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;
67 unsigned long done_secs = 0;
68 #ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP
69 pthread_mutex_t overlap_check = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
71 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
74 #define JOB_START_TIMEOUT (5 * 1000)
76 static void sig_int(int sig)
78 if (segments[0].threads) {
80 fio_server_got_signal(sig);
82 log_info("\nfio: terminating on signal %d\n", sig);
87 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
91 void sig_show_status(int sig)
93 show_running_run_stats();
96 static void set_sig_handlers(void)
100 memset(&act, 0, sizeof(act));
101 act.sa_handler = sig_int;
102 act.sa_flags = SA_RESTART;
103 sigaction(SIGINT, &act, NULL);
105 memset(&act, 0, sizeof(act));
106 act.sa_handler = sig_int;
107 act.sa_flags = SA_RESTART;
108 sigaction(SIGTERM, &act, NULL);
110 /* Windows uses SIGBREAK as a quit signal from other applications */
112 memset(&act, 0, sizeof(act));
113 act.sa_handler = sig_int;
114 act.sa_flags = SA_RESTART;
115 sigaction(SIGBREAK, &act, NULL);
118 memset(&act, 0, sizeof(act));
119 act.sa_handler = sig_show_status;
120 act.sa_flags = SA_RESTART;
121 sigaction(SIGUSR1, &act, NULL);
124 memset(&act, 0, sizeof(act));
125 act.sa_handler = sig_int;
126 act.sa_flags = SA_RESTART;
127 sigaction(SIGPIPE, &act, NULL);
132 * Check if we are above the minimum rate given.
134 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
137 unsigned long long bytes = 0;
138 unsigned long iops = 0;
140 unsigned long long rate;
141 unsigned long long ratemin = 0;
142 unsigned int rate_iops = 0;
143 unsigned int rate_iops_min = 0;
145 assert(ddir_rw(ddir));
147 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
151 * allow a 2 second settle period in the beginning
153 if (mtime_since(&td->start, now) < 2000)
156 iops += td->this_io_blocks[ddir];
157 bytes += td->this_io_bytes[ddir];
158 ratemin += td->o.ratemin[ddir];
159 rate_iops += td->o.rate_iops[ddir];
160 rate_iops_min += td->o.rate_iops_min[ddir];
163 * if rate blocks is set, sample is running
165 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
166 spent = mtime_since(&td->lastrate[ddir], now);
167 if (spent < td->o.ratecycle)
170 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
172 * check bandwidth specified rate
174 if (bytes < td->rate_bytes[ddir]) {
175 log_err("%s: rate_min=%lluB/s not met, only transferred %lluB\n",
176 td->o.name, ratemin, bytes);
180 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
184 if (rate < ratemin ||
185 bytes < td->rate_bytes[ddir]) {
186 log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
187 td->o.name, ratemin, rate);
193 * checks iops specified rate
195 if (iops < rate_iops) {
196 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
197 td->o.name, rate_iops, iops);
201 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
205 if (rate < rate_iops_min ||
206 iops < td->rate_blocks[ddir]) {
207 log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
208 td->o.name, rate_iops_min, rate);
215 td->rate_bytes[ddir] = bytes;
216 td->rate_blocks[ddir] = iops;
217 memcpy(&td->lastrate[ddir], now, sizeof(*now));
221 static bool check_min_rate(struct thread_data *td, struct timespec *now)
225 for_each_rw_ddir(ddir) {
226 if (td->bytes_done[ddir])
227 ret |= __check_min_rate(td, now, ddir);
234 * When job exits, we can cancel the in-flight IO if we are using async
235 * io. Attempt to do so.
237 static void cleanup_pending_aio(struct thread_data *td)
242 * get immediately available events, if any
244 r = io_u_queued_complete(td, 0);
247 * now cancel remaining active events
249 if (td->io_ops->cancel) {
253 io_u_qiter(&td->io_u_all, io_u, i) {
254 if (io_u->flags & IO_U_F_FLIGHT) {
255 r = td->io_ops->cancel(td, io_u);
263 r = io_u_queued_complete(td, td->cur_depth);
267 * Helper to handle the final sync of a file. Works just like the normal
268 * io path, just does everything sync.
270 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
272 struct io_u *io_u = __get_io_u(td);
273 enum fio_q_status ret;
278 io_u->ddir = DDIR_SYNC;
280 io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
282 if (td_io_prep(td, io_u)) {
288 ret = td_io_queue(td, io_u);
292 if (io_u_queued_complete(td, 1) < 0)
295 case FIO_Q_COMPLETED:
297 td_verror(td, io_u->error, "td_io_queue");
301 if (io_u_sync_complete(td, io_u) < 0)
312 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
316 if (fio_file_open(f))
317 return fio_io_sync(td, f);
319 if (td_io_open_file(td, f))
322 ret = fio_io_sync(td, f);
324 if (fio_file_open(f))
325 ret2 = td_io_close_file(td, f);
326 return (ret || ret2);
329 static inline void __update_ts_cache(struct thread_data *td)
331 fio_gettime(&td->ts_cache, NULL);
334 static inline void update_ts_cache(struct thread_data *td)
336 if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
337 __update_ts_cache(td);
340 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
342 if (in_ramp_time(td))
346 if (utime_since(&td->epoch, t) >= td->o.timeout)
353 * We need to update the runtime consistently in ms, but keep a running
354 * tally of the current elapsed time in microseconds for sub millisecond
357 static inline void update_runtime(struct thread_data *td,
358 unsigned long long *elapsed_us,
359 const enum fio_ddir ddir)
361 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
364 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
365 elapsed_us[ddir] += utime_since_now(&td->start);
366 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
369 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
374 if (ret < 0 || td->error) {
376 enum error_type_bit eb;
381 eb = td_error_type(ddir, err);
382 if (!(td->o.continue_on_error & (1 << eb)))
385 if (td_non_fatal_error(td, eb, err)) {
387 * Continue with the I/Os in case of
390 update_error_count(td, err);
394 } else if (td->o.fill_device && err == ENOSPC) {
396 * We expect to hit this error if
397 * fill_device option is set.
400 fio_mark_td_terminate(td);
404 * Stop the I/O in case of a fatal
407 update_error_count(td, err);
415 static void check_update_rusage(struct thread_data *td)
417 if (td->update_rusage) {
418 td->update_rusage = 0;
419 update_rusage_stat(td);
420 fio_sem_up(td->rusage_sem);
424 static int wait_for_completions(struct thread_data *td, struct timespec *time)
426 const int full = queue_full(td);
430 if (td->flags & TD_F_REGROW_LOGS)
431 return io_u_quiesce(td);
434 * if the queue is full, we MUST reap at least 1 event
436 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
437 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
440 if (time && __should_check_rate(td))
441 fio_gettime(time, NULL);
444 ret = io_u_queued_complete(td, min_evts);
447 } while (full && (td->cur_depth > td->o.iodepth_low));
452 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
453 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
454 struct timespec *comp_time)
457 case FIO_Q_COMPLETED:
460 clear_io_u(td, io_u);
461 } else if (io_u->resid) {
462 long long bytes = io_u->xfer_buflen - io_u->resid;
463 struct fio_file *f = io_u->file;
466 *bytes_issued += bytes;
476 unlog_io_piece(td, io_u);
477 td_verror(td, EIO, "full resid");
482 io_u->xfer_buflen = io_u->resid;
483 io_u->xfer_buf += bytes;
484 io_u->offset += bytes;
486 if (ddir_rw(io_u->ddir))
487 td->ts.short_io_u[io_u->ddir]++;
489 if (io_u->offset == f->real_file_size)
492 requeue_io_u(td, &io_u);
495 if (comp_time && __should_check_rate(td))
496 fio_gettime(comp_time, NULL);
498 *ret = io_u_sync_complete(td, io_u);
503 if (td->flags & TD_F_REGROW_LOGS)
507 * when doing I/O (not when verifying),
508 * check for any errors that are to be ignored
516 * if the engine doesn't have a commit hook,
517 * the io_u is really queued. if it does have such
518 * a hook, it has to call io_u_queued() itself.
520 if (td->io_ops->commit == NULL)
521 io_u_queued(td, io_u);
523 *bytes_issued += io_u->xfer_buflen;
527 unlog_io_piece(td, io_u);
528 requeue_io_u(td, &io_u);
533 td_verror(td, -(*ret), "td_io_queue");
537 if (break_on_this_error(td, ddir, ret))
543 static inline bool io_in_polling(struct thread_data *td)
545 return !td->o.iodepth_batch_complete_min &&
546 !td->o.iodepth_batch_complete_max;
549 * Unlinks files from thread data fio_file structure
551 static int unlink_all_files(struct thread_data *td)
557 for_each_file(td, f, i) {
558 if (f->filetype != FIO_TYPE_FILE)
560 ret = td_io_unlink_file(td, f);
566 td_verror(td, ret, "unlink_all_files");
572 * Check if io_u will overlap an in-flight IO in the queue
574 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
577 struct io_u *check_io_u;
578 unsigned long long x1, x2, y1, y2;
582 x2 = io_u->offset + io_u->buflen;
584 io_u_qiter(q, check_io_u, i) {
585 if (check_io_u->flags & IO_U_F_FLIGHT) {
586 y1 = check_io_u->offset;
587 y2 = check_io_u->offset + check_io_u->buflen;
589 if (x1 < y2 && y1 < x2) {
591 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
593 y1, check_io_u->buflen);
602 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
605 * Check for overlap if the user asked us to, and we have
606 * at least one IO in flight besides this one.
608 if (td->o.serialize_overlap && td->cur_depth > 1 &&
609 in_flight_overlap(&td->io_u_all, io_u))
612 return td_io_queue(td, io_u);
616 * The main verify engine. Runs over the writes we previously submitted,
617 * reads the blocks back in, and checks the crc/md5 of the data.
619 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
626 dprint(FD_VERIFY, "starting loop\n");
629 * sync io first and invalidate cache, to make sure we really
632 for_each_file(td, f, i) {
633 if (!fio_file_open(f))
635 if (fio_io_sync(td, f))
637 if (file_invalidate_cache(td, f))
641 check_update_rusage(td);
647 * verify_state needs to be reset before verification
648 * proceeds so that expected random seeds match actual
649 * random seeds in headers. The main loop will reset
650 * all random number generators if randrepeat is set.
652 if (!td->o.rand_repeatable)
653 td_fill_verify_state_seed(td);
655 td_set_runstate(td, TD_VERIFYING);
658 while (!td->terminate) {
663 check_update_rusage(td);
665 if (runtime_exceeded(td, &td->ts_cache)) {
666 __update_ts_cache(td);
667 if (runtime_exceeded(td, &td->ts_cache)) {
668 fio_mark_td_terminate(td);
673 if (flow_threshold_exceeded(td))
676 if (!td->o.experimental_verify) {
677 io_u = __get_io_u(td);
681 if (get_next_verify(td, io_u)) {
686 if (td_io_prep(td, io_u)) {
691 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
694 while ((io_u = get_io_u(td)) != NULL) {
695 if (IS_ERR_OR_NULL(io_u)) {
702 * We are only interested in the places where
703 * we wrote or trimmed IOs. Turn those into
704 * reads for verification purposes.
706 if (io_u->ddir == DDIR_READ) {
708 * Pretend we issued it for rwmix
711 td->io_issues[DDIR_READ]++;
714 } else if (io_u->ddir == DDIR_TRIM) {
715 io_u->ddir = DDIR_READ;
716 io_u_set(td, io_u, IO_U_F_TRIMMED);
718 } else if (io_u->ddir == DDIR_WRITE) {
719 io_u->ddir = DDIR_READ;
720 populate_verify_io_u(td, io_u);
732 if (verify_state_should_stop(td, io_u)) {
737 if (td->o.verify_async)
738 io_u->end_io = verify_io_u_async;
740 io_u->end_io = verify_io_u;
743 if (!td->o.disable_slat)
744 fio_gettime(&io_u->start_time, NULL);
746 ret = io_u_submit(td, io_u);
748 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
752 * if we can queue more, do so. but check if there are
753 * completed io_u's first. Note that we can get BUSY even
754 * without IO queued, if the system is resource starved.
757 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
758 if (full || io_in_polling(td))
759 ret = wait_for_completions(td, NULL);
765 check_update_rusage(td);
768 min_events = td->cur_depth;
771 ret = io_u_queued_complete(td, min_events);
773 cleanup_pending_aio(td);
775 td_set_runstate(td, TD_RUNNING);
777 dprint(FD_VERIFY, "exiting loop\n");
780 static bool exceeds_number_ios(struct thread_data *td)
782 unsigned long long number_ios;
784 if (!td->o.number_ios)
787 number_ios = ddir_rw_sum(td->io_blocks);
788 number_ios += td->io_u_queued + td->io_u_in_flight;
790 return number_ios >= (td->o.number_ios * td->loops);
793 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
795 unsigned long long bytes, limit;
798 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
799 else if (td_write(td))
800 bytes = this_bytes[DDIR_WRITE];
801 else if (td_read(td))
802 bytes = this_bytes[DDIR_READ];
804 bytes = this_bytes[DDIR_TRIM];
807 limit = td->o.io_size;
812 return bytes >= limit || exceeds_number_ios(td);
815 static bool io_issue_bytes_exceeded(struct thread_data *td)
817 return io_bytes_exceeded(td, td->io_issue_bytes);
820 static bool io_complete_bytes_exceeded(struct thread_data *td)
822 return io_bytes_exceeded(td, td->this_io_bytes);
826 * used to calculate the next io time for rate control
829 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
831 uint64_t bps = td->rate_bps[ddir];
833 assert(!(td->flags & TD_F_CHILD));
835 if (td->o.rate_process == RATE_PROCESS_POISSON) {
838 iops = bps / td->o.bs[ddir];
839 val = (int64_t) (1000000 / iops) *
840 -logf(__rand_0_1(&td->poisson_state[ddir]));
842 dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
843 (unsigned long long) 1000000 / val,
846 td->last_usec[ddir] += val;
847 return td->last_usec[ddir];
849 uint64_t bytes = td->rate_io_issue_bytes[ddir];
850 uint64_t secs = bytes / bps;
851 uint64_t remainder = bytes % bps;
853 return remainder * 1000000 / bps + secs * 1000000;
859 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
861 unsigned long long b;
865 b = ddir_rw_sum(td->io_blocks);
866 if (b % td->o.thinktime_blocks)
872 if (td->o.thinktime_spin)
873 total = usec_spin(td->o.thinktime_spin);
875 left = td->o.thinktime - total;
877 total += usec_sleep(td, left);
880 * If we're ignoring thinktime for the rate, add the number of bytes
881 * we would have done while sleeping, minus one block to ensure we
882 * start issuing immediately after the sleep.
884 if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
885 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
886 uint64_t bs = td->o.min_bs[ddir];
887 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
890 if (usperop <= total)
893 over = (usperop - total) / usperop * -bs;
895 td->rate_io_issue_bytes[ddir] += (missed - over);
896 /* adjust for rate_process=poisson */
897 td->last_usec[ddir] += total;
902 * Main IO worker function. It retrieves io_u's to process and queues
903 * and reaps them, checking for rate and errors along the way.
905 * Returns number of bytes written and trimmed.
907 static void do_io(struct thread_data *td, uint64_t *bytes_done)
911 uint64_t total_bytes, bytes_issued = 0;
913 for (i = 0; i < DDIR_RWDIR_CNT; i++)
914 bytes_done[i] = td->bytes_done[i];
916 if (in_ramp_time(td))
917 td_set_runstate(td, TD_RAMP);
919 td_set_runstate(td, TD_RUNNING);
923 total_bytes = td->o.size;
925 * Allow random overwrite workloads to write up to io_size
926 * before starting verification phase as 'size' doesn't apply.
928 if (td_write(td) && td_random(td) && td->o.norandommap)
929 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
931 * If verify_backlog is enabled, we'll run the verify in this
932 * handler as well. For that case, we may need up to twice the
935 if (td->o.verify != VERIFY_NONE &&
936 (td_write(td) && td->o.verify_backlog))
937 total_bytes += td->o.size;
939 /* In trimwrite mode, each byte is trimmed and then written, so
940 * allow total_bytes to be twice as big */
941 if (td_trimwrite(td))
942 total_bytes += td->total_io_size;
944 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
945 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
947 struct timespec comp_time;
952 check_update_rusage(td);
954 if (td->terminate || td->done)
959 if (runtime_exceeded(td, &td->ts_cache)) {
960 __update_ts_cache(td);
961 if (runtime_exceeded(td, &td->ts_cache)) {
962 fio_mark_td_terminate(td);
967 if (flow_threshold_exceeded(td))
971 * Break if we exceeded the bytes. The exception is time
972 * based runs, but we still need to break out of the loop
973 * for those to run verification, if enabled.
974 * Jobs read from iolog do not use this stop condition.
976 if (bytes_issued >= total_bytes &&
977 !td->o.read_iolog_file &&
978 (!td->o.time_based ||
979 (td->o.time_based && td->o.verify != VERIFY_NONE)))
983 if (IS_ERR_OR_NULL(io_u)) {
984 int err = PTR_ERR(io_u);
992 if (td->o.latency_target)
997 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
998 populate_verify_io_u(td, io_u);
1003 * Add verification end_io handler if:
1004 * - Asked to verify (!td_rw(td))
1005 * - Or the io_u is from our verify list (mixed write/ver)
1007 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1008 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1010 if (verify_state_should_stop(td, io_u)) {
1015 if (td->o.verify_async)
1016 io_u->end_io = verify_io_u_async;
1018 io_u->end_io = verify_io_u;
1019 td_set_runstate(td, TD_VERIFYING);
1020 } else if (in_ramp_time(td))
1021 td_set_runstate(td, TD_RAMP);
1023 td_set_runstate(td, TD_RUNNING);
1026 * Always log IO before it's issued, so we know the specific
1027 * order of it. The logged unit will track when the IO has
1030 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1032 td->o.verify != VERIFY_NONE &&
1033 !td->o.experimental_verify)
1034 log_io_piece(td, io_u);
1036 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1037 const unsigned long long blen = io_u->xfer_buflen;
1038 const enum fio_ddir __ddir = acct_ddir(io_u);
1043 workqueue_enqueue(&td->io_wq, &io_u->work);
1046 if (ddir_rw(__ddir)) {
1047 td->io_issues[__ddir]++;
1048 td->io_issue_bytes[__ddir] += blen;
1049 td->rate_io_issue_bytes[__ddir] += blen;
1052 if (should_check_rate(td))
1053 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1056 ret = io_u_submit(td, io_u);
1058 if (should_check_rate(td))
1059 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1061 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1065 * See if we need to complete some commands. Note that
1066 * we can get BUSY even without IO queued, if the
1067 * system is resource starved.
1070 full = queue_full(td) ||
1071 (ret == FIO_Q_BUSY && td->cur_depth);
1072 if (full || io_in_polling(td))
1073 ret = wait_for_completions(td, &comp_time);
1077 if (!ddir_rw_sum(td->bytes_done) &&
1078 !td_ioengine_flagged(td, FIO_NOIO))
1081 if (!in_ramp_time(td) && should_check_rate(td)) {
1082 if (check_min_rate(td, &comp_time)) {
1083 if (exitall_on_terminate || td->o.exitall_error)
1084 fio_terminate_threads(td->groupid, td->o.exit_what);
1085 td_verror(td, EIO, "check_min_rate");
1089 if (!in_ramp_time(td) && td->o.latency_target)
1090 lat_target_check(td);
1092 if (ddir_rw(ddir) && td->o.thinktime)
1093 handle_thinktime(td, ddir);
1096 check_update_rusage(td);
1098 if (td->trim_entries)
1099 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1101 if (td->o.fill_device && td->error == ENOSPC) {
1103 fio_mark_td_terminate(td);
1108 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1109 workqueue_flush(&td->io_wq);
1115 ret = io_u_queued_complete(td, i);
1116 if (td->o.fill_device && td->error == ENOSPC)
1120 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1121 td_set_runstate(td, TD_FSYNCING);
1123 for_each_file(td, f, i) {
1124 if (!fio_file_fsync(td, f))
1127 log_err("fio: end_fsync failed for file %s\n",
1132 cleanup_pending_aio(td);
1135 * stop job if we failed doing any IO
1137 if (!ddir_rw_sum(td->this_io_bytes))
1140 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1141 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1144 static void free_file_completion_logging(struct thread_data *td)
1149 for_each_file(td, f, i) {
1150 if (!f->last_write_comp)
1152 sfree(f->last_write_comp);
1156 static int init_file_completion_logging(struct thread_data *td,
1162 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1165 for_each_file(td, f, i) {
1166 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1167 if (!f->last_write_comp)
1174 free_file_completion_logging(td);
1175 log_err("fio: failed to alloc write comp data\n");
1179 static void cleanup_io_u(struct thread_data *td)
1183 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1185 if (td->io_ops->io_u_free)
1186 td->io_ops->io_u_free(td, io_u);
1188 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1193 io_u_rexit(&td->io_u_requeues);
1194 io_u_qexit(&td->io_u_freelist, false);
1195 io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1197 free_file_completion_logging(td);
1200 static int init_io_u(struct thread_data *td)
1203 int cl_align, i, max_units;
1206 max_units = td->o.iodepth;
1209 err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1210 err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1211 err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1214 log_err("fio: failed setting up IO queues\n");
1218 cl_align = os_cache_line_size();
1220 for (i = 0; i < max_units; i++) {
1226 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1228 log_err("fio: unable to allocate aligned memory\n");
1233 memset(io_u, 0, sizeof(*io_u));
1234 INIT_FLIST_HEAD(&io_u->verify_list);
1235 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1238 io_u->flags = IO_U_F_FREE;
1239 io_u_qpush(&td->io_u_freelist, io_u);
1242 * io_u never leaves this stack, used for iteration of all
1245 io_u_qpush(&td->io_u_all, io_u);
1247 if (td->io_ops->io_u_init) {
1248 int ret = td->io_ops->io_u_init(td, io_u);
1251 log_err("fio: failed to init engine data: %d\n", ret);
1257 init_io_u_buffers(td);
1259 if (init_file_completion_logging(td, max_units))
1265 int init_io_u_buffers(struct thread_data *td)
1268 unsigned long long max_bs, min_write;
1273 max_units = td->o.iodepth;
1274 max_bs = td_max_bs(td);
1275 min_write = td->o.min_bs[DDIR_WRITE];
1276 td->orig_buffer_size = (unsigned long long) max_bs
1277 * (unsigned long long) max_units;
1279 if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1283 * if we may later need to do address alignment, then add any
1284 * possible adjustment here so that we don't cause a buffer
1285 * overflow later. this adjustment may be too much if we get
1286 * lucky and the allocator gives us an aligned address.
1288 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1289 td_ioengine_flagged(td, FIO_RAWIO))
1290 td->orig_buffer_size += page_mask + td->o.mem_align;
1292 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1293 unsigned long long bs;
1295 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1296 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1299 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1300 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1304 if (data_xfer && allocate_io_mem(td))
1307 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1308 td_ioengine_flagged(td, FIO_RAWIO))
1309 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1311 p = td->orig_buffer;
1313 for (i = 0; i < max_units; i++) {
1314 io_u = td->io_u_all.io_us[i];
1315 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1319 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1322 io_u_fill_buffer(td, io_u, min_write, max_bs);
1323 if (td_write(td) && td->o.verify_pattern_bytes) {
1325 * Fill the buffer with the pattern if we are
1326 * going to be doing writes.
1328 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1338 * This function is Linux specific.
1339 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1341 static int switch_ioscheduler(struct thread_data *td)
1343 #ifdef FIO_HAVE_IOSCHED_SWITCH
1344 char tmp[256], tmp2[128], *p;
1348 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1351 assert(td->files && td->files[0]);
1352 sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1354 f = fopen(tmp, "r+");
1356 if (errno == ENOENT) {
1357 log_err("fio: os or kernel doesn't support IO scheduler"
1361 td_verror(td, errno, "fopen iosched");
1368 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1369 if (ferror(f) || ret != 1) {
1370 td_verror(td, errno, "fwrite");
1378 * Read back and check that the selected scheduler is now the default.
1380 ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1381 if (ferror(f) || ret < 0) {
1382 td_verror(td, errno, "fread");
1388 * either a list of io schedulers or "none\n" is expected. Strip the
1395 * Write to "none" entry doesn't fail, so check the result here.
1397 if (!strcmp(tmp, "none")) {
1398 log_err("fio: io scheduler is not tunable\n");
1403 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1404 if (!strstr(tmp, tmp2)) {
1405 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1406 td_verror(td, EINVAL, "iosched_switch");
1418 static bool keep_running(struct thread_data *td)
1420 unsigned long long limit;
1426 if (td->o.time_based)
1432 if (exceeds_number_ios(td))
1436 limit = td->o.io_size;
1440 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1444 * If the difference is less than the maximum IO size, we
1447 diff = limit - ddir_rw_sum(td->io_bytes);
1448 if (diff < td_max_bs(td))
1451 if (fio_files_done(td) && !td->o.io_size)
1460 static int exec_string(struct thread_options *o, const char *string,
1466 if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1469 log_info("%s : Saving output of %s in %s.%s.txt\n", 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, "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, "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);
2048 if (*nr_running == cputhreads && !pending && realthreads)
2049 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2052 static bool __check_trigger_file(void)
2059 if (stat(trigger_file, &sb))
2062 if (unlink(trigger_file) < 0)
2063 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2069 static bool trigger_timedout(void)
2071 if (trigger_timeout)
2072 if (time_since_genesis() >= trigger_timeout) {
2073 trigger_timeout = 0;
2080 void exec_trigger(const char *cmd)
2084 if (!cmd || cmd[0] == '\0')
2089 log_err("fio: failed executing %s trigger\n", cmd);
2092 void check_trigger_file(void)
2094 if (__check_trigger_file() || trigger_timedout()) {
2096 fio_clients_send_trigger(trigger_remote_cmd);
2098 verify_save_state(IO_LIST_ALL);
2099 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2100 exec_trigger(trigger_cmd);
2105 static int fio_verify_load_state(struct thread_data *td)
2109 if (!td->o.verify_state)
2115 ret = fio_server_get_verify_state(td->o.name,
2116 td->thread_number - 1, &data);
2118 verify_assign_state(td, data);
2120 char prefix[PATH_MAX];
2123 sprintf(prefix, "%s%clocal", aux_path,
2124 FIO_OS_PATH_SEPARATOR);
2126 strcpy(prefix, "local");
2127 ret = verify_load_state(td, prefix);
2133 static void do_usleep(unsigned int usecs)
2135 check_for_running_stats();
2136 check_trigger_file();
2140 static bool check_mount_writes(struct thread_data *td)
2145 if (!td_write(td) || td->o.allow_mounted_write)
2149 * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2150 * are mkfs'd and mounted.
2152 for_each_file(td, f, i) {
2153 #ifdef FIO_HAVE_CHARDEV_SIZE
2154 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2156 if (f->filetype != FIO_TYPE_BLOCK)
2159 if (device_is_mounted(f->file_name))
2165 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2169 static bool waitee_running(struct thread_data *me)
2171 const char *waitee = me->o.wait_for;
2172 const char *self = me->o.name;
2173 struct thread_data *td;
2179 for_each_td(td, i) {
2180 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2183 if (td->runstate < TD_EXITED) {
2184 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2186 runstate_to_name(td->runstate));
2191 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2196 * Main function for kicking off and reaping jobs, as needed.
2198 static void run_threads(struct sk_out *sk_out)
2200 struct thread_data *td;
2201 unsigned int i, todo, nr_running, nr_started;
2202 uint64_t m_rate, t_rate;
2205 if (fio_gtod_offload && fio_start_gtod_thread())
2208 fio_idle_prof_init();
2212 nr_thread = nr_process = 0;
2213 for_each_td(td, i) {
2214 if (check_mount_writes(td))
2216 if (td->o.use_thread)
2222 if (output_format & FIO_OUTPUT_NORMAL) {
2223 struct buf_output out;
2225 buf_output_init(&out);
2226 __log_buf(&out, "Starting ");
2228 __log_buf(&out, "%d thread%s", nr_thread,
2229 nr_thread > 1 ? "s" : "");
2232 __log_buf(&out, " and ");
2233 __log_buf(&out, "%d process%s", nr_process,
2234 nr_process > 1 ? "es" : "");
2236 __log_buf(&out, "\n");
2237 log_info_buf(out.buf, out.buflen);
2238 buf_output_free(&out);
2241 todo = thread_number;
2244 m_rate = t_rate = 0;
2246 for_each_td(td, i) {
2247 print_status_init(td->thread_number - 1);
2249 if (!td->o.create_serialize)
2252 if (fio_verify_load_state(td))
2256 * do file setup here so it happens sequentially,
2257 * we don't want X number of threads getting their
2258 * client data interspersed on disk
2260 if (setup_files(td)) {
2264 log_err("fio: pid=%d, err=%d/%s\n",
2265 (int) td->pid, td->error, td->verror);
2266 td_set_runstate(td, TD_REAPED);
2273 * for sharing to work, each job must always open
2274 * its own files. so close them, if we opened them
2277 for_each_file(td, f, j) {
2278 if (fio_file_open(f))
2279 td_io_close_file(td, f);
2284 /* start idle threads before io threads start to run */
2285 fio_idle_prof_start();
2290 struct thread_data *map[REAL_MAX_JOBS];
2291 struct timespec this_start;
2292 int this_jobs = 0, left;
2293 struct fork_data *fd;
2296 * create threads (TD_NOT_CREATED -> TD_CREATED)
2298 for_each_td(td, i) {
2299 if (td->runstate != TD_NOT_CREATED)
2303 * never got a chance to start, killed by other
2304 * thread for some reason
2306 if (td->terminate) {
2311 if (td->o.start_delay) {
2312 spent = utime_since_genesis();
2314 if (td->o.start_delay > spent)
2318 if (td->o.stonewall && (nr_started || nr_running)) {
2319 dprint(FD_PROCESS, "%s: stonewall wait\n",
2324 if (waitee_running(td)) {
2325 dprint(FD_PROCESS, "%s: waiting for %s\n",
2326 td->o.name, td->o.wait_for);
2332 td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2333 td->update_rusage = 0;
2336 * Set state to created. Thread will transition
2337 * to TD_INITIALIZED when it's done setting up.
2339 td_set_runstate(td, TD_CREATED);
2340 map[this_jobs++] = td;
2343 fd = calloc(1, sizeof(*fd));
2345 fd->sk_out = sk_out;
2347 if (td->o.use_thread) {
2350 dprint(FD_PROCESS, "will pthread_create\n");
2351 ret = pthread_create(&td->thread, NULL,
2354 log_err("pthread_create: %s\n",
2361 ret = pthread_detach(td->thread);
2363 log_err("pthread_detach: %s",
2367 dprint(FD_PROCESS, "will fork\n");
2372 ret = (int)(uintptr_t)thread_main(fd);
2374 } else if (i == fio_debug_jobno)
2375 *fio_debug_jobp = pid;
2377 dprint(FD_MUTEX, "wait on startup_sem\n");
2378 if (fio_sem_down_timeout(startup_sem, 10000)) {
2379 log_err("fio: job startup hung? exiting.\n");
2380 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2386 dprint(FD_MUTEX, "done waiting on startup_sem\n");
2390 * Wait for the started threads to transition to
2393 fio_gettime(&this_start, NULL);
2395 while (left && !fio_abort) {
2396 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2401 for (i = 0; i < this_jobs; i++) {
2405 if (td->runstate == TD_INITIALIZED) {
2408 } else if (td->runstate >= TD_EXITED) {
2412 nr_running++; /* work-around... */
2418 log_err("fio: %d job%s failed to start\n", left,
2419 left > 1 ? "s" : "");
2420 for (i = 0; i < this_jobs; i++) {
2424 kill(td->pid, SIGTERM);
2430 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2432 for_each_td(td, i) {
2433 if (td->runstate != TD_INITIALIZED)
2436 if (in_ramp_time(td))
2437 td_set_runstate(td, TD_RAMP);
2439 td_set_runstate(td, TD_RUNNING);
2442 m_rate += ddir_rw_sum(td->o.ratemin);
2443 t_rate += ddir_rw_sum(td->o.rate);
2445 fio_sem_up(td->sem);
2448 reap_threads(&nr_running, &t_rate, &m_rate);
2454 while (nr_running) {
2455 reap_threads(&nr_running, &t_rate, &m_rate);
2459 fio_idle_prof_stop();
2464 static void free_disk_util(void)
2466 disk_util_prune_entries();
2467 helper_thread_destroy();
2470 int fio_backend(struct sk_out *sk_out)
2472 struct thread_data *td;
2476 if (load_profile(exec_profile))
2479 exec_profile = NULL;
2485 struct log_params p = {
2486 .log_type = IO_LOG_TYPE_BW,
2489 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2490 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2491 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2494 startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2496 is_local_backend = true;
2497 if (startup_sem == NULL)
2502 if (helper_thread_create(startup_sem, sk_out))
2503 log_err("fio: failed to create helper thread\n");
2505 cgroup_list = smalloc(sizeof(*cgroup_list));
2507 INIT_FLIST_HEAD(cgroup_list);
2509 run_threads(sk_out);
2511 helper_thread_exit();
2516 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2517 struct io_log *log = agg_io_log[i];
2519 flush_log(log, false);
2525 for_each_td(td, i) {
2526 steadystate_free(td);
2527 fio_options_free(td);
2528 if (td->rusage_sem) {
2529 fio_sem_remove(td->rusage_sem);
2530 td->rusage_sem = NULL;
2532 fio_sem_remove(td->sem);
2538 cgroup_kill(cgroup_list);
2542 fio_sem_remove(startup_sem);