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.
40 #include "lib/memalign.h"
42 #include "lib/getrusage.h"
45 #include "workqueue.h"
46 #include "lib/mountcheck.h"
47 #include "rate-submit.h"
48 #include "helper_thread.h"
50 #include "zone-dist.h"
52 static struct fio_sem *startup_sem;
53 static struct flist_head *cgroup_list;
54 static struct cgroup_mnt *cgroup_mnt;
55 static int exit_value;
56 static volatile int fio_abort;
57 static unsigned int nr_process = 0;
58 static unsigned int nr_thread = 0;
60 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
63 unsigned int thread_number = 0;
64 unsigned int stat_number = 0;
67 unsigned long done_secs = 0;
69 #define JOB_START_TIMEOUT (5 * 1000)
71 static void sig_int(int sig)
75 fio_server_got_signal(sig);
77 log_info("\nfio: terminating on signal %d\n", sig);
82 fio_terminate_threads(TERMINATE_ALL);
86 void sig_show_status(int sig)
88 show_running_run_stats();
91 static void set_sig_handlers(void)
95 memset(&act, 0, sizeof(act));
96 act.sa_handler = sig_int;
97 act.sa_flags = SA_RESTART;
98 sigaction(SIGINT, &act, NULL);
100 memset(&act, 0, sizeof(act));
101 act.sa_handler = sig_int;
102 act.sa_flags = SA_RESTART;
103 sigaction(SIGTERM, &act, NULL);
105 /* Windows uses SIGBREAK as a quit signal from other applications */
107 memset(&act, 0, sizeof(act));
108 act.sa_handler = sig_int;
109 act.sa_flags = SA_RESTART;
110 sigaction(SIGBREAK, &act, NULL);
113 memset(&act, 0, sizeof(act));
114 act.sa_handler = sig_show_status;
115 act.sa_flags = SA_RESTART;
116 sigaction(SIGUSR1, &act, NULL);
119 memset(&act, 0, sizeof(act));
120 act.sa_handler = sig_int;
121 act.sa_flags = SA_RESTART;
122 sigaction(SIGPIPE, &act, NULL);
127 * Check if we are above the minimum rate given.
129 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
132 unsigned long long bytes = 0;
133 unsigned long iops = 0;
136 unsigned int ratemin = 0;
137 unsigned int rate_iops = 0;
138 unsigned int rate_iops_min = 0;
140 assert(ddir_rw(ddir));
142 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
146 * allow a 2 second settle period in the beginning
148 if (mtime_since(&td->start, now) < 2000)
151 iops += td->this_io_blocks[ddir];
152 bytes += td->this_io_bytes[ddir];
153 ratemin += td->o.ratemin[ddir];
154 rate_iops += td->o.rate_iops[ddir];
155 rate_iops_min += td->o.rate_iops_min[ddir];
158 * if rate blocks is set, sample is running
160 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
161 spent = mtime_since(&td->lastrate[ddir], now);
162 if (spent < td->o.ratecycle)
165 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
167 * check bandwidth specified rate
169 if (bytes < td->rate_bytes[ddir]) {
170 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
171 td->o.name, ratemin, bytes);
175 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
179 if (rate < ratemin ||
180 bytes < td->rate_bytes[ddir]) {
181 log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
182 td->o.name, ratemin, rate);
188 * checks iops specified rate
190 if (iops < rate_iops) {
191 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
192 td->o.name, rate_iops, iops);
196 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
200 if (rate < rate_iops_min ||
201 iops < td->rate_blocks[ddir]) {
202 log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
203 td->o.name, rate_iops_min, rate);
210 td->rate_bytes[ddir] = bytes;
211 td->rate_blocks[ddir] = iops;
212 memcpy(&td->lastrate[ddir], now, sizeof(*now));
216 static bool check_min_rate(struct thread_data *td, struct timespec *now)
220 if (td->bytes_done[DDIR_READ])
221 ret |= __check_min_rate(td, now, DDIR_READ);
222 if (td->bytes_done[DDIR_WRITE])
223 ret |= __check_min_rate(td, now, DDIR_WRITE);
224 if (td->bytes_done[DDIR_TRIM])
225 ret |= __check_min_rate(td, now, DDIR_TRIM);
231 * When job exits, we can cancel the in-flight IO if we are using async
232 * io. Attempt to do so.
234 static void cleanup_pending_aio(struct thread_data *td)
239 * get immediately available events, if any
241 r = io_u_queued_complete(td, 0);
246 * now cancel remaining active events
248 if (td->io_ops->cancel) {
252 io_u_qiter(&td->io_u_all, io_u, i) {
253 if (io_u->flags & IO_U_F_FLIGHT) {
254 r = td->io_ops->cancel(td, io_u);
262 r = io_u_queued_complete(td, td->cur_depth);
266 * Helper to handle the final sync of a file. Works just like the normal
267 * io path, just does everything sync.
269 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
271 struct io_u *io_u = __get_io_u(td);
272 enum fio_q_status ret;
277 io_u->ddir = DDIR_SYNC;
280 if (td_io_prep(td, io_u)) {
286 ret = td_io_queue(td, io_u);
290 if (io_u_queued_complete(td, 1) < 0)
293 case FIO_Q_COMPLETED:
295 td_verror(td, io_u->error, "td_io_queue");
299 if (io_u_sync_complete(td, io_u) < 0)
310 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
314 if (fio_file_open(f))
315 return fio_io_sync(td, f);
317 if (td_io_open_file(td, f))
320 ret = fio_io_sync(td, f);
321 td_io_close_file(td, f);
325 static inline void __update_ts_cache(struct thread_data *td)
327 fio_gettime(&td->ts_cache, NULL);
330 static inline void update_ts_cache(struct thread_data *td)
332 if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
333 __update_ts_cache(td);
336 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
338 if (in_ramp_time(td))
342 if (utime_since(&td->epoch, t) >= td->o.timeout)
349 * We need to update the runtime consistently in ms, but keep a running
350 * tally of the current elapsed time in microseconds for sub millisecond
353 static inline void update_runtime(struct thread_data *td,
354 unsigned long long *elapsed_us,
355 const enum fio_ddir ddir)
357 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
360 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
361 elapsed_us[ddir] += utime_since_now(&td->start);
362 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
365 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
370 if (ret < 0 || td->error) {
372 enum error_type_bit eb;
377 eb = td_error_type(ddir, err);
378 if (!(td->o.continue_on_error & (1 << eb)))
381 if (td_non_fatal_error(td, eb, err)) {
383 * Continue with the I/Os in case of
386 update_error_count(td, err);
390 } else if (td->o.fill_device && err == ENOSPC) {
392 * We expect to hit this error if
393 * fill_device option is set.
396 fio_mark_td_terminate(td);
400 * Stop the I/O in case of a fatal
403 update_error_count(td, err);
411 static void check_update_rusage(struct thread_data *td)
413 if (td->update_rusage) {
414 td->update_rusage = 0;
415 update_rusage_stat(td);
416 fio_sem_up(td->rusage_sem);
420 static int wait_for_completions(struct thread_data *td, struct timespec *time)
422 const int full = queue_full(td);
426 if (td->flags & TD_F_REGROW_LOGS)
427 return io_u_quiesce(td);
430 * if the queue is full, we MUST reap at least 1 event
432 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
433 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
436 if (time && __should_check_rate(td))
437 fio_gettime(time, NULL);
440 ret = io_u_queued_complete(td, min_evts);
443 } while (full && (td->cur_depth > td->o.iodepth_low));
448 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
449 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
450 struct timespec *comp_time)
453 case FIO_Q_COMPLETED:
456 clear_io_u(td, io_u);
457 } else if (io_u->resid) {
458 long long bytes = io_u->xfer_buflen - io_u->resid;
459 struct fio_file *f = io_u->file;
462 *bytes_issued += bytes;
472 unlog_io_piece(td, io_u);
473 td_verror(td, EIO, "full resid");
478 io_u->xfer_buflen = io_u->resid;
479 io_u->xfer_buf += bytes;
480 io_u->offset += bytes;
482 if (ddir_rw(io_u->ddir))
483 td->ts.short_io_u[io_u->ddir]++;
485 if (io_u->offset == f->real_file_size)
488 requeue_io_u(td, &io_u);
491 if (comp_time && __should_check_rate(td))
492 fio_gettime(comp_time, NULL);
494 *ret = io_u_sync_complete(td, io_u);
499 if (td->flags & TD_F_REGROW_LOGS)
503 * when doing I/O (not when verifying),
504 * check for any errors that are to be ignored
512 * if the engine doesn't have a commit hook,
513 * the io_u is really queued. if it does have such
514 * a hook, it has to call io_u_queued() itself.
516 if (td->io_ops->commit == NULL)
517 io_u_queued(td, io_u);
519 *bytes_issued += io_u->xfer_buflen;
523 unlog_io_piece(td, io_u);
524 requeue_io_u(td, &io_u);
529 td_verror(td, -(*ret), "td_io_queue");
533 if (break_on_this_error(td, ddir, ret))
539 static inline bool io_in_polling(struct thread_data *td)
541 return !td->o.iodepth_batch_complete_min &&
542 !td->o.iodepth_batch_complete_max;
545 * Unlinks files from thread data fio_file structure
547 static int unlink_all_files(struct thread_data *td)
553 for_each_file(td, f, i) {
554 if (f->filetype != FIO_TYPE_FILE)
556 ret = td_io_unlink_file(td, f);
562 td_verror(td, ret, "unlink_all_files");
568 * Check if io_u will overlap an in-flight IO in the queue
570 static bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
573 struct io_u *check_io_u;
574 unsigned long long x1, x2, y1, y2;
578 x2 = io_u->offset + io_u->buflen;
580 io_u_qiter(q, check_io_u, i) {
581 if (check_io_u->flags & IO_U_F_FLIGHT) {
582 y1 = check_io_u->offset;
583 y2 = check_io_u->offset + check_io_u->buflen;
585 if (x1 < y2 && y1 < x2) {
587 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
589 y1, check_io_u->buflen);
598 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
601 * Check for overlap if the user asked us to, and we have
602 * at least one IO in flight besides this one.
604 if (td->o.serialize_overlap && td->cur_depth > 1 &&
605 in_flight_overlap(&td->io_u_all, io_u))
608 return td_io_queue(td, io_u);
612 * The main verify engine. Runs over the writes we previously submitted,
613 * reads the blocks back in, and checks the crc/md5 of the data.
615 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
622 dprint(FD_VERIFY, "starting loop\n");
625 * sync io first and invalidate cache, to make sure we really
628 for_each_file(td, f, i) {
629 if (!fio_file_open(f))
631 if (fio_io_sync(td, f))
633 if (file_invalidate_cache(td, f))
637 check_update_rusage(td);
643 * verify_state needs to be reset before verification
644 * proceeds so that expected random seeds match actual
645 * random seeds in headers. The main loop will reset
646 * all random number generators if randrepeat is set.
648 if (!td->o.rand_repeatable)
649 td_fill_verify_state_seed(td);
651 td_set_runstate(td, TD_VERIFYING);
654 while (!td->terminate) {
659 check_update_rusage(td);
661 if (runtime_exceeded(td, &td->ts_cache)) {
662 __update_ts_cache(td);
663 if (runtime_exceeded(td, &td->ts_cache)) {
664 fio_mark_td_terminate(td);
669 if (flow_threshold_exceeded(td))
672 if (!td->o.experimental_verify) {
673 io_u = __get_io_u(td);
677 if (get_next_verify(td, io_u)) {
682 if (td_io_prep(td, io_u)) {
687 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
690 while ((io_u = get_io_u(td)) != NULL) {
691 if (IS_ERR_OR_NULL(io_u)) {
698 * We are only interested in the places where
699 * we wrote or trimmed IOs. Turn those into
700 * reads for verification purposes.
702 if (io_u->ddir == DDIR_READ) {
704 * Pretend we issued it for rwmix
707 td->io_issues[DDIR_READ]++;
710 } else if (io_u->ddir == DDIR_TRIM) {
711 io_u->ddir = DDIR_READ;
712 io_u_set(td, io_u, IO_U_F_TRIMMED);
714 } else if (io_u->ddir == DDIR_WRITE) {
715 io_u->ddir = DDIR_READ;
716 populate_verify_io_u(td, io_u);
728 if (verify_state_should_stop(td, io_u)) {
733 if (td->o.verify_async)
734 io_u->end_io = verify_io_u_async;
736 io_u->end_io = verify_io_u;
739 if (!td->o.disable_slat)
740 fio_gettime(&io_u->start_time, NULL);
742 ret = io_u_submit(td, io_u);
744 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
748 * if we can queue more, do so. but check if there are
749 * completed io_u's first. Note that we can get BUSY even
750 * without IO queued, if the system is resource starved.
753 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
754 if (full || io_in_polling(td))
755 ret = wait_for_completions(td, NULL);
761 check_update_rusage(td);
764 min_events = td->cur_depth;
767 ret = io_u_queued_complete(td, min_events);
769 cleanup_pending_aio(td);
771 td_set_runstate(td, TD_RUNNING);
773 dprint(FD_VERIFY, "exiting loop\n");
776 static bool exceeds_number_ios(struct thread_data *td)
778 unsigned long long number_ios;
780 if (!td->o.number_ios)
783 number_ios = ddir_rw_sum(td->io_blocks);
784 number_ios += td->io_u_queued + td->io_u_in_flight;
786 return number_ios >= (td->o.number_ios * td->loops);
789 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
791 unsigned long long bytes, limit;
794 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
795 else if (td_write(td))
796 bytes = this_bytes[DDIR_WRITE];
797 else if (td_read(td))
798 bytes = this_bytes[DDIR_READ];
800 bytes = this_bytes[DDIR_TRIM];
803 limit = td->o.io_size;
808 return bytes >= limit || exceeds_number_ios(td);
811 static bool io_issue_bytes_exceeded(struct thread_data *td)
813 return io_bytes_exceeded(td, td->io_issue_bytes);
816 static bool io_complete_bytes_exceeded(struct thread_data *td)
818 return io_bytes_exceeded(td, td->this_io_bytes);
822 * used to calculate the next io time for rate control
825 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
827 uint64_t bps = td->rate_bps[ddir];
829 assert(!(td->flags & TD_F_CHILD));
831 if (td->o.rate_process == RATE_PROCESS_POISSON) {
834 iops = bps / td->o.bs[ddir];
835 val = (int64_t) (1000000 / iops) *
836 -logf(__rand_0_1(&td->poisson_state[ddir]));
838 dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
839 (unsigned long long) 1000000 / val,
842 td->last_usec[ddir] += val;
843 return td->last_usec[ddir];
845 uint64_t bytes = td->rate_io_issue_bytes[ddir];
846 uint64_t secs = bytes / bps;
847 uint64_t remainder = bytes % bps;
849 return remainder * 1000000 / bps + secs * 1000000;
855 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
857 unsigned long long b;
861 b = ddir_rw_sum(td->io_blocks);
862 if (b % td->o.thinktime_blocks)
868 if (td->o.thinktime_spin)
869 total = usec_spin(td->o.thinktime_spin);
871 left = td->o.thinktime - total;
873 total += usec_sleep(td, left);
876 * If we're ignoring thinktime for the rate, add the number of bytes
877 * we would have done while sleeping, minus one block to ensure we
878 * start issuing immediately after the sleep.
880 if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
881 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
882 uint64_t bs = td->o.min_bs[ddir];
883 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
886 if (usperop <= total)
889 over = (usperop - total) / usperop * -bs;
891 td->rate_io_issue_bytes[ddir] += (missed - over);
892 /* adjust for rate_process=poisson */
893 td->last_usec[ddir] += total;
898 * Main IO worker function. It retrieves io_u's to process and queues
899 * and reaps them, checking for rate and errors along the way.
901 * Returns number of bytes written and trimmed.
903 static void do_io(struct thread_data *td, uint64_t *bytes_done)
907 uint64_t total_bytes, bytes_issued = 0;
909 for (i = 0; i < DDIR_RWDIR_CNT; i++)
910 bytes_done[i] = td->bytes_done[i];
912 if (in_ramp_time(td))
913 td_set_runstate(td, TD_RAMP);
915 td_set_runstate(td, TD_RUNNING);
919 total_bytes = td->o.size;
921 * Allow random overwrite workloads to write up to io_size
922 * before starting verification phase as 'size' doesn't apply.
924 if (td_write(td) && td_random(td) && td->o.norandommap)
925 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
927 * If verify_backlog is enabled, we'll run the verify in this
928 * handler as well. For that case, we may need up to twice the
931 if (td->o.verify != VERIFY_NONE &&
932 (td_write(td) && td->o.verify_backlog))
933 total_bytes += td->o.size;
935 /* In trimwrite mode, each byte is trimmed and then written, so
936 * allow total_bytes to be twice as big */
937 if (td_trimwrite(td))
938 total_bytes += td->total_io_size;
940 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
941 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
943 struct timespec comp_time;
948 check_update_rusage(td);
950 if (td->terminate || td->done)
955 if (runtime_exceeded(td, &td->ts_cache)) {
956 __update_ts_cache(td);
957 if (runtime_exceeded(td, &td->ts_cache)) {
958 fio_mark_td_terminate(td);
963 if (flow_threshold_exceeded(td))
967 * Break if we exceeded the bytes. The exception is time
968 * based runs, but we still need to break out of the loop
969 * for those to run verification, if enabled.
970 * Jobs read from iolog do not use this stop condition.
972 if (bytes_issued >= total_bytes &&
973 !td->o.read_iolog_file &&
974 (!td->o.time_based ||
975 (td->o.time_based && td->o.verify != VERIFY_NONE)))
979 if (IS_ERR_OR_NULL(io_u)) {
980 int err = PTR_ERR(io_u);
988 if (td->o.latency_target)
993 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
994 populate_verify_io_u(td, io_u);
999 * Add verification end_io handler if:
1000 * - Asked to verify (!td_rw(td))
1001 * - Or the io_u is from our verify list (mixed write/ver)
1003 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1004 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1006 if (!td->o.verify_pattern_bytes) {
1007 io_u->rand_seed = __rand(&td->verify_state);
1008 if (sizeof(int) != sizeof(long *))
1009 io_u->rand_seed *= __rand(&td->verify_state);
1012 if (verify_state_should_stop(td, io_u)) {
1017 if (td->o.verify_async)
1018 io_u->end_io = verify_io_u_async;
1020 io_u->end_io = verify_io_u;
1021 td_set_runstate(td, TD_VERIFYING);
1022 } else if (in_ramp_time(td))
1023 td_set_runstate(td, TD_RAMP);
1025 td_set_runstate(td, TD_RUNNING);
1028 * Always log IO before it's issued, so we know the specific
1029 * order of it. The logged unit will track when the IO has
1032 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1034 td->o.verify != VERIFY_NONE &&
1035 !td->o.experimental_verify)
1036 log_io_piece(td, io_u);
1038 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1039 const unsigned long long blen = io_u->xfer_buflen;
1040 const enum fio_ddir __ddir = acct_ddir(io_u);
1045 workqueue_enqueue(&td->io_wq, &io_u->work);
1048 if (ddir_rw(__ddir)) {
1049 td->io_issues[__ddir]++;
1050 td->io_issue_bytes[__ddir] += blen;
1051 td->rate_io_issue_bytes[__ddir] += blen;
1054 if (should_check_rate(td))
1055 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1058 ret = io_u_submit(td, io_u);
1060 if (should_check_rate(td))
1061 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1063 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1067 * See if we need to complete some commands. Note that
1068 * we can get BUSY even without IO queued, if the
1069 * system is resource starved.
1072 full = queue_full(td) ||
1073 (ret == FIO_Q_BUSY && td->cur_depth);
1074 if (full || io_in_polling(td))
1075 ret = wait_for_completions(td, &comp_time);
1079 if (!ddir_rw_sum(td->bytes_done) &&
1080 !td_ioengine_flagged(td, FIO_NOIO))
1083 if (!in_ramp_time(td) && should_check_rate(td)) {
1084 if (check_min_rate(td, &comp_time)) {
1085 if (exitall_on_terminate || td->o.exitall_error)
1086 fio_terminate_threads(td->groupid);
1087 td_verror(td, EIO, "check_min_rate");
1091 if (!in_ramp_time(td) && td->o.latency_target)
1092 lat_target_check(td);
1094 if (ddir_rw(ddir) && td->o.thinktime)
1095 handle_thinktime(td, ddir);
1098 check_update_rusage(td);
1100 if (td->trim_entries)
1101 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1103 if (td->o.fill_device && td->error == ENOSPC) {
1105 fio_mark_td_terminate(td);
1110 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1111 workqueue_flush(&td->io_wq);
1117 ret = io_u_queued_complete(td, i);
1118 if (td->o.fill_device && td->error == ENOSPC)
1122 if (should_fsync(td) && td->o.end_fsync) {
1123 td_set_runstate(td, TD_FSYNCING);
1125 for_each_file(td, f, i) {
1126 if (!fio_file_fsync(td, f))
1129 log_err("fio: end_fsync failed for file %s\n",
1134 cleanup_pending_aio(td);
1137 * stop job if we failed doing any IO
1139 if (!ddir_rw_sum(td->this_io_bytes))
1142 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1143 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1146 static void free_file_completion_logging(struct thread_data *td)
1151 for_each_file(td, f, i) {
1152 if (!f->last_write_comp)
1154 sfree(f->last_write_comp);
1158 static int init_file_completion_logging(struct thread_data *td,
1164 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1167 for_each_file(td, f, i) {
1168 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1169 if (!f->last_write_comp)
1176 free_file_completion_logging(td);
1177 log_err("fio: failed to alloc write comp data\n");
1181 static void cleanup_io_u(struct thread_data *td)
1185 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1187 if (td->io_ops->io_u_free)
1188 td->io_ops->io_u_free(td, io_u);
1190 fio_memfree(io_u, sizeof(*io_u));
1195 io_u_rexit(&td->io_u_requeues);
1196 io_u_qexit(&td->io_u_freelist);
1197 io_u_qexit(&td->io_u_all);
1199 free_file_completion_logging(td);
1202 static int init_io_u(struct thread_data *td)
1205 int cl_align, i, max_units;
1208 max_units = td->o.iodepth;
1211 err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1212 err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1213 err += !io_u_qinit(&td->io_u_all, td->o.iodepth);
1216 log_err("fio: failed setting up IO queues\n");
1220 cl_align = os_cache_line_size();
1222 for (i = 0; i < max_units; i++) {
1228 ptr = fio_memalign(cl_align, sizeof(*io_u));
1230 log_err("fio: unable to allocate aligned memory\n");
1235 memset(io_u, 0, sizeof(*io_u));
1236 INIT_FLIST_HEAD(&io_u->verify_list);
1237 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1240 io_u->flags = IO_U_F_FREE;
1241 io_u_qpush(&td->io_u_freelist, io_u);
1244 * io_u never leaves this stack, used for iteration of all
1247 io_u_qpush(&td->io_u_all, io_u);
1249 if (td->io_ops->io_u_init) {
1250 int ret = td->io_ops->io_u_init(td, io_u);
1253 log_err("fio: failed to init engine data: %d\n", ret);
1259 init_io_u_buffers(td);
1261 if (init_file_completion_logging(td, max_units))
1267 int init_io_u_buffers(struct thread_data *td)
1270 unsigned long long max_bs, min_write;
1275 max_units = td->o.iodepth;
1276 max_bs = td_max_bs(td);
1277 min_write = td->o.min_bs[DDIR_WRITE];
1278 td->orig_buffer_size = (unsigned long long) max_bs
1279 * (unsigned long long) max_units;
1281 if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1285 * if we may later need to do address alignment, then add any
1286 * possible adjustment here so that we don't cause a buffer
1287 * overflow later. this adjustment may be too much if we get
1288 * lucky and the allocator gives us an aligned address.
1290 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1291 td_ioengine_flagged(td, FIO_RAWIO))
1292 td->orig_buffer_size += page_mask + td->o.mem_align;
1294 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1295 unsigned long long bs;
1297 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1298 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1301 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1302 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1306 if (data_xfer && allocate_io_mem(td))
1309 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1310 td_ioengine_flagged(td, FIO_RAWIO))
1311 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1313 p = td->orig_buffer;
1315 for (i = 0; i < max_units; i++) {
1316 io_u = td->io_u_all.io_us[i];
1317 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1321 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1324 io_u_fill_buffer(td, io_u, min_write, max_bs);
1325 if (td_write(td) && td->o.verify_pattern_bytes) {
1327 * Fill the buffer with the pattern if we are
1328 * going to be doing writes.
1330 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1340 * This function is Linux specific.
1341 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1343 static int switch_ioscheduler(struct thread_data *td)
1345 #ifdef FIO_HAVE_IOSCHED_SWITCH
1346 char tmp[256], tmp2[128], *p;
1350 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1353 assert(td->files && td->files[0]);
1354 sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1356 f = fopen(tmp, "r+");
1358 if (errno == ENOENT) {
1359 log_err("fio: os or kernel doesn't support IO scheduler"
1363 td_verror(td, errno, "fopen iosched");
1370 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1371 if (ferror(f) || ret != 1) {
1372 td_verror(td, errno, "fwrite");
1380 * Read back and check that the selected scheduler is now the default.
1382 ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1383 if (ferror(f) || ret < 0) {
1384 td_verror(td, errno, "fread");
1390 * either a list of io schedulers or "none\n" is expected. Strip the
1397 * Write to "none" entry doesn't fail, so check the result here.
1399 if (!strcmp(tmp, "none")) {
1400 log_err("fio: io scheduler is not tunable\n");
1405 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1406 if (!strstr(tmp, tmp2)) {
1407 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1408 td_verror(td, EINVAL, "iosched_switch");
1420 static bool keep_running(struct thread_data *td)
1422 unsigned long long limit;
1428 if (td->o.time_based)
1434 if (exceeds_number_ios(td))
1438 limit = td->o.io_size;
1442 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1446 * If the difference is less than the maximum IO size, we
1449 diff = limit - ddir_rw_sum(td->io_bytes);
1450 if (diff < td_max_bs(td))
1453 if (fio_files_done(td) && !td->o.io_size)
1462 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1464 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1468 str = malloc(newlen);
1469 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1471 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1474 log_err("fio: exec of cmd <%s> failed\n", str);
1481 * Dry run to compute correct state of numberio for verification.
1483 static uint64_t do_dry_run(struct thread_data *td)
1485 td_set_runstate(td, TD_RUNNING);
1487 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1488 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1492 if (td->terminate || td->done)
1495 io_u = get_io_u(td);
1496 if (IS_ERR_OR_NULL(io_u))
1499 io_u_set(td, io_u, IO_U_F_FLIGHT);
1502 if (ddir_rw(acct_ddir(io_u)))
1503 td->io_issues[acct_ddir(io_u)]++;
1504 if (ddir_rw(io_u->ddir)) {
1505 io_u_mark_depth(td, 1);
1506 td->ts.total_io_u[io_u->ddir]++;
1509 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1511 td->o.verify != VERIFY_NONE &&
1512 !td->o.experimental_verify)
1513 log_io_piece(td, io_u);
1515 ret = io_u_sync_complete(td, io_u);
1519 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1523 struct thread_data *td;
1524 struct sk_out *sk_out;
1528 * Entry point for the thread based jobs. The process based jobs end up
1529 * here as well, after a little setup.
1531 static void *thread_main(void *data)
1533 struct fork_data *fd = data;
1534 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1535 struct thread_data *td = fd->td;
1536 struct thread_options *o = &td->o;
1537 struct sk_out *sk_out = fd->sk_out;
1538 uint64_t bytes_done[DDIR_RWDIR_CNT];
1539 int deadlock_loop_cnt;
1540 bool clear_state, did_some_io;
1543 sk_out_assign(sk_out);
1546 if (!o->use_thread) {
1552 fio_local_clock_init();
1554 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1557 fio_server_send_start(td);
1559 INIT_FLIST_HEAD(&td->io_log_list);
1560 INIT_FLIST_HEAD(&td->io_hist_list);
1561 INIT_FLIST_HEAD(&td->verify_list);
1562 INIT_FLIST_HEAD(&td->trim_list);
1563 td->io_hist_tree = RB_ROOT;
1565 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1567 td_verror(td, ret, "mutex_cond_init_pshared");
1570 ret = cond_init_pshared(&td->verify_cond);
1572 td_verror(td, ret, "mutex_cond_pshared");
1576 td_set_runstate(td, TD_INITIALIZED);
1577 dprint(FD_MUTEX, "up startup_sem\n");
1578 fio_sem_up(startup_sem);
1579 dprint(FD_MUTEX, "wait on td->sem\n");
1580 fio_sem_down(td->sem);
1581 dprint(FD_MUTEX, "done waiting on td->sem\n");
1584 * A new gid requires privilege, so we need to do this before setting
1587 if (o->gid != -1U && setgid(o->gid)) {
1588 td_verror(td, errno, "setgid");
1591 if (o->uid != -1U && setuid(o->uid)) {
1592 td_verror(td, errno, "setuid");
1596 td_zone_gen_index(td);
1599 * Do this early, we don't want the compress threads to be limited
1600 * to the same CPUs as the IO workers. So do this before we set
1601 * any potential CPU affinity
1603 if (iolog_compress_init(td, sk_out))
1607 * If we have a gettimeofday() thread, make sure we exclude that
1608 * thread from this job
1611 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1614 * Set affinity first, in case it has an impact on the memory
1617 if (fio_option_is_set(o, cpumask)) {
1618 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1619 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1621 log_err("fio: no CPUs set\n");
1622 log_err("fio: Try increasing number of available CPUs\n");
1623 td_verror(td, EINVAL, "cpus_split");
1627 ret = fio_setaffinity(td->pid, o->cpumask);
1629 td_verror(td, errno, "cpu_set_affinity");
1634 #ifdef CONFIG_LIBNUMA
1635 /* numa node setup */
1636 if (fio_option_is_set(o, numa_cpunodes) ||
1637 fio_option_is_set(o, numa_memnodes)) {
1638 struct bitmask *mask;
1640 if (numa_available() < 0) {
1641 td_verror(td, errno, "Does not support NUMA API\n");
1645 if (fio_option_is_set(o, numa_cpunodes)) {
1646 mask = numa_parse_nodestring(o->numa_cpunodes);
1647 ret = numa_run_on_node_mask(mask);
1648 numa_free_nodemask(mask);
1650 td_verror(td, errno, \
1651 "numa_run_on_node_mask failed\n");
1656 if (fio_option_is_set(o, numa_memnodes)) {
1658 if (o->numa_memnodes)
1659 mask = numa_parse_nodestring(o->numa_memnodes);
1661 switch (o->numa_mem_mode) {
1662 case MPOL_INTERLEAVE:
1663 numa_set_interleave_mask(mask);
1666 numa_set_membind(mask);
1669 numa_set_localalloc();
1671 case MPOL_PREFERRED:
1672 numa_set_preferred(o->numa_mem_prefer_node);
1680 numa_free_nodemask(mask);
1686 if (fio_pin_memory(td))
1690 * May alter parameters that init_io_u() will use, so we need to
1693 if (!init_iolog(td))
1699 if (o->verify_async && verify_async_init(td))
1702 if (fio_option_is_set(o, ioprio) ||
1703 fio_option_is_set(o, ioprio_class)) {
1704 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1706 td_verror(td, errno, "ioprio_set");
1711 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1715 if (nice(o->nice) == -1 && errno != 0) {
1716 td_verror(td, errno, "nice");
1720 if (o->ioscheduler && switch_ioscheduler(td))
1723 if (!o->create_serialize && setup_files(td))
1729 if (!init_random_map(td))
1732 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1735 if (o->pre_read && !pre_read_files(td))
1738 fio_verify_init(td);
1740 if (rate_submit_init(td, sk_out))
1743 set_epoch_time(td, o->log_unix_epoch);
1744 fio_getrusage(&td->ru_start);
1745 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1746 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1747 memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1749 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1750 o->ratemin[DDIR_TRIM]) {
1751 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1752 sizeof(td->bw_sample_time));
1753 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1754 sizeof(td->bw_sample_time));
1755 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1756 sizeof(td->bw_sample_time));
1759 memset(bytes_done, 0, sizeof(bytes_done));
1760 clear_state = false;
1761 did_some_io = false;
1763 while (keep_running(td)) {
1764 uint64_t verify_bytes;
1766 fio_gettime(&td->start, NULL);
1767 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1770 clear_io_state(td, 0);
1772 if (o->unlink_each_loop && unlink_all_files(td))
1776 prune_io_piece_log(td);
1778 if (td->o.verify_only && td_write(td))
1779 verify_bytes = do_dry_run(td);
1781 do_io(td, bytes_done);
1783 if (!ddir_rw_sum(bytes_done)) {
1784 fio_mark_td_terminate(td);
1787 verify_bytes = bytes_done[DDIR_WRITE] +
1788 bytes_done[DDIR_TRIM];
1793 * If we took too long to shut down, the main thread could
1794 * already consider us reaped/exited. If that happens, break
1797 if (td->runstate >= TD_EXITED)
1803 * Make sure we've successfully updated the rusage stats
1804 * before waiting on the stat mutex. Otherwise we could have
1805 * the stat thread holding stat mutex and waiting for
1806 * the rusage_sem, which would never get upped because
1807 * this thread is waiting for the stat mutex.
1809 deadlock_loop_cnt = 0;
1811 check_update_rusage(td);
1812 if (!fio_sem_down_trylock(stat_sem))
1815 if (deadlock_loop_cnt++ > 5000) {
1816 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1817 td->error = EDEADLK;
1822 if (td_read(td) && td->io_bytes[DDIR_READ])
1823 update_runtime(td, elapsed_us, DDIR_READ);
1824 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1825 update_runtime(td, elapsed_us, DDIR_WRITE);
1826 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1827 update_runtime(td, elapsed_us, DDIR_TRIM);
1828 fio_gettime(&td->start, NULL);
1829 fio_sem_up(stat_sem);
1831 if (td->error || td->terminate)
1834 if (!o->do_verify ||
1835 o->verify == VERIFY_NONE ||
1836 td_ioengine_flagged(td, FIO_UNIDIR))
1839 if (ddir_rw_sum(bytes_done))
1842 clear_io_state(td, 0);
1844 fio_gettime(&td->start, NULL);
1846 do_verify(td, verify_bytes);
1849 * See comment further up for why this is done here.
1851 check_update_rusage(td);
1853 fio_sem_down(stat_sem);
1854 update_runtime(td, elapsed_us, DDIR_READ);
1855 fio_gettime(&td->start, NULL);
1856 fio_sem_up(stat_sem);
1858 if (td->error || td->terminate)
1863 * If td ended up with no I/O when it should have had,
1864 * then something went wrong unless FIO_NOIO or FIO_DISKLESSIO.
1865 * (Are we not missing other flags that can be ignored ?)
1867 if ((td->o.size || td->o.io_size) && !ddir_rw_sum(bytes_done) &&
1868 !did_some_io && !td->o.create_only &&
1869 !(td_ioengine_flagged(td, FIO_NOIO) ||
1870 td_ioengine_flagged(td, FIO_DISKLESSIO)))
1871 log_err("%s: No I/O performed by %s, "
1872 "perhaps try --debug=io option for details?\n",
1873 td->o.name, td->io_ops->name);
1875 td_set_runstate(td, TD_FINISHING);
1877 update_rusage_stat(td);
1878 td->ts.total_run_time = mtime_since_now(&td->epoch);
1879 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1880 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1881 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1883 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1884 (td->o.verify != VERIFY_NONE && td_write(td)))
1885 verify_save_state(td->thread_number);
1887 fio_unpin_memory(td);
1889 td_writeout_logs(td, true);
1891 iolog_compress_exit(td);
1892 rate_submit_exit(td);
1894 if (o->exec_postrun)
1895 exec_string(o, o->exec_postrun, (const char *)"postrun");
1897 if (exitall_on_terminate || (o->exitall_error && td->error))
1898 fio_terminate_threads(td->groupid);
1902 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1905 if (o->verify_async)
1906 verify_async_exit(td);
1908 close_and_free_files(td);
1911 cgroup_shutdown(td, cgroup_mnt);
1912 verify_free_state(td);
1913 td_zone_free_index(td);
1915 if (fio_option_is_set(o, cpumask)) {
1916 ret = fio_cpuset_exit(&o->cpumask);
1918 td_verror(td, ret, "fio_cpuset_exit");
1922 * do this very late, it will log file closing as well
1924 if (o->write_iolog_file)
1925 write_iolog_close(td);
1926 if (td->io_log_rfile)
1927 fclose(td->io_log_rfile);
1929 td_set_runstate(td, TD_EXITED);
1932 * Do this last after setting our runstate to exited, so we
1933 * know that the stat thread is signaled.
1935 check_update_rusage(td);
1938 return (void *) (uintptr_t) td->error;
1942 * Run over the job map and reap the threads that have exited, if any.
1944 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1947 struct thread_data *td;
1948 unsigned int cputhreads, realthreads, pending;
1952 * reap exited threads (TD_EXITED -> TD_REAPED)
1954 realthreads = pending = cputhreads = 0;
1955 for_each_td(td, i) {
1958 if (!strcmp(td->o.ioengine, "cpuio"))
1967 if (td->runstate == TD_REAPED)
1969 if (td->o.use_thread) {
1970 if (td->runstate == TD_EXITED) {
1971 td_set_runstate(td, TD_REAPED);
1978 if (td->runstate == TD_EXITED)
1982 * check if someone quit or got killed in an unusual way
1984 ret = waitpid(td->pid, &status, flags);
1986 if (errno == ECHILD) {
1987 log_err("fio: pid=%d disappeared %d\n",
1988 (int) td->pid, td->runstate);
1990 td_set_runstate(td, TD_REAPED);
1994 } else if (ret == td->pid) {
1995 if (WIFSIGNALED(status)) {
1996 int sig = WTERMSIG(status);
1998 if (sig != SIGTERM && sig != SIGUSR2)
1999 log_err("fio: pid=%d, got signal=%d\n",
2000 (int) td->pid, sig);
2002 td_set_runstate(td, TD_REAPED);
2005 if (WIFEXITED(status)) {
2006 if (WEXITSTATUS(status) && !td->error)
2007 td->error = WEXITSTATUS(status);
2009 td_set_runstate(td, TD_REAPED);
2015 * If the job is stuck, do a forceful timeout of it and
2018 if (td->terminate &&
2019 td->runstate < TD_FSYNCING &&
2020 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2021 log_err("fio: job '%s' (state=%d) hasn't exited in "
2022 "%lu seconds, it appears to be stuck. Doing "
2023 "forceful exit of this job.\n",
2024 td->o.name, td->runstate,
2025 (unsigned long) time_since_now(&td->terminate_time));
2026 td_set_runstate(td, TD_REAPED);
2031 * thread is not dead, continue
2037 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2038 (*t_rate) -= ddir_rw_sum(td->o.rate);
2045 done_secs += mtime_since_now(&td->epoch) / 1000;
2046 profile_td_exit(td);
2049 if (*nr_running == cputhreads && !pending && realthreads)
2050 fio_terminate_threads(TERMINATE_ALL);
2053 static bool __check_trigger_file(void)
2060 if (stat(trigger_file, &sb))
2063 if (unlink(trigger_file) < 0)
2064 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2070 static bool trigger_timedout(void)
2072 if (trigger_timeout)
2073 if (time_since_genesis() >= trigger_timeout) {
2074 trigger_timeout = 0;
2081 void exec_trigger(const char *cmd)
2085 if (!cmd || cmd[0] == '\0')
2090 log_err("fio: failed executing %s trigger\n", cmd);
2093 void check_trigger_file(void)
2095 if (__check_trigger_file() || trigger_timedout()) {
2097 fio_clients_send_trigger(trigger_remote_cmd);
2099 verify_save_state(IO_LIST_ALL);
2100 fio_terminate_threads(TERMINATE_ALL);
2101 exec_trigger(trigger_cmd);
2106 static int fio_verify_load_state(struct thread_data *td)
2110 if (!td->o.verify_state)
2116 ret = fio_server_get_verify_state(td->o.name,
2117 td->thread_number - 1, &data);
2119 verify_assign_state(td, data);
2121 ret = verify_load_state(td, "local");
2126 static void do_usleep(unsigned int usecs)
2128 check_for_running_stats();
2129 check_trigger_file();
2133 static bool check_mount_writes(struct thread_data *td)
2138 if (!td_write(td) || td->o.allow_mounted_write)
2142 * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2143 * are mkfs'd and mounted.
2145 for_each_file(td, f, i) {
2146 #ifdef FIO_HAVE_CHARDEV_SIZE
2147 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2149 if (f->filetype != FIO_TYPE_BLOCK)
2152 if (device_is_mounted(f->file_name))
2158 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2162 static bool waitee_running(struct thread_data *me)
2164 const char *waitee = me->o.wait_for;
2165 const char *self = me->o.name;
2166 struct thread_data *td;
2172 for_each_td(td, i) {
2173 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2176 if (td->runstate < TD_EXITED) {
2177 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2179 runstate_to_name(td->runstate));
2184 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2189 * Main function for kicking off and reaping jobs, as needed.
2191 static void run_threads(struct sk_out *sk_out)
2193 struct thread_data *td;
2194 unsigned int i, todo, nr_running, nr_started;
2195 uint64_t m_rate, t_rate;
2198 if (fio_gtod_offload && fio_start_gtod_thread())
2201 fio_idle_prof_init();
2205 nr_thread = nr_process = 0;
2206 for_each_td(td, i) {
2207 if (check_mount_writes(td))
2209 if (td->o.use_thread)
2215 if (output_format & FIO_OUTPUT_NORMAL) {
2216 struct buf_output out;
2218 buf_output_init(&out);
2219 __log_buf(&out, "Starting ");
2221 __log_buf(&out, "%d thread%s", nr_thread,
2222 nr_thread > 1 ? "s" : "");
2225 __log_buf(&out, " and ");
2226 __log_buf(&out, "%d process%s", nr_process,
2227 nr_process > 1 ? "es" : "");
2229 __log_buf(&out, "\n");
2230 log_info_buf(out.buf, out.buflen);
2231 buf_output_free(&out);
2234 todo = thread_number;
2237 m_rate = t_rate = 0;
2239 for_each_td(td, i) {
2240 print_status_init(td->thread_number - 1);
2242 if (!td->o.create_serialize)
2245 if (fio_verify_load_state(td))
2249 * do file setup here so it happens sequentially,
2250 * we don't want X number of threads getting their
2251 * client data interspersed on disk
2253 if (setup_files(td)) {
2257 log_err("fio: pid=%d, err=%d/%s\n",
2258 (int) td->pid, td->error, td->verror);
2259 td_set_runstate(td, TD_REAPED);
2266 * for sharing to work, each job must always open
2267 * its own files. so close them, if we opened them
2270 for_each_file(td, f, j) {
2271 if (fio_file_open(f))
2272 td_io_close_file(td, f);
2277 /* start idle threads before io threads start to run */
2278 fio_idle_prof_start();
2283 struct thread_data *map[REAL_MAX_JOBS];
2284 struct timespec this_start;
2285 int this_jobs = 0, left;
2286 struct fork_data *fd;
2289 * create threads (TD_NOT_CREATED -> TD_CREATED)
2291 for_each_td(td, i) {
2292 if (td->runstate != TD_NOT_CREATED)
2296 * never got a chance to start, killed by other
2297 * thread for some reason
2299 if (td->terminate) {
2304 if (td->o.start_delay) {
2305 spent = utime_since_genesis();
2307 if (td->o.start_delay > spent)
2311 if (td->o.stonewall && (nr_started || nr_running)) {
2312 dprint(FD_PROCESS, "%s: stonewall wait\n",
2317 if (waitee_running(td)) {
2318 dprint(FD_PROCESS, "%s: waiting for %s\n",
2319 td->o.name, td->o.wait_for);
2325 td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2326 td->update_rusage = 0;
2329 * Set state to created. Thread will transition
2330 * to TD_INITIALIZED when it's done setting up.
2332 td_set_runstate(td, TD_CREATED);
2333 map[this_jobs++] = td;
2336 fd = calloc(1, sizeof(*fd));
2338 fd->sk_out = sk_out;
2340 if (td->o.use_thread) {
2343 dprint(FD_PROCESS, "will pthread_create\n");
2344 ret = pthread_create(&td->thread, NULL,
2347 log_err("pthread_create: %s\n",
2354 ret = pthread_detach(td->thread);
2356 log_err("pthread_detach: %s",
2360 dprint(FD_PROCESS, "will fork\n");
2365 ret = (int)(uintptr_t)thread_main(fd);
2367 } else if (i == fio_debug_jobno)
2368 *fio_debug_jobp = pid;
2370 dprint(FD_MUTEX, "wait on startup_sem\n");
2371 if (fio_sem_down_timeout(startup_sem, 10000)) {
2372 log_err("fio: job startup hung? exiting.\n");
2373 fio_terminate_threads(TERMINATE_ALL);
2379 dprint(FD_MUTEX, "done waiting on startup_sem\n");
2383 * Wait for the started threads to transition to
2386 fio_gettime(&this_start, NULL);
2388 while (left && !fio_abort) {
2389 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2394 for (i = 0; i < this_jobs; i++) {
2398 if (td->runstate == TD_INITIALIZED) {
2401 } else if (td->runstate >= TD_EXITED) {
2405 nr_running++; /* work-around... */
2411 log_err("fio: %d job%s failed to start\n", left,
2412 left > 1 ? "s" : "");
2413 for (i = 0; i < this_jobs; i++) {
2417 kill(td->pid, SIGTERM);
2423 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2425 for_each_td(td, i) {
2426 if (td->runstate != TD_INITIALIZED)
2429 if (in_ramp_time(td))
2430 td_set_runstate(td, TD_RAMP);
2432 td_set_runstate(td, TD_RUNNING);
2435 m_rate += ddir_rw_sum(td->o.ratemin);
2436 t_rate += ddir_rw_sum(td->o.rate);
2438 fio_sem_up(td->sem);
2441 reap_threads(&nr_running, &t_rate, &m_rate);
2447 while (nr_running) {
2448 reap_threads(&nr_running, &t_rate, &m_rate);
2452 fio_idle_prof_stop();
2457 static void free_disk_util(void)
2459 disk_util_prune_entries();
2460 helper_thread_destroy();
2463 int fio_backend(struct sk_out *sk_out)
2465 struct thread_data *td;
2469 if (load_profile(exec_profile))
2472 exec_profile = NULL;
2478 struct log_params p = {
2479 .log_type = IO_LOG_TYPE_BW,
2482 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2483 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2484 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2487 startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2489 is_local_backend = true;
2490 if (startup_sem == NULL)
2495 helper_thread_create(startup_sem, sk_out);
2497 cgroup_list = smalloc(sizeof(*cgroup_list));
2499 INIT_FLIST_HEAD(cgroup_list);
2501 run_threads(sk_out);
2503 helper_thread_exit();
2508 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2509 struct io_log *log = agg_io_log[i];
2511 flush_log(log, false);
2517 for_each_td(td, i) {
2518 steadystate_free(td);
2519 fio_options_free(td);
2520 if (td->rusage_sem) {
2521 fio_sem_remove(td->rusage_sem);
2522 td->rusage_sem = NULL;
2524 fio_sem_remove(td->sem);
2530 cgroup_kill(cgroup_list);
2534 fio_sem_remove(startup_sem);