bytes_done) &&
1008 !td_ioengine_flagged(td, FIO_NOIO))
1011 if (!in_ramp_time(td) && should_check_rate(td)) {
1012 if (check_min_rate(td, &comp_time)) {
1013 if (exitall_on_terminate || td->o.exitall_error)
1014 fio_terminate_threads(td->groupid);
1015 td_verror(td, EIO, "check_min_rate");
1019 if (!in_ramp_time(td) && td->o.latency_target)
1020 lat_target_check(td);
1022 if (td->o.thinktime) {
1023 unsigned long long b;
1025 b = ddir_rw_sum(td->io_blocks);
1026 if (!(b % td->o.thinktime_blocks)) {
1031 if (td->o.thinktime_spin)
1032 usec_spin(td->o.thinktime_spin);
1034 left = td->o.thinktime - td->o.thinktime_spin;
1036 usec_sleep(td, left);
1041 check_update_rusage(td);
1043 if (td->trim_entries)
1044 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1046 if (td->o.fill_device && td->error == ENOSPC) {
1048 fio_mark_td_terminate(td);
1053 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1054 workqueue_flush(&td->io_wq);
1060 ret = io_u_queued_complete(td, i);
1061 if (td->o.fill_device && td->error == ENOSPC)
1065 if (should_fsync(td) && td->o.end_fsync) {
1066 td_set_runstate(td, TD_FSYNCING);
1068 for_each_file(td, f, i) {
1069 if (!fio_file_fsync(td, f))
1072 log_err("fio: end_fsync failed for file %s\n",
1077 cleanup_pending_aio(td);
1080 * stop job if we failed doing any IO
1082 if (!ddir_rw_sum(td->this_io_bytes))
1085 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1086 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1089 static void free_file_completion_logging(struct thread_data *td)
1094 for_each_file(td, f, i) {
1095 if (!f->last_write_comp)
1097 sfree(f->last_write_comp);
1101 static int init_file_completion_logging(struct thread_data *td,
1107 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1110 for_each_file(td, f, i) {
1111 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1112 if (!f->last_write_comp)
1119 free_file_completion_logging(td);
1120 log_err("fio: failed to alloc write comp data\n");
1124 static void cleanup_io_u(struct thread_data *td)
1128 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1130 if (td->io_ops->io_u_free)
1131 td->io_ops->io_u_free(td, io_u);
1133 fio_memfree(io_u, sizeof(*io_u));
1138 io_u_rexit(&td->io_u_requeues);
1139 io_u_qexit(&td->io_u_freelist);
1140 io_u_qexit(&td->io_u_all);
1142 free_file_completion_logging(td);
1145 static int init_io_u(struct thread_data *td)
1148 unsigned int max_bs, min_write;
1149 int cl_align, i, max_units;
1150 int data_xfer = 1, err;
1153 max_units = td->o.iodepth;
1154 max_bs = td_max_bs(td);
1155 min_write = td->o.min_bs[DDIR_WRITE];
1156 td->orig_buffer_size = (unsigned long long) max_bs
1157 * (unsigned long long) max_units;
1159 if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1163 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1164 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1165 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1168 log_err("fio: failed setting up IO queues\n");
1173 * if we may later need to do address alignment, then add any
1174 * possible adjustment here so that we don't cause a buffer
1175 * overflow later. this adjustment may be too much if we get
1176 * lucky and the allocator gives us an aligned address.
1178 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1179 td_ioengine_flagged(td, FIO_RAWIO))
1180 td->orig_buffer_size += page_mask + td->o.mem_align;
1182 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1185 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1186 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1189 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1190 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1194 if (data_xfer && allocate_io_mem(td))
1197 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1198 td_ioengine_flagged(td, FIO_RAWIO))
1199 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1201 p = td->orig_buffer;
1203 cl_align = os_cache_line_size();
1205 for (i = 0; i < max_units; i++) {
1211 ptr = fio_memalign(cl_align, sizeof(*io_u));
1213 log_err("fio: unable to allocate aligned memory\n");
1218 memset(io_u, 0, sizeof(*io_u));
1219 INIT_FLIST_HEAD(&io_u->verify_list);
1220 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1224 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1227 io_u_fill_buffer(td, io_u, min_write, max_bs);
1228 if (td_write(td) && td->o.verify_pattern_bytes) {
1230 * Fill the buffer with the pattern if we are
1231 * going to be doing writes.
1233 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1238 io_u->flags = IO_U_F_FREE;
1239 io_u_qpush(&td->io_u_freelist, io_u);
1242 * io_u never leaves this stack, used for iteration of all
1245 io_u_qpush(&td->io_u_all, io_u);
1247 if (td->io_ops->io_u_init) {
1248 int ret = td->io_ops->io_u_init(td, io_u);
1251 log_err("fio: failed to init engine data: %d\n", ret);
1259 if (init_file_completion_logging(td, max_units))
1266 * This function is Linux specific.
1267 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1269 static int switch_ioscheduler(struct thread_data *td)
1271 #ifdef FIO_HAVE_IOSCHED_SWITCH
1272 char tmp[256], tmp2[128];
1276 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1279 assert(td->files && td->files[0]);
1280 sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1282 f = fopen(tmp, "r+");
1284 if (errno == ENOENT) {
1285 log_err("fio: os or kernel doesn't support IO scheduler"
1289 td_verror(td, errno, "fopen iosched");
1296 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1297 if (ferror(f) || ret != 1) {
1298 td_verror(td, errno, "fwrite");
1306 * Read back and check that the selected scheduler is now the default.
1308 memset(tmp, 0, sizeof(tmp));
1309 ret = fread(tmp, sizeof(tmp), 1, f);
1310 if (ferror(f) || ret < 0) {
1311 td_verror(td, errno, "fread");
1316 * either a list of io schedulers or "none\n" is expected.
1318 tmp[strlen(tmp) - 1] = '\0';
1321 * Write to "none" entry doesn't fail, so check the result here.
1323 if (!strcmp(tmp, "none")) {
1324 log_err("fio: io scheduler is not tunable\n");
1329 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1330 if (!strstr(tmp, tmp2)) {
1331 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1332 td_verror(td, EINVAL, "iosched_switch");
1344 static bool keep_running(struct thread_data *td)
1346 unsigned long long limit;
1350 if (td->o.time_based)
1356 if (exceeds_number_ios(td))
1360 limit = td->o.io_size;
1364 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1368 * If the difference is less than the maximum IO size, we
1371 diff = limit - ddir_rw_sum(td->io_bytes);
1372 if (diff < td_max_bs(td))
1375 if (fio_files_done(td) && !td->o.io_size)
1384 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1386 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1390 str = malloc(newlen);
1391 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1393 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1396 log_err("fio: exec of cmd <%s> failed\n", str);
1403 * Dry run to compute correct state of numberio for verification.
1405 static uint64_t do_dry_run(struct thread_data *td)
1407 td_set_runstate(td, TD_RUNNING);
1409 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1410 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1414 if (td->terminate || td->done)
1417 io_u = get_io_u(td);
1418 if (IS_ERR_OR_NULL(io_u))
1421 io_u_set(td, io_u, IO_U_F_FLIGHT);
1424 if (ddir_rw(acct_ddir(io_u)))
1425 td->io_issues[acct_ddir(io_u)]++;
1426 if (ddir_rw(io_u->ddir)) {
1427 io_u_mark_depth(td, 1);
1428 td->ts.total_io_u[io_u->ddir]++;
1431 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1433 td->o.verify != VERIFY_NONE &&
1434 !td->o.experimental_verify)
1435 log_io_piece(td, io_u);
1437 ret = io_u_sync_complete(td, io_u);
1441 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1445 struct thread_data *td;
1446 struct sk_out *sk_out;
1450 * Entry point for the thread based jobs. The process based jobs end up
1451 * here as well, after a little setup.
1453 static void *thread_main(void *data)
1455 struct fork_data *fd = data;
1456 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1457 struct thread_data *td = fd->td;
1458 struct thread_options *o = &td->o;
1459 struct sk_out *sk_out = fd->sk_out;
1460 uint64_t bytes_done[DDIR_RWDIR_CNT];
1461 int deadlock_loop_cnt;
1465 sk_out_assign(sk_out);
1468 if (!o->use_thread) {
1474 fio_local_clock_init(o->use_thread);
1476 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1479 fio_server_send_start(td);
1481 INIT_FLIST_HEAD(&td->io_log_list);
1482 INIT_FLIST_HEAD(&td->io_hist_list);
1483 INIT_FLIST_HEAD(&td->verify_list);
1484 INIT_FLIST_HEAD(&td->trim_list);
1485 INIT_FLIST_HEAD(&td->next_rand_list);
1486 td->io_hist_tree = RB_ROOT;
1488 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1490 td_verror(td, ret, "mutex_cond_init_pshared");
1493 ret = cond_init_pshared(&td->verify_cond);
1495 td_verror(td, ret, "mutex_cond_pshared");
1499 td_set_runstate(td, TD_INITIALIZED);
1500 dprint(FD_MUTEX, "up startup_mutex\n");
1501 fio_mutex_up(startup_mutex);
1502 dprint(FD_MUTEX, "wait on td->mutex\n");
1503 fio_mutex_down(td->mutex);
1504 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1507 * A new gid requires privilege, so we need to do this before setting
1510 if (o->gid != -1U && setgid(o->gid)) {
1511 td_verror(td, errno, "setgid");
1514 if (o->uid != -1U && setuid(o->uid)) {
1515 td_verror(td, errno, "setuid");
1520 * Do this early, we don't want the compress threads to be limited
1521 * to the same CPUs as the IO workers. So do this before we set
1522 * any potential CPU affinity
1524 if (iolog_compress_init(td, sk_out))
1528 * If we have a gettimeofday() thread, make sure we exclude that
1529 * thread from this job
1532 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1535 * Set affinity first, in case it has an impact on the memory
1538 if (fio_option_is_set(o, cpumask)) {
1539 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1540 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1542 log_err("fio: no CPUs set\n");
1543 log_err("fio: Try increasing number of available CPUs\n");
1544 td_verror(td, EINVAL, "cpus_split");
1548 ret = fio_setaffinity(td->pid, o->cpumask);
1550 td_verror(td, errno, "cpu_set_affinity");
1555 #ifdef CONFIG_LIBNUMA
1556 /* numa node setup */
1557 if (fio_option_is_set(o, numa_cpunodes) ||
1558 fio_option_is_set(o, numa_memnodes)) {
1559 struct bitmask *mask;
1561 if (numa_available() < 0) {
1562 td_verror(td, errno, "Does not support NUMA API\n");
1566 if (fio_option_is_set(o, numa_cpunodes)) {
1567 mask = numa_parse_nodestring(o->numa_cpunodes);
1568 ret = numa_run_on_node_mask(mask);
1569 numa_free_nodemask(mask);
1571 td_verror(td, errno, \
1572 "numa_run_on_node_mask failed\n");
1577 if (fio_option_is_set(o, numa_memnodes)) {
1579 if (o->numa_memnodes)
1580 mask = numa_parse_nodestring(o->numa_memnodes);
1582 switch (o->numa_mem_mode) {
1583 case MPOL_INTERLEAVE:
1584 numa_set_interleave_mask(mask);
1587 numa_set_membind(mask);
1590 numa_set_localalloc();
1592 case MPOL_PREFERRED:
1593 numa_set_preferred(o->numa_mem_prefer_node);
1601 numa_free_nodemask(mask);
1607 if (fio_pin_memory(td))
1611 * May alter parameters that init_io_u() will use, so we need to
1620 if (o->verify_async && verify_async_init(td))
1623 if (fio_option_is_set(o, ioprio) ||
1624 fio_option_is_set(o, ioprio_class)) {
1625 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1627 td_verror(td, errno, "ioprio_set");
1632 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1636 if (nice(o->nice) == -1 && errno != 0) {
1637 td_verror(td, errno, "nice");
1641 if (o->ioscheduler && switch_ioscheduler(td))
1644 if (!o->create_serialize && setup_files(td))
1650 if (init_random_map(td))
1653 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1657 if (pre_read_files(td) < 0)
1661 fio_verify_init(td);
1663 if (rate_submit_init(td, sk_out))
1666 set_epoch_time(td, o->log_unix_epoch);
1667 fio_getrusage(&td->ru_start);
1668 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1669 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1670 memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1672 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1673 o->ratemin[DDIR_TRIM]) {
1674 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1675 sizeof(td->bw_sample_time));
1676 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1677 sizeof(td->bw_sample_time));
1678 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1679 sizeof(td->bw_sample_time));
1682 memset(bytes_done, 0, sizeof(bytes_done));
1685 while (keep_running(td)) {
1686 uint64_t verify_bytes;
1688 fio_gettime(&td->start, NULL);
1689 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1692 clear_io_state(td, 0);
1694 if (o->unlink_each_loop && unlink_all_files(td))
1698 prune_io_piece_log(td);
1700 if (td->o.verify_only && td_write(td))
1701 verify_bytes = do_dry_run(td);
1703 do_io(td, bytes_done);
1705 if (!ddir_rw_sum(bytes_done)) {
1706 fio_mark_td_terminate(td);
1709 verify_bytes = bytes_done[DDIR_WRITE] +
1710 bytes_done[DDIR_TRIM];
1715 * If we took too long to shut down, the main thread could
1716 * already consider us reaped/exited. If that happens, break
1719 if (td->runstate >= TD_EXITED)
1725 * Make sure we've successfully updated the rusage stats
1726 * before waiting on the stat mutex. Otherwise we could have
1727 * the stat thread holding stat mutex and waiting for
1728 * the rusage_sem, which would never get upped because
1729 * this thread is waiting for the stat mutex.
1731 deadlock_loop_cnt = 0;
1733 check_update_rusage(td);
1734 if (!fio_mutex_down_trylock(stat_mutex))
1737 if (deadlock_loop_cnt++ > 5000) {
1738 log_err("fio seems to be stuck grabbing stat_mutex, forcibly exiting\n");
1739 td->error = EDEADLK;
1744 if (td_read(td) && td->io_bytes[DDIR_READ])
1745 update_runtime(td, elapsed_us, DDIR_READ);
1746 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1747 update_runtime(td, elapsed_us, DDIR_WRITE);
1748 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1749 update_runtime(td, elapsed_us, DDIR_TRIM);
1750 fio_gettime(&td->start, NULL);
1751 fio_mutex_up(stat_mutex);
1753 if (td->error || td->terminate)
1756 if (!o->do_verify ||
1757 o->verify == VERIFY_NONE ||
1758 td_ioengine_flagged(td, FIO_UNIDIR))
1761 clear_io_state(td, 0);
1763 fio_gettime(&td->start, NULL);
1765 do_verify(td, verify_bytes);
1768 * See comment further up for why this is done here.
1770 check_update_rusage(td);
1772 fio_mutex_down(stat_mutex);
1773 update_runtime(td, elapsed_us, DDIR_READ);
1774 fio_gettime(&td->start, NULL);
1775 fio_mutex_up(stat_mutex);
1777 if (td->error || td->terminate)
1782 * If td ended up with no I/O when it should have had,
1783 * then something went wrong unless FIO_NOIO or FIO_DISKLESSIO.
1784 * (Are we not missing other flags that can be ignored ?)
1786 if ((td->o.size || td->o.io_size) && !ddir_rw_sum(bytes_done) &&
1787 !(td_ioengine_flagged(td, FIO_NOIO) ||
1788 td_ioengine_flagged(td, FIO_DISKLESSIO)))
1789 log_err("%s: No I/O performed by %s, "
1790 "perhaps try --debug=io option for details?\n",
1791 td->o.name, td->io_ops->name);
1793 td_set_runstate(td, TD_FINISHING);
1795 update_rusage_stat(td);
1796 td->ts.total_run_time = mtime_since_now(&td->epoch);
1797 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1798 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1799 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1801 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1802 (td->o.verify != VERIFY_NONE && td_write(td)))
1803 verify_save_state(td->thread_number);
1805 fio_unpin_memory(td);
1807 td_writeout_logs(td, true);
1809 iolog_compress_exit(td);
1810 rate_submit_exit(td);
1812 if (o->exec_postrun)
1813 exec_string(o, o->exec_postrun, (const char *)"postrun");
1815 if (exitall_on_terminate || (o->exitall_error && td->error))
1816 fio_terminate_threads(td->groupid);
1820 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1823 if (o->verify_async)
1824 verify_async_exit(td);
1826 close_and_free_files(td);
1829 cgroup_shutdown(td, &cgroup_mnt);
1830 verify_free_state(td);
1832 if (td->zone_state_index) {
1835 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1836 free(td->zone_state_index[i]);
1837 free(td->zone_state_index);
1838 td->zone_state_index = NULL;
1841 if (fio_option_is_set(o, cpumask)) {
1842 ret = fio_cpuset_exit(&o->cpumask);
1844 td_verror(td, ret, "fio_cpuset_exit");
1848 * do this very late, it will log file closing as well
1850 if (o->write_iolog_file)
1851 write_iolog_close(td);
1853 td_set_runstate(td, TD_EXITED);
1856 * Do this last after setting our runstate to exited, so we
1857 * know that the stat thread is signaled.
1859 check_update_rusage(td);
1862 return (void *) (uintptr_t) td->error;
1866 * Run over the job map and reap the threads that have exited, if any.
1868 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1871 struct thread_data *td;
1872 unsigned int cputhreads, realthreads, pending;
1876 * reap exited threads (TD_EXITED -> TD_REAPED)
1878 realthreads = pending = cputhreads = 0;
1879 for_each_td(td, i) {
1883 * ->io_ops is NULL for a thread that has closed its
1886 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1895 if (td->runstate == TD_REAPED)
1897 if (td->o.use_thread) {
1898 if (td->runstate == TD_EXITED) {
1899 td_set_runstate(td, TD_REAPED);
1906 if (td->runstate == TD_EXITED)
1910 * check if someone quit or got killed in an unusual way
1912 ret = waitpid(td->pid, &status, flags);
1914 if (errno == ECHILD) {
1915 log_err("fio: pid=%d disappeared %d\n",
1916 (int) td->pid, td->runstate);
1918 td_set_runstate(td, TD_REAPED);
1922 } else if (ret == td->pid) {
1923 if (WIFSIGNALED(status)) {
1924 int sig = WTERMSIG(status);
1926 if (sig != SIGTERM && sig != SIGUSR2)
1927 log_err("fio: pid=%d, got signal=%d\n",
1928 (int) td->pid, sig);
1930 td_set_runstate(td, TD_REAPED);
1933 if (WIFEXITED(status)) {
1934 if (WEXITSTATUS(status) && !td->error)
1935 td->error = WEXITSTATUS(status);
1937 td_set_runstate(td, TD_REAPED);
1943 * If the job is stuck, do a forceful timeout of it and
1946 if (td->terminate &&
1947 td->runstate < TD_FSYNCING &&
1948 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1949 log_err("fio: job '%s' (state=%d) hasn't exited in "
1950 "%lu seconds, it appears to be stuck. Doing "
1951 "forceful exit of this job.\n",
1952 td->o.name, td->runstate,
1953 (unsigned long) time_since_now(&td->terminate_time));
1954 td_set_runstate(td, TD_REAPED);
1959 * thread is not dead, continue
1965 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1966 (*t_rate) -= ddir_rw_sum(td->o.rate);
1973 done_secs += mtime_since_now(&td->epoch) / 1000;
1974 profile_td_exit(td);
1977 if (*nr_running == cputhreads && !pending && realthreads)
1978 fio_terminate_threads(TERMINATE_ALL);
1981 static bool __check_trigger_file(void)
1988 if (stat(trigger_file, &sb))
1991 if (unlink(trigger_file) < 0)
1992 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1998 static bool trigger_timedout(void)
2000 if (trigger_timeout)
2001 return time_since_genesis() >= trigger_timeout;
2006 void exec_trigger(const char *cmd)
2015 log_err("fio: failed executing %s trigger\n", cmd);
2018 void check_trigger_file(void)
2020 if (__check_trigger_file() || trigger_timedout()) {
2022 fio_clients_send_trigger(trigger_remote_cmd);
2024 verify_save_state(IO_LIST_ALL);
2025 fio_terminate_threads(TERMINATE_ALL);
2026 exec_trigger(trigger_cmd);
2031 static int fio_verify_load_state(struct thread_data *td)
2035 if (!td->o.verify_state)
2041 ret = fio_server_get_verify_state(td->o.name,
2042 td->thread_number - 1, &data);
2044 verify_assign_state(td, data);
2046 ret = verify_load_state(td, "local");
2051 static void do_usleep(unsigned int usecs)
2053 check_for_running_stats();
2054 check_trigger_file();
2058 static bool check_mount_writes(struct thread_data *td)
2063 if (!td_write(td) || td->o.allow_mounted_write)
2067 * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2068 * are mkfs'd and mounted.
2070 for_each_file(td, f, i) {
2071 #ifdef FIO_HAVE_CHARDEV_SIZE
2072 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2074 if (f->filetype != FIO_TYPE_BLOCK)
2077 if (device_is_mounted(f->file_name))
2083 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2087 static bool waitee_running(struct thread_data *me)
2089 const char *waitee = me->o.wait_for;
2090 const char *self = me->o.name;
2091 struct thread_data *td;
2097 for_each_td(td, i) {
2098 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2101 if (td->runstate < TD_EXITED) {
2102 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2104 runstate_to_name(td->runstate));
2109 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2114 * Main function for kicking off and reaping jobs, as needed.
2116 static void run_threads(struct sk_out *sk_out)
2118 struct thread_data *td;
2119 unsigned int i, todo, nr_running, nr_started;
2120 uint64_t m_rate, t_rate;
2123 if (fio_gtod_offload && fio_start_gtod_thread())
2126 fio_idle_prof_init();
2130 nr_thread = nr_process = 0;
2131 for_each_td(td, i) {
2132 if (check_mount_writes(td))
2134 if (td->o.use_thread)
2140 if (output_format & FIO_OUTPUT_NORMAL) {
2141 log_info("Starting ");
2143 log_info("%d thread%s", nr_thread,
2144 nr_thread > 1 ? "s" : "");
2148 log_info("%d process%s", nr_process,
2149 nr_process > 1 ? "es" : "");
2155 todo = thread_number;
2158 m_rate = t_rate = 0;
2160 for_each_td(td, i) {
2161 print_status_init(td->thread_number - 1);
2163 if (!td->o.create_serialize)
2166 if (fio_verify_load_state(td))
2170 * do file setup here so it happens sequentially,
2171 * we don't want X number of threads getting their
2172 * client data interspersed on disk
2174 if (setup_files(td)) {
2178 log_err("fio: pid=%d, err=%d/%s\n",
2179 (int) td->pid, td->error, td->verror);
2180 td_set_runstate(td, TD_REAPED);
2187 * for sharing to work, each job must always open
2188 * its own files. so close them, if we opened them
2191 for_each_file(td, f, j) {
2192 if (fio_file_open(f))
2193 td_io_close_file(td, f);
2198 /* start idle threads before io threads start to run */
2199 fio_idle_prof_start();
2204 struct thread_data *map[REAL_MAX_JOBS];
2205 struct timespec this_start;
2206 int this_jobs = 0, left;
2207 struct fork_data *fd;
2210 * create threads (TD_NOT_CREATED -> TD_CREATED)
2212 for_each_td(td, i) {
2213 if (td->runstate != TD_NOT_CREATED)
2217 * never got a chance to start, killed by other
2218 * thread for some reason
2220 if (td->terminate) {
2225 if (td->o.start_delay) {
2226 spent = utime_since_genesis();
2228 if (td->o.start_delay > spent)
2232 if (td->o.stonewall && (nr_started || nr_running)) {
2233 dprint(FD_PROCESS, "%s: stonewall wait\n",
2238 if (waitee_running(td)) {
2239 dprint(FD_PROCESS, "%s: waiting for %s\n",
2240 td->o.name, td->o.wait_for);
2246 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2247 td->update_rusage = 0;
2250 * Set state to created. Thread will transition
2251 * to TD_INITIALIZED when it's done setting up.
2253 td_set_runstate(td, TD_CREATED);
2254 map[this_jobs++] = td;
2257 fd = calloc(1, sizeof(*fd));
2259 fd->sk_out = sk_out;
2261 if (td->o.use_thread) {
2264 dprint(FD_PROCESS, "will pthread_create\n");
2265 ret = pthread_create(&td->thread, NULL,
2268 log_err("pthread_create: %s\n",
2274 ret = pthread_detach(td->thread);
2276 log_err("pthread_detach: %s",
2280 dprint(FD_PROCESS, "will fork\n");
2285 ret = (int)(uintptr_t)thread_main(fd);
2287 } else if (i == fio_debug_jobno)
2288 *fio_debug_jobp = pid;
2290 dprint(FD_MUTEX, "wait on startup_mutex\n");
2291 if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2292 log_err("fio: job startup hung? exiting.\n");
2293 fio_terminate_threads(TERMINATE_ALL);
2298 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2302 * Wait for the started threads to transition to
2305 fio_gettime(&this_start, NULL);
2307 while (left && !fio_abort) {
2308 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2313 for (i = 0; i < this_jobs; i++) {
2317 if (td->runstate == TD_INITIALIZED) {
2320 } else if (td->runstate >= TD_EXITED) {
2324 nr_running++; /* work-around... */
2330 log_err("fio: %d job%s failed to start\n", left,
2331 left > 1 ? "s" : "");
2332 for (i = 0; i < this_jobs; i++) {
2336 kill(td->pid, SIGTERM);
2342 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2344 for_each_td(td, i) {
2345 if (td->runstate != TD_INITIALIZED)
2348 if (in_ramp_time(td))
2349 td_set_runstate(td, TD_RAMP);
2351 td_set_runstate(td, TD_RUNNING);
2354 m_rate += ddir_rw_sum(td->o.ratemin);
2355 t_rate += ddir_rw_sum(td->o.rate);
2357 fio_mutex_up(td->mutex);
2360 reap_threads(&nr_running, &t_rate, &m_rate);
2366 while (nr_running) {
2367 reap_threads(&nr_running, &t_rate, &m_rate);
2371 fio_idle_prof_stop();
2376 static void free_disk_util(void)
2378 disk_util_prune_entries();
2379 helper_thread_destroy();
2382 int fio_backend(struct sk_out *sk_out)
2384 struct thread_data *td;
2388 if (load_profile(exec_profile))
2391 exec_profile = NULL;
2397 struct log_params p = {
2398 .log_type = IO_LOG_TYPE_BW,
2401 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2402 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2403 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2406 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2407 if (startup_mutex == NULL)
2412 helper_thread_create(startup_mutex, sk_out);
2414 cgroup_list = smalloc(sizeof(*cgroup_list));
2415 INIT_FLIST_HEAD(cgroup_list);
2417 run_threads(sk_out);
2419 helper_thread_exit();
2424 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2425 struct io_log *log = agg_io_log[i];
2427 flush_log(log, false);
2433 for_each_td(td, i) {
2435 if (td->ss.iops_data != NULL) {
2436 free(td->ss.iops_data);
2437 free(td->ss.bw_data);
2440 fio_options_free(td);
2441 if (td->rusage_sem) {
2442 fio_mutex_remove(td->rusage_sem);
2443 td->rusage_sem = NULL;
2445 fio_mutex_remove(td->mutex);
2450 cgroup_kill(cgroup_list);
2454 fio_mutex_remove(startup_mutex);