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 #ifndef FIO_NO_HAVE_SHM_H
52 #include "lib/memalign.h"
54 #include "lib/getrusage.h"
57 #include "workqueue.h"
58 #include "lib/mountcheck.h"
59 #include "rate-submit.h"
60 #include "helper_thread.h"
62 static struct fio_mutex *startup_mutex;
63 static struct flist_head *cgroup_list;
64 static char *cgroup_mnt;
65 static int exit_value;
66 static volatile int fio_abort;
67 static unsigned int nr_process = 0;
68 static unsigned int nr_thread = 0;
70 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
73 unsigned int thread_number = 0;
74 unsigned int stat_number = 0;
77 unsigned long done_secs = 0;
79 #define JOB_START_TIMEOUT (5 * 1000)
81 static void sig_int(int sig)
85 fio_server_got_signal(sig);
87 log_info("\nfio: terminating on signal %d\n", sig);
92 fio_terminate_threads(TERMINATE_ALL);
96 void sig_show_status(int sig)
98 show_running_run_stats();
101 static void set_sig_handlers(void)
103 struct sigaction act;
105 memset(&act, 0, sizeof(act));
106 act.sa_handler = sig_int;
107 act.sa_flags = SA_RESTART;
108 sigaction(SIGINT, &act, NULL);
110 memset(&act, 0, sizeof(act));
111 act.sa_handler = sig_int;
112 act.sa_flags = SA_RESTART;
113 sigaction(SIGTERM, &act, NULL);
115 /* Windows uses SIGBREAK as a quit signal from other applications */
117 memset(&act, 0, sizeof(act));
118 act.sa_handler = sig_int;
119 act.sa_flags = SA_RESTART;
120 sigaction(SIGBREAK, &act, NULL);
123 memset(&act, 0, sizeof(act));
124 act.sa_handler = sig_show_status;
125 act.sa_flags = SA_RESTART;
126 sigaction(SIGUSR1, &act, NULL);
129 memset(&act, 0, sizeof(act));
130 act.sa_handler = sig_int;
131 act.sa_flags = SA_RESTART;
132 sigaction(SIGPIPE, &act, NULL);
137 * Check if we are above the minimum rate given.
139 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
142 unsigned long long bytes = 0;
143 unsigned long iops = 0;
146 unsigned int ratemin = 0;
147 unsigned int rate_iops = 0;
148 unsigned int rate_iops_min = 0;
150 assert(ddir_rw(ddir));
152 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
156 * allow a 2 second settle period in the beginning
158 if (mtime_since(&td->start, now) < 2000)
161 iops += td->this_io_blocks[ddir];
162 bytes += td->this_io_bytes[ddir];
163 ratemin += td->o.ratemin[ddir];
164 rate_iops += td->o.rate_iops[ddir];
165 rate_iops_min += td->o.rate_iops_min[ddir];
168 * if rate blocks is set, sample is running
170 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
171 spent = mtime_since(&td->lastrate[ddir], now);
172 if (spent < td->o.ratecycle)
175 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
177 * check bandwidth specified rate
179 if (bytes < td->rate_bytes[ddir]) {
180 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
181 td->o.name, ratemin, bytes);
185 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
189 if (rate < ratemin ||
190 bytes < td->rate_bytes[ddir]) {
191 log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
192 td->o.name, ratemin, rate);
198 * checks iops specified rate
200 if (iops < rate_iops) {
201 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
202 td->o.name, rate_iops, iops);
206 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
210 if (rate < rate_iops_min ||
211 iops < td->rate_blocks[ddir]) {
212 log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
213 td->o.name, rate_iops_min, rate);
220 td->rate_bytes[ddir] = bytes;
221 td->rate_blocks[ddir] = iops;
222 memcpy(&td->lastrate[ddir], now, sizeof(*now));
226 static bool check_min_rate(struct thread_data *td, struct timespec *now)
230 if (td->bytes_done[DDIR_READ])
231 ret |= __check_min_rate(td, now, DDIR_READ);
232 if (td->bytes_done[DDIR_WRITE])
233 ret |= __check_min_rate(td, now, DDIR_WRITE);
234 if (td->bytes_done[DDIR_TRIM])
235 ret |= __check_min_rate(td, now, DDIR_TRIM);
241 * When job exits, we can cancel the in-flight IO if we are using async
242 * io. Attempt to do so.
244 static void cleanup_pending_aio(struct thread_data *td)
249 * get immediately available events, if any
251 r = io_u_queued_complete(td, 0);
256 * now cancel remaining active events
258 if (td->io_ops->cancel) {
262 io_u_qiter(&td->io_u_all, io_u, i) {
263 if (io_u->flags & IO_U_F_FLIGHT) {
264 r = td->io_ops->cancel(td, io_u);
272 r = io_u_queued_complete(td, td->cur_depth);
276 * Helper to handle the final sync of a file. Works just like the normal
277 * io path, just does everything sync.
279 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
281 struct io_u *io_u = __get_io_u(td);
287 io_u->ddir = DDIR_SYNC;
290 if (td_io_prep(td, io_u)) {
296 ret = td_io_queue(td, io_u);
298 td_verror(td, io_u->error, "td_io_queue");
301 } else if (ret == FIO_Q_QUEUED) {
302 if (td_io_commit(td))
304 if (io_u_queued_complete(td, 1) < 0)
306 } else if (ret == FIO_Q_COMPLETED) {
308 td_verror(td, io_u->error, "td_io_queue");
312 if (io_u_sync_complete(td, io_u) < 0)
314 } else if (ret == FIO_Q_BUSY) {
315 if (td_io_commit(td))
323 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
327 if (fio_file_open(f))
328 return fio_io_sync(td, f);
330 if (td_io_open_file(td, f))
333 ret = fio_io_sync(td, f);
334 td_io_close_file(td, f);
338 static inline void __update_ts_cache(struct thread_data *td)
340 fio_gettime(&td->ts_cache, NULL);
343 static inline void update_ts_cache(struct thread_data *td)
345 if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
346 __update_ts_cache(td);
349 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
351 if (in_ramp_time(td))
355 if (utime_since(&td->epoch, t) >= td->o.timeout)
362 * We need to update the runtime consistently in ms, but keep a running
363 * tally of the current elapsed time in microseconds for sub millisecond
366 static inline void update_runtime(struct thread_data *td,
367 unsigned long long *elapsed_us,
368 const enum fio_ddir ddir)
370 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
373 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
374 elapsed_us[ddir] += utime_since_now(&td->start);
375 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
378 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
383 if (ret < 0 || td->error) {
385 enum error_type_bit eb;
390 eb = td_error_type(ddir, err);
391 if (!(td->o.continue_on_error & (1 << eb)))
394 if (td_non_fatal_error(td, eb, err)) {
396 * Continue with the I/Os in case of
399 update_error_count(td, err);
403 } else if (td->o.fill_device && err == ENOSPC) {
405 * We expect to hit this error if
406 * fill_device option is set.
409 fio_mark_td_terminate(td);
413 * Stop the I/O in case of a fatal
416 update_error_count(td, err);
424 static void check_update_rusage(struct thread_data *td)
426 if (td->update_rusage) {
427 td->update_rusage = 0;
428 update_rusage_stat(td);
429 fio_mutex_up(td->rusage_sem);
433 static int wait_for_completions(struct thread_data *td, struct timespec *time)
435 const int full = queue_full(td);
439 if (td->flags & TD_F_REGROW_LOGS)
440 return io_u_quiesce(td);
443 * if the queue is full, we MUST reap at least 1 event
445 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
446 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
449 if (time && (__should_check_rate(td, DDIR_READ) ||
450 __should_check_rate(td, DDIR_WRITE) ||
451 __should_check_rate(td, DDIR_TRIM)))
452 fio_gettime(time, NULL);
455 ret = io_u_queued_complete(td, min_evts);
458 } while (full && (td->cur_depth > td->o.iodepth_low));
463 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
464 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
465 struct timespec *comp_time)
470 case FIO_Q_COMPLETED:
473 clear_io_u(td, io_u);
474 } else if (io_u->resid) {
475 int bytes = io_u->xfer_buflen - io_u->resid;
476 struct fio_file *f = io_u->file;
479 *bytes_issued += bytes;
482 trim_io_piece(td, io_u);
489 unlog_io_piece(td, io_u);
490 td_verror(td, EIO, "full resid");
495 io_u->xfer_buflen = io_u->resid;
496 io_u->xfer_buf += bytes;
497 io_u->offset += bytes;
499 if (ddir_rw(io_u->ddir))
500 td->ts.short_io_u[io_u->ddir]++;
502 if (io_u->offset == f->real_file_size)
505 requeue_io_u(td, &io_u);
508 if (comp_time && (__should_check_rate(td, DDIR_READ) ||
509 __should_check_rate(td, DDIR_WRITE) ||
510 __should_check_rate(td, DDIR_TRIM)))
511 fio_gettime(comp_time, NULL);
513 *ret = io_u_sync_complete(td, io_u);
518 if (td->flags & TD_F_REGROW_LOGS)
522 * when doing I/O (not when verifying),
523 * check for any errors that are to be ignored
531 * if the engine doesn't have a commit hook,
532 * the io_u is really queued. if it does have such
533 * a hook, it has to call io_u_queued() itself.
535 if (td->io_ops->commit == NULL)
536 io_u_queued(td, io_u);
538 *bytes_issued += io_u->xfer_buflen;
542 unlog_io_piece(td, io_u);
543 requeue_io_u(td, &io_u);
544 ret2 = td_io_commit(td);
550 td_verror(td, -(*ret), "td_io_queue");
554 if (break_on_this_error(td, ddir, ret))
560 static inline bool io_in_polling(struct thread_data *td)
562 return !td->o.iodepth_batch_complete_min &&
563 !td->o.iodepth_batch_complete_max;
566 * Unlinks files from thread data fio_file structure
568 static int unlink_all_files(struct thread_data *td)
574 for_each_file(td, f, i) {
575 if (f->filetype != FIO_TYPE_FILE)
577 ret = td_io_unlink_file(td, f);
583 td_verror(td, ret, "unlink_all_files");
589 * Check if io_u will overlap an in-flight IO in the queue
591 static bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
594 struct io_u *check_io_u;
595 unsigned long long x1, x2, y1, y2;
599 x2 = io_u->offset + io_u->buflen;
601 io_u_qiter(q, check_io_u, i) {
602 if (check_io_u->flags & IO_U_F_FLIGHT) {
603 y1 = check_io_u->offset;
604 y2 = check_io_u->offset + check_io_u->buflen;
606 if (x1 < y2 && y1 < x2) {
608 dprint(FD_IO, "in-flight overlap: %llu/%lu, %llu/%lu\n",
610 y1, check_io_u->buflen);
619 static int io_u_submit(struct thread_data *td, struct io_u *io_u)
622 * Check for overlap if the user asked us to, and we have
623 * at least one IO in flight besides this one.
625 if (td->o.serialize_overlap && td->cur_depth > 1 &&
626 in_flight_overlap(&td->io_u_all, io_u))
629 return td_io_queue(td, io_u);
633 * The main verify engine. Runs over the writes we previously submitted,
634 * reads the blocks back in, and checks the crc/md5 of the data.
636 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
643 dprint(FD_VERIFY, "starting loop\n");
646 * sync io first and invalidate cache, to make sure we really
649 for_each_file(td, f, i) {
650 if (!fio_file_open(f))
652 if (fio_io_sync(td, f))
654 if (file_invalidate_cache(td, f))
658 check_update_rusage(td);
664 * verify_state needs to be reset before verification
665 * proceeds so that expected random seeds match actual
666 * random seeds in headers. The main loop will reset
667 * all random number generators if randrepeat is set.
669 if (!td->o.rand_repeatable)
670 td_fill_verify_state_seed(td);
672 td_set_runstate(td, TD_VERIFYING);
675 while (!td->terminate) {
680 check_update_rusage(td);
682 if (runtime_exceeded(td, &td->ts_cache)) {
683 __update_ts_cache(td);
684 if (runtime_exceeded(td, &td->ts_cache)) {
685 fio_mark_td_terminate(td);
690 if (flow_threshold_exceeded(td))
693 if (!td->o.experimental_verify) {
694 io_u = __get_io_u(td);
698 if (get_next_verify(td, io_u)) {
703 if (td_io_prep(td, io_u)) {
708 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
711 while ((io_u = get_io_u(td)) != NULL) {
712 if (IS_ERR_OR_NULL(io_u)) {
719 * We are only interested in the places where
720 * we wrote or trimmed IOs. Turn those into
721 * reads for verification purposes.
723 if (io_u->ddir == DDIR_READ) {
725 * Pretend we issued it for rwmix
728 td->io_issues[DDIR_READ]++;
731 } else if (io_u->ddir == DDIR_TRIM) {
732 io_u->ddir = DDIR_READ;
733 io_u_set(td, io_u, IO_U_F_TRIMMED);
735 } else if (io_u->ddir == DDIR_WRITE) {
736 io_u->ddir = DDIR_READ;
748 if (verify_state_should_stop(td, io_u)) {
753 if (td->o.verify_async)
754 io_u->end_io = verify_io_u_async;
756 io_u->end_io = verify_io_u;
759 if (!td->o.disable_slat)
760 fio_gettime(&io_u->start_time, NULL);
762 ret = io_u_submit(td, io_u);
764 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
768 * if we can queue more, do so. but check if there are
769 * completed io_u's first. Note that we can get BUSY even
770 * without IO queued, if the system is resource starved.
773 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
774 if (full || io_in_polling(td))
775 ret = wait_for_completions(td, NULL);
781 check_update_rusage(td);
784 min_events = td->cur_depth;
787 ret = io_u_queued_complete(td, min_events);
789 cleanup_pending_aio(td);
791 td_set_runstate(td, TD_RUNNING);
793 dprint(FD_VERIFY, "exiting loop\n");
796 static bool exceeds_number_ios(struct thread_data *td)
798 unsigned long long number_ios;
800 if (!td->o.number_ios)
803 number_ios = ddir_rw_sum(td->io_blocks);
804 number_ios += td->io_u_queued + td->io_u_in_flight;
806 return number_ios >= (td->o.number_ios * td->loops);
809 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
811 unsigned long long bytes, limit;
814 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
815 else if (td_write(td))
816 bytes = this_bytes[DDIR_WRITE];
817 else if (td_read(td))
818 bytes = this_bytes[DDIR_READ];
820 bytes = this_bytes[DDIR_TRIM];
823 limit = td->o.io_size;
828 return bytes >= limit || exceeds_number_ios(td);
831 static bool io_issue_bytes_exceeded(struct thread_data *td)
833 return io_bytes_exceeded(td, td->io_issue_bytes);
836 static bool io_complete_bytes_exceeded(struct thread_data *td)
838 return io_bytes_exceeded(td, td->this_io_bytes);
842 * used to calculate the next io time for rate control
845 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
847 uint64_t secs, remainder, bps, bytes, iops;
849 assert(!(td->flags & TD_F_CHILD));
850 bytes = td->rate_io_issue_bytes[ddir];
851 bps = td->rate_bps[ddir];
853 if (td->o.rate_process == RATE_PROCESS_POISSON) {
855 iops = bps / td->o.bs[ddir];
856 val = (int64_t) (1000000 / iops) *
857 -logf(__rand_0_1(&td->poisson_state[ddir]));
859 dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
860 (unsigned long long) 1000000 / val,
863 td->last_usec[ddir] += val;
864 return td->last_usec[ddir];
867 remainder = bytes % bps;
868 return remainder * 1000000 / bps + secs * 1000000;
875 * Main IO worker function. It retrieves io_u's to process and queues
876 * and reaps them, checking for rate and errors along the way.
878 * Returns number of bytes written and trimmed.
880 static void do_io(struct thread_data *td, uint64_t *bytes_done)
884 uint64_t total_bytes, bytes_issued = 0;
886 for (i = 0; i < DDIR_RWDIR_CNT; i++)
887 bytes_done[i] = td->bytes_done[i];
889 if (in_ramp_time(td))
890 td_set_runstate(td, TD_RAMP);
892 td_set_runstate(td, TD_RUNNING);
896 total_bytes = td->o.size;
898 * Allow random overwrite workloads to write up to io_size
899 * before starting verification phase as 'size' doesn't apply.
901 if (td_write(td) && td_random(td) && td->o.norandommap)
902 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
904 * If verify_backlog is enabled, we'll run the verify in this
905 * handler as well. For that case, we may need up to twice the
908 if (td->o.verify != VERIFY_NONE &&
909 (td_write(td) && td->o.verify_backlog))
910 total_bytes += td->o.size;
912 /* In trimwrite mode, each byte is trimmed and then written, so
913 * allow total_bytes to be twice as big */
914 if (td_trimwrite(td))
915 total_bytes += td->total_io_size;
917 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
918 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
920 struct timespec comp_time;
925 check_update_rusage(td);
927 if (td->terminate || td->done)
932 if (runtime_exceeded(td, &td->ts_cache)) {
933 __update_ts_cache(td);
934 if (runtime_exceeded(td, &td->ts_cache)) {
935 fio_mark_td_terminate(td);
940 if (flow_threshold_exceeded(td))
944 * Break if we exceeded the bytes. The exception is time
945 * based runs, but we still need to break out of the loop
946 * for those to run verification, if enabled.
948 if (bytes_issued >= total_bytes &&
949 (!td->o.time_based ||
950 (td->o.time_based && td->o.verify != VERIFY_NONE)))
954 if (IS_ERR_OR_NULL(io_u)) {
955 int err = PTR_ERR(io_u);
962 if (td->o.latency_target)
970 * Add verification end_io handler if:
971 * - Asked to verify (!td_rw(td))
972 * - Or the io_u is from our verify list (mixed write/ver)
974 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
975 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
977 if (!td->o.verify_pattern_bytes) {
978 io_u->rand_seed = __rand(&td->verify_state);
979 if (sizeof(int) != sizeof(long *))
980 io_u->rand_seed *= __rand(&td->verify_state);
983 if (verify_state_should_stop(td, io_u)) {
988 if (td->o.verify_async)
989 io_u->end_io = verify_io_u_async;
991 io_u->end_io = verify_io_u;
992 td_set_runstate(td, TD_VERIFYING);
993 } else if (in_ramp_time(td))
994 td_set_runstate(td, TD_RAMP);
996 td_set_runstate(td, TD_RUNNING);
999 * Always log IO before it's issued, so we know the specific
1000 * order of it. The logged unit will track when the IO has
1003 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1005 td->o.verify != VERIFY_NONE &&
1006 !td->o.experimental_verify)
1007 log_io_piece(td, io_u);
1009 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1010 const unsigned long blen = io_u->xfer_buflen;
1011 const enum fio_ddir ddir = acct_ddir(io_u);
1016 workqueue_enqueue(&td->io_wq, &io_u->work);
1019 if (ddir_rw(ddir)) {
1020 td->io_issues[ddir]++;
1021 td->io_issue_bytes[ddir] += blen;
1022 td->rate_io_issue_bytes[ddir] += blen;
1025 if (should_check_rate(td))
1026 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1029 ret = io_u_submit(td, io_u);
1031 if (should_check_rate(td))
1032 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1034 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1038 * See if we need to complete some commands. Note that
1039 * we can get BUSY even without IO queued, if the
1040 * system is resource starved.
1043 full = queue_full(td) ||
1044 (ret == FIO_Q_BUSY && td->cur_depth);
1045 if (full || io_in_polling(td))
1046 ret = wait_for_completions(td, &comp_time);
1050 if (!ddir_rw_sum(td->bytes_done) &&
1051 !td_ioengine_flagged(td, FIO_NOIO))
1054 if (!in_ramp_time(td) && should_check_rate(td)) {
1055 if (check_min_rate(td, &comp_time)) {
1056 if (exitall_on_terminate || td->o.exitall_error)
1057 fio_terminate_threads(td->groupid);
1058 td_verror(td, EIO, "check_min_rate");
1062 if (!in_ramp_time(td) && td->o.latency_target)
1063 lat_target_check(td);
1065 if (td->o.thinktime) {
1066 unsigned long long b;
1068 b = ddir_rw_sum(td->io_blocks);
1069 if (!(b % td->o.thinktime_blocks)) {
1074 if (td->o.thinktime_spin)
1075 usec_spin(td->o.thinktime_spin);
1077 left = td->o.thinktime - td->o.thinktime_spin;
1079 usec_sleep(td, left);
1084 check_update_rusage(td);
1086 if (td->trim_entries)
1087 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1089 if (td->o.fill_device && td->error == ENOSPC) {
1091 fio_mark_td_terminate(td);
1096 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1097 workqueue_flush(&td->io_wq);
1103 ret = io_u_queued_complete(td, i);
1104 if (td->o.fill_device && td->error == ENOSPC)
1108 if (should_fsync(td) && td->o.end_fsync) {
1109 td_set_runstate(td, TD_FSYNCING);
1111 for_each_file(td, f, i) {
1112 if (!fio_file_fsync(td, f))
1115 log_err("fio: end_fsync failed for file %s\n",
1120 cleanup_pending_aio(td);
1123 * stop job if we failed doing any IO
1125 if (!ddir_rw_sum(td->this_io_bytes))
1128 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1129 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1132 static void free_file_completion_logging(struct thread_data *td)
1137 for_each_file(td, f, i) {
1138 if (!f->last_write_comp)
1140 sfree(f->last_write_comp);
1144 static int init_file_completion_logging(struct thread_data *td,
1150 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1153 for_each_file(td, f, i) {
1154 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1155 if (!f->last_write_comp)
1162 free_file_completion_logging(td);
1163 log_err("fio: failed to alloc write comp data\n");
1167 static void cleanup_io_u(struct thread_data *td)
1171 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1173 if (td->io_ops->io_u_free)
1174 td->io_ops->io_u_free(td, io_u);
1176 fio_memfree(io_u, sizeof(*io_u));
1181 io_u_rexit(&td->io_u_requeues);
1182 io_u_qexit(&td->io_u_freelist);
1183 io_u_qexit(&td->io_u_all);
1185 free_file_completion_logging(td);
1188 static int init_io_u(struct thread_data *td)
1191 unsigned int max_bs, min_write;
1192 int cl_align, i, max_units;
1193 int data_xfer = 1, err;
1196 max_units = td->o.iodepth;
1197 max_bs = td_max_bs(td);
1198 min_write = td->o.min_bs[DDIR_WRITE];
1199 td->orig_buffer_size = (unsigned long long) max_bs
1200 * (unsigned long long) max_units;
1202 if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1206 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1207 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1208 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1211 log_err("fio: failed setting up IO queues\n");
1216 * if we may later need to do address alignment, then add any
1217 * possible adjustment here so that we don't cause a buffer
1218 * overflow later. this adjustment may be too much if we get
1219 * lucky and the allocator gives us an aligned address.
1221 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1222 td_ioengine_flagged(td, FIO_RAWIO))
1223 td->orig_buffer_size += page_mask + td->o.mem_align;
1225 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1228 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1229 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1232 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1233 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1237 if (data_xfer && allocate_io_mem(td))
1240 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1241 td_ioengine_flagged(td, FIO_RAWIO))
1242 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1244 p = td->orig_buffer;
1246 cl_align = os_cache_line_size();
1248 for (i = 0; i < max_units; i++) {
1254 ptr = fio_memalign(cl_align, sizeof(*io_u));
1256 log_err("fio: unable to allocate aligned memory\n");
1261 memset(io_u, 0, sizeof(*io_u));
1262 INIT_FLIST_HEAD(&io_u->verify_list);
1263 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1267 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1270 io_u_fill_buffer(td, io_u, min_write, max_bs);
1271 if (td_write(td) && td->o.verify_pattern_bytes) {
1273 * Fill the buffer with the pattern if we are
1274 * going to be doing writes.
1276 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1281 io_u->flags = IO_U_F_FREE;
1282 io_u_qpush(&td->io_u_freelist, io_u);
1285 * io_u never leaves this stack, used for iteration of all
1288 io_u_qpush(&td->io_u_all, io_u);
1290 if (td->io_ops->io_u_init) {
1291 int ret = td->io_ops->io_u_init(td, io_u);
1294 log_err("fio: failed to init engine data: %d\n", ret);
1302 if (init_file_completion_logging(td, max_units))
1309 * This function is Linux specific.
1310 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1312 static int switch_ioscheduler(struct thread_data *td)
1314 #ifdef FIO_HAVE_IOSCHED_SWITCH
1315 char tmp[256], tmp2[128];
1319 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1322 assert(td->files && td->files[0]);
1323 sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1325 f = fopen(tmp, "r+");
1327 if (errno == ENOENT) {
1328 log_err("fio: os or kernel doesn't support IO scheduler"
1332 td_verror(td, errno, "fopen iosched");
1339 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1340 if (ferror(f) || ret != 1) {
1341 td_verror(td, errno, "fwrite");
1349 * Read back and check that the selected scheduler is now the default.
1351 memset(tmp, 0, sizeof(tmp));
1352 ret = fread(tmp, sizeof(tmp), 1, f);
1353 if (ferror(f) || ret < 0) {
1354 td_verror(td, errno, "fread");
1359 * either a list of io schedulers or "none\n" is expected.
1361 tmp[strlen(tmp) - 1] = '\0';
1364 * Write to "none" entry doesn't fail, so check the result here.
1366 if (!strcmp(tmp, "none")) {
1367 log_err("fio: io scheduler is not tunable\n");
1372 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1373 if (!strstr(tmp, tmp2)) {
1374 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1375 td_verror(td, EINVAL, "iosched_switch");
1387 static bool keep_running(struct thread_data *td)
1389 unsigned long long limit;
1395 if (td->o.time_based)
1401 if (exceeds_number_ios(td))
1405 limit = td->o.io_size;
1409 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1413 * If the difference is less than the maximum IO size, we
1416 diff = limit - ddir_rw_sum(td->io_bytes);
1417 if (diff < td_max_bs(td))
1420 if (fio_files_done(td) && !td->o.io_size)
1429 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1431 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1435 str = malloc(newlen);
1436 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1438 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1441 log_err("fio: exec of cmd <%s> failed\n", str);
1448 * Dry run to compute correct state of numberio for verification.
1450 static uint64_t do_dry_run(struct thread_data *td)
1452 td_set_runstate(td, TD_RUNNING);
1454 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1455 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1459 if (td->terminate || td->done)
1462 io_u = get_io_u(td);
1463 if (IS_ERR_OR_NULL(io_u))
1466 io_u_set(td, io_u, IO_U_F_FLIGHT);
1469 if (ddir_rw(acct_ddir(io_u)))
1470 td->io_issues[acct_ddir(io_u)]++;
1471 if (ddir_rw(io_u->ddir)) {
1472 io_u_mark_depth(td, 1);
1473 td->ts.total_io_u[io_u->ddir]++;
1476 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1478 td->o.verify != VERIFY_NONE &&
1479 !td->o.experimental_verify)
1480 log_io_piece(td, io_u);
1482 ret = io_u_sync_complete(td, io_u);
1486 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1490 struct thread_data *td;
1491 struct sk_out *sk_out;
1495 * Entry point for the thread based jobs. The process based jobs end up
1496 * here as well, after a little setup.
1498 static void *thread_main(void *data)
1500 struct fork_data *fd = data;
1501 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1502 struct thread_data *td = fd->td;
1503 struct thread_options *o = &td->o;
1504 struct sk_out *sk_out = fd->sk_out;
1505 uint64_t bytes_done[DDIR_RWDIR_CNT];
1506 int deadlock_loop_cnt;
1507 bool clear_state, did_some_io;
1510 sk_out_assign(sk_out);
1513 if (!o->use_thread) {
1519 fio_local_clock_init(o->use_thread);
1521 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1524 fio_server_send_start(td);
1526 INIT_FLIST_HEAD(&td->io_log_list);
1527 INIT_FLIST_HEAD(&td->io_hist_list);
1528 INIT_FLIST_HEAD(&td->verify_list);
1529 INIT_FLIST_HEAD(&td->trim_list);
1530 INIT_FLIST_HEAD(&td->next_rand_list);
1531 td->io_hist_tree = RB_ROOT;
1533 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1535 td_verror(td, ret, "mutex_cond_init_pshared");
1538 ret = cond_init_pshared(&td->verify_cond);
1540 td_verror(td, ret, "mutex_cond_pshared");
1544 td_set_runstate(td, TD_INITIALIZED);
1545 dprint(FD_MUTEX, "up startup_mutex\n");
1546 fio_mutex_up(startup_mutex);
1547 dprint(FD_MUTEX, "wait on td->mutex\n");
1548 fio_mutex_down(td->mutex);
1549 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1552 * A new gid requires privilege, so we need to do this before setting
1555 if (o->gid != -1U && setgid(o->gid)) {
1556 td_verror(td, errno, "setgid");
1559 if (o->uid != -1U && setuid(o->uid)) {
1560 td_verror(td, errno, "setuid");
1565 * Do this early, we don't want the compress threads to be limited
1566 * to the same CPUs as the IO workers. So do this before we set
1567 * any potential CPU affinity
1569 if (iolog_compress_init(td, sk_out))
1573 * If we have a gettimeofday() thread, make sure we exclude that
1574 * thread from this job
1577 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1580 * Set affinity first, in case it has an impact on the memory
1583 if (fio_option_is_set(o, cpumask)) {
1584 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1585 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1587 log_err("fio: no CPUs set\n");
1588 log_err("fio: Try increasing number of available CPUs\n");
1589 td_verror(td, EINVAL, "cpus_split");
1593 ret = fio_setaffinity(td->pid, o->cpumask);
1595 td_verror(td, errno, "cpu_set_affinity");
1600 #ifdef CONFIG_LIBNUMA
1601 /* numa node setup */
1602 if (fio_option_is_set(o, numa_cpunodes) ||
1603 fio_option_is_set(o, numa_memnodes)) {
1604 struct bitmask *mask;
1606 if (numa_available() < 0) {
1607 td_verror(td, errno, "Does not support NUMA API\n");
1611 if (fio_option_is_set(o, numa_cpunodes)) {
1612 mask = numa_parse_nodestring(o->numa_cpunodes);
1613 ret = numa_run_on_node_mask(mask);
1614 numa_free_nodemask(mask);
1616 td_verror(td, errno, \
1617 "numa_run_on_node_mask failed\n");
1622 if (fio_option_is_set(o, numa_memnodes)) {
1624 if (o->numa_memnodes)
1625 mask = numa_parse_nodestring(o->numa_memnodes);
1627 switch (o->numa_mem_mode) {
1628 case MPOL_INTERLEAVE:
1629 numa_set_interleave_mask(mask);
1632 numa_set_membind(mask);
1635 numa_set_localalloc();
1637 case MPOL_PREFERRED:
1638 numa_set_preferred(o->numa_mem_prefer_node);
1646 numa_free_nodemask(mask);
1652 if (fio_pin_memory(td))
1656 * May alter parameters that init_io_u() will use, so we need to
1665 if (o->verify_async && verify_async_init(td))
1668 if (fio_option_is_set(o, ioprio) ||
1669 fio_option_is_set(o, ioprio_class)) {
1670 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1672 td_verror(td, errno, "ioprio_set");
1677 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1681 if (nice(o->nice) == -1 && errno != 0) {
1682 td_verror(td, errno, "nice");
1686 if (o->ioscheduler && switch_ioscheduler(td))
1689 if (!o->create_serialize && setup_files(td))
1695 if (init_random_map(td))
1698 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1702 if (pre_read_files(td) < 0)
1706 fio_verify_init(td);
1708 if (rate_submit_init(td, sk_out))
1711 set_epoch_time(td, o->log_unix_epoch);
1712 fio_getrusage(&td->ru_start);
1713 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1714 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1715 memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1717 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1718 o->ratemin[DDIR_TRIM]) {
1719 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1720 sizeof(td->bw_sample_time));
1721 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1722 sizeof(td->bw_sample_time));
1723 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1724 sizeof(td->bw_sample_time));
1727 memset(bytes_done, 0, sizeof(bytes_done));
1728 clear_state = false;
1729 did_some_io = false;
1731 while (keep_running(td)) {
1732 uint64_t verify_bytes;
1734 fio_gettime(&td->start, NULL);
1735 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1738 clear_io_state(td, 0);
1740 if (o->unlink_each_loop && unlink_all_files(td))
1744 prune_io_piece_log(td);
1746 if (td->o.verify_only && td_write(td))
1747 verify_bytes = do_dry_run(td);
1749 do_io(td, bytes_done);
1751 if (!ddir_rw_sum(bytes_done)) {
1752 fio_mark_td_terminate(td);
1755 verify_bytes = bytes_done[DDIR_WRITE] +
1756 bytes_done[DDIR_TRIM];
1761 * If we took too long to shut down, the main thread could
1762 * already consider us reaped/exited. If that happens, break
1765 if (td->runstate >= TD_EXITED)
1771 * Make sure we've successfully updated the rusage stats
1772 * before waiting on the stat mutex. Otherwise we could have
1773 * the stat thread holding stat mutex and waiting for
1774 * the rusage_sem, which would never get upped because
1775 * this thread is waiting for the stat mutex.
1777 deadlock_loop_cnt = 0;
1779 check_update_rusage(td);
1780 if (!fio_mutex_down_trylock(stat_mutex))
1783 if (deadlock_loop_cnt++ > 5000) {
1784 log_err("fio seems to be stuck grabbing stat_mutex, forcibly exiting\n");
1785 td->error = EDEADLK;
1790 if (td_read(td) && td->io_bytes[DDIR_READ])
1791 update_runtime(td, elapsed_us, DDIR_READ);
1792 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1793 update_runtime(td, elapsed_us, DDIR_WRITE);
1794 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1795 update_runtime(td, elapsed_us, DDIR_TRIM);
1796 fio_gettime(&td->start, NULL);
1797 fio_mutex_up(stat_mutex);
1799 if (td->error || td->terminate)
1802 if (!o->do_verify ||
1803 o->verify == VERIFY_NONE ||
1804 td_ioengine_flagged(td, FIO_UNIDIR))
1807 if (ddir_rw_sum(bytes_done))
1810 clear_io_state(td, 0);
1812 fio_gettime(&td->start, NULL);
1814 do_verify(td, verify_bytes);
1817 * See comment further up for why this is done here.
1819 check_update_rusage(td);
1821 fio_mutex_down(stat_mutex);
1822 update_runtime(td, elapsed_us, DDIR_READ);
1823 fio_gettime(&td->start, NULL);
1824 fio_mutex_up(stat_mutex);
1826 if (td->error || td->terminate)
1831 * If td ended up with no I/O when it should have had,
1832 * then something went wrong unless FIO_NOIO or FIO_DISKLESSIO.
1833 * (Are we not missing other flags that can be ignored ?)
1835 if ((td->o.size || td->o.io_size) && !ddir_rw_sum(bytes_done) &&
1836 !did_some_io && !td->o.create_only &&
1837 !(td_ioengine_flagged(td, FIO_NOIO) ||
1838 td_ioengine_flagged(td, FIO_DISKLESSIO)))
1839 log_err("%s: No I/O performed by %s, "
1840 "perhaps try --debug=io option for details?\n",
1841 td->o.name, td->io_ops->name);
1843 td_set_runstate(td, TD_FINISHING);
1845 update_rusage_stat(td);
1846 td->ts.total_run_time = mtime_since_now(&td->epoch);
1847 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1848 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1849 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1851 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1852 (td->o.verify != VERIFY_NONE && td_write(td)))
1853 verify_save_state(td->thread_number);
1855 fio_unpin_memory(td);
1857 td_writeout_logs(td, true);
1859 iolog_compress_exit(td);
1860 rate_submit_exit(td);
1862 if (o->exec_postrun)
1863 exec_string(o, o->exec_postrun, (const char *)"postrun");
1865 if (exitall_on_terminate || (o->exitall_error && td->error))
1866 fio_terminate_threads(td->groupid);
1870 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1873 if (o->verify_async)
1874 verify_async_exit(td);
1876 close_and_free_files(td);
1879 cgroup_shutdown(td, &cgroup_mnt);
1880 verify_free_state(td);
1882 if (td->zone_state_index) {
1885 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1886 free(td->zone_state_index[i]);
1887 free(td->zone_state_index);
1888 td->zone_state_index = NULL;
1891 if (fio_option_is_set(o, cpumask)) {
1892 ret = fio_cpuset_exit(&o->cpumask);
1894 td_verror(td, ret, "fio_cpuset_exit");
1898 * do this very late, it will log file closing as well
1900 if (o->write_iolog_file)
1901 write_iolog_close(td);
1903 td_set_runstate(td, TD_EXITED);
1906 * Do this last after setting our runstate to exited, so we
1907 * know that the stat thread is signaled.
1909 check_update_rusage(td);
1912 return (void *) (uintptr_t) td->error;
1916 * Run over the job map and reap the threads that have exited, if any.
1918 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1921 struct thread_data *td;
1922 unsigned int cputhreads, realthreads, pending;
1926 * reap exited threads (TD_EXITED -> TD_REAPED)
1928 realthreads = pending = cputhreads = 0;
1929 for_each_td(td, i) {
1932 if (!strcmp(td->o.ioengine, "cpuio"))
1941 if (td->runstate == TD_REAPED)
1943 if (td->o.use_thread) {
1944 if (td->runstate == TD_EXITED) {
1945 td_set_runstate(td, TD_REAPED);
1952 if (td->runstate == TD_EXITED)
1956 * check if someone quit or got killed in an unusual way
1958 ret = waitpid(td->pid, &status, flags);
1960 if (errno == ECHILD) {
1961 log_err("fio: pid=%d disappeared %d\n",
1962 (int) td->pid, td->runstate);
1964 td_set_runstate(td, TD_REAPED);
1968 } else if (ret == td->pid) {
1969 if (WIFSIGNALED(status)) {
1970 int sig = WTERMSIG(status);
1972 if (sig != SIGTERM && sig != SIGUSR2)
1973 log_err("fio: pid=%d, got signal=%d\n",
1974 (int) td->pid, sig);
1976 td_set_runstate(td, TD_REAPED);
1979 if (WIFEXITED(status)) {
1980 if (WEXITSTATUS(status) && !td->error)
1981 td->error = WEXITSTATUS(status);
1983 td_set_runstate(td, TD_REAPED);
1989 * If the job is stuck, do a forceful timeout of it and
1992 if (td->terminate &&
1993 td->runstate < TD_FSYNCING &&
1994 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1995 log_err("fio: job '%s' (state=%d) hasn't exited in "
1996 "%lu seconds, it appears to be stuck. Doing "
1997 "forceful exit of this job.\n",
1998 td->o.name, td->runstate,
1999 (unsigned long) time_since_now(&td->terminate_time));
2000 td_set_runstate(td, TD_REAPED);
2005 * thread is not dead, continue
2011 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2012 (*t_rate) -= ddir_rw_sum(td->o.rate);
2019 done_secs += mtime_since_now(&td->epoch) / 1000;
2020 profile_td_exit(td);
2023 if (*nr_running == cputhreads && !pending && realthreads)
2024 fio_terminate_threads(TERMINATE_ALL);
2027 static bool __check_trigger_file(void)
2034 if (stat(trigger_file, &sb))
2037 if (unlink(trigger_file) < 0)
2038 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2044 static bool trigger_timedout(void)
2046 if (trigger_timeout)
2047 if (time_since_genesis() >= trigger_timeout) {
2048 trigger_timeout = 0;
2055 void exec_trigger(const char *cmd)
2059 if (!cmd || cmd[0] == '\0')
2064 log_err("fio: failed executing %s trigger\n", cmd);
2067 void check_trigger_file(void)
2069 if (__check_trigger_file() || trigger_timedout()) {
2071 fio_clients_send_trigger(trigger_remote_cmd);
2073 verify_save_state(IO_LIST_ALL);
2074 fio_terminate_threads(TERMINATE_ALL);
2075 exec_trigger(trigger_cmd);
2080 static int fio_verify_load_state(struct thread_data *td)
2084 if (!td->o.verify_state)
2090 ret = fio_server_get_verify_state(td->o.name,
2091 td->thread_number - 1, &data);
2093 verify_assign_state(td, data);
2095 ret = verify_load_state(td, "local");
2100 static void do_usleep(unsigned int usecs)
2102 check_for_running_stats();
2103 check_trigger_file();
2107 static bool check_mount_writes(struct thread_data *td)
2112 if (!td_write(td) || td->o.allow_mounted_write)
2116 * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2117 * are mkfs'd and mounted.
2119 for_each_file(td, f, i) {
2120 #ifdef FIO_HAVE_CHARDEV_SIZE
2121 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2123 if (f->filetype != FIO_TYPE_BLOCK)
2126 if (device_is_mounted(f->file_name))
2132 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2136 static bool waitee_running(struct thread_data *me)
2138 const char *waitee = me->o.wait_for;
2139 const char *self = me->o.name;
2140 struct thread_data *td;
2146 for_each_td(td, i) {
2147 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2150 if (td->runstate < TD_EXITED) {
2151 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2153 runstate_to_name(td->runstate));
2158 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2163 * Main function for kicking off and reaping jobs, as needed.
2165 static void run_threads(struct sk_out *sk_out)
2167 struct thread_data *td;
2168 unsigned int i, todo, nr_running, nr_started;
2169 uint64_t m_rate, t_rate;
2172 if (fio_gtod_offload && fio_start_gtod_thread())
2175 fio_idle_prof_init();
2179 nr_thread = nr_process = 0;
2180 for_each_td(td, i) {
2181 if (check_mount_writes(td))
2183 if (td->o.use_thread)
2189 if (output_format & FIO_OUTPUT_NORMAL) {
2190 log_info("Starting ");
2192 log_info("%d thread%s", nr_thread,
2193 nr_thread > 1 ? "s" : "");
2197 log_info("%d process%s", nr_process,
2198 nr_process > 1 ? "es" : "");
2204 todo = thread_number;
2207 m_rate = t_rate = 0;
2209 for_each_td(td, i) {
2210 print_status_init(td->thread_number - 1);
2212 if (!td->o.create_serialize)
2215 if (fio_verify_load_state(td))
2219 * do file setup here so it happens sequentially,
2220 * we don't want X number of threads getting their
2221 * client data interspersed on disk
2223 if (setup_files(td)) {
2227 log_err("fio: pid=%d, err=%d/%s\n",
2228 (int) td->pid, td->error, td->verror);
2229 td_set_runstate(td, TD_REAPED);
2236 * for sharing to work, each job must always open
2237 * its own files. so close them, if we opened them
2240 for_each_file(td, f, j) {
2241 if (fio_file_open(f))
2242 td_io_close_file(td, f);
2247 /* start idle threads before io threads start to run */
2248 fio_idle_prof_start();
2253 struct thread_data *map[REAL_MAX_JOBS];
2254 struct timespec this_start;
2255 int this_jobs = 0, left;
2256 struct fork_data *fd;
2259 * create threads (TD_NOT_CREATED -> TD_CREATED)
2261 for_each_td(td, i) {
2262 if (td->runstate != TD_NOT_CREATED)
2266 * never got a chance to start, killed by other
2267 * thread for some reason
2269 if (td->terminate) {
2274 if (td->o.start_delay) {
2275 spent = utime_since_genesis();
2277 if (td->o.start_delay > spent)
2281 if (td->o.stonewall && (nr_started || nr_running)) {
2282 dprint(FD_PROCESS, "%s: stonewall wait\n",
2287 if (waitee_running(td)) {
2288 dprint(FD_PROCESS, "%s: waiting for %s\n",
2289 td->o.name, td->o.wait_for);
2295 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2296 td->update_rusage = 0;
2299 * Set state to created. Thread will transition
2300 * to TD_INITIALIZED when it's done setting up.
2302 td_set_runstate(td, TD_CREATED);
2303 map[this_jobs++] = td;
2306 fd = calloc(1, sizeof(*fd));
2308 fd->sk_out = sk_out;
2310 if (td->o.use_thread) {
2313 dprint(FD_PROCESS, "will pthread_create\n");
2314 ret = pthread_create(&td->thread, NULL,
2317 log_err("pthread_create: %s\n",
2323 ret = pthread_detach(td->thread);
2325 log_err("pthread_detach: %s",
2329 dprint(FD_PROCESS, "will fork\n");
2334 ret = (int)(uintptr_t)thread_main(fd);
2336 } else if (i == fio_debug_jobno)
2337 *fio_debug_jobp = pid;
2339 dprint(FD_MUTEX, "wait on startup_mutex\n");
2340 if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2341 log_err("fio: job startup hung? exiting.\n");
2342 fio_terminate_threads(TERMINATE_ALL);
2348 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2352 * Wait for the started threads to transition to
2355 fio_gettime(&this_start, NULL);
2357 while (left && !fio_abort) {
2358 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2363 for (i = 0; i < this_jobs; i++) {
2367 if (td->runstate == TD_INITIALIZED) {
2370 } else if (td->runstate >= TD_EXITED) {
2374 nr_running++; /* work-around... */
2380 log_err("fio: %d job%s failed to start\n", left,
2381 left > 1 ? "s" : "");
2382 for (i = 0; i < this_jobs; i++) {
2386 kill(td->pid, SIGTERM);
2392 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2394 for_each_td(td, i) {
2395 if (td->runstate != TD_INITIALIZED)
2398 if (in_ramp_time(td))
2399 td_set_runstate(td, TD_RAMP);
2401 td_set_runstate(td, TD_RUNNING);
2404 m_rate += ddir_rw_sum(td->o.ratemin);
2405 t_rate += ddir_rw_sum(td->o.rate);
2407 fio_mutex_up(td->mutex);
2410 reap_threads(&nr_running, &t_rate, &m_rate);
2416 while (nr_running) {
2417 reap_threads(&nr_running, &t_rate, &m_rate);
2421 fio_idle_prof_stop();
2426 static void free_disk_util(void)
2428 disk_util_prune_entries();
2429 helper_thread_destroy();
2432 int fio_backend(struct sk_out *sk_out)
2434 struct thread_data *td;
2438 if (load_profile(exec_profile))
2441 exec_profile = NULL;
2447 struct log_params p = {
2448 .log_type = IO_LOG_TYPE_BW,
2451 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2452 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2453 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2456 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2457 if (startup_mutex == NULL)
2462 helper_thread_create(startup_mutex, sk_out);
2464 cgroup_list = smalloc(sizeof(*cgroup_list));
2465 INIT_FLIST_HEAD(cgroup_list);
2467 run_threads(sk_out);
2469 helper_thread_exit();
2474 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2475 struct io_log *log = agg_io_log[i];
2477 flush_log(log, false);
2483 for_each_td(td, i) {
2485 if (td->ss.iops_data != NULL) {
2486 free(td->ss.iops_data);
2487 free(td->ss.bw_data);
2490 fio_options_free(td);
2491 if (td->rusage_sem) {
2492 fio_mutex_remove(td->rusage_sem);
2493 td->rusage_sem = NULL;
2495 fio_mutex_remove(td->mutex);
2500 cgroup_kill(cgroup_list);
2504 fio_mutex_remove(startup_mutex);