2 * fio - the flexible io tester
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
7 * The license below covers all files distributed with fio unless otherwise
8 * noted in the file itself.
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License version 2 as
12 * published by the Free Software Foundation.
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
41 #include "lib/memalign.h"
43 #include "lib/getrusage.h"
46 #include "workqueue.h"
47 #include "lib/mountcheck.h"
48 #include "rate-submit.h"
49 #include "helper_thread.h"
51 #include "zone-dist.h"
53 static struct fio_sem *startup_sem;
54 static struct flist_head *cgroup_list;
55 static struct cgroup_mnt *cgroup_mnt;
56 static int exit_value;
57 static volatile bool fio_abort;
58 static unsigned int nr_process = 0;
59 static unsigned int nr_thread = 0;
61 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
64 unsigned int thread_number = 0;
65 unsigned int stat_number = 0;
68 unsigned long done_secs = 0;
69 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
71 #define JOB_START_TIMEOUT (5 * 1000)
73 static void sig_int(int sig)
77 fio_server_got_signal(sig);
79 log_info("\nfio: terminating on signal %d\n", sig);
84 fio_terminate_threads(TERMINATE_ALL);
88 void sig_show_status(int sig)
90 show_running_run_stats();
93 static void set_sig_handlers(void)
97 memset(&act, 0, sizeof(act));
98 act.sa_handler = sig_int;
99 act.sa_flags = SA_RESTART;
100 sigaction(SIGINT, &act, NULL);
102 memset(&act, 0, sizeof(act));
103 act.sa_handler = sig_int;
104 act.sa_flags = SA_RESTART;
105 sigaction(SIGTERM, &act, NULL);
107 /* Windows uses SIGBREAK as a quit signal from other applications */
109 memset(&act, 0, sizeof(act));
110 act.sa_handler = sig_int;
111 act.sa_flags = SA_RESTART;
112 sigaction(SIGBREAK, &act, NULL);
115 memset(&act, 0, sizeof(act));
116 act.sa_handler = sig_show_status;
117 act.sa_flags = SA_RESTART;
118 sigaction(SIGUSR1, &act, NULL);
121 memset(&act, 0, sizeof(act));
122 act.sa_handler = sig_int;
123 act.sa_flags = SA_RESTART;
124 sigaction(SIGPIPE, &act, NULL);
129 * Check if we are above the minimum rate given.
131 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
134 unsigned long long bytes = 0;
135 unsigned long iops = 0;
138 unsigned int ratemin = 0;
139 unsigned int rate_iops = 0;
140 unsigned int rate_iops_min = 0;
142 assert(ddir_rw(ddir));
144 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
148 * allow a 2 second settle period in the beginning
150 if (mtime_since(&td->start, now) < 2000)
153 iops += td->this_io_blocks[ddir];
154 bytes += td->this_io_bytes[ddir];
155 ratemin += td->o.ratemin[ddir];
156 rate_iops += td->o.rate_iops[ddir];
157 rate_iops_min += td->o.rate_iops_min[ddir];
160 * if rate blocks is set, sample is running
162 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
163 spent = mtime_since(&td->lastrate[ddir], now);
164 if (spent < td->o.ratecycle)
167 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
169 * check bandwidth specified rate
171 if (bytes < td->rate_bytes[ddir]) {
172 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
173 td->o.name, ratemin, bytes);
177 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
181 if (rate < ratemin ||
182 bytes < td->rate_bytes[ddir]) {
183 log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
184 td->o.name, ratemin, rate);
190 * checks iops specified rate
192 if (iops < rate_iops) {
193 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
194 td->o.name, rate_iops, iops);
198 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
202 if (rate < rate_iops_min ||
203 iops < td->rate_blocks[ddir]) {
204 log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
205 td->o.name, rate_iops_min, rate);
212 td->rate_bytes[ddir] = bytes;
213 td->rate_blocks[ddir] = iops;
214 memcpy(&td->lastrate[ddir], now, sizeof(*now));
218 static bool check_min_rate(struct thread_data *td, struct timespec *now)
222 if (td->bytes_done[DDIR_READ])
223 ret |= __check_min_rate(td, now, DDIR_READ);
224 if (td->bytes_done[DDIR_WRITE])
225 ret |= __check_min_rate(td, now, DDIR_WRITE);
226 if (td->bytes_done[DDIR_TRIM])
227 ret |= __check_min_rate(td, now, DDIR_TRIM);
233 * When job exits, we can cancel the in-flight IO if we are using async
234 * io. Attempt to do so.
236 static void cleanup_pending_aio(struct thread_data *td)
244 * get immediately available events, if any
246 r = io_u_queued_complete(td, 0);
251 * now cancel remaining active events
253 if (td->io_ops->cancel) {
257 io_u_qiter(&td->io_u_all, io_u, i) {
258 if (io_u->flags & IO_U_F_FLIGHT) {
259 r = td->io_ops->cancel(td, io_u);
267 r = io_u_queued_complete(td, td->cur_depth);
271 * Helper to handle the final sync of a file. Works just like the normal
272 * io path, just does everything sync.
274 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
276 struct io_u *io_u = __get_io_u(td);
277 enum fio_q_status ret;
282 io_u->ddir = DDIR_SYNC;
285 if (td_io_prep(td, io_u)) {
291 ret = td_io_queue(td, io_u);
295 if (io_u_queued_complete(td, 1) < 0)
298 case FIO_Q_COMPLETED:
300 td_verror(td, io_u->error, "td_io_queue");
304 if (io_u_sync_complete(td, io_u) < 0)
315 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
319 if (fio_file_open(f))
320 return fio_io_sync(td, f);
322 if (td_io_open_file(td, f))
325 ret = fio_io_sync(td, f);
326 td_io_close_file(td, f);
330 static inline void __update_ts_cache(struct thread_data *td)
332 fio_gettime(&td->ts_cache, NULL);
335 static inline void update_ts_cache(struct thread_data *td)
337 if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
338 __update_ts_cache(td);
341 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
343 if (in_ramp_time(td))
347 if (utime_since(&td->epoch, t) >= td->o.timeout)
354 * We need to update the runtime consistently in ms, but keep a running
355 * tally of the current elapsed time in microseconds for sub millisecond
358 static inline void update_runtime(struct thread_data *td,
359 unsigned long long *elapsed_us,
360 const enum fio_ddir ddir)
362 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
365 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
366 elapsed_us[ddir] += utime_since_now(&td->start);
367 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
370 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
375 if (ret < 0 || td->error) {
377 enum error_type_bit eb;
382 eb = td_error_type(ddir, err);
383 if (!(td->o.continue_on_error & (1 << eb)))
386 if (td_non_fatal_error(td, eb, err)) {
388 * Continue with the I/Os in case of
391 update_error_count(td, err);
395 } else if (td->o.fill_device && err == ENOSPC) {
397 * We expect to hit this error if
398 * fill_device option is set.
401 fio_mark_td_terminate(td);
405 * Stop the I/O in case of a fatal
408 update_error_count(td, err);
416 static void check_update_rusage(struct thread_data *td)
418 if (td->update_rusage) {
419 td->update_rusage = 0;
420 update_rusage_stat(td);
421 fio_sem_up(td->rusage_sem);
425 static int wait_for_completions(struct thread_data *td, struct timespec *time)
427 const int full = queue_full(td);
431 if (td->flags & TD_F_REGROW_LOGS)
432 return io_u_quiesce(td);
435 * if the queue is full, we MUST reap at least 1 event
437 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
438 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
441 if (time && __should_check_rate(td))
442 fio_gettime(time, NULL);
445 ret = io_u_queued_complete(td, min_evts);
448 } while (full && (td->cur_depth > td->o.iodepth_low));
453 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
454 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
455 struct timespec *comp_time)
458 case FIO_Q_COMPLETED:
461 clear_io_u(td, io_u);
462 } else if (io_u->resid) {
463 long long bytes = io_u->xfer_buflen - io_u->resid;
464 struct fio_file *f = io_u->file;
467 *bytes_issued += bytes;
477 unlog_io_piece(td, io_u);
478 td_verror(td, EIO, "full resid");
483 io_u->xfer_buflen = io_u->resid;
484 io_u->xfer_buf += bytes;
485 io_u->offset += bytes;
487 if (ddir_rw(io_u->ddir))
488 td->ts.short_io_u[io_u->ddir]++;
490 if (io_u->offset == f->real_file_size)
493 requeue_io_u(td, &io_u);
496 if (comp_time && __should_check_rate(td))
497 fio_gettime(comp_time, NULL);
499 *ret = io_u_sync_complete(td, io_u);
504 if (td->flags & TD_F_REGROW_LOGS)
508 * when doing I/O (not when verifying),
509 * check for any errors that are to be ignored
517 * if the engine doesn't have a commit hook,
518 * the io_u is really queued. if it does have such
519 * a hook, it has to call io_u_queued() itself.
521 if (td->io_ops->commit == NULL)
522 io_u_queued(td, io_u);
524 *bytes_issued += io_u->xfer_buflen;
528 unlog_io_piece(td, io_u);
529 requeue_io_u(td, &io_u);
534 td_verror(td, -(*ret), "td_io_queue");
538 if (break_on_this_error(td, ddir, ret))
544 static inline bool io_in_polling(struct thread_data *td)
546 return !td->o.iodepth_batch_complete_min &&
547 !td->o.iodepth_batch_complete_max;
550 * Unlinks files from thread data fio_file structure
552 static int unlink_all_files(struct thread_data *td)
558 for_each_file(td, f, i) {
559 if (f->filetype != FIO_TYPE_FILE)
561 ret = td_io_unlink_file(td, f);
567 td_verror(td, ret, "unlink_all_files");
573 * Check if io_u will overlap an in-flight IO in the queue
575 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
578 struct io_u *check_io_u;
579 unsigned long long x1, x2, y1, y2;
583 x2 = io_u->offset + io_u->buflen;
585 io_u_qiter(q, check_io_u, i) {
586 if (check_io_u->flags & IO_U_F_FLIGHT) {
587 y1 = check_io_u->offset;
588 y2 = check_io_u->offset + check_io_u->buflen;
590 if (x1 < y2 && y1 < x2) {
592 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
594 y1, check_io_u->buflen);
603 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
606 * Check for overlap if the user asked us to, and we have
607 * at least one IO in flight besides this one.
609 if (td->o.serialize_overlap && td->cur_depth > 1 &&
610 in_flight_overlap(&td->io_u_all, io_u))
613 return td_io_queue(td, io_u);
617 * The main verify engine. Runs over the writes we previously submitted,
618 * reads the blocks back in, and checks the crc/md5 of the data.
620 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
627 dprint(FD_VERIFY, "starting loop\n");
630 * sync io first and invalidate cache, to make sure we really
633 for_each_file(td, f, i) {
634 if (!fio_file_open(f))
636 if (fio_io_sync(td, f))
638 if (file_invalidate_cache(td, f))
642 check_update_rusage(td);
648 * verify_state needs to be reset before verification
649 * proceeds so that expected random seeds match actual
650 * random seeds in headers. The main loop will reset
651 * all random number generators if randrepeat is set.
653 if (!td->o.rand_repeatable)
654 td_fill_verify_state_seed(td);
656 td_set_runstate(td, TD_VERIFYING);
659 while (!td->terminate) {
664 check_update_rusage(td);
666 if (runtime_exceeded(td, &td->ts_cache)) {
667 __update_ts_cache(td);
668 if (runtime_exceeded(td, &td->ts_cache)) {
669 fio_mark_td_terminate(td);
674 if (flow_threshold_exceeded(td))
677 if (!td->o.experimental_verify) {
678 io_u = __get_io_u(td);
682 if (get_next_verify(td, io_u)) {
687 if (td_io_prep(td, io_u)) {
692 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
695 while ((io_u = get_io_u(td)) != NULL) {
696 if (IS_ERR_OR_NULL(io_u)) {
703 * We are only interested in the places where
704 * we wrote or trimmed IOs. Turn those into
705 * reads for verification purposes.
707 if (io_u->ddir == DDIR_READ) {
709 * Pretend we issued it for rwmix
712 td->io_issues[DDIR_READ]++;
715 } else if (io_u->ddir == DDIR_TRIM) {
716 io_u->ddir = DDIR_READ;
717 io_u_set(td, io_u, IO_U_F_TRIMMED);
719 } else if (io_u->ddir == DDIR_WRITE) {
720 io_u->ddir = DDIR_READ;
721 populate_verify_io_u(td, io_u);
733 if (verify_state_should_stop(td, io_u)) {
738 if (td->o.verify_async)
739 io_u->end_io = verify_io_u_async;
741 io_u->end_io = verify_io_u;
744 if (!td->o.disable_slat)
745 fio_gettime(&io_u->start_time, NULL);
747 ret = io_u_submit(td, io_u);
749 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
753 * if we can queue more, do so. but check if there are
754 * completed io_u's first. Note that we can get BUSY even
755 * without IO queued, if the system is resource starved.
758 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
759 if (full || io_in_polling(td))
760 ret = wait_for_completions(td, NULL);
766 check_update_rusage(td);
769 min_events = td->cur_depth;
772 ret = io_u_queued_complete(td, min_events);
774 cleanup_pending_aio(td);
776 td_set_runstate(td, TD_RUNNING);
778 dprint(FD_VERIFY, "exiting loop\n");
781 static bool exceeds_number_ios(struct thread_data *td)
783 unsigned long long number_ios;
785 if (!td->o.number_ios)
788 number_ios = ddir_rw_sum(td->io_blocks);
789 number_ios += td->io_u_queued + td->io_u_in_flight;
791 return number_ios >= (td->o.number_ios * td->loops);
794 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
796 unsigned long long bytes, limit;
799 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
800 else if (td_write(td))
801 bytes = this_bytes[DDIR_WRITE];
802 else if (td_read(td))
803 bytes = this_bytes[DDIR_READ];
805 bytes = this_bytes[DDIR_TRIM];
808 limit = td->o.io_size;
813 return bytes >= limit || exceeds_number_ios(td);
816 static bool io_issue_bytes_exceeded(struct thread_data *td)
818 return io_bytes_exceeded(td, td->io_issue_bytes);
821 static bool io_complete_bytes_exceeded(struct thread_data *td)
823 return io_bytes_exceeded(td, td->this_io_bytes);
827 * used to calculate the next io time for rate control
830 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
832 uint64_t bps = td->rate_bps[ddir];
834 assert(!(td->flags & TD_F_CHILD));
836 if (td->o.rate_process == RATE_PROCESS_POISSON) {
839 iops = bps / td->o.bs[ddir];
840 val = (int64_t) (1000000 / iops) *
841 -logf(__rand_0_1(&td->poisson_state[ddir]));
843 dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
844 (unsigned long long) 1000000 / val,
847 td->last_usec[ddir] += val;
848 return td->last_usec[ddir];
850 uint64_t bytes = td->rate_io_issue_bytes[ddir];
851 uint64_t secs = bytes / bps;
852 uint64_t remainder = bytes % bps;
854 return remainder * 1000000 / bps + secs * 1000000;
860 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
862 unsigned long long b;
866 b = ddir_rw_sum(td->io_blocks);
867 if (b % td->o.thinktime_blocks)
873 if (td->o.thinktime_spin)
874 total = usec_spin(td->o.thinktime_spin);
876 left = td->o.thinktime - total;
878 total += usec_sleep(td, left);
881 * If we're ignoring thinktime for the rate, add the number of bytes
882 * we would have done while sleeping, minus one block to ensure we
883 * start issuing immediately after the sleep.
885 if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
886 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
887 uint64_t bs = td->o.min_bs[ddir];
888 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
891 if (usperop <= total)
894 over = (usperop - total) / usperop * -bs;
896 td->rate_io_issue_bytes[ddir] += (missed - over);
897 /* adjust for rate_process=poisson */
898 td->last_usec[ddir] += total;
903 * Main IO worker function. It retrieves io_u's to process and queues
904 * and reaps them, checking for rate and errors along the way.
906 * Returns number of bytes written and trimmed.
908 static void do_io(struct thread_data *td, uint64_t *bytes_done)
912 uint64_t total_bytes, bytes_issued = 0;
914 for (i = 0; i < DDIR_RWDIR_CNT; i++)
915 bytes_done[i] = td->bytes_done[i];
917 if (in_ramp_time(td))
918 td_set_runstate(td, TD_RAMP);
920 td_set_runstate(td, TD_RUNNING);
924 total_bytes = td->o.size;
926 * Allow random overwrite workloads to write up to io_size
927 * before starting verification phase as 'size' doesn't apply.
929 if (td_write(td) && td_random(td) && td->o.norandommap)
930 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
932 * If verify_backlog is enabled, we'll run the verify in this
933 * handler as well. For that case, we may need up to twice the
936 if (td->o.verify != VERIFY_NONE &&
937 (td_write(td) && td->o.verify_backlog))
938 total_bytes += td->o.size;
940 /* In trimwrite mode, each byte is trimmed and then written, so
941 * allow total_bytes to be twice as big */
942 if (td_trimwrite(td))
943 total_bytes += td->total_io_size;
945 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
946 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
948 struct timespec comp_time;
953 check_update_rusage(td);
955 if (td->terminate || td->done)
960 if (runtime_exceeded(td, &td->ts_cache)) {
961 __update_ts_cache(td);
962 if (runtime_exceeded(td, &td->ts_cache)) {
963 fio_mark_td_terminate(td);
968 if (flow_threshold_exceeded(td))
972 * Break if we exceeded the bytes. The exception is time
973 * based runs, but we still need to break out of the loop
974 * for those to run verification, if enabled.
975 * Jobs read from iolog do not use this stop condition.
977 if (bytes_issued >= total_bytes &&
978 !td->o.read_iolog_file &&
979 (!td->o.time_based ||
980 (td->o.time_based && td->o.verify != VERIFY_NONE)))
984 if (IS_ERR_OR_NULL(io_u)) {
985 int err = PTR_ERR(io_u);
993 if (td->o.latency_target)
998 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
999 populate_verify_io_u(td, io_u);
1004 * Add verification end_io handler if:
1005 * - Asked to verify (!td_rw(td))
1006 * - Or the io_u is from our verify list (mixed write/ver)
1008 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1009 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1011 if (!td->o.verify_pattern_bytes) {
1012 io_u->rand_seed = __rand(&td->verify_state);
1013 if (sizeof(int) != sizeof(long *))
1014 io_u->rand_seed *= __rand(&td->verify_state);
1017 if (verify_state_should_stop(td, io_u)) {
1022 if (td->o.verify_async)
1023 io_u->end_io = verify_io_u_async;
1025 io_u->end_io = verify_io_u;
1026 td_set_runstate(td, TD_VERIFYING);
1027 } else if (in_ramp_time(td))
1028 td_set_runstate(td, TD_RAMP);
1030 td_set_runstate(td, TD_RUNNING);
1033 * Always log IO before it's issued, so we know the specific
1034 * order of it. The logged unit will track when the IO has
1037 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1039 td->o.verify != VERIFY_NONE &&
1040 !td->o.experimental_verify)
1041 log_io_piece(td, io_u);
1043 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1044 const unsigned long long blen = io_u->xfer_buflen;
1045 const enum fio_ddir __ddir = acct_ddir(io_u);
1050 workqueue_enqueue(&td->io_wq, &io_u->work);
1053 if (ddir_rw(__ddir)) {
1054 td->io_issues[__ddir]++;
1055 td->io_issue_bytes[__ddir] += blen;
1056 td->rate_io_issue_bytes[__ddir] += blen;
1059 if (should_check_rate(td))
1060 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1063 ret = io_u_submit(td, io_u);
1065 if (should_check_rate(td))
1066 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1068 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1072 * See if we need to complete some commands. Note that
1073 * we can get BUSY even without IO queued, if the
1074 * system is resource starved.
1077 full = queue_full(td) ||
1078 (ret == FIO_Q_BUSY && td->cur_depth);
1079 if (full || io_in_polling(td))
1080 ret = wait_for_completions(td, &comp_time);
1084 if (!ddir_rw_sum(td->bytes_done) &&
1085 !td_ioengine_flagged(td, FIO_NOIO))
1088 if (!in_ramp_time(td) && should_check_rate(td)) {
1089 if (check_min_rate(td, &comp_time)) {
1090 if (exitall_on_terminate || td->o.exitall_error)
1091 fio_terminate_threads(td->groupid);
1092 td_verror(td, EIO, "check_min_rate");
1096 if (!in_ramp_time(td) && td->o.latency_target)
1097 lat_target_check(td);
1099 if (ddir_rw(ddir) && td->o.thinktime)
1100 handle_thinktime(td, ddir);
1103 check_update_rusage(td);
1105 if (td->trim_entries)
1106 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1108 if (td->o.fill_device && td->error == ENOSPC) {
1110 fio_mark_td_terminate(td);
1115 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1116 workqueue_flush(&td->io_wq);
1122 ret = io_u_queued_complete(td, i);
1123 if (td->o.fill_device && td->error == ENOSPC)
1127 if (should_fsync(td) && td->o.end_fsync) {
1128 td_set_runstate(td, TD_FSYNCING);
1130 for_each_file(td, f, i) {
1131 if (!fio_file_fsync(td, f))
1134 log_err("fio: end_fsync failed for file %s\n",
1139 cleanup_pending_aio(td);
1142 * stop job if we failed doing any IO
1144 if (!ddir_rw_sum(td->this_io_bytes))
1147 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1148 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1151 static void free_file_completion_logging(struct thread_data *td)
1156 for_each_file(td, f, i) {
1157 if (!f->last_write_comp)
1159 sfree(f->last_write_comp);
1163 static int init_file_completion_logging(struct thread_data *td,
1169 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1172 for_each_file(td, f, i) {
1173 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1174 if (!f->last_write_comp)
1181 free_file_completion_logging(td);
1182 log_err("fio: failed to alloc write comp data\n");
1186 static void cleanup_io_u(struct thread_data *td)
1190 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1192 if (td->io_ops->io_u_free)
1193 td->io_ops->io_u_free(td, io_u);
1195 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1200 io_u_rexit(&td->io_u_requeues);
1201 io_u_qexit(&td->io_u_freelist, false);
1202 io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1204 free_file_completion_logging(td);
1207 static int init_io_u(struct thread_data *td)
1210 int cl_align, i, max_units;
1213 max_units = td->o.iodepth;
1216 err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1217 err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1218 err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1221 log_err("fio: failed setting up IO queues\n");
1225 cl_align = os_cache_line_size();
1227 for (i = 0; i < max_units; i++) {
1233 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1235 log_err("fio: unable to allocate aligned memory\n");
1240 memset(io_u, 0, sizeof(*io_u));
1241 INIT_FLIST_HEAD(&io_u->verify_list);
1242 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1245 io_u->flags = IO_U_F_FREE;
1246 io_u_qpush(&td->io_u_freelist, io_u);
1249 * io_u never leaves this stack, used for iteration of all
1252 io_u_qpush(&td->io_u_all, io_u);
1254 if (td->io_ops->io_u_init) {
1255 int ret = td->io_ops->io_u_init(td, io_u);
1258 log_err("fio: failed to init engine data: %d\n", ret);
1264 init_io_u_buffers(td);
1266 if (init_file_completion_logging(td, max_units))
1272 int init_io_u_buffers(struct thread_data *td)
1275 unsigned long long max_bs, min_write;
1280 max_units = td->o.iodepth;
1281 max_bs = td_max_bs(td);
1282 min_write = td->o.min_bs[DDIR_WRITE];
1283 td->orig_buffer_size = (unsigned long long) max_bs
1284 * (unsigned long long) max_units;
1286 if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1290 * if we may later need to do address alignment, then add any
1291 * possible adjustment here so that we don't cause a buffer
1292 * overflow later. this adjustment may be too much if we get
1293 * lucky and the allocator gives us an aligned address.
1295 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1296 td_ioengine_flagged(td, FIO_RAWIO))
1297 td->orig_buffer_size += page_mask + td->o.mem_align;
1299 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1300 unsigned long long bs;
1302 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1303 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1306 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1307 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1311 if (data_xfer && allocate_io_mem(td))
1314 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1315 td_ioengine_flagged(td, FIO_RAWIO))
1316 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1318 p = td->orig_buffer;
1320 for (i = 0; i < max_units; i++) {
1321 io_u = td->io_u_all.io_us[i];
1322 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1326 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1329 io_u_fill_buffer(td, io_u, min_write, max_bs);
1330 if (td_write(td) && td->o.verify_pattern_bytes) {
1332 * Fill the buffer with the pattern if we are
1333 * going to be doing writes.
1335 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1345 * This function is Linux specific.
1346 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1348 static int switch_ioscheduler(struct thread_data *td)
1350 #ifdef FIO_HAVE_IOSCHED_SWITCH
1351 char tmp[256], tmp2[128], *p;
1355 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1358 assert(td->files && td->files[0]);
1359 sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1361 f = fopen(tmp, "r+");
1363 if (errno == ENOENT) {
1364 log_err("fio: os or kernel doesn't support IO scheduler"
1368 td_verror(td, errno, "fopen iosched");
1375 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1376 if (ferror(f) || ret != 1) {
1377 td_verror(td, errno, "fwrite");
1385 * Read back and check that the selected scheduler is now the default.
1387 ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1388 if (ferror(f) || ret < 0) {
1389 td_verror(td, errno, "fread");
1395 * either a list of io schedulers or "none\n" is expected. Strip the
1402 * Write to "none" entry doesn't fail, so check the result here.
1404 if (!strcmp(tmp, "none")) {
1405 log_err("fio: io scheduler is not tunable\n");
1410 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1411 if (!strstr(tmp, tmp2)) {
1412 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1413 td_verror(td, EINVAL, "iosched_switch");
1425 static bool keep_running(struct thread_data *td)
1427 unsigned long long limit;
1433 if (td->o.time_based)
1439 if (exceeds_number_ios(td))
1443 limit = td->o.io_size;
1447 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1451 * If the difference is less than the maximum IO size, we
1454 diff = limit - ddir_rw_sum(td->io_bytes);
1455 if (diff < td_max_bs(td))
1458 if (fio_files_done(td) && !td->o.io_size)
1467 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1469 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1473 str = malloc(newlen);
1474 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1476 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1479 log_err("fio: exec of cmd <%s> failed\n", str);
1486 * Dry run to compute correct state of numberio for verification.
1488 static uint64_t do_dry_run(struct thread_data *td)
1490 td_set_runstate(td, TD_RUNNING);
1492 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1493 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1497 if (td->terminate || td->done)
1500 io_u = get_io_u(td);
1501 if (IS_ERR_OR_NULL(io_u))
1504 io_u_set(td, io_u, IO_U_F_FLIGHT);
1507 if (ddir_rw(acct_ddir(io_u)))
1508 td->io_issues[acct_ddir(io_u)]++;
1509 if (ddir_rw(io_u->ddir)) {
1510 io_u_mark_depth(td, 1);
1511 td->ts.total_io_u[io_u->ddir]++;
1514 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1516 td->o.verify != VERIFY_NONE &&
1517 !td->o.experimental_verify)
1518 log_io_piece(td, io_u);
1520 ret = io_u_sync_complete(td, io_u);
1524 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1528 struct thread_data *td;
1529 struct sk_out *sk_out;
1533 * Entry point for the thread based jobs. The process based jobs end up
1534 * here as well, after a little setup.
1536 static void *thread_main(void *data)
1538 struct fork_data *fd = data;
1539 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1540 struct thread_data *td = fd->td;
1541 struct thread_options *o = &td->o;
1542 struct sk_out *sk_out = fd->sk_out;
1543 uint64_t bytes_done[DDIR_RWDIR_CNT];
1544 int deadlock_loop_cnt;
1548 sk_out_assign(sk_out);
1551 if (!o->use_thread) {
1557 fio_local_clock_init();
1559 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1562 fio_server_send_start(td);
1564 INIT_FLIST_HEAD(&td->io_log_list);
1565 INIT_FLIST_HEAD(&td->io_hist_list);
1566 INIT_FLIST_HEAD(&td->verify_list);
1567 INIT_FLIST_HEAD(&td->trim_list);
1568 td->io_hist_tree = RB_ROOT;
1570 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1572 td_verror(td, ret, "mutex_cond_init_pshared");
1575 ret = cond_init_pshared(&td->verify_cond);
1577 td_verror(td, ret, "mutex_cond_pshared");
1581 td_set_runstate(td, TD_INITIALIZED);
1582 dprint(FD_MUTEX, "up startup_sem\n");
1583 fio_sem_up(startup_sem);
1584 dprint(FD_MUTEX, "wait on td->sem\n");
1585 fio_sem_down(td->sem);
1586 dprint(FD_MUTEX, "done waiting on td->sem\n");
1589 * A new gid requires privilege, so we need to do this before setting
1592 if (o->gid != -1U && setgid(o->gid)) {
1593 td_verror(td, errno, "setgid");
1596 if (o->uid != -1U && setuid(o->uid)) {
1597 td_verror(td, errno, "setuid");
1601 td_zone_gen_index(td);
1604 * Do this early, we don't want the compress threads to be limited
1605 * to the same CPUs as the IO workers. So do this before we set
1606 * any potential CPU affinity
1608 if (iolog_compress_init(td, sk_out))
1612 * If we have a gettimeofday() thread, make sure we exclude that
1613 * thread from this job
1616 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1619 * Set affinity first, in case it has an impact on the memory
1622 if (fio_option_is_set(o, cpumask)) {
1623 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1624 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1626 log_err("fio: no CPUs set\n");
1627 log_err("fio: Try increasing number of available CPUs\n");
1628 td_verror(td, EINVAL, "cpus_split");
1632 ret = fio_setaffinity(td->pid, o->cpumask);
1634 td_verror(td, errno, "cpu_set_affinity");
1639 #ifdef CONFIG_LIBNUMA
1640 /* numa node setup */
1641 if (fio_option_is_set(o, numa_cpunodes) ||
1642 fio_option_is_set(o, numa_memnodes)) {
1643 struct bitmask *mask;
1645 if (numa_available() < 0) {
1646 td_verror(td, errno, "Does not support NUMA API\n");
1650 if (fio_option_is_set(o, numa_cpunodes)) {
1651 mask = numa_parse_nodestring(o->numa_cpunodes);
1652 ret = numa_run_on_node_mask(mask);
1653 numa_free_nodemask(mask);
1655 td_verror(td, errno, \
1656 "numa_run_on_node_mask failed\n");
1661 if (fio_option_is_set(o, numa_memnodes)) {
1663 if (o->numa_memnodes)
1664 mask = numa_parse_nodestring(o->numa_memnodes);
1666 switch (o->numa_mem_mode) {
1667 case MPOL_INTERLEAVE:
1668 numa_set_interleave_mask(mask);
1671 numa_set_membind(mask);
1674 numa_set_localalloc();
1676 case MPOL_PREFERRED:
1677 numa_set_preferred(o->numa_mem_prefer_node);
1685 numa_free_nodemask(mask);
1691 if (fio_pin_memory(td))
1695 * May alter parameters that init_io_u() will use, so we need to
1698 if (!init_iolog(td))
1707 if (td->io_ops->post_init && td->io_ops->post_init(td))
1710 if (o->verify_async && verify_async_init(td))
1713 if (fio_option_is_set(o, ioprio) ||
1714 fio_option_is_set(o, ioprio_class)) {
1715 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1717 td_verror(td, errno, "ioprio_set");
1722 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1726 if (nice(o->nice) == -1 && errno != 0) {
1727 td_verror(td, errno, "nice");
1731 if (o->ioscheduler && switch_ioscheduler(td))
1734 if (!o->create_serialize && setup_files(td))
1737 if (!init_random_map(td))
1740 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1743 if (o->pre_read && !pre_read_files(td))
1746 fio_verify_init(td);
1748 if (rate_submit_init(td, sk_out))
1751 set_epoch_time(td, o->log_unix_epoch);
1752 fio_getrusage(&td->ru_start);
1753 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1754 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1755 memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1757 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1758 o->ratemin[DDIR_TRIM]) {
1759 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1760 sizeof(td->bw_sample_time));
1761 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1762 sizeof(td->bw_sample_time));
1763 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1764 sizeof(td->bw_sample_time));
1767 memset(bytes_done, 0, sizeof(bytes_done));
1768 clear_state = false;
1770 while (keep_running(td)) {
1771 uint64_t verify_bytes;
1773 fio_gettime(&td->start, NULL);
1774 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1777 clear_io_state(td, 0);
1779 if (o->unlink_each_loop && unlink_all_files(td))
1783 prune_io_piece_log(td);
1785 if (td->o.verify_only && td_write(td))
1786 verify_bytes = do_dry_run(td);
1788 do_io(td, bytes_done);
1790 if (!ddir_rw_sum(bytes_done)) {
1791 fio_mark_td_terminate(td);
1794 verify_bytes = bytes_done[DDIR_WRITE] +
1795 bytes_done[DDIR_TRIM];
1800 * If we took too long to shut down, the main thread could
1801 * already consider us reaped/exited. If that happens, break
1804 if (td->runstate >= TD_EXITED)
1810 * Make sure we've successfully updated the rusage stats
1811 * before waiting on the stat mutex. Otherwise we could have
1812 * the stat thread holding stat mutex and waiting for
1813 * the rusage_sem, which would never get upped because
1814 * this thread is waiting for the stat mutex.
1816 deadlock_loop_cnt = 0;
1818 check_update_rusage(td);
1819 if (!fio_sem_down_trylock(stat_sem))
1822 if (deadlock_loop_cnt++ > 5000) {
1823 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1824 td->error = EDEADLK;
1829 if (td_read(td) && td->io_bytes[DDIR_READ])
1830 update_runtime(td, elapsed_us, DDIR_READ);
1831 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1832 update_runtime(td, elapsed_us, DDIR_WRITE);
1833 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1834 update_runtime(td, elapsed_us, DDIR_TRIM);
1835 fio_gettime(&td->start, NULL);
1836 fio_sem_up(stat_sem);
1838 if (td->error || td->terminate)
1841 if (!o->do_verify ||
1842 o->verify == VERIFY_NONE ||
1843 td_ioengine_flagged(td, FIO_UNIDIR))
1846 clear_io_state(td, 0);
1848 fio_gettime(&td->start, NULL);
1850 do_verify(td, verify_bytes);
1853 * See comment further up for why this is done here.
1855 check_update_rusage(td);
1857 fio_sem_down(stat_sem);
1858 update_runtime(td, elapsed_us, DDIR_READ);
1859 fio_gettime(&td->start, NULL);
1860 fio_sem_up(stat_sem);
1862 if (td->error || td->terminate)
1867 * Acquire this lock if we were doing overlap checking in
1868 * offload mode so that we don't clean up this job while
1869 * another thread is checking its io_u's for overlap
1871 if (td_offload_overlap(td))
1872 pthread_mutex_lock(&overlap_check);
1873 td_set_runstate(td, TD_FINISHING);
1874 if (td_offload_overlap(td))
1875 pthread_mutex_unlock(&overlap_check);
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);