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., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
41 #include "lib/memalign.h"
43 #include "lib/getrusage.h"
46 #include "workqueue.h"
47 #include "lib/mountcheck.h"
48 #include "rate-submit.h"
49 #include "helper_thread.h"
51 #include "zone-dist.h"
53 static struct fio_sem *startup_sem;
54 static struct flist_head *cgroup_list;
55 static struct cgroup_mnt *cgroup_mnt;
56 static int exit_value;
57 static volatile bool fio_abort;
58 static unsigned int nr_process = 0;
59 static unsigned int nr_thread = 0;
61 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
64 unsigned int thread_number = 0;
65 unsigned int stat_number = 0;
68 unsigned long done_secs = 0;
69 #ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP
70 pthread_mutex_t overlap_check = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
72 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
75 #define JOB_START_TIMEOUT (5 * 1000)
77 static void sig_int(int sig)
81 fio_server_got_signal(sig);
83 log_info("\nfio: terminating on signal %d\n", sig);
88 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
92 void sig_show_status(int sig)
94 show_running_run_stats();
97 static void set_sig_handlers(void)
101 memset(&act, 0, sizeof(act));
102 act.sa_handler = sig_int;
103 act.sa_flags = SA_RESTART;
104 sigaction(SIGINT, &act, NULL);
106 memset(&act, 0, sizeof(act));
107 act.sa_handler = sig_int;
108 act.sa_flags = SA_RESTART;
109 sigaction(SIGTERM, &act, NULL);
111 /* Windows uses SIGBREAK as a quit signal from other applications */
113 memset(&act, 0, sizeof(act));
114 act.sa_handler = sig_int;
115 act.sa_flags = SA_RESTART;
116 sigaction(SIGBREAK, &act, NULL);
119 memset(&act, 0, sizeof(act));
120 act.sa_handler = sig_show_status;
121 act.sa_flags = SA_RESTART;
122 sigaction(SIGUSR1, &act, NULL);
125 memset(&act, 0, sizeof(act));
126 act.sa_handler = sig_int;
127 act.sa_flags = SA_RESTART;
128 sigaction(SIGPIPE, &act, NULL);
133 * Check if we are above the minimum rate given.
135 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
138 unsigned long long bytes = 0;
139 unsigned long iops = 0;
141 unsigned long long rate;
142 unsigned long long ratemin = 0;
143 unsigned int rate_iops = 0;
144 unsigned int rate_iops_min = 0;
146 assert(ddir_rw(ddir));
148 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
152 * allow a 2 second settle period in the beginning
154 if (mtime_since(&td->start, now) < 2000)
157 iops += td->this_io_blocks[ddir];
158 bytes += td->this_io_bytes[ddir];
159 ratemin += td->o.ratemin[ddir];
160 rate_iops += td->o.rate_iops[ddir];
161 rate_iops_min += td->o.rate_iops_min[ddir];
164 * if rate blocks is set, sample is running
166 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
167 spent = mtime_since(&td->lastrate[ddir], now);
168 if (spent < td->o.ratecycle)
171 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
173 * check bandwidth specified rate
175 if (bytes < td->rate_bytes[ddir]) {
176 log_err("%s: rate_min=%lluB/s not met, only transferred %lluB\n",
177 td->o.name, ratemin, bytes);
181 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
185 if (rate < ratemin ||
186 bytes < td->rate_bytes[ddir]) {
187 log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
188 td->o.name, ratemin, rate);
194 * checks iops specified rate
196 if (iops < rate_iops) {
197 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
198 td->o.name, rate_iops, iops);
202 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
206 if (rate < rate_iops_min ||
207 iops < td->rate_blocks[ddir]) {
208 log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
209 td->o.name, rate_iops_min, rate);
216 td->rate_bytes[ddir] = bytes;
217 td->rate_blocks[ddir] = iops;
218 memcpy(&td->lastrate[ddir], now, sizeof(*now));
222 static bool check_min_rate(struct thread_data *td, struct timespec *now)
226 for_each_rw_ddir(ddir) {
227 if (td->bytes_done[ddir])
228 ret |= __check_min_rate(td, now, ddir);
235 * When job exits, we can cancel the in-flight IO if we are using async
236 * io. Attempt to do so.
238 static void cleanup_pending_aio(struct thread_data *td)
243 * get immediately available events, if any
245 r = io_u_queued_complete(td, 0);
248 * now cancel remaining active events
250 if (td->io_ops->cancel) {
254 io_u_qiter(&td->io_u_all, io_u, i) {
255 if (io_u->flags & IO_U_F_FLIGHT) {
256 r = td->io_ops->cancel(td, io_u);
264 r = io_u_queued_complete(td, td->cur_depth);
268 * Helper to handle the final sync of a file. Works just like the normal
269 * io path, just does everything sync.
271 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
273 struct io_u *io_u = __get_io_u(td);
274 enum fio_q_status ret;
279 io_u->ddir = DDIR_SYNC;
281 io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
283 if (td_io_prep(td, io_u)) {
289 ret = td_io_queue(td, io_u);
293 if (io_u_queued_complete(td, 1) < 0)
296 case FIO_Q_COMPLETED:
298 td_verror(td, io_u->error, "td_io_queue");
302 if (io_u_sync_complete(td, io_u) < 0)
313 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
317 if (fio_file_open(f))
318 return fio_io_sync(td, f);
320 if (td_io_open_file(td, f))
323 ret = fio_io_sync(td, f);
325 if (fio_file_open(f))
326 ret2 = td_io_close_file(td, f);
327 return (ret || ret2);
330 static inline void __update_ts_cache(struct thread_data *td)
332 fio_gettime(&td->ts_cache, NULL);
335 static inline void update_ts_cache(struct thread_data *td)
337 if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
338 __update_ts_cache(td);
341 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
343 if (in_ramp_time(td))
347 if (utime_since(&td->epoch, t) >= td->o.timeout)
354 * We need to update the runtime consistently in ms, but keep a running
355 * tally of the current elapsed time in microseconds for sub millisecond
358 static inline void update_runtime(struct thread_data *td,
359 unsigned long long *elapsed_us,
360 const enum fio_ddir ddir)
362 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
365 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
366 elapsed_us[ddir] += utime_since_now(&td->start);
367 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
370 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
375 if (ret < 0 || td->error) {
377 enum error_type_bit eb;
382 eb = td_error_type(ddir, err);
383 if (!(td->o.continue_on_error & (1 << eb)))
386 if (td_non_fatal_error(td, eb, err)) {
388 * Continue with the I/Os in case of
391 update_error_count(td, err);
395 } else if (td->o.fill_device && err == ENOSPC) {
397 * We expect to hit this error if
398 * fill_device option is set.
401 fio_mark_td_terminate(td);
405 * Stop the I/O in case of a fatal
408 update_error_count(td, err);
416 static void check_update_rusage(struct thread_data *td)
418 if (td->update_rusage) {
419 td->update_rusage = 0;
420 update_rusage_stat(td);
421 fio_sem_up(td->rusage_sem);
425 static int wait_for_completions(struct thread_data *td, struct timespec *time)
427 const int full = queue_full(td);
431 if (td->flags & TD_F_REGROW_LOGS)
432 return io_u_quiesce(td);
435 * if the queue is full, we MUST reap at least 1 event
437 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
438 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
441 if (time && __should_check_rate(td))
442 fio_gettime(time, NULL);
445 ret = io_u_queued_complete(td, min_evts);
448 } while (full && (td->cur_depth > td->o.iodepth_low));
453 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
454 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
455 struct timespec *comp_time)
458 case FIO_Q_COMPLETED:
461 clear_io_u(td, io_u);
462 } else if (io_u->resid) {
463 long long bytes = io_u->xfer_buflen - io_u->resid;
464 struct fio_file *f = io_u->file;
467 *bytes_issued += bytes;
477 unlog_io_piece(td, io_u);
478 td_verror(td, EIO, "full resid");
483 io_u->xfer_buflen = io_u->resid;
484 io_u->xfer_buf += bytes;
485 io_u->offset += bytes;
487 if (ddir_rw(io_u->ddir))
488 td->ts.short_io_u[io_u->ddir]++;
490 if (io_u->offset == f->real_file_size)
493 requeue_io_u(td, &io_u);
496 if (comp_time && __should_check_rate(td))
497 fio_gettime(comp_time, NULL);
499 *ret = io_u_sync_complete(td, io_u);
504 if (td->flags & TD_F_REGROW_LOGS)
508 * when doing I/O (not when verifying),
509 * check for any errors that are to be ignored
517 * if the engine doesn't have a commit hook,
518 * the io_u is really queued. if it does have such
519 * a hook, it has to call io_u_queued() itself.
521 if (td->io_ops->commit == NULL)
522 io_u_queued(td, io_u);
524 *bytes_issued += io_u->xfer_buflen;
528 unlog_io_piece(td, io_u);
529 requeue_io_u(td, &io_u);
534 td_verror(td, -(*ret), "td_io_queue");
538 if (break_on_this_error(td, ddir, ret))
544 static inline bool io_in_polling(struct thread_data *td)
546 return !td->o.iodepth_batch_complete_min &&
547 !td->o.iodepth_batch_complete_max;
550 * Unlinks files from thread data fio_file structure
552 static int unlink_all_files(struct thread_data *td)
558 for_each_file(td, f, i) {
559 if (f->filetype != FIO_TYPE_FILE)
561 ret = td_io_unlink_file(td, f);
567 td_verror(td, ret, "unlink_all_files");
573 * Check if io_u will overlap an in-flight IO in the queue
575 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
578 struct io_u *check_io_u;
579 unsigned long long x1, x2, y1, y2;
583 x2 = io_u->offset + io_u->buflen;
585 io_u_qiter(q, check_io_u, i) {
586 if (check_io_u->flags & IO_U_F_FLIGHT) {
587 y1 = check_io_u->offset;
588 y2 = check_io_u->offset + check_io_u->buflen;
590 if (x1 < y2 && y1 < x2) {
592 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
594 y1, check_io_u->buflen);
603 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
606 * Check for overlap if the user asked us to, and we have
607 * at least one IO in flight besides this one.
609 if (td->o.serialize_overlap && td->cur_depth > 1 &&
610 in_flight_overlap(&td->io_u_all, io_u))
613 return td_io_queue(td, io_u);
617 * The main verify engine. Runs over the writes we previously submitted,
618 * reads the blocks back in, and checks the crc/md5 of the data.
620 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
627 dprint(FD_VERIFY, "starting loop\n");
630 * sync io first and invalidate cache, to make sure we really
633 for_each_file(td, f, i) {
634 if (!fio_file_open(f))
636 if (fio_io_sync(td, f))
638 if (file_invalidate_cache(td, f))
642 check_update_rusage(td);
648 * verify_state needs to be reset before verification
649 * proceeds so that expected random seeds match actual
650 * random seeds in headers. The main loop will reset
651 * all random number generators if randrepeat is set.
653 if (!td->o.rand_repeatable)
654 td_fill_verify_state_seed(td);
656 td_set_runstate(td, TD_VERIFYING);
659 while (!td->terminate) {
664 check_update_rusage(td);
666 if (runtime_exceeded(td, &td->ts_cache)) {
667 __update_ts_cache(td);
668 if (runtime_exceeded(td, &td->ts_cache)) {
669 fio_mark_td_terminate(td);
674 if (flow_threshold_exceeded(td))
677 if (!td->o.experimental_verify) {
678 io_u = __get_io_u(td);
682 if (get_next_verify(td, io_u)) {
687 if (td_io_prep(td, io_u)) {
692 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
695 while ((io_u = get_io_u(td)) != NULL) {
696 if (IS_ERR_OR_NULL(io_u)) {
703 * We are only interested in the places where
704 * we wrote or trimmed IOs. Turn those into
705 * reads for verification purposes.
707 if (io_u->ddir == DDIR_READ) {
709 * Pretend we issued it for rwmix
712 td->io_issues[DDIR_READ]++;
715 } else if (io_u->ddir == DDIR_TRIM) {
716 io_u->ddir = DDIR_READ;
717 io_u_set(td, io_u, IO_U_F_TRIMMED);
719 } else if (io_u->ddir == DDIR_WRITE) {
720 io_u->ddir = DDIR_READ;
721 populate_verify_io_u(td, io_u);
733 if (verify_state_should_stop(td, io_u)) {
738 if (td->o.verify_async)
739 io_u->end_io = verify_io_u_async;
741 io_u->end_io = verify_io_u;
744 if (!td->o.disable_slat)
745 fio_gettime(&io_u->start_time, NULL);
747 ret = io_u_submit(td, io_u);
749 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
753 * if we can queue more, do so. but check if there are
754 * completed io_u's first. Note that we can get BUSY even
755 * without IO queued, if the system is resource starved.
758 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
759 if (full || io_in_polling(td))
760 ret = wait_for_completions(td, NULL);
766 check_update_rusage(td);
769 min_events = td->cur_depth;
772 ret = io_u_queued_complete(td, min_events);
774 cleanup_pending_aio(td);
776 td_set_runstate(td, TD_RUNNING);
778 dprint(FD_VERIFY, "exiting loop\n");
781 static bool exceeds_number_ios(struct thread_data *td)
783 unsigned long long number_ios;
785 if (!td->o.number_ios)
788 number_ios = ddir_rw_sum(td->io_blocks);
789 number_ios += td->io_u_queued + td->io_u_in_flight;
791 return number_ios >= (td->o.number_ios * td->loops);
794 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
796 unsigned long long bytes, limit;
799 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
800 else if (td_write(td))
801 bytes = this_bytes[DDIR_WRITE];
802 else if (td_read(td))
803 bytes = this_bytes[DDIR_READ];
805 bytes = this_bytes[DDIR_TRIM];
808 limit = td->o.io_size;
813 return bytes >= limit || exceeds_number_ios(td);
816 static bool io_issue_bytes_exceeded(struct thread_data *td)
818 return io_bytes_exceeded(td, td->io_issue_bytes);
821 static bool io_complete_bytes_exceeded(struct thread_data *td)
823 return io_bytes_exceeded(td, td->this_io_bytes);
827 * used to calculate the next io time for rate control
830 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
832 uint64_t bps = td->rate_bps[ddir];
834 assert(!(td->flags & TD_F_CHILD));
836 if (td->o.rate_process == RATE_PROCESS_POISSON) {
839 iops = bps / td->o.bs[ddir];
840 val = (int64_t) (1000000 / iops) *
841 -logf(__rand_0_1(&td->poisson_state[ddir]));
843 dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
844 (unsigned long long) 1000000 / val,
847 td->last_usec[ddir] += val;
848 return td->last_usec[ddir];
850 uint64_t bytes = td->rate_io_issue_bytes[ddir];
851 uint64_t secs = bytes / bps;
852 uint64_t remainder = bytes % bps;
854 return remainder * 1000000 / bps + secs * 1000000;
860 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
862 unsigned long long b;
866 b = ddir_rw_sum(td->io_blocks);
867 if (b % td->o.thinktime_blocks)
873 if (td->o.thinktime_spin)
874 total = usec_spin(td->o.thinktime_spin);
876 left = td->o.thinktime - total;
878 total += usec_sleep(td, left);
881 * If we're ignoring thinktime for the rate, add the number of bytes
882 * we would have done while sleeping, minus one block to ensure we
883 * start issuing immediately after the sleep.
885 if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
886 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
887 uint64_t bs = td->o.min_bs[ddir];
888 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
891 if (usperop <= total)
894 over = (usperop - total) / usperop * -bs;
896 td->rate_io_issue_bytes[ddir] += (missed - over);
897 /* adjust for rate_process=poisson */
898 td->last_usec[ddir] += total;
903 * Main IO worker function. It retrieves io_u's to process and queues
904 * and reaps them, checking for rate and errors along the way.
906 * Returns number of bytes written and trimmed.
908 static void do_io(struct thread_data *td, uint64_t *bytes_done)
912 uint64_t total_bytes, bytes_issued = 0;
914 for (i = 0; i < DDIR_RWDIR_CNT; i++)
915 bytes_done[i] = td->bytes_done[i];
917 if (in_ramp_time(td))
918 td_set_runstate(td, TD_RAMP);
920 td_set_runstate(td, TD_RUNNING);
924 total_bytes = td->o.size;
926 * Allow random overwrite workloads to write up to io_size
927 * before starting verification phase as 'size' doesn't apply.
929 if (td_write(td) && td_random(td) && td->o.norandommap)
930 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
932 * If verify_backlog is enabled, we'll run the verify in this
933 * handler as well. For that case, we may need up to twice the
936 if (td->o.verify != VERIFY_NONE &&
937 (td_write(td) && td->o.verify_backlog))
938 total_bytes += td->o.size;
940 /* In trimwrite mode, each byte is trimmed and then written, so
941 * allow total_bytes to be twice as big */
942 if (td_trimwrite(td))
943 total_bytes += td->total_io_size;
945 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
946 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
948 struct timespec comp_time;
953 check_update_rusage(td);
955 if (td->terminate || td->done)
960 if (runtime_exceeded(td, &td->ts_cache)) {
961 __update_ts_cache(td);
962 if (runtime_exceeded(td, &td->ts_cache)) {
963 fio_mark_td_terminate(td);
968 if (flow_threshold_exceeded(td))
972 * Break if we exceeded the bytes. The exception is time
973 * based runs, but we still need to break out of the loop
974 * for those to run verification, if enabled.
975 * Jobs read from iolog do not use this stop condition.
977 if (bytes_issued >= total_bytes &&
978 !td->o.read_iolog_file &&
979 (!td->o.time_based ||
980 (td->o.time_based && td->o.verify != VERIFY_NONE)))
984 if (IS_ERR_OR_NULL(io_u)) {
985 int err = PTR_ERR(io_u);
993 if (td->o.latency_target)
998 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
999 populate_verify_io_u(td, io_u);
1004 * Add verification end_io handler if:
1005 * - Asked to verify (!td_rw(td))
1006 * - Or the io_u is from our verify list (mixed write/ver)
1008 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1009 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1011 if (verify_state_should_stop(td, io_u)) {
1016 if (td->o.verify_async)
1017 io_u->end_io = verify_io_u_async;
1019 io_u->end_io = verify_io_u;
1020 td_set_runstate(td, TD_VERIFYING);
1021 } else if (in_ramp_time(td))
1022 td_set_runstate(td, TD_RAMP);
1024 td_set_runstate(td, TD_RUNNING);
1027 * Always log IO before it's issued, so we know the specific
1028 * order of it. The logged unit will track when the IO has
1031 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1033 td->o.verify != VERIFY_NONE &&
1034 !td->o.experimental_verify)
1035 log_io_piece(td, io_u);
1037 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1038 const unsigned long long blen = io_u->xfer_buflen;
1039 const enum fio_ddir __ddir = acct_ddir(io_u);
1044 workqueue_enqueue(&td->io_wq, &io_u->work);
1047 if (ddir_rw(__ddir)) {
1048 td->io_issues[__ddir]++;
1049 td->io_issue_bytes[__ddir] += blen;
1050 td->rate_io_issue_bytes[__ddir] += blen;
1053 if (should_check_rate(td))
1054 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1057 ret = io_u_submit(td, io_u);
1059 if (should_check_rate(td))
1060 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1062 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1066 * See if we need to complete some commands. Note that
1067 * we can get BUSY even without IO queued, if the
1068 * system is resource starved.
1071 full = queue_full(td) ||
1072 (ret == FIO_Q_BUSY && td->cur_depth);
1073 if (full || io_in_polling(td))
1074 ret = wait_for_completions(td, &comp_time);
1078 if (!ddir_rw_sum(td->bytes_done) &&
1079 !td_ioengine_flagged(td, FIO_NOIO))
1082 if (!in_ramp_time(td) && should_check_rate(td)) {
1083 if (check_min_rate(td, &comp_time)) {
1084 if (exitall_on_terminate || td->o.exitall_error)
1085 fio_terminate_threads(td->groupid, td->o.exit_what);
1086 td_verror(td, EIO, "check_min_rate");
1090 if (!in_ramp_time(td) && td->o.latency_target)
1091 lat_target_check(td);
1093 if (ddir_rw(ddir) && td->o.thinktime)
1094 handle_thinktime(td, ddir);
1097 check_update_rusage(td);
1099 if (td->trim_entries)
1100 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1102 if (td->o.fill_device && td->error == ENOSPC) {
1104 fio_mark_td_terminate(td);
1109 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1110 workqueue_flush(&td->io_wq);
1116 ret = io_u_queued_complete(td, i);
1117 if (td->o.fill_device && td->error == ENOSPC)
1121 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1122 td_set_runstate(td, TD_FSYNCING);
1124 for_each_file(td, f, i) {
1125 if (!fio_file_fsync(td, f))
1128 log_err("fio: end_fsync failed for file %s\n",
1133 cleanup_pending_aio(td);
1136 * stop job if we failed doing any IO
1138 if (!ddir_rw_sum(td->this_io_bytes))
1141 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1142 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1145 static void free_file_completion_logging(struct thread_data *td)
1150 for_each_file(td, f, i) {
1151 if (!f->last_write_comp)
1153 sfree(f->last_write_comp);
1157 static int init_file_completion_logging(struct thread_data *td,
1163 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1166 for_each_file(td, f, i) {
1167 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1168 if (!f->last_write_comp)
1175 free_file_completion_logging(td);
1176 log_err("fio: failed to alloc write comp data\n");
1180 static void cleanup_io_u(struct thread_data *td)
1184 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1186 if (td->io_ops->io_u_free)
1187 td->io_ops->io_u_free(td, io_u);
1189 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1194 io_u_rexit(&td->io_u_requeues);
1195 io_u_qexit(&td->io_u_freelist, false);
1196 io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1198 free_file_completion_logging(td);
1201 static int init_io_u(struct thread_data *td)
1204 int cl_align, i, max_units;
1207 max_units = td->o.iodepth;
1210 err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1211 err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1212 err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1215 log_err("fio: failed setting up IO queues\n");
1219 cl_align = os_cache_line_size();
1221 for (i = 0; i < max_units; i++) {
1227 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1229 log_err("fio: unable to allocate aligned memory\n");
1234 memset(io_u, 0, sizeof(*io_u));
1235 INIT_FLIST_HEAD(&io_u->verify_list);
1236 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1239 io_u->flags = IO_U_F_FREE;
1240 io_u_qpush(&td->io_u_freelist, io_u);
1243 * io_u never leaves this stack, used for iteration of all
1246 io_u_qpush(&td->io_u_all, io_u);
1248 if (td->io_ops->io_u_init) {
1249 int ret = td->io_ops->io_u_init(td, io_u);
1252 log_err("fio: failed to init engine data: %d\n", ret);
1258 init_io_u_buffers(td);
1260 if (init_file_completion_logging(td, max_units))
1266 int init_io_u_buffers(struct thread_data *td)
1269 unsigned long long max_bs, min_write;
1274 max_units = td->o.iodepth;
1275 max_bs = td_max_bs(td);
1276 min_write = td->o.min_bs[DDIR_WRITE];
1277 td->orig_buffer_size = (unsigned long long) max_bs
1278 * (unsigned long long) max_units;
1280 if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1284 * if we may later need to do address alignment, then add any
1285 * possible adjustment here so that we don't cause a buffer
1286 * overflow later. this adjustment may be too much if we get
1287 * lucky and the allocator gives us an aligned address.
1289 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1290 td_ioengine_flagged(td, FIO_RAWIO))
1291 td->orig_buffer_size += page_mask + td->o.mem_align;
1293 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1294 unsigned long long bs;
1296 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1297 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1300 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1301 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1305 if (data_xfer && allocate_io_mem(td))
1308 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1309 td_ioengine_flagged(td, FIO_RAWIO))
1310 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1312 p = td->orig_buffer;
1314 for (i = 0; i < max_units; i++) {
1315 io_u = td->io_u_all.io_us[i];
1316 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1320 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1323 io_u_fill_buffer(td, io_u, min_write, max_bs);
1324 if (td_write(td) && td->o.verify_pattern_bytes) {
1326 * Fill the buffer with the pattern if we are
1327 * going to be doing writes.
1329 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1339 * This function is Linux specific.
1340 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1342 static int switch_ioscheduler(struct thread_data *td)
1344 #ifdef FIO_HAVE_IOSCHED_SWITCH
1345 char tmp[256], tmp2[128], *p;
1349 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1352 assert(td->files && td->files[0]);
1353 sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1355 f = fopen(tmp, "r+");
1357 if (errno == ENOENT) {
1358 log_err("fio: os or kernel doesn't support IO scheduler"
1362 td_verror(td, errno, "fopen iosched");
1369 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1370 if (ferror(f) || ret != 1) {
1371 td_verror(td, errno, "fwrite");
1379 * Read back and check that the selected scheduler is now the default.
1381 ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1382 if (ferror(f) || ret < 0) {
1383 td_verror(td, errno, "fread");
1389 * either a list of io schedulers or "none\n" is expected. Strip the
1396 * Write to "none" entry doesn't fail, so check the result here.
1398 if (!strcmp(tmp, "none")) {
1399 log_err("fio: io scheduler is not tunable\n");
1404 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1405 if (!strstr(tmp, tmp2)) {
1406 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1407 td_verror(td, EINVAL, "iosched_switch");
1419 static bool keep_running(struct thread_data *td)
1421 unsigned long long limit;
1427 if (td->o.time_based)
1433 if (exceeds_number_ios(td))
1437 limit = td->o.io_size;
1441 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1445 * If the difference is less than the maximum IO size, we
1448 diff = limit - ddir_rw_sum(td->io_bytes);
1449 if (diff < td_max_bs(td))
1452 if (fio_files_done(td) && !td->o.io_size)
1461 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1463 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 13 + 1;
1467 str = malloc(newlen);
1468 sprintf(str, "%s > %s.%s.txt 2>&1", string, o->name, mode);
1470 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1473 log_err("fio: exec of cmd <%s> failed\n", str);
1480 * Dry run to compute correct state of numberio for verification.
1482 static uint64_t do_dry_run(struct thread_data *td)
1484 td_set_runstate(td, TD_RUNNING);
1486 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1487 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1491 if (td->terminate || td->done)
1494 io_u = get_io_u(td);
1495 if (IS_ERR_OR_NULL(io_u))
1498 io_u_set(td, io_u, IO_U_F_FLIGHT);
1501 if (ddir_rw(acct_ddir(io_u)))
1502 td->io_issues[acct_ddir(io_u)]++;
1503 if (ddir_rw(io_u->ddir)) {
1504 io_u_mark_depth(td, 1);
1505 td->ts.total_io_u[io_u->ddir]++;
1508 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1510 td->o.verify != VERIFY_NONE &&
1511 !td->o.experimental_verify)
1512 log_io_piece(td, io_u);
1514 ret = io_u_sync_complete(td, io_u);
1518 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1522 struct thread_data *td;
1523 struct sk_out *sk_out;
1527 * Entry point for the thread based jobs. The process based jobs end up
1528 * here as well, after a little setup.
1530 static void *thread_main(void *data)
1532 struct fork_data *fd = data;
1533 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1534 struct thread_data *td = fd->td;
1535 struct thread_options *o = &td->o;
1536 struct sk_out *sk_out = fd->sk_out;
1537 uint64_t bytes_done[DDIR_RWDIR_CNT];
1538 int deadlock_loop_cnt;
1542 sk_out_assign(sk_out);
1545 if (!o->use_thread) {
1551 fio_local_clock_init();
1553 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1556 fio_server_send_start(td);
1558 INIT_FLIST_HEAD(&td->io_log_list);
1559 INIT_FLIST_HEAD(&td->io_hist_list);
1560 INIT_FLIST_HEAD(&td->verify_list);
1561 INIT_FLIST_HEAD(&td->trim_list);
1562 td->io_hist_tree = RB_ROOT;
1564 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1566 td_verror(td, ret, "mutex_cond_init_pshared");
1569 ret = cond_init_pshared(&td->verify_cond);
1571 td_verror(td, ret, "mutex_cond_pshared");
1575 td_set_runstate(td, TD_INITIALIZED);
1576 dprint(FD_MUTEX, "up startup_sem\n");
1577 fio_sem_up(startup_sem);
1578 dprint(FD_MUTEX, "wait on td->sem\n");
1579 fio_sem_down(td->sem);
1580 dprint(FD_MUTEX, "done waiting on td->sem\n");
1583 * A new gid requires privilege, so we need to do this before setting
1586 if (o->gid != -1U && setgid(o->gid)) {
1587 td_verror(td, errno, "setgid");
1590 if (o->uid != -1U && setuid(o->uid)) {
1591 td_verror(td, errno, "setuid");
1595 td_zone_gen_index(td);
1598 * Do this early, we don't want the compress threads to be limited
1599 * to the same CPUs as the IO workers. So do this before we set
1600 * any potential CPU affinity
1602 if (iolog_compress_init(td, sk_out))
1606 * If we have a gettimeofday() thread, make sure we exclude that
1607 * thread from this job
1610 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1613 * Set affinity first, in case it has an impact on the memory
1616 if (fio_option_is_set(o, cpumask)) {
1617 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1618 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1620 log_err("fio: no CPUs set\n");
1621 log_err("fio: Try increasing number of available CPUs\n");
1622 td_verror(td, EINVAL, "cpus_split");
1626 ret = fio_setaffinity(td->pid, o->cpumask);
1628 td_verror(td, errno, "cpu_set_affinity");
1633 #ifdef CONFIG_LIBNUMA
1634 /* numa node setup */
1635 if (fio_option_is_set(o, numa_cpunodes) ||
1636 fio_option_is_set(o, numa_memnodes)) {
1637 struct bitmask *mask;
1639 if (numa_available() < 0) {
1640 td_verror(td, errno, "Does not support NUMA API\n");
1644 if (fio_option_is_set(o, numa_cpunodes)) {
1645 mask = numa_parse_nodestring(o->numa_cpunodes);
1646 ret = numa_run_on_node_mask(mask);
1647 numa_free_nodemask(mask);
1649 td_verror(td, errno, \
1650 "numa_run_on_node_mask failed\n");
1655 if (fio_option_is_set(o, numa_memnodes)) {
1657 if (o->numa_memnodes)
1658 mask = numa_parse_nodestring(o->numa_memnodes);
1660 switch (o->numa_mem_mode) {
1661 case MPOL_INTERLEAVE:
1662 numa_set_interleave_mask(mask);
1665 numa_set_membind(mask);
1668 numa_set_localalloc();
1670 case MPOL_PREFERRED:
1671 numa_set_preferred(o->numa_mem_prefer_node);
1679 numa_free_nodemask(mask);
1685 if (fio_pin_memory(td))
1689 * May alter parameters that init_io_u() will use, so we need to
1692 if (!init_iolog(td))
1701 if (td->io_ops->post_init && td->io_ops->post_init(td))
1704 if (o->verify_async && verify_async_init(td))
1707 if (fio_option_is_set(o, ioprio) ||
1708 fio_option_is_set(o, ioprio_class)) {
1709 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1711 td_verror(td, errno, "ioprio_set");
1716 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1720 if (nice(o->nice) == -1 && errno != 0) {
1721 td_verror(td, errno, "nice");
1725 if (o->ioscheduler && switch_ioscheduler(td))
1728 if (!o->create_serialize && setup_files(td))
1731 if (!init_random_map(td))
1734 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1737 if (o->pre_read && !pre_read_files(td))
1740 fio_verify_init(td);
1742 if (rate_submit_init(td, sk_out))
1745 set_epoch_time(td, o->log_unix_epoch);
1746 fio_getrusage(&td->ru_start);
1747 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1748 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1749 memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1751 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1752 o->ratemin[DDIR_TRIM]) {
1753 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1754 sizeof(td->bw_sample_time));
1755 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1756 sizeof(td->bw_sample_time));
1757 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1758 sizeof(td->bw_sample_time));
1761 memset(bytes_done, 0, sizeof(bytes_done));
1762 clear_state = false;
1764 while (keep_running(td)) {
1765 uint64_t verify_bytes;
1767 fio_gettime(&td->start, NULL);
1768 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1771 clear_io_state(td, 0);
1773 if (o->unlink_each_loop && unlink_all_files(td))
1777 prune_io_piece_log(td);
1779 if (td->o.verify_only && td_write(td))
1780 verify_bytes = do_dry_run(td);
1782 do_io(td, bytes_done);
1784 if (!ddir_rw_sum(bytes_done)) {
1785 fio_mark_td_terminate(td);
1788 verify_bytes = bytes_done[DDIR_WRITE] +
1789 bytes_done[DDIR_TRIM];
1794 * If we took too long to shut down, the main thread could
1795 * already consider us reaped/exited. If that happens, break
1798 if (td->runstate >= TD_EXITED)
1804 * Make sure we've successfully updated the rusage stats
1805 * before waiting on the stat mutex. Otherwise we could have
1806 * the stat thread holding stat mutex and waiting for
1807 * the rusage_sem, which would never get upped because
1808 * this thread is waiting for the stat mutex.
1810 deadlock_loop_cnt = 0;
1812 check_update_rusage(td);
1813 if (!fio_sem_down_trylock(stat_sem))
1816 if (deadlock_loop_cnt++ > 5000) {
1817 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1818 td->error = EDEADLK;
1823 if (td_read(td) && td->io_bytes[DDIR_READ])
1824 update_runtime(td, elapsed_us, DDIR_READ);
1825 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1826 update_runtime(td, elapsed_us, DDIR_WRITE);
1827 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1828 update_runtime(td, elapsed_us, DDIR_TRIM);
1829 fio_gettime(&td->start, NULL);
1830 fio_sem_up(stat_sem);
1832 if (td->error || td->terminate)
1835 if (!o->do_verify ||
1836 o->verify == VERIFY_NONE ||
1837 td_ioengine_flagged(td, FIO_UNIDIR))
1840 clear_io_state(td, 0);
1842 fio_gettime(&td->start, NULL);
1844 do_verify(td, verify_bytes);
1847 * See comment further up for why this is done here.
1849 check_update_rusage(td);
1851 fio_sem_down(stat_sem);
1852 update_runtime(td, elapsed_us, DDIR_READ);
1853 fio_gettime(&td->start, NULL);
1854 fio_sem_up(stat_sem);
1856 if (td->error || td->terminate)
1861 * Acquire this lock if we were doing overlap checking in
1862 * offload mode so that we don't clean up this job while
1863 * another thread is checking its io_u's for overlap
1865 if (td_offload_overlap(td)) {
1866 int res = pthread_mutex_lock(&overlap_check);
1869 td_set_runstate(td, TD_FINISHING);
1870 if (td_offload_overlap(td)) {
1871 res = pthread_mutex_unlock(&overlap_check);
1875 update_rusage_stat(td);
1876 td->ts.total_run_time = mtime_since_now(&td->epoch);
1877 for_each_rw_ddir(ddir) {
1878 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
1881 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1882 (td->o.verify != VERIFY_NONE && td_write(td)))
1883 verify_save_state(td->thread_number);
1885 fio_unpin_memory(td);
1887 td_writeout_logs(td, true);
1889 iolog_compress_exit(td);
1890 rate_submit_exit(td);
1892 if (o->exec_postrun)
1893 exec_string(o, o->exec_postrun, (const char *)"postrun");
1895 if (exitall_on_terminate || (o->exitall_error && td->error))
1896 fio_terminate_threads(td->groupid, td->o.exit_what);
1900 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1903 if (o->verify_async)
1904 verify_async_exit(td);
1906 close_and_free_files(td);
1909 cgroup_shutdown(td, cgroup_mnt);
1910 verify_free_state(td);
1911 td_zone_free_index(td);
1913 if (fio_option_is_set(o, cpumask)) {
1914 ret = fio_cpuset_exit(&o->cpumask);
1916 td_verror(td, ret, "fio_cpuset_exit");
1920 * do this very late, it will log file closing as well
1922 if (o->write_iolog_file)
1923 write_iolog_close(td);
1924 if (td->io_log_rfile)
1925 fclose(td->io_log_rfile);
1927 td_set_runstate(td, TD_EXITED);
1930 * Do this last after setting our runstate to exited, so we
1931 * know that the stat thread is signaled.
1933 check_update_rusage(td);
1936 return (void *) (uintptr_t) td->error;
1940 * Run over the job map and reap the threads that have exited, if any.
1942 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1945 struct thread_data *td;
1946 unsigned int cputhreads, realthreads, pending;
1950 * reap exited threads (TD_EXITED -> TD_REAPED)
1952 realthreads = pending = cputhreads = 0;
1953 for_each_td(td, i) {
1956 if (!strcmp(td->o.ioengine, "cpuio"))
1965 if (td->runstate == TD_REAPED)
1967 if (td->o.use_thread) {
1968 if (td->runstate == TD_EXITED) {
1969 td_set_runstate(td, TD_REAPED);
1976 if (td->runstate == TD_EXITED)
1980 * check if someone quit or got killed in an unusual way
1982 ret = waitpid(td->pid, &status, flags);
1984 if (errno == ECHILD) {
1985 log_err("fio: pid=%d disappeared %d\n",
1986 (int) td->pid, td->runstate);
1988 td_set_runstate(td, TD_REAPED);
1992 } else if (ret == td->pid) {
1993 if (WIFSIGNALED(status)) {
1994 int sig = WTERMSIG(status);
1996 if (sig != SIGTERM && sig != SIGUSR2)
1997 log_err("fio: pid=%d, got signal=%d\n",
1998 (int) td->pid, sig);
2000 td_set_runstate(td, TD_REAPED);
2003 if (WIFEXITED(status)) {
2004 if (WEXITSTATUS(status) && !td->error)
2005 td->error = WEXITSTATUS(status);
2007 td_set_runstate(td, TD_REAPED);
2013 * If the job is stuck, do a forceful timeout of it and
2016 if (td->terminate &&
2017 td->runstate < TD_FSYNCING &&
2018 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2019 log_err("fio: job '%s' (state=%d) hasn't exited in "
2020 "%lu seconds, it appears to be stuck. Doing "
2021 "forceful exit of this job.\n",
2022 td->o.name, td->runstate,
2023 (unsigned long) time_since_now(&td->terminate_time));
2024 td_set_runstate(td, TD_REAPED);
2029 * thread is not dead, continue
2035 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2036 (*t_rate) -= ddir_rw_sum(td->o.rate);
2043 done_secs += mtime_since_now(&td->epoch) / 1000;
2044 profile_td_exit(td);
2047 if (*nr_running == cputhreads && !pending && realthreads)
2048 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2051 static bool __check_trigger_file(void)
2058 if (stat(trigger_file, &sb))
2061 if (unlink(trigger_file) < 0)
2062 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2068 static bool trigger_timedout(void)
2070 if (trigger_timeout)
2071 if (time_since_genesis() >= trigger_timeout) {
2072 trigger_timeout = 0;
2079 void exec_trigger(const char *cmd)
2083 if (!cmd || cmd[0] == '\0')
2088 log_err("fio: failed executing %s trigger\n", cmd);
2091 void check_trigger_file(void)
2093 if (__check_trigger_file() || trigger_timedout()) {
2095 fio_clients_send_trigger(trigger_remote_cmd);
2097 verify_save_state(IO_LIST_ALL);
2098 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2099 exec_trigger(trigger_cmd);
2104 static int fio_verify_load_state(struct thread_data *td)
2108 if (!td->o.verify_state)
2114 ret = fio_server_get_verify_state(td->o.name,
2115 td->thread_number - 1, &data);
2117 verify_assign_state(td, data);
2119 char prefix[PATH_MAX];
2122 sprintf(prefix, "%s%clocal", aux_path,
2123 FIO_OS_PATH_SEPARATOR);
2125 strcpy(prefix, "local");
2126 ret = verify_load_state(td, prefix);
2132 static void do_usleep(unsigned int usecs)
2134 check_for_running_stats();
2135 check_trigger_file();
2139 static bool check_mount_writes(struct thread_data *td)
2144 if (!td_write(td) || td->o.allow_mounted_write)
2148 * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2149 * are mkfs'd and mounted.
2151 for_each_file(td, f, i) {
2152 #ifdef FIO_HAVE_CHARDEV_SIZE
2153 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2155 if (f->filetype != FIO_TYPE_BLOCK)
2158 if (device_is_mounted(f->file_name))
2164 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2168 static bool waitee_running(struct thread_data *me)
2170 const char *waitee = me->o.wait_for;
2171 const char *self = me->o.name;
2172 struct thread_data *td;
2178 for_each_td(td, i) {
2179 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2182 if (td->runstate < TD_EXITED) {
2183 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2185 runstate_to_name(td->runstate));
2190 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2195 * Main function for kicking off and reaping jobs, as needed.
2197 static void run_threads(struct sk_out *sk_out)
2199 struct thread_data *td;
2200 unsigned int i, todo, nr_running, nr_started;
2201 uint64_t m_rate, t_rate;
2204 if (fio_gtod_offload && fio_start_gtod_thread())
2207 fio_idle_prof_init();
2211 nr_thread = nr_process = 0;
2212 for_each_td(td, i) {
2213 if (check_mount_writes(td))
2215 if (td->o.use_thread)
2221 if (output_format & FIO_OUTPUT_NORMAL) {
2222 struct buf_output out;
2224 buf_output_init(&out);
2225 __log_buf(&out, "Starting ");
2227 __log_buf(&out, "%d thread%s", nr_thread,
2228 nr_thread > 1 ? "s" : "");
2231 __log_buf(&out, " and ");
2232 __log_buf(&out, "%d process%s", nr_process,
2233 nr_process > 1 ? "es" : "");
2235 __log_buf(&out, "\n");
2236 log_info_buf(out.buf, out.buflen);
2237 buf_output_free(&out);
2240 todo = thread_number;
2243 m_rate = t_rate = 0;
2245 for_each_td(td, i) {
2246 print_status_init(td->thread_number - 1);
2248 if (!td->o.create_serialize)
2251 if (fio_verify_load_state(td))
2255 * do file setup here so it happens sequentially,
2256 * we don't want X number of threads getting their
2257 * client data interspersed on disk
2259 if (setup_files(td)) {
2263 log_err("fio: pid=%d, err=%d/%s\n",
2264 (int) td->pid, td->error, td->verror);
2265 td_set_runstate(td, TD_REAPED);
2272 * for sharing to work, each job must always open
2273 * its own files. so close them, if we opened them
2276 for_each_file(td, f, j) {
2277 if (fio_file_open(f))
2278 td_io_close_file(td, f);
2283 /* start idle threads before io threads start to run */
2284 fio_idle_prof_start();
2289 struct thread_data *map[REAL_MAX_JOBS];
2290 struct timespec this_start;
2291 int this_jobs = 0, left;
2292 struct fork_data *fd;
2295 * create threads (TD_NOT_CREATED -> TD_CREATED)
2297 for_each_td(td, i) {
2298 if (td->runstate != TD_NOT_CREATED)
2302 * never got a chance to start, killed by other
2303 * thread for some reason
2305 if (td->terminate) {
2310 if (td->o.start_delay) {
2311 spent = utime_since_genesis();
2313 if (td->o.start_delay > spent)
2317 if (td->o.stonewall && (nr_started || nr_running)) {
2318 dprint(FD_PROCESS, "%s: stonewall wait\n",
2323 if (waitee_running(td)) {
2324 dprint(FD_PROCESS, "%s: waiting for %s\n",
2325 td->o.name, td->o.wait_for);
2331 td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2332 td->update_rusage = 0;
2335 * Set state to created. Thread will transition
2336 * to TD_INITIALIZED when it's done setting up.
2338 td_set_runstate(td, TD_CREATED);
2339 map[this_jobs++] = td;
2342 fd = calloc(1, sizeof(*fd));
2344 fd->sk_out = sk_out;
2346 if (td->o.use_thread) {
2349 dprint(FD_PROCESS, "will pthread_create\n");
2350 ret = pthread_create(&td->thread, NULL,
2353 log_err("pthread_create: %s\n",
2360 ret = pthread_detach(td->thread);
2362 log_err("pthread_detach: %s",
2366 dprint(FD_PROCESS, "will fork\n");
2371 ret = (int)(uintptr_t)thread_main(fd);
2373 } else if (i == fio_debug_jobno)
2374 *fio_debug_jobp = pid;
2376 dprint(FD_MUTEX, "wait on startup_sem\n");
2377 if (fio_sem_down_timeout(startup_sem, 10000)) {
2378 log_err("fio: job startup hung? exiting.\n");
2379 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2385 dprint(FD_MUTEX, "done waiting on startup_sem\n");
2389 * Wait for the started threads to transition to
2392 fio_gettime(&this_start, NULL);
2394 while (left && !fio_abort) {
2395 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2400 for (i = 0; i < this_jobs; i++) {
2404 if (td->runstate == TD_INITIALIZED) {
2407 } else if (td->runstate >= TD_EXITED) {
2411 nr_running++; /* work-around... */
2417 log_err("fio: %d job%s failed to start\n", left,
2418 left > 1 ? "s" : "");
2419 for (i = 0; i < this_jobs; i++) {
2423 kill(td->pid, SIGTERM);
2429 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2431 for_each_td(td, i) {
2432 if (td->runstate != TD_INITIALIZED)
2435 if (in_ramp_time(td))
2436 td_set_runstate(td, TD_RAMP);
2438 td_set_runstate(td, TD_RUNNING);
2441 m_rate += ddir_rw_sum(td->o.ratemin);
2442 t_rate += ddir_rw_sum(td->o.rate);
2444 fio_sem_up(td->sem);
2447 reap_threads(&nr_running, &t_rate, &m_rate);
2453 while (nr_running) {
2454 reap_threads(&nr_running, &t_rate, &m_rate);
2458 fio_idle_prof_stop();
2463 static void free_disk_util(void)
2465 disk_util_prune_entries();
2466 helper_thread_destroy();
2469 int fio_backend(struct sk_out *sk_out)
2471 struct thread_data *td;
2475 if (load_profile(exec_profile))
2478 exec_profile = NULL;
2484 struct log_params p = {
2485 .log_type = IO_LOG_TYPE_BW,
2488 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2489 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2490 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2493 startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2495 is_local_backend = true;
2496 if (startup_sem == NULL)
2501 if (helper_thread_create(startup_sem, sk_out))
2502 log_err("fio: failed to create helper thread\n");
2504 cgroup_list = smalloc(sizeof(*cgroup_list));
2506 INIT_FLIST_HEAD(cgroup_list);
2508 run_threads(sk_out);
2510 helper_thread_exit();
2515 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2516 struct io_log *log = agg_io_log[i];
2518 flush_log(log, false);
2524 for_each_td(td, i) {
2525 steadystate_free(td);
2526 fio_options_free(td);
2527 if (td->rusage_sem) {
2528 fio_sem_remove(td->rusage_sem);
2529 td->rusage_sem = NULL;
2531 fio_sem_remove(td->sem);
2537 cgroup_kill(cgroup_list);
2541 fio_sem_remove(startup_sem);