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"
54 static struct fio_sem *startup_sem;
55 static struct flist_head *cgroup_list;
56 static struct cgroup_mnt *cgroup_mnt;
57 static int exit_value;
58 static volatile bool fio_abort;
59 static unsigned int nr_process = 0;
60 static unsigned int nr_thread = 0;
62 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
65 unsigned int thread_number = 0;
66 unsigned int nr_segments = 0;
67 unsigned int cur_segment = 0;
68 unsigned int stat_number = 0;
70 unsigned long done_secs = 0;
71 #ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP
72 pthread_mutex_t overlap_check = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
74 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
77 #define JOB_START_TIMEOUT (5 * 1000)
79 static void sig_int(int sig)
83 fio_server_got_signal(sig);
85 log_info("\nfio: terminating on signal %d\n", sig);
90 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
95 static void sig_break(int sig)
100 * Windows terminates all job processes on SIGBREAK after the handler
101 * returns, so give them time to wrap-up and give stats
104 while (td->runstate < TD_EXITED)
110 void sig_show_status(int sig)
112 show_running_run_stats();
115 static void set_sig_handlers(void)
117 struct sigaction act;
119 memset(&act, 0, sizeof(act));
120 act.sa_handler = sig_int;
121 act.sa_flags = SA_RESTART;
122 sigaction(SIGINT, &act, NULL);
124 memset(&act, 0, sizeof(act));
125 act.sa_handler = sig_int;
126 act.sa_flags = SA_RESTART;
127 sigaction(SIGTERM, &act, NULL);
129 /* Windows uses SIGBREAK as a quit signal from other applications */
131 memset(&act, 0, sizeof(act));
132 act.sa_handler = sig_break;
133 act.sa_flags = SA_RESTART;
134 sigaction(SIGBREAK, &act, NULL);
137 memset(&act, 0, sizeof(act));
138 act.sa_handler = sig_show_status;
139 act.sa_flags = SA_RESTART;
140 sigaction(SIGUSR1, &act, NULL);
143 memset(&act, 0, sizeof(act));
144 act.sa_handler = sig_int;
145 act.sa_flags = SA_RESTART;
146 sigaction(SIGPIPE, &act, NULL);
151 * Check if we are above the minimum rate given.
153 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
156 unsigned long long current_rate_check_bytes = td->this_io_bytes[ddir];
157 unsigned long current_rate_check_blocks = td->this_io_blocks[ddir];
158 unsigned long long option_rate_bytes_min = td->o.ratemin[ddir];
159 unsigned int option_rate_iops_min = td->o.rate_iops_min[ddir];
161 assert(ddir_rw(ddir));
163 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
167 * allow a 2 second settle period in the beginning
169 if (mtime_since(&td->start, now) < 2000)
173 * if last_rate_check_blocks or last_rate_check_bytes is set,
174 * we can compute a rate per ratecycle
176 if (td->last_rate_check_bytes[ddir] || td->last_rate_check_blocks[ddir]) {
177 unsigned long spent = mtime_since(&td->last_rate_check_time[ddir], now);
178 if (spent < td->o.ratecycle || spent==0)
181 if (td->o.ratemin[ddir]) {
183 * check bandwidth specified rate
185 unsigned long long current_rate_bytes =
186 ((current_rate_check_bytes - td->last_rate_check_bytes[ddir]) * 1000) / spent;
187 if (current_rate_bytes < option_rate_bytes_min) {
188 log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
189 td->o.name, option_rate_bytes_min, current_rate_bytes);
194 * checks iops specified rate
196 unsigned long long current_rate_iops =
197 ((current_rate_check_blocks - td->last_rate_check_blocks[ddir]) * 1000) / spent;
199 if (current_rate_iops < option_rate_iops_min) {
200 log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
201 td->o.name, option_rate_iops_min, current_rate_iops);
207 td->last_rate_check_bytes[ddir] = current_rate_check_bytes;
208 td->last_rate_check_blocks[ddir] = current_rate_check_blocks;
209 memcpy(&td->last_rate_check_time[ddir], now, sizeof(*now));
213 static bool check_min_rate(struct thread_data *td, struct timespec *now)
217 for_each_rw_ddir(ddir) {
218 if (td->bytes_done[ddir])
219 ret |= __check_min_rate(td, now, ddir);
226 * When job exits, we can cancel the in-flight IO if we are using async
227 * io. Attempt to do so.
229 static void cleanup_pending_aio(struct thread_data *td)
234 * get immediately available events, if any
236 r = io_u_queued_complete(td, 0);
239 * now cancel remaining active events
241 if (td->io_ops->cancel) {
245 io_u_qiter(&td->io_u_all, io_u, i) {
246 if (io_u->flags & IO_U_F_FLIGHT) {
247 r = td->io_ops->cancel(td, io_u);
255 r = io_u_queued_complete(td, td->cur_depth);
259 * Helper to handle the final sync of a file. Works just like the normal
260 * io path, just does everything sync.
262 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
264 struct io_u *io_u = __get_io_u(td);
265 enum fio_q_status ret;
270 io_u->ddir = DDIR_SYNC;
272 io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
274 if (td_io_prep(td, io_u)) {
280 ret = td_io_queue(td, io_u);
284 if (io_u_queued_complete(td, 1) < 0)
287 case FIO_Q_COMPLETED:
289 td_verror(td, io_u->error, "td_io_queue");
293 if (io_u_sync_complete(td, io_u) < 0)
304 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
308 if (fio_file_open(f))
309 return fio_io_sync(td, f);
311 if (td_io_open_file(td, f))
314 ret = fio_io_sync(td, f);
316 if (fio_file_open(f))
317 ret2 = td_io_close_file(td, f);
318 return (ret || ret2);
321 static inline void __update_ts_cache(struct thread_data *td)
323 fio_gettime(&td->ts_cache, NULL);
326 static inline void update_ts_cache(struct thread_data *td)
328 if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
329 __update_ts_cache(td);
332 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
334 if (in_ramp_time(td))
338 if (utime_since(&td->epoch, t) >= td->o.timeout)
345 * We need to update the runtime consistently in ms, but keep a running
346 * tally of the current elapsed time in microseconds for sub millisecond
349 static inline void update_runtime(struct thread_data *td,
350 unsigned long long *elapsed_us,
351 const enum fio_ddir ddir)
353 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
356 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
357 elapsed_us[ddir] += utime_since_now(&td->start);
358 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
361 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
366 if (ret < 0 || td->error) {
368 enum error_type_bit eb;
373 eb = td_error_type(ddir, err);
374 if (!(td->o.continue_on_error & (1 << eb)))
377 if (td_non_fatal_error(td, eb, err)) {
379 * Continue with the I/Os in case of
382 update_error_count(td, err);
386 } else if (td->o.fill_device && (err == ENOSPC || err == EDQUOT)) {
388 * We expect to hit this error if
389 * fill_device option is set.
392 fio_mark_td_terminate(td);
396 * Stop the I/O in case of a fatal
399 update_error_count(td, err);
407 static void check_update_rusage(struct thread_data *td)
409 if (td->update_rusage) {
410 td->update_rusage = 0;
411 update_rusage_stat(td);
412 fio_sem_up(td->rusage_sem);
416 static int wait_for_completions(struct thread_data *td, struct timespec *time)
418 const int full = queue_full(td);
422 if (td->flags & TD_F_REGROW_LOGS)
423 return io_u_quiesce(td);
426 * if the queue is full, we MUST reap at least 1 event
428 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
429 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
432 if (time && should_check_rate(td))
433 fio_gettime(time, NULL);
436 ret = io_u_queued_complete(td, min_evts);
439 } while (full && (td->cur_depth > td->o.iodepth_low));
444 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
445 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
446 struct timespec *comp_time)
449 case FIO_Q_COMPLETED:
452 clear_io_u(td, io_u);
453 } else if (io_u->resid) {
454 long long bytes = io_u->xfer_buflen - io_u->resid;
455 struct fio_file *f = io_u->file;
458 *bytes_issued += bytes;
468 unlog_io_piece(td, io_u);
469 td_verror(td, EIO, "full resid");
470 clear_io_u(td, io_u);
474 io_u->xfer_buflen = io_u->resid;
475 io_u->xfer_buf += bytes;
476 io_u->offset += bytes;
478 if (ddir_rw(io_u->ddir))
479 td->ts.short_io_u[io_u->ddir]++;
481 if (io_u->offset == f->real_file_size)
484 requeue_io_u(td, &io_u);
487 if (comp_time && should_check_rate(td))
488 fio_gettime(comp_time, NULL);
490 *ret = io_u_sync_complete(td, io_u);
495 if (td->flags & TD_F_REGROW_LOGS)
499 * when doing I/O (not when verifying),
500 * check for any errors that are to be ignored
508 * if the engine doesn't have a commit hook,
509 * the io_u is really queued. if it does have such
510 * a hook, it has to call io_u_queued() itself.
512 if (td->io_ops->commit == NULL)
513 io_u_queued(td, io_u);
515 *bytes_issued += io_u->xfer_buflen;
519 unlog_io_piece(td, io_u);
520 requeue_io_u(td, &io_u);
525 td_verror(td, -(*ret), "td_io_queue");
529 if (break_on_this_error(td, ddir, ret))
535 static inline bool io_in_polling(struct thread_data *td)
537 return !td->o.iodepth_batch_complete_min &&
538 !td->o.iodepth_batch_complete_max;
541 * Unlinks files from thread data fio_file structure
543 static int unlink_all_files(struct thread_data *td)
549 for_each_file(td, f, i) {
550 if (f->filetype != FIO_TYPE_FILE)
552 ret = td_io_unlink_file(td, f);
558 td_verror(td, ret, "unlink_all_files");
564 * Check if io_u will overlap an in-flight IO in the queue
566 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
569 struct io_u *check_io_u;
570 unsigned long long x1, x2, y1, y2;
574 x2 = io_u->offset + io_u->buflen;
576 io_u_qiter(q, check_io_u, i) {
577 if (check_io_u->flags & IO_U_F_FLIGHT) {
578 y1 = check_io_u->offset;
579 y2 = check_io_u->offset + check_io_u->buflen;
581 if (x1 < y2 && y1 < x2) {
583 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
585 y1, check_io_u->buflen);
594 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
597 * Check for overlap if the user asked us to, and we have
598 * at least one IO in flight besides this one.
600 if (td->o.serialize_overlap && td->cur_depth > 1 &&
601 in_flight_overlap(&td->io_u_all, io_u))
604 return td_io_queue(td, io_u);
608 * The main verify engine. Runs over the writes we previously submitted,
609 * reads the blocks back in, and checks the crc/md5 of the data.
611 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
618 dprint(FD_VERIFY, "starting loop\n");
621 * sync io first and invalidate cache, to make sure we really
624 for_each_file(td, f, i) {
625 if (!fio_file_open(f))
627 if (fio_io_sync(td, f))
629 if (file_invalidate_cache(td, f))
633 check_update_rusage(td);
638 td_set_runstate(td, TD_VERIFYING);
641 while (!td->terminate) {
646 check_update_rusage(td);
648 if (runtime_exceeded(td, &td->ts_cache)) {
649 __update_ts_cache(td);
650 if (runtime_exceeded(td, &td->ts_cache)) {
651 fio_mark_td_terminate(td);
656 if (flow_threshold_exceeded(td))
659 if (!td->o.experimental_verify) {
660 io_u = __get_io_u(td);
664 if (get_next_verify(td, io_u)) {
669 if (td_io_prep(td, io_u)) {
674 if (td->bytes_verified + td->o.rw_min_bs > verify_bytes)
677 while ((io_u = get_io_u(td)) != NULL) {
678 if (IS_ERR_OR_NULL(io_u)) {
685 * We are only interested in the places where
686 * we wrote or trimmed IOs. Turn those into
687 * reads for verification purposes.
689 if (io_u->ddir == DDIR_READ) {
691 * Pretend we issued it for rwmix
694 td->io_issues[DDIR_READ]++;
697 } else if (io_u->ddir == DDIR_TRIM) {
698 io_u->ddir = DDIR_READ;
699 io_u_set(td, io_u, IO_U_F_TRIMMED);
701 } else if (io_u->ddir == DDIR_WRITE) {
702 io_u->ddir = DDIR_READ;
703 io_u->numberio = td->verify_read_issues;
704 td->verify_read_issues++;
705 populate_verify_io_u(td, io_u);
717 if (verify_state_should_stop(td, io_u)) {
722 if (td->o.verify_async)
723 io_u->end_io = verify_io_u_async;
725 io_u->end_io = verify_io_u;
728 if (!td->o.disable_slat)
729 fio_gettime(&io_u->start_time, NULL);
731 ret = io_u_submit(td, io_u);
733 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
737 * if we can queue more, do so. but check if there are
738 * completed io_u's first. Note that we can get BUSY even
739 * without IO queued, if the system is resource starved.
742 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
743 if (full || io_in_polling(td))
744 ret = wait_for_completions(td, NULL);
750 check_update_rusage(td);
753 min_events = td->cur_depth;
756 ret = io_u_queued_complete(td, min_events);
758 cleanup_pending_aio(td);
760 td_set_runstate(td, TD_RUNNING);
762 dprint(FD_VERIFY, "exiting loop\n");
765 static bool exceeds_number_ios(struct thread_data *td)
767 unsigned long long number_ios;
769 if (!td->o.number_ios)
772 number_ios = ddir_rw_sum(td->io_blocks);
773 number_ios += td->io_u_queued + td->io_u_in_flight;
775 return number_ios >= (td->o.number_ios * td->loops);
778 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
780 unsigned long long bytes, limit;
783 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
784 else if (td_write(td))
785 bytes = this_bytes[DDIR_WRITE];
786 else if (td_read(td))
787 bytes = this_bytes[DDIR_READ];
789 bytes = this_bytes[DDIR_TRIM];
792 limit = td->o.io_size;
797 return bytes >= limit || exceeds_number_ios(td);
800 static bool io_issue_bytes_exceeded(struct thread_data *td)
802 return io_bytes_exceeded(td, td->io_issue_bytes);
805 static bool io_complete_bytes_exceeded(struct thread_data *td)
807 return io_bytes_exceeded(td, td->this_io_bytes);
811 * used to calculate the next io time for rate control
814 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
816 uint64_t bps = td->rate_bps[ddir];
818 assert(!(td->flags & TD_F_CHILD));
820 if (td->o.rate_process == RATE_PROCESS_POISSON) {
823 iops = bps / td->o.min_bs[ddir];
824 val = (int64_t) (1000000 / iops) *
825 -logf(__rand_0_1(&td->poisson_state[ddir]));
827 dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
828 (unsigned long long) 1000000 / val,
831 td->last_usec[ddir] += val;
832 return td->last_usec[ddir];
834 uint64_t bytes = td->rate_io_issue_bytes[ddir];
835 uint64_t secs = bytes / bps;
836 uint64_t remainder = bytes % bps;
838 return remainder * 1000000 / bps + secs * 1000000;
844 static void init_thinktime(struct thread_data *td)
846 if (td->o.thinktime_blocks_type == THINKTIME_BLOCKS_TYPE_COMPLETE)
847 td->thinktime_blocks_counter = td->io_blocks;
849 td->thinktime_blocks_counter = td->io_issues;
850 td->last_thinktime = td->epoch;
851 td->last_thinktime_blocks = 0;
854 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir,
855 struct timespec *time)
857 unsigned long long b;
858 unsigned long long runtime_left;
864 if (td->o.thinktime_iotime) {
865 fio_gettime(&now, NULL);
866 if (utime_since(&td->last_thinktime, &now)
867 >= td->o.thinktime_iotime) {
869 } else if (!fio_option_is_set(&td->o, thinktime_blocks)) {
871 * When thinktime_iotime is set and thinktime_blocks is
872 * not set, skip the thinktime_blocks check, since
873 * thinktime_blocks default value 1 does not work
874 * together with thinktime_iotime.
881 b = ddir_rw_sum(td->thinktime_blocks_counter);
882 if (b >= td->last_thinktime_blocks + td->o.thinktime_blocks)
890 left = td->o.thinktime_spin;
892 runtime_left = td->o.timeout - utime_since_now(&td->epoch);
893 if (runtime_left < (unsigned long long)left)
899 total = usec_spin(left);
902 * usec_spin() might run for slightly longer than intended in a VM
903 * where the vCPU could get descheduled or the hypervisor could steal
904 * CPU time. Ensure "left" doesn't become negative.
906 if (total < td->o.thinktime)
907 left = td->o.thinktime - total;
912 runtime_left = td->o.timeout - utime_since_now(&td->epoch);
913 if (runtime_left < (unsigned long long)left)
918 total += usec_sleep(td, left);
921 * If we're ignoring thinktime for the rate, add the number of bytes
922 * we would have done while sleeping, minus one block to ensure we
923 * start issuing immediately after the sleep.
925 if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
926 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
927 uint64_t bs = td->o.min_bs[ddir];
928 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
931 if (usperop <= total)
934 over = (usperop - total) / usperop * -bs;
936 td->rate_io_issue_bytes[ddir] += (missed - over);
937 /* adjust for rate_process=poisson */
938 td->last_usec[ddir] += total;
941 if (time && should_check_rate(td))
942 fio_gettime(time, NULL);
944 td->last_thinktime_blocks = b;
945 if (td->o.thinktime_iotime) {
946 fio_gettime(&now, NULL);
947 td->last_thinktime = now;
952 * Main IO worker function. It retrieves io_u's to process and queues
953 * and reaps them, checking for rate and errors along the way.
955 * Returns number of bytes written and trimmed.
957 static void do_io(struct thread_data *td, uint64_t *bytes_done)
961 uint64_t total_bytes, bytes_issued = 0;
963 for (i = 0; i < DDIR_RWDIR_CNT; i++)
964 bytes_done[i] = td->bytes_done[i];
966 if (in_ramp_time(td))
967 td_set_runstate(td, TD_RAMP);
969 td_set_runstate(td, TD_RUNNING);
973 total_bytes = td->o.size;
975 * Allow random overwrite workloads to write up to io_size
976 * before starting verification phase as 'size' doesn't apply.
978 if (td_write(td) && td_random(td) && td->o.norandommap)
979 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
981 * If verify_backlog is enabled, we'll run the verify in this
982 * handler as well. For that case, we may need up to twice the
985 if (td->o.verify != VERIFY_NONE &&
986 (td_write(td) && td->o.verify_backlog))
987 total_bytes += td->o.size;
989 /* In trimwrite mode, each byte is trimmed and then written, so
990 * allow total_bytes or number of ios to be twice as big */
991 if (td_trimwrite(td)) {
992 total_bytes += td->total_io_size;
993 td->o.number_ios *= 2;
996 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
997 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
999 struct timespec comp_time;
1004 check_update_rusage(td);
1006 if (td->terminate || td->done)
1009 update_ts_cache(td);
1011 if (runtime_exceeded(td, &td->ts_cache)) {
1012 __update_ts_cache(td);
1013 if (runtime_exceeded(td, &td->ts_cache)) {
1014 fio_mark_td_terminate(td);
1019 if (flow_threshold_exceeded(td))
1023 * Break if we exceeded the bytes. The exception is time
1024 * based runs, but we still need to break out of the loop
1025 * for those to run verification, if enabled.
1026 * Jobs read from iolog do not use this stop condition.
1028 if (bytes_issued >= total_bytes &&
1029 !td->o.read_iolog_file &&
1030 (!td->o.time_based ||
1031 (td->o.time_based && td->o.verify != VERIFY_NONE)))
1034 io_u = get_io_u(td);
1035 if (IS_ERR_OR_NULL(io_u)) {
1036 int err = PTR_ERR(io_u);
1040 if (err == -EBUSY) {
1044 if (td->o.latency_target)
1049 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY) {
1050 if (!(io_u->flags & IO_U_F_PATTERN_DONE)) {
1051 io_u_set(td, io_u, IO_U_F_PATTERN_DONE);
1052 io_u->numberio = td->io_issues[io_u->ddir];
1053 populate_verify_io_u(td, io_u);
1060 * Add verification end_io handler if:
1061 * - Asked to verify (!td_rw(td))
1062 * - Or the io_u is from our verify list (mixed write/ver)
1064 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1065 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1067 if (verify_state_should_stop(td, io_u)) {
1072 if (td->o.verify_async)
1073 io_u->end_io = verify_io_u_async;
1075 io_u->end_io = verify_io_u;
1076 td_set_runstate(td, TD_VERIFYING);
1077 } else if (in_ramp_time(td))
1078 td_set_runstate(td, TD_RAMP);
1080 td_set_runstate(td, TD_RUNNING);
1083 * Always log IO before it's issued, so we know the specific
1084 * order of it. The logged unit will track when the IO has
1087 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1089 td->o.verify != VERIFY_NONE &&
1090 !td->o.experimental_verify)
1091 log_io_piece(td, io_u);
1093 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1094 const unsigned long long blen = io_u->xfer_buflen;
1095 const enum fio_ddir __ddir = acct_ddir(io_u);
1100 workqueue_enqueue(&td->io_wq, &io_u->work);
1103 if (ddir_rw(__ddir)) {
1104 td->io_issues[__ddir]++;
1105 td->io_issue_bytes[__ddir] += blen;
1106 td->rate_io_issue_bytes[__ddir] += blen;
1109 if (should_check_rate(td)) {
1110 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1111 fio_gettime(&comp_time, NULL);
1115 ret = io_u_submit(td, io_u);
1117 if (should_check_rate(td))
1118 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1120 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1124 * See if we need to complete some commands. Note that
1125 * we can get BUSY even without IO queued, if the
1126 * system is resource starved.
1129 full = queue_full(td) ||
1130 (ret == FIO_Q_BUSY && td->cur_depth);
1131 if (full || io_in_polling(td))
1132 ret = wait_for_completions(td, &comp_time);
1137 if (ddir_rw(ddir) && td->o.thinkcycles)
1138 cycles_spin(td->o.thinkcycles);
1140 if (ddir_rw(ddir) && td->o.thinktime)
1141 handle_thinktime(td, ddir, &comp_time);
1143 if (!ddir_rw_sum(td->bytes_done) &&
1144 !td_ioengine_flagged(td, FIO_NOIO))
1147 if (!in_ramp_time(td) && should_check_rate(td)) {
1148 if (check_min_rate(td, &comp_time)) {
1149 if (exitall_on_terminate || td->o.exitall_error)
1150 fio_terminate_threads(td->groupid, td->o.exit_what);
1151 td_verror(td, EIO, "check_min_rate");
1155 if (!in_ramp_time(td) && td->o.latency_target)
1156 lat_target_check(td);
1159 check_update_rusage(td);
1161 if (td->trim_entries)
1162 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1164 if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
1166 fio_mark_td_terminate(td);
1171 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1172 workqueue_flush(&td->io_wq);
1178 ret = io_u_queued_complete(td, i);
1179 if (td->o.fill_device &&
1180 (td->error == ENOSPC || td->error == EDQUOT))
1184 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1185 td_set_runstate(td, TD_FSYNCING);
1187 for_each_file(td, f, i) {
1188 if (!fio_file_fsync(td, f))
1191 log_err("fio: end_fsync failed for file %s\n",
1196 if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
1197 workqueue_flush(&td->io_wq);
1198 cleanup_pending_aio(td);
1202 * stop job if we failed doing any IO
1204 if (!ddir_rw_sum(td->this_io_bytes))
1207 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1208 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1211 static void free_file_completion_logging(struct thread_data *td)
1216 for_each_file(td, f, i) {
1217 if (!f->last_write_comp)
1219 sfree(f->last_write_comp);
1223 static int init_file_completion_logging(struct thread_data *td,
1229 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1232 for_each_file(td, f, i) {
1233 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1234 if (!f->last_write_comp)
1241 free_file_completion_logging(td);
1242 log_err("fio: failed to alloc write comp data\n");
1246 static void cleanup_io_u(struct thread_data *td)
1250 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1252 if (td->io_ops->io_u_free)
1253 td->io_ops->io_u_free(td, io_u);
1255 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1260 io_u_rexit(&td->io_u_requeues);
1261 io_u_qexit(&td->io_u_freelist, false);
1262 io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1264 free_file_completion_logging(td);
1267 static int init_io_u(struct thread_data *td)
1270 int cl_align, i, max_units;
1273 max_units = td->o.iodepth;
1276 err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1277 err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1278 err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1281 log_err("fio: failed setting up IO queues\n");
1285 cl_align = os_cache_line_size();
1287 for (i = 0; i < max_units; i++) {
1293 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1295 log_err("fio: unable to allocate aligned memory\n");
1300 memset(io_u, 0, sizeof(*io_u));
1301 INIT_FLIST_HEAD(&io_u->verify_list);
1302 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1305 io_u->flags = IO_U_F_FREE;
1306 io_u_qpush(&td->io_u_freelist, io_u);
1309 * io_u never leaves this stack, used for iteration of all
1312 io_u_qpush(&td->io_u_all, io_u);
1314 if (td->io_ops->io_u_init) {
1315 int ret = td->io_ops->io_u_init(td, io_u);
1318 log_err("fio: failed to init engine data: %d\n", ret);
1324 if (init_io_u_buffers(td))
1327 if (init_file_completion_logging(td, max_units))
1333 int init_io_u_buffers(struct thread_data *td)
1336 unsigned long long max_bs, min_write, trim_bs = 0;
1341 max_units = td->o.iodepth;
1342 max_bs = td_max_bs(td);
1343 min_write = td->o.min_bs[DDIR_WRITE];
1344 td->orig_buffer_size = (unsigned long long) max_bs
1345 * (unsigned long long) max_units;
1347 if (td_trim(td) && td->o.num_range > 1) {
1348 trim_bs = td->o.num_range * sizeof(struct trim_range);
1349 td->orig_buffer_size = trim_bs
1350 * (unsigned long long) max_units;
1354 * For reads, writes, and multi-range trim operations we need a
1357 if (td_ioengine_flagged(td, FIO_NOIO) ||
1358 !(td_read(td) || td_write(td) || (td_trim(td) && td->o.num_range > 1)))
1362 * if we may later need to do address alignment, then add any
1363 * possible adjustment here so that we don't cause a buffer
1364 * overflow later. this adjustment may be too much if we get
1365 * lucky and the allocator gives us an aligned address.
1367 if (td->o.odirect || td->o.mem_align ||
1368 td_ioengine_flagged(td, FIO_RAWIO))
1369 td->orig_buffer_size += page_mask + td->o.mem_align;
1371 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1372 unsigned long long bs;
1374 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1375 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1378 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1379 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1383 if (data_xfer && allocate_io_mem(td))
1386 if (td->o.odirect || td->o.mem_align ||
1387 td_ioengine_flagged(td, FIO_RAWIO))
1388 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1390 p = td->orig_buffer;
1392 for (i = 0; i < max_units; i++) {
1393 io_u = td->io_u_all.io_us[i];
1394 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1398 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1401 io_u_fill_buffer(td, io_u, min_write, max_bs);
1402 if (td_write(td) && td->o.verify_pattern_bytes) {
1404 * Fill the buffer with the pattern if we are
1405 * going to be doing writes.
1407 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1410 if (td_trim(td) && td->o.num_range > 1)
1419 #ifdef FIO_HAVE_IOSCHED_SWITCH
1421 * These functions are Linux specific.
1422 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1424 static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
1426 char tmp[256], tmp2[128], *p;
1430 assert(file->du && file->du->sysfs_root);
1431 sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);
1433 f = fopen(tmp, "r+");
1435 if (errno == ENOENT) {
1436 log_err("fio: os or kernel doesn't support IO scheduler"
1440 td_verror(td, errno, "fopen iosched");
1447 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1448 if (ferror(f) || ret != 1) {
1449 td_verror(td, errno, "fwrite");
1457 * Read back and check that the selected scheduler is now the default.
1459 ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1460 if (ferror(f) || ret < 0) {
1461 td_verror(td, errno, "fread");
1467 * either a list of io schedulers or "none\n" is expected. Strip the
1474 * Write to "none" entry doesn't fail, so check the result here.
1476 if (!strcmp(tmp, "none")) {
1477 log_err("fio: io scheduler is not tunable\n");
1482 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1483 if (!strstr(tmp, tmp2)) {
1484 log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
1485 td_verror(td, EINVAL, "iosched_switch");
1494 static int switch_ioscheduler(struct thread_data *td)
1500 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1503 assert(td->files && td->files[0]);
1505 for_each_file(td, f, i) {
1507 /* Only consider regular files and block device files */
1508 switch (f->filetype) {
1510 case FIO_TYPE_BLOCK:
1512 * Make sure that the device hosting the file could
1524 ret = set_ioscheduler(td, f);
1534 static int switch_ioscheduler(struct thread_data *td)
1539 #endif /* FIO_HAVE_IOSCHED_SWITCH */
1541 static bool keep_running(struct thread_data *td)
1543 unsigned long long limit;
1549 if (td->o.time_based)
1555 if (exceeds_number_ios(td))
1559 limit = td->o.io_size;
1563 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1567 * If the difference is less than the maximum IO size, we
1570 diff = limit - ddir_rw_sum(td->io_bytes);
1571 if (diff < td_max_bs(td))
1574 if (fio_files_done(td) && !td->o.io_size)
1583 static int exec_string(struct thread_options *o, const char *string,
1589 if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1592 log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1596 log_err("fio: exec of cmd <%s> failed\n", str);
1603 * Dry run to compute correct state of numberio for verification.
1605 static uint64_t do_dry_run(struct thread_data *td)
1607 td_set_runstate(td, TD_RUNNING);
1609 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1610 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1614 if (td->terminate || td->done)
1617 io_u = get_io_u(td);
1618 if (IS_ERR_OR_NULL(io_u))
1621 io_u_set(td, io_u, IO_U_F_FLIGHT);
1624 if (ddir_rw(acct_ddir(io_u)))
1625 td->io_issues[acct_ddir(io_u)]++;
1626 if (ddir_rw(io_u->ddir)) {
1627 io_u_mark_depth(td, 1);
1628 td->ts.total_io_u[io_u->ddir]++;
1631 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1633 td->o.verify != VERIFY_NONE &&
1634 !td->o.experimental_verify)
1635 log_io_piece(td, io_u);
1637 ret = io_u_sync_complete(td, io_u);
1641 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1645 struct thread_data *td;
1646 struct sk_out *sk_out;
1650 * Entry point for the thread based jobs. The process based jobs end up
1651 * here as well, after a little setup.
1653 static void *thread_main(void *data)
1655 struct fork_data *fd = data;
1656 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1657 struct thread_data *td = fd->td;
1658 struct thread_options *o = &td->o;
1659 struct sk_out *sk_out = fd->sk_out;
1660 uint64_t bytes_done[DDIR_RWDIR_CNT];
1661 int deadlock_loop_cnt;
1665 sk_out_assign(sk_out);
1668 if (!o->use_thread) {
1674 fio_local_clock_init();
1676 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1679 fio_server_send_start(td);
1681 INIT_FLIST_HEAD(&td->io_log_list);
1682 INIT_FLIST_HEAD(&td->io_hist_list);
1683 INIT_FLIST_HEAD(&td->verify_list);
1684 INIT_FLIST_HEAD(&td->trim_list);
1685 td->io_hist_tree = RB_ROOT;
1687 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1689 td_verror(td, ret, "mutex_cond_init_pshared");
1692 ret = cond_init_pshared(&td->verify_cond);
1694 td_verror(td, ret, "mutex_cond_pshared");
1698 td_set_runstate(td, TD_INITIALIZED);
1699 dprint(FD_MUTEX, "up startup_sem\n");
1700 fio_sem_up(startup_sem);
1701 dprint(FD_MUTEX, "wait on td->sem\n");
1702 fio_sem_down(td->sem);
1703 dprint(FD_MUTEX, "done waiting on td->sem\n");
1706 * A new gid requires privilege, so we need to do this before setting
1709 if (o->gid != -1U && setgid(o->gid)) {
1710 td_verror(td, errno, "setgid");
1713 if (o->uid != -1U && setuid(o->uid)) {
1714 td_verror(td, errno, "setuid");
1718 td_zone_gen_index(td);
1721 * Do this early, we don't want the compress threads to be limited
1722 * to the same CPUs as the IO workers. So do this before we set
1723 * any potential CPU affinity
1725 if (iolog_compress_init(td, sk_out))
1729 * If we have a gettimeofday() thread, make sure we exclude that
1730 * thread from this job
1733 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1736 * Set affinity first, in case it has an impact on the memory
1739 if (fio_option_is_set(o, cpumask)) {
1740 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1741 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1743 log_err("fio: no CPUs set\n");
1744 log_err("fio: Try increasing number of available CPUs\n");
1745 td_verror(td, EINVAL, "cpus_split");
1749 ret = fio_setaffinity(td->pid, o->cpumask);
1751 td_verror(td, errno, "cpu_set_affinity");
1756 #ifdef CONFIG_LIBNUMA
1757 /* numa node setup */
1758 if (fio_option_is_set(o, numa_cpunodes) ||
1759 fio_option_is_set(o, numa_memnodes)) {
1760 struct bitmask *mask;
1762 if (numa_available() < 0) {
1763 td_verror(td, errno, "Does not support NUMA API\n");
1767 if (fio_option_is_set(o, numa_cpunodes)) {
1768 mask = numa_parse_nodestring(o->numa_cpunodes);
1769 ret = numa_run_on_node_mask(mask);
1770 numa_free_nodemask(mask);
1772 td_verror(td, errno, \
1773 "numa_run_on_node_mask failed\n");
1778 if (fio_option_is_set(o, numa_memnodes)) {
1780 if (o->numa_memnodes)
1781 mask = numa_parse_nodestring(o->numa_memnodes);
1783 switch (o->numa_mem_mode) {
1784 case MPOL_INTERLEAVE:
1785 numa_set_interleave_mask(mask);
1788 numa_set_membind(mask);
1791 numa_set_localalloc();
1793 case MPOL_PREFERRED:
1794 numa_set_preferred(o->numa_mem_prefer_node);
1802 numa_free_nodemask(mask);
1808 if (fio_pin_memory(td))
1812 * May alter parameters that init_io_u() will use, so we need to
1815 if (!init_iolog(td))
1818 /* ioprio_set() has to be done before td_io_init() */
1819 if (fio_option_is_set(o, ioprio) ||
1820 fio_option_is_set(o, ioprio_class) ||
1821 fio_option_is_set(o, ioprio_hint)) {
1822 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class,
1823 o->ioprio, o->ioprio_hint);
1825 td_verror(td, errno, "ioprio_set");
1828 td->ioprio = ioprio_value(o->ioprio_class, o->ioprio,
1830 td->ts.ioprio = td->ioprio;
1836 if (td_ioengine_flagged(td, FIO_SYNCIO) && td->o.iodepth > 1 && td->o.io_submit_mode != IO_MODE_OFFLOAD) {
1837 log_info("note: both iodepth >= 1 and synchronous I/O engine "
1838 "are selected, queue depth will be capped at 1\n");
1844 if (td->io_ops->post_init && td->io_ops->post_init(td))
1847 if (o->verify_async && verify_async_init(td))
1850 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1854 if (nice(o->nice) == -1 && errno != 0) {
1855 td_verror(td, errno, "nice");
1859 if (o->ioscheduler && switch_ioscheduler(td))
1862 if (!o->create_serialize && setup_files(td))
1865 if (!init_random_map(td))
1868 if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1871 if (o->pre_read && !pre_read_files(td))
1874 fio_verify_init(td);
1876 if (rate_submit_init(td, sk_out))
1879 set_epoch_time(td, o->log_alternate_epoch_clock_id, o->job_start_clock_id);
1880 fio_getrusage(&td->ru_start);
1881 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1882 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1883 memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1887 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1888 o->ratemin[DDIR_TRIM]) {
1889 memcpy(&td->last_rate_check_time[DDIR_READ], &td->bw_sample_time,
1890 sizeof(td->bw_sample_time));
1891 memcpy(&td->last_rate_check_time[DDIR_WRITE], &td->bw_sample_time,
1892 sizeof(td->bw_sample_time));
1893 memcpy(&td->last_rate_check_time[DDIR_TRIM], &td->bw_sample_time,
1894 sizeof(td->bw_sample_time));
1897 memset(bytes_done, 0, sizeof(bytes_done));
1898 clear_state = false;
1900 while (keep_running(td)) {
1901 uint64_t verify_bytes;
1903 fio_gettime(&td->start, NULL);
1904 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1907 clear_io_state(td, 0);
1909 if (o->unlink_each_loop && unlink_all_files(td))
1913 prune_io_piece_log(td);
1915 if (td->o.verify_only && td_write(td))
1916 verify_bytes = do_dry_run(td);
1918 if (!td->o.rand_repeatable)
1919 /* save verify rand state to replay hdr seeds later at verify */
1920 frand_copy(&td->verify_state_last_do_io, &td->verify_state);
1921 do_io(td, bytes_done);
1922 if (!td->o.rand_repeatable)
1923 frand_copy(&td->verify_state, &td->verify_state_last_do_io);
1924 if (!ddir_rw_sum(bytes_done)) {
1925 fio_mark_td_terminate(td);
1928 verify_bytes = bytes_done[DDIR_WRITE] +
1929 bytes_done[DDIR_TRIM];
1934 * If we took too long to shut down, the main thread could
1935 * already consider us reaped/exited. If that happens, break
1938 if (td->runstate >= TD_EXITED)
1944 * Make sure we've successfully updated the rusage stats
1945 * before waiting on the stat mutex. Otherwise we could have
1946 * the stat thread holding stat mutex and waiting for
1947 * the rusage_sem, which would never get upped because
1948 * this thread is waiting for the stat mutex.
1950 deadlock_loop_cnt = 0;
1952 check_update_rusage(td);
1953 if (!fio_sem_down_trylock(stat_sem))
1956 if (deadlock_loop_cnt++ > 5000) {
1957 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1958 td->error = EDEADLK;
1963 if (td->io_bytes[DDIR_READ] && (td_read(td) ||
1964 ((td->flags & TD_F_VER_BACKLOG) && td_write(td))))
1965 update_runtime(td, elapsed_us, DDIR_READ);
1966 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1967 update_runtime(td, elapsed_us, DDIR_WRITE);
1968 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1969 update_runtime(td, elapsed_us, DDIR_TRIM);
1970 fio_gettime(&td->start, NULL);
1971 fio_sem_up(stat_sem);
1973 if (td->error || td->terminate)
1976 if (!o->do_verify ||
1977 o->verify == VERIFY_NONE ||
1978 td_ioengine_flagged(td, FIO_UNIDIR))
1981 clear_io_state(td, 0);
1983 fio_gettime(&td->start, NULL);
1985 do_verify(td, verify_bytes);
1988 * See comment further up for why this is done here.
1990 check_update_rusage(td);
1992 fio_sem_down(stat_sem);
1993 update_runtime(td, elapsed_us, DDIR_READ);
1994 fio_gettime(&td->start, NULL);
1995 fio_sem_up(stat_sem);
1997 if (td->error || td->terminate)
2002 * Acquire this lock if we were doing overlap checking in
2003 * offload mode so that we don't clean up this job while
2004 * another thread is checking its io_u's for overlap
2006 if (td_offload_overlap(td)) {
2009 res = pthread_mutex_lock(&overlap_check);
2015 td_set_runstate(td, TD_FINISHING);
2016 if (td_offload_overlap(td)) {
2019 res = pthread_mutex_unlock(&overlap_check);
2026 update_rusage_stat(td);
2027 td->ts.total_run_time = mtime_since_now(&td->epoch);
2028 for_each_rw_ddir(ddir) {
2029 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2032 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
2033 (td->o.verify != VERIFY_NONE && td_write(td)))
2034 verify_save_state(td->thread_number);
2036 fio_unpin_memory(td);
2038 td_writeout_logs(td, true);
2040 iolog_compress_exit(td);
2041 rate_submit_exit(td);
2043 if (o->exec_postrun)
2044 exec_string(o, o->exec_postrun, "postrun");
2046 if (exitall_on_terminate || (o->exitall_error && td->error))
2047 fio_terminate_threads(td->groupid, td->o.exit_what);
2051 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
2054 if (o->verify_async)
2055 verify_async_exit(td);
2057 close_and_free_files(td);
2060 cgroup_shutdown(td, cgroup_mnt);
2061 verify_free_state(td);
2062 td_zone_free_index(td);
2064 if (fio_option_is_set(o, cpumask)) {
2065 ret = fio_cpuset_exit(&o->cpumask);
2067 td_verror(td, ret, "fio_cpuset_exit");
2071 * do this very late, it will log file closing as well
2073 if (o->write_iolog_file)
2074 write_iolog_close(td);
2075 if (td->io_log_rfile)
2076 fclose(td->io_log_rfile);
2078 td_set_runstate(td, TD_EXITED);
2081 * Do this last after setting our runstate to exited, so we
2082 * know that the stat thread is signaled.
2084 check_update_rusage(td);
2087 return (void *) (uintptr_t) td->error;
2091 * Run over the job map and reap the threads that have exited, if any.
2093 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
2096 unsigned int cputhreads, realthreads, pending;
2100 * reap exited threads (TD_EXITED -> TD_REAPED)
2102 realthreads = pending = cputhreads = 0;
2104 int flags = 0, status;
2106 if (!strcmp(td->o.ioengine, "cpuio"))
2115 if (td->runstate == TD_REAPED)
2117 if (td->o.use_thread) {
2118 if (td->runstate == TD_EXITED) {
2119 td_set_runstate(td, TD_REAPED);
2126 if (td->runstate == TD_EXITED)
2130 * check if someone quit or got killed in an unusual way
2132 ret = waitpid(td->pid, &status, flags);
2134 if (errno == ECHILD) {
2135 log_err("fio: pid=%d disappeared %d\n",
2136 (int) td->pid, td->runstate);
2138 td_set_runstate(td, TD_REAPED);
2142 } else if (ret == td->pid) {
2143 if (WIFSIGNALED(status)) {
2144 int sig = WTERMSIG(status);
2146 if (sig != SIGTERM && sig != SIGUSR2)
2147 log_err("fio: pid=%d, got signal=%d\n",
2148 (int) td->pid, sig);
2150 td_set_runstate(td, TD_REAPED);
2153 if (WIFEXITED(status)) {
2154 if (WEXITSTATUS(status) && !td->error)
2155 td->error = WEXITSTATUS(status);
2157 td_set_runstate(td, TD_REAPED);
2163 * If the job is stuck, do a forceful timeout of it and
2166 if (td->terminate &&
2167 td->runstate < TD_FSYNCING &&
2168 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2169 log_err("fio: job '%s' (state=%d) hasn't exited in "
2170 "%lu seconds, it appears to be stuck. Doing "
2171 "forceful exit of this job.\n",
2172 td->o.name, td->runstate,
2173 (unsigned long) time_since_now(&td->terminate_time));
2174 td_set_runstate(td, TD_REAPED);
2179 * thread is not dead, continue
2185 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2186 (*t_rate) -= ddir_rw_sum(td->o.rate);
2193 done_secs += mtime_since_now(&td->epoch) / 1000;
2194 profile_td_exit(td);
2198 if (*nr_running == cputhreads && !pending && realthreads)
2199 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2202 static bool __check_trigger_file(void)
2209 if (stat(trigger_file, &sb))
2212 if (unlink(trigger_file) < 0)
2213 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2219 static bool trigger_timedout(void)
2221 if (trigger_timeout)
2222 if (time_since_genesis() >= trigger_timeout) {
2223 trigger_timeout = 0;
2230 void exec_trigger(const char *cmd)
2234 if (!cmd || cmd[0] == '\0')
2239 log_err("fio: failed executing %s trigger\n", cmd);
2242 void check_trigger_file(void)
2244 if (__check_trigger_file() || trigger_timedout()) {
2246 fio_clients_send_trigger(trigger_remote_cmd);
2248 verify_save_state(IO_LIST_ALL);
2249 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2250 exec_trigger(trigger_cmd);
2255 static int fio_verify_load_state(struct thread_data *td)
2259 if (!td->o.verify_state)
2265 ret = fio_server_get_verify_state(td->o.name,
2266 td->thread_number - 1, &data);
2268 verify_assign_state(td, data);
2270 char prefix[PATH_MAX];
2273 sprintf(prefix, "%s%clocal", aux_path,
2274 FIO_OS_PATH_SEPARATOR);
2276 strcpy(prefix, "local");
2277 ret = verify_load_state(td, prefix);
2283 static void do_usleep(unsigned int usecs)
2285 check_for_running_stats();
2286 check_trigger_file();
2290 static bool check_mount_writes(struct thread_data *td)
2295 if (!td_write(td) || td->o.allow_mounted_write)
2299 * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2300 * are mkfs'd and mounted.
2302 for_each_file(td, f, i) {
2303 #ifdef FIO_HAVE_CHARDEV_SIZE
2304 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2306 if (f->filetype != FIO_TYPE_BLOCK)
2309 if (device_is_mounted(f->file_name))
2315 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2319 static bool waitee_running(struct thread_data *me)
2321 const char *waitee = me->o.wait_for;
2322 const char *self = me->o.name;
2328 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2331 if (td->runstate < TD_EXITED) {
2332 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2334 runstate_to_name(td->runstate));
2339 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2344 * Main function for kicking off and reaping jobs, as needed.
2346 static void run_threads(struct sk_out *sk_out)
2348 struct thread_data *td;
2349 unsigned int i, todo, nr_running, nr_started;
2350 uint64_t m_rate, t_rate;
2353 if (fio_gtod_offload && fio_start_gtod_thread())
2356 fio_idle_prof_init();
2360 nr_thread = nr_process = 0;
2362 if (check_mount_writes(td))
2364 if (td->o.use_thread)
2370 if (output_format & FIO_OUTPUT_NORMAL) {
2371 struct buf_output out;
2373 buf_output_init(&out);
2374 __log_buf(&out, "Starting ");
2376 __log_buf(&out, "%d thread%s", nr_thread,
2377 nr_thread > 1 ? "s" : "");
2380 __log_buf(&out, " and ");
2381 __log_buf(&out, "%d process%s", nr_process,
2382 nr_process > 1 ? "es" : "");
2384 __log_buf(&out, "\n");
2385 log_info_buf(out.buf, out.buflen);
2386 buf_output_free(&out);
2389 todo = thread_number;
2392 m_rate = t_rate = 0;
2395 print_status_init(td->thread_number - 1);
2397 if (!td->o.create_serialize)
2400 if (fio_verify_load_state(td))
2404 * do file setup here so it happens sequentially,
2405 * we don't want X number of threads getting their
2406 * client data interspersed on disk
2408 if (setup_files(td)) {
2412 log_err("fio: pid=%d, err=%d/%s\n",
2413 (int) td->pid, td->error, td->verror);
2414 td_set_runstate(td, TD_REAPED);
2421 * for sharing to work, each job must always open
2422 * its own files. so close them, if we opened them
2425 for_each_file(td, f, j) {
2426 if (fio_file_open(f))
2427 td_io_close_file(td, f);
2432 /* start idle threads before io threads start to run */
2433 fio_idle_prof_start();
2438 struct thread_data *map[REAL_MAX_JOBS];
2439 struct timespec this_start;
2440 int this_jobs = 0, left;
2441 struct fork_data *fd;
2444 * create threads (TD_NOT_CREATED -> TD_CREATED)
2447 if (td->runstate != TD_NOT_CREATED)
2451 * never got a chance to start, killed by other
2452 * thread for some reason
2454 if (td->terminate) {
2459 if (td->o.start_delay) {
2460 spent = utime_since_genesis();
2462 if (td->o.start_delay > spent)
2466 if (td->o.stonewall && (nr_started || nr_running)) {
2467 dprint(FD_PROCESS, "%s: stonewall wait\n",
2472 if (waitee_running(td)) {
2473 dprint(FD_PROCESS, "%s: waiting for %s\n",
2474 td->o.name, td->o.wait_for);
2480 td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2481 td->update_rusage = 0;
2484 * Set state to created. Thread will transition
2485 * to TD_INITIALIZED when it's done setting up.
2487 td_set_runstate(td, TD_CREATED);
2488 map[this_jobs++] = td;
2491 fd = calloc(1, sizeof(*fd));
2493 fd->sk_out = sk_out;
2495 if (td->o.use_thread) {
2498 dprint(FD_PROCESS, "will pthread_create\n");
2499 ret = pthread_create(&td->thread, NULL,
2502 log_err("pthread_create: %s\n",
2509 ret = pthread_detach(td->thread);
2511 log_err("pthread_detach: %s",
2516 dprint(FD_PROCESS, "will fork\n");
2523 ret = (int)(uintptr_t)thread_main(fd);
2525 } else if (__td_index == fio_debug_jobno)
2526 *fio_debug_jobp = pid;
2531 dprint(FD_MUTEX, "wait on startup_sem\n");
2532 if (fio_sem_down_timeout(startup_sem, 10000)) {
2533 log_err("fio: job startup hung? exiting.\n");
2534 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2540 dprint(FD_MUTEX, "done waiting on startup_sem\n");
2544 * Wait for the started threads to transition to
2547 fio_gettime(&this_start, NULL);
2549 while (left && !fio_abort) {
2550 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2555 for (i = 0; i < this_jobs; i++) {
2559 if (td->runstate == TD_INITIALIZED) {
2562 } else if (td->runstate >= TD_EXITED) {
2566 nr_running++; /* work-around... */
2572 log_err("fio: %d job%s failed to start\n", left,
2573 left > 1 ? "s" : "");
2574 for (i = 0; i < this_jobs; i++) {
2578 kill(td->pid, SIGTERM);
2584 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2587 if (td->runstate != TD_INITIALIZED)
2590 if (in_ramp_time(td))
2591 td_set_runstate(td, TD_RAMP);
2593 td_set_runstate(td, TD_RUNNING);
2596 m_rate += ddir_rw_sum(td->o.ratemin);
2597 t_rate += ddir_rw_sum(td->o.rate);
2599 fio_sem_up(td->sem);
2602 reap_threads(&nr_running, &t_rate, &m_rate);
2608 while (nr_running) {
2609 reap_threads(&nr_running, &t_rate, &m_rate);
2613 fio_idle_prof_stop();
2618 static void free_disk_util(void)
2620 disk_util_prune_entries();
2621 helper_thread_destroy();
2624 int fio_backend(struct sk_out *sk_out)
2628 if (load_profile(exec_profile))
2631 exec_profile = NULL;
2637 struct log_params p = {
2638 .log_type = IO_LOG_TYPE_BW,
2641 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2642 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2643 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2646 if (init_global_dedupe_working_set_seeds()) {
2647 log_err("fio: failed to initialize global dedupe working set\n");
2651 startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2653 is_local_backend = true;
2654 if (startup_sem == NULL)
2659 if (helper_thread_create(startup_sem, sk_out))
2660 log_err("fio: failed to create helper thread\n");
2662 cgroup_list = smalloc(sizeof(*cgroup_list));
2664 INIT_FLIST_HEAD(cgroup_list);
2666 run_threads(sk_out);
2668 helper_thread_exit();
2673 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2674 struct io_log *log = agg_io_log[i];
2676 flush_log(log, false);
2683 struct thread_stat *ts = &td->ts;
2685 free_clat_prio_stats(ts);
2686 steadystate_free(td);
2687 fio_options_free(td);
2688 fio_dump_options_free(td);
2689 if (td->rusage_sem) {
2690 fio_sem_remove(td->rusage_sem);
2691 td->rusage_sem = NULL;
2693 fio_sem_remove(td->sem);
2699 cgroup_kill(cgroup_list);
2703 fio_sem_remove(startup_sem);