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 nr_segments = 0;
66 unsigned int cur_segment = 0;
67 unsigned int stat_number = 0;
69 unsigned long done_secs = 0;
70 #ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP
71 pthread_mutex_t overlap_check = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
73 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
76 #define JOB_START_TIMEOUT (5 * 1000)
78 static void sig_int(int sig)
82 fio_server_got_signal(sig);
84 log_info("\nfio: terminating on signal %d\n", sig);
89 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
94 static void sig_break(int sig)
96 struct thread_data *td;
102 * Windows terminates all job processes on SIGBREAK after the handler
103 * returns, so give them time to wrap-up and give stats
106 while (td->runstate < TD_EXITED)
112 void sig_show_status(int sig)
114 show_running_run_stats();
117 static void set_sig_handlers(void)
119 struct sigaction act;
121 memset(&act, 0, sizeof(act));
122 act.sa_handler = sig_int;
123 act.sa_flags = SA_RESTART;
124 sigaction(SIGINT, &act, NULL);
126 memset(&act, 0, sizeof(act));
127 act.sa_handler = sig_int;
128 act.sa_flags = SA_RESTART;
129 sigaction(SIGTERM, &act, NULL);
131 /* Windows uses SIGBREAK as a quit signal from other applications */
133 memset(&act, 0, sizeof(act));
134 act.sa_handler = sig_break;
135 act.sa_flags = SA_RESTART;
136 sigaction(SIGBREAK, &act, NULL);
139 memset(&act, 0, sizeof(act));
140 act.sa_handler = sig_show_status;
141 act.sa_flags = SA_RESTART;
142 sigaction(SIGUSR1, &act, NULL);
145 memset(&act, 0, sizeof(act));
146 act.sa_handler = sig_int;
147 act.sa_flags = SA_RESTART;
148 sigaction(SIGPIPE, &act, NULL);
153 * Check if we are above the minimum rate given.
155 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
158 unsigned long long current_rate_check_bytes = td->this_io_bytes[ddir];
159 unsigned long current_rate_check_blocks = td->this_io_blocks[ddir];
160 unsigned long long option_rate_bytes_min = td->o.ratemin[ddir];
161 unsigned int option_rate_iops_min = td->o.rate_iops_min[ddir];
163 assert(ddir_rw(ddir));
165 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
169 * allow a 2 second settle period in the beginning
171 if (mtime_since(&td->start, now) < 2000)
175 * if last_rate_check_blocks or last_rate_check_bytes is set,
176 * we can compute a rate per ratecycle
178 if (td->last_rate_check_bytes[ddir] || td->last_rate_check_blocks[ddir]) {
179 unsigned long spent = mtime_since(&td->last_rate_check_time[ddir], now);
180 if (spent < td->o.ratecycle || spent==0)
183 if (td->o.ratemin[ddir]) {
185 * check bandwidth specified rate
187 unsigned long long current_rate_bytes =
188 ((current_rate_check_bytes - td->last_rate_check_bytes[ddir]) * 1000) / spent;
189 if (current_rate_bytes < option_rate_bytes_min) {
190 log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
191 td->o.name, option_rate_bytes_min, current_rate_bytes);
196 * checks iops specified rate
198 unsigned long long current_rate_iops =
199 ((current_rate_check_blocks - td->last_rate_check_blocks[ddir]) * 1000) / spent;
201 if (current_rate_iops < option_rate_iops_min) {
202 log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
203 td->o.name, option_rate_iops_min, current_rate_iops);
209 td->last_rate_check_bytes[ddir] = current_rate_check_bytes;
210 td->last_rate_check_blocks[ddir] = current_rate_check_blocks;
211 memcpy(&td->last_rate_check_time[ddir], now, sizeof(*now));
215 static bool check_min_rate(struct thread_data *td, struct timespec *now)
219 for_each_rw_ddir(ddir) {
220 if (td->bytes_done[ddir])
221 ret |= __check_min_rate(td, now, ddir);
228 * When job exits, we can cancel the in-flight IO if we are using async
229 * io. Attempt to do so.
231 static void cleanup_pending_aio(struct thread_data *td)
236 * get immediately available events, if any
238 r = io_u_queued_complete(td, 0);
241 * now cancel remaining active events
243 if (td->io_ops->cancel) {
247 io_u_qiter(&td->io_u_all, io_u, i) {
248 if (io_u->flags & IO_U_F_FLIGHT) {
249 r = td->io_ops->cancel(td, io_u);
257 r = io_u_queued_complete(td, td->cur_depth);
261 * Helper to handle the final sync of a file. Works just like the normal
262 * io path, just does everything sync.
264 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
266 struct io_u *io_u = __get_io_u(td);
267 enum fio_q_status ret;
272 io_u->ddir = DDIR_SYNC;
274 io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
276 if (td_io_prep(td, io_u)) {
282 ret = td_io_queue(td, io_u);
286 if (io_u_queued_complete(td, 1) < 0)
289 case FIO_Q_COMPLETED:
291 td_verror(td, io_u->error, "td_io_queue");
295 if (io_u_sync_complete(td, io_u) < 0)
306 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
310 if (fio_file_open(f))
311 return fio_io_sync(td, f);
313 if (td_io_open_file(td, f))
316 ret = fio_io_sync(td, f);
318 if (fio_file_open(f))
319 ret2 = td_io_close_file(td, f);
320 return (ret || ret2);
323 static inline void __update_ts_cache(struct thread_data *td)
325 fio_gettime(&td->ts_cache, NULL);
328 static inline void update_ts_cache(struct thread_data *td)
330 if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
331 __update_ts_cache(td);
334 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
336 if (in_ramp_time(td))
340 if (utime_since(&td->epoch, t) >= td->o.timeout)
347 * We need to update the runtime consistently in ms, but keep a running
348 * tally of the current elapsed time in microseconds for sub millisecond
351 static inline void update_runtime(struct thread_data *td,
352 unsigned long long *elapsed_us,
353 const enum fio_ddir ddir)
355 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
358 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
359 elapsed_us[ddir] += utime_since_now(&td->start);
360 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
363 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
368 if (ret < 0 || td->error) {
370 enum error_type_bit eb;
375 eb = td_error_type(ddir, err);
376 if (!(td->o.continue_on_error & (1 << eb)))
379 if (td_non_fatal_error(td, eb, err)) {
381 * Continue with the I/Os in case of
384 update_error_count(td, err);
388 } else if (td->o.fill_device && (err == ENOSPC || err == EDQUOT)) {
390 * We expect to hit this error if
391 * fill_device option is set.
394 fio_mark_td_terminate(td);
398 * Stop the I/O in case of a fatal
401 update_error_count(td, err);
409 static void check_update_rusage(struct thread_data *td)
411 if (td->update_rusage) {
412 td->update_rusage = 0;
413 update_rusage_stat(td);
414 fio_sem_up(td->rusage_sem);
418 static int wait_for_completions(struct thread_data *td, struct timespec *time)
420 const int full = queue_full(td);
424 if (td->flags & TD_F_REGROW_LOGS)
425 return io_u_quiesce(td);
428 * if the queue is full, we MUST reap at least 1 event
430 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
431 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
434 if (time && should_check_rate(td))
435 fio_gettime(time, NULL);
438 ret = io_u_queued_complete(td, min_evts);
441 } while (full && (td->cur_depth > td->o.iodepth_low));
446 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
447 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
448 struct timespec *comp_time)
451 case FIO_Q_COMPLETED:
454 clear_io_u(td, io_u);
455 } else if (io_u->resid) {
456 long long bytes = io_u->xfer_buflen - io_u->resid;
457 struct fio_file *f = io_u->file;
460 *bytes_issued += bytes;
470 unlog_io_piece(td, io_u);
471 td_verror(td, EIO, "full resid");
476 io_u->xfer_buflen = io_u->resid;
477 io_u->xfer_buf += bytes;
478 io_u->offset += bytes;
480 if (ddir_rw(io_u->ddir))
481 td->ts.short_io_u[io_u->ddir]++;
483 if (io_u->offset == f->real_file_size)
486 requeue_io_u(td, &io_u);
489 if (comp_time && should_check_rate(td))
490 fio_gettime(comp_time, NULL);
492 *ret = io_u_sync_complete(td, io_u);
497 if (td->flags & TD_F_REGROW_LOGS)
501 * when doing I/O (not when verifying),
502 * check for any errors that are to be ignored
510 * if the engine doesn't have a commit hook,
511 * the io_u is really queued. if it does have such
512 * a hook, it has to call io_u_queued() itself.
514 if (td->io_ops->commit == NULL)
515 io_u_queued(td, io_u);
517 *bytes_issued += io_u->xfer_buflen;
521 unlog_io_piece(td, io_u);
522 requeue_io_u(td, &io_u);
527 td_verror(td, -(*ret), "td_io_queue");
531 if (break_on_this_error(td, ddir, ret))
537 static inline bool io_in_polling(struct thread_data *td)
539 return !td->o.iodepth_batch_complete_min &&
540 !td->o.iodepth_batch_complete_max;
543 * Unlinks files from thread data fio_file structure
545 static int unlink_all_files(struct thread_data *td)
551 for_each_file(td, f, i) {
552 if (f->filetype != FIO_TYPE_FILE)
554 ret = td_io_unlink_file(td, f);
560 td_verror(td, ret, "unlink_all_files");
566 * Check if io_u will overlap an in-flight IO in the queue
568 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
571 struct io_u *check_io_u;
572 unsigned long long x1, x2, y1, y2;
576 x2 = io_u->offset + io_u->buflen;
578 io_u_qiter(q, check_io_u, i) {
579 if (check_io_u->flags & IO_U_F_FLIGHT) {
580 y1 = check_io_u->offset;
581 y2 = check_io_u->offset + check_io_u->buflen;
583 if (x1 < y2 && y1 < x2) {
585 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
587 y1, check_io_u->buflen);
596 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
599 * Check for overlap if the user asked us to, and we have
600 * at least one IO in flight besides this one.
602 if (td->o.serialize_overlap && td->cur_depth > 1 &&
603 in_flight_overlap(&td->io_u_all, io_u))
606 return td_io_queue(td, io_u);
610 * The main verify engine. Runs over the writes we previously submitted,
611 * reads the blocks back in, and checks the crc/md5 of the data.
613 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
620 dprint(FD_VERIFY, "starting loop\n");
623 * sync io first and invalidate cache, to make sure we really
626 for_each_file(td, f, i) {
627 if (!fio_file_open(f))
629 if (fio_io_sync(td, f))
631 if (file_invalidate_cache(td, f))
635 check_update_rusage(td);
640 td_set_runstate(td, TD_VERIFYING);
643 while (!td->terminate) {
648 check_update_rusage(td);
650 if (runtime_exceeded(td, &td->ts_cache)) {
651 __update_ts_cache(td);
652 if (runtime_exceeded(td, &td->ts_cache)) {
653 fio_mark_td_terminate(td);
658 if (flow_threshold_exceeded(td))
661 if (!td->o.experimental_verify) {
662 io_u = __get_io_u(td);
666 if (get_next_verify(td, io_u)) {
671 if (td_io_prep(td, io_u)) {
676 if (td->bytes_verified + td->o.rw_min_bs > verify_bytes)
679 while ((io_u = get_io_u(td)) != NULL) {
680 if (IS_ERR_OR_NULL(io_u)) {
687 * We are only interested in the places where
688 * we wrote or trimmed IOs. Turn those into
689 * reads for verification purposes.
691 if (io_u->ddir == DDIR_READ) {
693 * Pretend we issued it for rwmix
696 td->io_issues[DDIR_READ]++;
699 } else if (io_u->ddir == DDIR_TRIM) {
700 io_u->ddir = DDIR_READ;
701 io_u_set(td, io_u, IO_U_F_TRIMMED);
703 } else if (io_u->ddir == DDIR_WRITE) {
704 io_u->ddir = DDIR_READ;
705 io_u->numberio = td->verify_read_issues;
706 td->verify_read_issues++;
707 populate_verify_io_u(td, io_u);
719 if (verify_state_should_stop(td, io_u)) {
724 if (td->o.verify_async)
725 io_u->end_io = verify_io_u_async;
727 io_u->end_io = verify_io_u;
730 if (!td->o.disable_slat)
731 fio_gettime(&io_u->start_time, NULL);
733 ret = io_u_submit(td, io_u);
735 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
739 * if we can queue more, do so. but check if there are
740 * completed io_u's first. Note that we can get BUSY even
741 * without IO queued, if the system is resource starved.
744 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
745 if (full || io_in_polling(td))
746 ret = wait_for_completions(td, NULL);
752 check_update_rusage(td);
755 min_events = td->cur_depth;
758 ret = io_u_queued_complete(td, min_events);
760 cleanup_pending_aio(td);
762 td_set_runstate(td, TD_RUNNING);
764 dprint(FD_VERIFY, "exiting loop\n");
767 static bool exceeds_number_ios(struct thread_data *td)
769 unsigned long long number_ios;
771 if (!td->o.number_ios)
774 number_ios = ddir_rw_sum(td->io_blocks);
775 number_ios += td->io_u_queued + td->io_u_in_flight;
777 return number_ios >= (td->o.number_ios * td->loops);
780 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
782 unsigned long long bytes, limit;
785 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
786 else if (td_write(td))
787 bytes = this_bytes[DDIR_WRITE];
788 else if (td_read(td))
789 bytes = this_bytes[DDIR_READ];
791 bytes = this_bytes[DDIR_TRIM];
794 limit = td->o.io_size;
799 return bytes >= limit || exceeds_number_ios(td);
802 static bool io_issue_bytes_exceeded(struct thread_data *td)
804 return io_bytes_exceeded(td, td->io_issue_bytes);
807 static bool io_complete_bytes_exceeded(struct thread_data *td)
809 return io_bytes_exceeded(td, td->this_io_bytes);
813 * used to calculate the next io time for rate control
816 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
818 uint64_t bps = td->rate_bps[ddir];
820 assert(!(td->flags & TD_F_CHILD));
822 if (td->o.rate_process == RATE_PROCESS_POISSON) {
825 iops = bps / td->o.min_bs[ddir];
826 val = (int64_t) (1000000 / iops) *
827 -logf(__rand_0_1(&td->poisson_state[ddir]));
829 dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
830 (unsigned long long) 1000000 / val,
833 td->last_usec[ddir] += val;
834 return td->last_usec[ddir];
836 uint64_t bytes = td->rate_io_issue_bytes[ddir];
837 uint64_t secs = bytes / bps;
838 uint64_t remainder = bytes % bps;
840 return remainder * 1000000 / bps + secs * 1000000;
846 static void init_thinktime(struct thread_data *td)
848 if (td->o.thinktime_blocks_type == THINKTIME_BLOCKS_TYPE_COMPLETE)
849 td->thinktime_blocks_counter = td->io_blocks;
851 td->thinktime_blocks_counter = td->io_issues;
852 td->last_thinktime = td->epoch;
853 td->last_thinktime_blocks = 0;
856 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir,
857 struct timespec *time)
859 unsigned long long b;
860 unsigned long long runtime_left;
866 if (td->o.thinktime_iotime) {
867 fio_gettime(&now, NULL);
868 if (utime_since(&td->last_thinktime, &now)
869 >= td->o.thinktime_iotime) {
871 } else if (!fio_option_is_set(&td->o, thinktime_blocks)) {
873 * When thinktime_iotime is set and thinktime_blocks is
874 * not set, skip the thinktime_blocks check, since
875 * thinktime_blocks default value 1 does not work
876 * together with thinktime_iotime.
883 b = ddir_rw_sum(td->thinktime_blocks_counter);
884 if (b >= td->last_thinktime_blocks + td->o.thinktime_blocks)
892 left = td->o.thinktime_spin;
894 runtime_left = td->o.timeout - utime_since_now(&td->epoch);
895 if (runtime_left < (unsigned long long)left)
901 total = usec_spin(left);
903 left = td->o.thinktime - total;
905 runtime_left = td->o.timeout - utime_since_now(&td->epoch);
906 if (runtime_left < (unsigned long long)left)
911 total += usec_sleep(td, left);
914 * If we're ignoring thinktime for the rate, add the number of bytes
915 * we would have done while sleeping, minus one block to ensure we
916 * start issuing immediately after the sleep.
918 if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
919 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
920 uint64_t bs = td->o.min_bs[ddir];
921 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
924 if (usperop <= total)
927 over = (usperop - total) / usperop * -bs;
929 td->rate_io_issue_bytes[ddir] += (missed - over);
930 /* adjust for rate_process=poisson */
931 td->last_usec[ddir] += total;
934 if (time && should_check_rate(td))
935 fio_gettime(time, NULL);
937 td->last_thinktime_blocks = b;
938 if (td->o.thinktime_iotime) {
939 fio_gettime(&now, NULL);
940 td->last_thinktime = now;
945 * Main IO worker function. It retrieves io_u's to process and queues
946 * and reaps them, checking for rate and errors along the way.
948 * Returns number of bytes written and trimmed.
950 static void do_io(struct thread_data *td, uint64_t *bytes_done)
954 uint64_t total_bytes, bytes_issued = 0;
956 for (i = 0; i < DDIR_RWDIR_CNT; i++)
957 bytes_done[i] = td->bytes_done[i];
959 if (in_ramp_time(td))
960 td_set_runstate(td, TD_RAMP);
962 td_set_runstate(td, TD_RUNNING);
966 total_bytes = td->o.size;
968 * Allow random overwrite workloads to write up to io_size
969 * before starting verification phase as 'size' doesn't apply.
971 if (td_write(td) && td_random(td) && td->o.norandommap)
972 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
974 * If verify_backlog is enabled, we'll run the verify in this
975 * handler as well. For that case, we may need up to twice the
978 if (td->o.verify != VERIFY_NONE &&
979 (td_write(td) && td->o.verify_backlog))
980 total_bytes += td->o.size;
982 /* In trimwrite mode, each byte is trimmed and then written, so
983 * allow total_bytes or number of ios to be twice as big */
984 if (td_trimwrite(td)) {
985 total_bytes += td->total_io_size;
986 td->o.number_ios *= 2;
989 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
990 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
992 struct timespec comp_time;
997 check_update_rusage(td);
999 if (td->terminate || td->done)
1002 update_ts_cache(td);
1004 if (runtime_exceeded(td, &td->ts_cache)) {
1005 __update_ts_cache(td);
1006 if (runtime_exceeded(td, &td->ts_cache)) {
1007 fio_mark_td_terminate(td);
1012 if (flow_threshold_exceeded(td))
1016 * Break if we exceeded the bytes. The exception is time
1017 * based runs, but we still need to break out of the loop
1018 * for those to run verification, if enabled.
1019 * Jobs read from iolog do not use this stop condition.
1021 if (bytes_issued >= total_bytes &&
1022 !td->o.read_iolog_file &&
1023 (!td->o.time_based ||
1024 (td->o.time_based && td->o.verify != VERIFY_NONE)))
1027 io_u = get_io_u(td);
1028 if (IS_ERR_OR_NULL(io_u)) {
1029 int err = PTR_ERR(io_u);
1033 if (err == -EBUSY) {
1037 if (td->o.latency_target)
1042 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY) {
1043 io_u->numberio = td->io_issues[io_u->ddir];
1044 populate_verify_io_u(td, io_u);
1050 * Add verification end_io handler if:
1051 * - Asked to verify (!td_rw(td))
1052 * - Or the io_u is from our verify list (mixed write/ver)
1054 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1055 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1057 if (verify_state_should_stop(td, io_u)) {
1062 if (td->o.verify_async)
1063 io_u->end_io = verify_io_u_async;
1065 io_u->end_io = verify_io_u;
1066 td_set_runstate(td, TD_VERIFYING);
1067 } else if (in_ramp_time(td))
1068 td_set_runstate(td, TD_RAMP);
1070 td_set_runstate(td, TD_RUNNING);
1073 * Always log IO before it's issued, so we know the specific
1074 * order of it. The logged unit will track when the IO has
1077 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1079 td->o.verify != VERIFY_NONE &&
1080 !td->o.experimental_verify)
1081 log_io_piece(td, io_u);
1083 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1084 const unsigned long long blen = io_u->xfer_buflen;
1085 const enum fio_ddir __ddir = acct_ddir(io_u);
1090 workqueue_enqueue(&td->io_wq, &io_u->work);
1093 if (ddir_rw(__ddir)) {
1094 td->io_issues[__ddir]++;
1095 td->io_issue_bytes[__ddir] += blen;
1096 td->rate_io_issue_bytes[__ddir] += blen;
1099 if (should_check_rate(td)) {
1100 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1101 fio_gettime(&comp_time, NULL);
1105 ret = io_u_submit(td, io_u);
1107 if (should_check_rate(td))
1108 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1110 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1114 * See if we need to complete some commands. Note that
1115 * we can get BUSY even without IO queued, if the
1116 * system is resource starved.
1119 full = queue_full(td) ||
1120 (ret == FIO_Q_BUSY && td->cur_depth);
1121 if (full || io_in_polling(td))
1122 ret = wait_for_completions(td, &comp_time);
1127 if (ddir_rw(ddir) && td->o.thinktime)
1128 handle_thinktime(td, ddir, &comp_time);
1130 if (!ddir_rw_sum(td->bytes_done) &&
1131 !td_ioengine_flagged(td, FIO_NOIO))
1134 if (!in_ramp_time(td) && should_check_rate(td)) {
1135 if (check_min_rate(td, &comp_time)) {
1136 if (exitall_on_terminate || td->o.exitall_error)
1137 fio_terminate_threads(td->groupid, td->o.exit_what);
1138 td_verror(td, EIO, "check_min_rate");
1142 if (!in_ramp_time(td) && td->o.latency_target)
1143 lat_target_check(td);
1146 check_update_rusage(td);
1148 if (td->trim_entries)
1149 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1151 if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
1153 fio_mark_td_terminate(td);
1158 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1159 workqueue_flush(&td->io_wq);
1165 ret = io_u_queued_complete(td, i);
1166 if (td->o.fill_device &&
1167 (td->error == ENOSPC || td->error == EDQUOT))
1171 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1172 td_set_runstate(td, TD_FSYNCING);
1174 for_each_file(td, f, i) {
1175 if (!fio_file_fsync(td, f))
1178 log_err("fio: end_fsync failed for file %s\n",
1183 if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
1184 workqueue_flush(&td->io_wq);
1185 cleanup_pending_aio(td);
1189 * stop job if we failed doing any IO
1191 if (!ddir_rw_sum(td->this_io_bytes))
1194 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1195 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1198 static void free_file_completion_logging(struct thread_data *td)
1203 for_each_file(td, f, i) {
1204 if (!f->last_write_comp)
1206 sfree(f->last_write_comp);
1210 static int init_file_completion_logging(struct thread_data *td,
1216 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1219 for_each_file(td, f, i) {
1220 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1221 if (!f->last_write_comp)
1228 free_file_completion_logging(td);
1229 log_err("fio: failed to alloc write comp data\n");
1233 static void cleanup_io_u(struct thread_data *td)
1237 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1239 if (td->io_ops->io_u_free)
1240 td->io_ops->io_u_free(td, io_u);
1242 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1247 io_u_rexit(&td->io_u_requeues);
1248 io_u_qexit(&td->io_u_freelist, false);
1249 io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1251 free_file_completion_logging(td);
1254 static int init_io_u(struct thread_data *td)
1257 int cl_align, i, max_units;
1260 max_units = td->o.iodepth;
1263 err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1264 err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1265 err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1268 log_err("fio: failed setting up IO queues\n");
1272 cl_align = os_cache_line_size();
1274 for (i = 0; i < max_units; i++) {
1280 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1282 log_err("fio: unable to allocate aligned memory\n");
1287 memset(io_u, 0, sizeof(*io_u));
1288 INIT_FLIST_HEAD(&io_u->verify_list);
1289 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1292 io_u->flags = IO_U_F_FREE;
1293 io_u_qpush(&td->io_u_freelist, io_u);
1296 * io_u never leaves this stack, used for iteration of all
1299 io_u_qpush(&td->io_u_all, io_u);
1301 if (td->io_ops->io_u_init) {
1302 int ret = td->io_ops->io_u_init(td, io_u);
1305 log_err("fio: failed to init engine data: %d\n", ret);
1311 if (init_io_u_buffers(td))
1314 if (init_file_completion_logging(td, max_units))
1320 int init_io_u_buffers(struct thread_data *td)
1323 unsigned long long max_bs, min_write;
1328 max_units = td->o.iodepth;
1329 max_bs = td_max_bs(td);
1330 min_write = td->o.min_bs[DDIR_WRITE];
1331 td->orig_buffer_size = (unsigned long long) max_bs
1332 * (unsigned long long) max_units;
1334 if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1338 * if we may later need to do address alignment, then add any
1339 * possible adjustment here so that we don't cause a buffer
1340 * overflow later. this adjustment may be too much if we get
1341 * lucky and the allocator gives us an aligned address.
1343 if (td->o.odirect || td->o.mem_align ||
1344 td_ioengine_flagged(td, FIO_RAWIO))
1345 td->orig_buffer_size += page_mask + td->o.mem_align;
1347 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1348 unsigned long long bs;
1350 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1351 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1354 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1355 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1359 if (data_xfer && allocate_io_mem(td))
1362 if (td->o.odirect || td->o.mem_align ||
1363 td_ioengine_flagged(td, FIO_RAWIO))
1364 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1366 p = td->orig_buffer;
1368 for (i = 0; i < max_units; i++) {
1369 io_u = td->io_u_all.io_us[i];
1370 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1374 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1377 io_u_fill_buffer(td, io_u, min_write, max_bs);
1378 if (td_write(td) && td->o.verify_pattern_bytes) {
1380 * Fill the buffer with the pattern if we are
1381 * going to be doing writes.
1383 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1392 #ifdef FIO_HAVE_IOSCHED_SWITCH
1394 * These functions are Linux specific.
1395 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1397 static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
1399 char tmp[256], tmp2[128], *p;
1403 assert(file->du && file->du->sysfs_root);
1404 sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);
1406 f = fopen(tmp, "r+");
1408 if (errno == ENOENT) {
1409 log_err("fio: os or kernel doesn't support IO scheduler"
1413 td_verror(td, errno, "fopen iosched");
1420 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1421 if (ferror(f) || ret != 1) {
1422 td_verror(td, errno, "fwrite");
1430 * Read back and check that the selected scheduler is now the default.
1432 ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1433 if (ferror(f) || ret < 0) {
1434 td_verror(td, errno, "fread");
1440 * either a list of io schedulers or "none\n" is expected. Strip the
1447 * Write to "none" entry doesn't fail, so check the result here.
1449 if (!strcmp(tmp, "none")) {
1450 log_err("fio: io scheduler is not tunable\n");
1455 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1456 if (!strstr(tmp, tmp2)) {
1457 log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
1458 td_verror(td, EINVAL, "iosched_switch");
1467 static int switch_ioscheduler(struct thread_data *td)
1473 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1476 assert(td->files && td->files[0]);
1478 for_each_file(td, f, i) {
1480 /* Only consider regular files and block device files */
1481 switch (f->filetype) {
1483 case FIO_TYPE_BLOCK:
1485 * Make sure that the device hosting the file could
1497 ret = set_ioscheduler(td, f);
1507 static int switch_ioscheduler(struct thread_data *td)
1512 #endif /* FIO_HAVE_IOSCHED_SWITCH */
1514 static bool keep_running(struct thread_data *td)
1516 unsigned long long limit;
1522 if (td->o.time_based)
1528 if (exceeds_number_ios(td))
1532 limit = td->o.io_size;
1536 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1540 * If the difference is less than the maximum IO size, we
1543 diff = limit - ddir_rw_sum(td->io_bytes);
1544 if (diff < td_max_bs(td))
1547 if (fio_files_done(td) && !td->o.io_size)
1556 static int exec_string(struct thread_options *o, const char *string,
1562 if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1565 log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1569 log_err("fio: exec of cmd <%s> failed\n", str);
1576 * Dry run to compute correct state of numberio for verification.
1578 static uint64_t do_dry_run(struct thread_data *td)
1580 td_set_runstate(td, TD_RUNNING);
1582 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1583 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1587 if (td->terminate || td->done)
1590 io_u = get_io_u(td);
1591 if (IS_ERR_OR_NULL(io_u))
1594 io_u_set(td, io_u, IO_U_F_FLIGHT);
1597 if (ddir_rw(acct_ddir(io_u)))
1598 td->io_issues[acct_ddir(io_u)]++;
1599 if (ddir_rw(io_u->ddir)) {
1600 io_u_mark_depth(td, 1);
1601 td->ts.total_io_u[io_u->ddir]++;
1604 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1606 td->o.verify != VERIFY_NONE &&
1607 !td->o.experimental_verify)
1608 log_io_piece(td, io_u);
1610 ret = io_u_sync_complete(td, io_u);
1614 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1618 struct thread_data *td;
1619 struct sk_out *sk_out;
1623 * Entry point for the thread based jobs. The process based jobs end up
1624 * here as well, after a little setup.
1626 static void *thread_main(void *data)
1628 struct fork_data *fd = data;
1629 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1630 struct thread_data *td = fd->td;
1631 struct thread_options *o = &td->o;
1632 struct sk_out *sk_out = fd->sk_out;
1633 uint64_t bytes_done[DDIR_RWDIR_CNT];
1634 int deadlock_loop_cnt;
1638 sk_out_assign(sk_out);
1641 if (!o->use_thread) {
1647 fio_local_clock_init();
1649 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1652 fio_server_send_start(td);
1654 INIT_FLIST_HEAD(&td->io_log_list);
1655 INIT_FLIST_HEAD(&td->io_hist_list);
1656 INIT_FLIST_HEAD(&td->verify_list);
1657 INIT_FLIST_HEAD(&td->trim_list);
1658 td->io_hist_tree = RB_ROOT;
1660 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1662 td_verror(td, ret, "mutex_cond_init_pshared");
1665 ret = cond_init_pshared(&td->verify_cond);
1667 td_verror(td, ret, "mutex_cond_pshared");
1671 td_set_runstate(td, TD_INITIALIZED);
1672 dprint(FD_MUTEX, "up startup_sem\n");
1673 fio_sem_up(startup_sem);
1674 dprint(FD_MUTEX, "wait on td->sem\n");
1675 fio_sem_down(td->sem);
1676 dprint(FD_MUTEX, "done waiting on td->sem\n");
1679 * A new gid requires privilege, so we need to do this before setting
1682 if (o->gid != -1U && setgid(o->gid)) {
1683 td_verror(td, errno, "setgid");
1686 if (o->uid != -1U && setuid(o->uid)) {
1687 td_verror(td, errno, "setuid");
1691 td_zone_gen_index(td);
1694 * Do this early, we don't want the compress threads to be limited
1695 * to the same CPUs as the IO workers. So do this before we set
1696 * any potential CPU affinity
1698 if (iolog_compress_init(td, sk_out))
1702 * If we have a gettimeofday() thread, make sure we exclude that
1703 * thread from this job
1706 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1709 * Set affinity first, in case it has an impact on the memory
1712 if (fio_option_is_set(o, cpumask)) {
1713 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1714 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1716 log_err("fio: no CPUs set\n");
1717 log_err("fio: Try increasing number of available CPUs\n");
1718 td_verror(td, EINVAL, "cpus_split");
1722 ret = fio_setaffinity(td->pid, o->cpumask);
1724 td_verror(td, errno, "cpu_set_affinity");
1729 #ifdef CONFIG_LIBNUMA
1730 /* numa node setup */
1731 if (fio_option_is_set(o, numa_cpunodes) ||
1732 fio_option_is_set(o, numa_memnodes)) {
1733 struct bitmask *mask;
1735 if (numa_available() < 0) {
1736 td_verror(td, errno, "Does not support NUMA API\n");
1740 if (fio_option_is_set(o, numa_cpunodes)) {
1741 mask = numa_parse_nodestring(o->numa_cpunodes);
1742 ret = numa_run_on_node_mask(mask);
1743 numa_free_nodemask(mask);
1745 td_verror(td, errno, \
1746 "numa_run_on_node_mask failed\n");
1751 if (fio_option_is_set(o, numa_memnodes)) {
1753 if (o->numa_memnodes)
1754 mask = numa_parse_nodestring(o->numa_memnodes);
1756 switch (o->numa_mem_mode) {
1757 case MPOL_INTERLEAVE:
1758 numa_set_interleave_mask(mask);
1761 numa_set_membind(mask);
1764 numa_set_localalloc();
1766 case MPOL_PREFERRED:
1767 numa_set_preferred(o->numa_mem_prefer_node);
1775 numa_free_nodemask(mask);
1781 if (fio_pin_memory(td))
1785 * May alter parameters that init_io_u() will use, so we need to
1788 if (!init_iolog(td))
1791 /* ioprio_set() has to be done before td_io_init() */
1792 if (fio_option_is_set(o, ioprio) ||
1793 fio_option_is_set(o, ioprio_class)) {
1794 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1796 td_verror(td, errno, "ioprio_set");
1799 td->ioprio = ioprio_value(o->ioprio_class, o->ioprio);
1800 td->ts.ioprio = td->ioprio;
1806 if (td_ioengine_flagged(td, FIO_SYNCIO) && td->o.iodepth > 1 && td->o.io_submit_mode != IO_MODE_OFFLOAD) {
1807 log_info("note: both iodepth >= 1 and synchronous I/O engine "
1808 "are selected, queue depth will be capped at 1\n");
1814 if (td->io_ops->post_init && td->io_ops->post_init(td))
1817 if (o->verify_async && verify_async_init(td))
1820 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1824 if (nice(o->nice) == -1 && errno != 0) {
1825 td_verror(td, errno, "nice");
1829 if (o->ioscheduler && switch_ioscheduler(td))
1832 if (!o->create_serialize && setup_files(td))
1835 if (!init_random_map(td))
1838 if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1841 if (o->pre_read && !pre_read_files(td))
1844 fio_verify_init(td);
1846 if (rate_submit_init(td, sk_out))
1849 set_epoch_time(td, o->log_unix_epoch | o->log_alternate_epoch, o->log_alternate_epoch_clock_id);
1850 fio_getrusage(&td->ru_start);
1851 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1852 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1853 memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1857 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1858 o->ratemin[DDIR_TRIM]) {
1859 memcpy(&td->last_rate_check_time[DDIR_READ], &td->bw_sample_time,
1860 sizeof(td->bw_sample_time));
1861 memcpy(&td->last_rate_check_time[DDIR_WRITE], &td->bw_sample_time,
1862 sizeof(td->bw_sample_time));
1863 memcpy(&td->last_rate_check_time[DDIR_TRIM], &td->bw_sample_time,
1864 sizeof(td->bw_sample_time));
1867 memset(bytes_done, 0, sizeof(bytes_done));
1868 clear_state = false;
1870 while (keep_running(td)) {
1871 uint64_t verify_bytes;
1873 fio_gettime(&td->start, NULL);
1874 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1877 clear_io_state(td, 0);
1879 if (o->unlink_each_loop && unlink_all_files(td))
1883 prune_io_piece_log(td);
1885 if (td->o.verify_only && td_write(td))
1886 verify_bytes = do_dry_run(td);
1888 if (!td->o.rand_repeatable)
1889 /* save verify rand state to replay hdr seeds later at verify */
1890 frand_copy(&td->verify_state_last_do_io, &td->verify_state);
1891 do_io(td, bytes_done);
1892 if (!td->o.rand_repeatable)
1893 frand_copy(&td->verify_state, &td->verify_state_last_do_io);
1894 if (!ddir_rw_sum(bytes_done)) {
1895 fio_mark_td_terminate(td);
1898 verify_bytes = bytes_done[DDIR_WRITE] +
1899 bytes_done[DDIR_TRIM];
1904 * If we took too long to shut down, the main thread could
1905 * already consider us reaped/exited. If that happens, break
1908 if (td->runstate >= TD_EXITED)
1914 * Make sure we've successfully updated the rusage stats
1915 * before waiting on the stat mutex. Otherwise we could have
1916 * the stat thread holding stat mutex and waiting for
1917 * the rusage_sem, which would never get upped because
1918 * this thread is waiting for the stat mutex.
1920 deadlock_loop_cnt = 0;
1922 check_update_rusage(td);
1923 if (!fio_sem_down_trylock(stat_sem))
1926 if (deadlock_loop_cnt++ > 5000) {
1927 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1928 td->error = EDEADLK;
1933 if (td->io_bytes[DDIR_READ] && (td_read(td) ||
1934 ((td->flags & TD_F_VER_BACKLOG) && td_write(td))))
1935 update_runtime(td, elapsed_us, DDIR_READ);
1936 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1937 update_runtime(td, elapsed_us, DDIR_WRITE);
1938 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1939 update_runtime(td, elapsed_us, DDIR_TRIM);
1940 fio_gettime(&td->start, NULL);
1941 fio_sem_up(stat_sem);
1943 if (td->error || td->terminate)
1946 if (!o->do_verify ||
1947 o->verify == VERIFY_NONE ||
1948 td_ioengine_flagged(td, FIO_UNIDIR))
1951 clear_io_state(td, 0);
1953 fio_gettime(&td->start, NULL);
1955 do_verify(td, verify_bytes);
1958 * See comment further up for why this is done here.
1960 check_update_rusage(td);
1962 fio_sem_down(stat_sem);
1963 update_runtime(td, elapsed_us, DDIR_READ);
1964 fio_gettime(&td->start, NULL);
1965 fio_sem_up(stat_sem);
1967 if (td->error || td->terminate)
1972 * Acquire this lock if we were doing overlap checking in
1973 * offload mode so that we don't clean up this job while
1974 * another thread is checking its io_u's for overlap
1976 if (td_offload_overlap(td)) {
1977 int res = pthread_mutex_lock(&overlap_check);
1980 td_set_runstate(td, TD_FINISHING);
1981 if (td_offload_overlap(td)) {
1982 res = pthread_mutex_unlock(&overlap_check);
1986 update_rusage_stat(td);
1987 td->ts.total_run_time = mtime_since_now(&td->epoch);
1988 for_each_rw_ddir(ddir) {
1989 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
1992 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1993 (td->o.verify != VERIFY_NONE && td_write(td)))
1994 verify_save_state(td->thread_number);
1996 fio_unpin_memory(td);
1998 td_writeout_logs(td, true);
2000 iolog_compress_exit(td);
2001 rate_submit_exit(td);
2003 if (o->exec_postrun)
2004 exec_string(o, o->exec_postrun, "postrun");
2006 if (exitall_on_terminate || (o->exitall_error && td->error))
2007 fio_terminate_threads(td->groupid, td->o.exit_what);
2011 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
2014 if (o->verify_async)
2015 verify_async_exit(td);
2017 close_and_free_files(td);
2020 cgroup_shutdown(td, cgroup_mnt);
2021 verify_free_state(td);
2022 td_zone_free_index(td);
2024 if (fio_option_is_set(o, cpumask)) {
2025 ret = fio_cpuset_exit(&o->cpumask);
2027 td_verror(td, ret, "fio_cpuset_exit");
2031 * do this very late, it will log file closing as well
2033 if (o->write_iolog_file)
2034 write_iolog_close(td);
2035 if (td->io_log_rfile)
2036 fclose(td->io_log_rfile);
2038 td_set_runstate(td, TD_EXITED);
2041 * Do this last after setting our runstate to exited, so we
2042 * know that the stat thread is signaled.
2044 check_update_rusage(td);
2047 return (void *) (uintptr_t) td->error;
2051 * Run over the job map and reap the threads that have exited, if any.
2053 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
2056 struct thread_data *td;
2057 unsigned int cputhreads, realthreads, pending;
2061 * reap exited threads (TD_EXITED -> TD_REAPED)
2063 realthreads = pending = cputhreads = 0;
2064 for_each_td(td, i) {
2067 if (!strcmp(td->o.ioengine, "cpuio"))
2076 if (td->runstate == TD_REAPED)
2078 if (td->o.use_thread) {
2079 if (td->runstate == TD_EXITED) {
2080 td_set_runstate(td, TD_REAPED);
2087 if (td->runstate == TD_EXITED)
2091 * check if someone quit or got killed in an unusual way
2093 ret = waitpid(td->pid, &status, flags);
2095 if (errno == ECHILD) {
2096 log_err("fio: pid=%d disappeared %d\n",
2097 (int) td->pid, td->runstate);
2099 td_set_runstate(td, TD_REAPED);
2103 } else if (ret == td->pid) {
2104 if (WIFSIGNALED(status)) {
2105 int sig = WTERMSIG(status);
2107 if (sig != SIGTERM && sig != SIGUSR2)
2108 log_err("fio: pid=%d, got signal=%d\n",
2109 (int) td->pid, sig);
2111 td_set_runstate(td, TD_REAPED);
2114 if (WIFEXITED(status)) {
2115 if (WEXITSTATUS(status) && !td->error)
2116 td->error = WEXITSTATUS(status);
2118 td_set_runstate(td, TD_REAPED);
2124 * If the job is stuck, do a forceful timeout of it and
2127 if (td->terminate &&
2128 td->runstate < TD_FSYNCING &&
2129 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2130 log_err("fio: job '%s' (state=%d) hasn't exited in "
2131 "%lu seconds, it appears to be stuck. Doing "
2132 "forceful exit of this job.\n",
2133 td->o.name, td->runstate,
2134 (unsigned long) time_since_now(&td->terminate_time));
2135 td_set_runstate(td, TD_REAPED);
2140 * thread is not dead, continue
2146 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2147 (*t_rate) -= ddir_rw_sum(td->o.rate);
2154 done_secs += mtime_since_now(&td->epoch) / 1000;
2155 profile_td_exit(td);
2159 if (*nr_running == cputhreads && !pending && realthreads)
2160 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2163 static bool __check_trigger_file(void)
2170 if (stat(trigger_file, &sb))
2173 if (unlink(trigger_file) < 0)
2174 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2180 static bool trigger_timedout(void)
2182 if (trigger_timeout)
2183 if (time_since_genesis() >= trigger_timeout) {
2184 trigger_timeout = 0;
2191 void exec_trigger(const char *cmd)
2195 if (!cmd || cmd[0] == '\0')
2200 log_err("fio: failed executing %s trigger\n", cmd);
2203 void check_trigger_file(void)
2205 if (__check_trigger_file() || trigger_timedout()) {
2207 fio_clients_send_trigger(trigger_remote_cmd);
2209 verify_save_state(IO_LIST_ALL);
2210 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2211 exec_trigger(trigger_cmd);
2216 static int fio_verify_load_state(struct thread_data *td)
2220 if (!td->o.verify_state)
2226 ret = fio_server_get_verify_state(td->o.name,
2227 td->thread_number - 1, &data);
2229 verify_assign_state(td, data);
2231 char prefix[PATH_MAX];
2234 sprintf(prefix, "%s%clocal", aux_path,
2235 FIO_OS_PATH_SEPARATOR);
2237 strcpy(prefix, "local");
2238 ret = verify_load_state(td, prefix);
2244 static void do_usleep(unsigned int usecs)
2246 check_for_running_stats();
2247 check_trigger_file();
2251 static bool check_mount_writes(struct thread_data *td)
2256 if (!td_write(td) || td->o.allow_mounted_write)
2260 * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2261 * are mkfs'd and mounted.
2263 for_each_file(td, f, i) {
2264 #ifdef FIO_HAVE_CHARDEV_SIZE
2265 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2267 if (f->filetype != FIO_TYPE_BLOCK)
2270 if (device_is_mounted(f->file_name))
2276 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2280 static bool waitee_running(struct thread_data *me)
2282 const char *waitee = me->o.wait_for;
2283 const char *self = me->o.name;
2284 struct thread_data *td;
2290 for_each_td(td, i) {
2291 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2294 if (td->runstate < TD_EXITED) {
2295 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2297 runstate_to_name(td->runstate));
2302 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2307 * Main function for kicking off and reaping jobs, as needed.
2309 static void run_threads(struct sk_out *sk_out)
2311 struct thread_data *td;
2312 unsigned int i, todo, nr_running, nr_started;
2313 uint64_t m_rate, t_rate;
2316 if (fio_gtod_offload && fio_start_gtod_thread())
2319 fio_idle_prof_init();
2323 nr_thread = nr_process = 0;
2324 for_each_td(td, i) {
2325 if (check_mount_writes(td))
2327 if (td->o.use_thread)
2333 if (output_format & FIO_OUTPUT_NORMAL) {
2334 struct buf_output out;
2336 buf_output_init(&out);
2337 __log_buf(&out, "Starting ");
2339 __log_buf(&out, "%d thread%s", nr_thread,
2340 nr_thread > 1 ? "s" : "");
2343 __log_buf(&out, " and ");
2344 __log_buf(&out, "%d process%s", nr_process,
2345 nr_process > 1 ? "es" : "");
2347 __log_buf(&out, "\n");
2348 log_info_buf(out.buf, out.buflen);
2349 buf_output_free(&out);
2352 todo = thread_number;
2355 m_rate = t_rate = 0;
2357 for_each_td(td, i) {
2358 print_status_init(td->thread_number - 1);
2360 if (!td->o.create_serialize)
2363 if (fio_verify_load_state(td))
2367 * do file setup here so it happens sequentially,
2368 * we don't want X number of threads getting their
2369 * client data interspersed on disk
2371 if (setup_files(td)) {
2375 log_err("fio: pid=%d, err=%d/%s\n",
2376 (int) td->pid, td->error, td->verror);
2377 td_set_runstate(td, TD_REAPED);
2384 * for sharing to work, each job must always open
2385 * its own files. so close them, if we opened them
2388 for_each_file(td, f, j) {
2389 if (fio_file_open(f))
2390 td_io_close_file(td, f);
2395 /* start idle threads before io threads start to run */
2396 fio_idle_prof_start();
2401 struct thread_data *map[REAL_MAX_JOBS];
2402 struct timespec this_start;
2403 int this_jobs = 0, left;
2404 struct fork_data *fd;
2407 * create threads (TD_NOT_CREATED -> TD_CREATED)
2409 for_each_td(td, i) {
2410 if (td->runstate != TD_NOT_CREATED)
2414 * never got a chance to start, killed by other
2415 * thread for some reason
2417 if (td->terminate) {
2422 if (td->o.start_delay) {
2423 spent = utime_since_genesis();
2425 if (td->o.start_delay > spent)
2429 if (td->o.stonewall && (nr_started || nr_running)) {
2430 dprint(FD_PROCESS, "%s: stonewall wait\n",
2435 if (waitee_running(td)) {
2436 dprint(FD_PROCESS, "%s: waiting for %s\n",
2437 td->o.name, td->o.wait_for);
2443 td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2444 td->update_rusage = 0;
2447 * Set state to created. Thread will transition
2448 * to TD_INITIALIZED when it's done setting up.
2450 td_set_runstate(td, TD_CREATED);
2451 map[this_jobs++] = td;
2454 fd = calloc(1, sizeof(*fd));
2456 fd->sk_out = sk_out;
2458 if (td->o.use_thread) {
2461 dprint(FD_PROCESS, "will pthread_create\n");
2462 ret = pthread_create(&td->thread, NULL,
2465 log_err("pthread_create: %s\n",
2472 ret = pthread_detach(td->thread);
2474 log_err("pthread_detach: %s",
2479 dprint(FD_PROCESS, "will fork\n");
2486 ret = (int)(uintptr_t)thread_main(fd);
2488 } else if (i == fio_debug_jobno)
2489 *fio_debug_jobp = pid;
2494 dprint(FD_MUTEX, "wait on startup_sem\n");
2495 if (fio_sem_down_timeout(startup_sem, 10000)) {
2496 log_err("fio: job startup hung? exiting.\n");
2497 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2503 dprint(FD_MUTEX, "done waiting on startup_sem\n");
2507 * Wait for the started threads to transition to
2510 fio_gettime(&this_start, NULL);
2512 while (left && !fio_abort) {
2513 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2518 for (i = 0; i < this_jobs; i++) {
2522 if (td->runstate == TD_INITIALIZED) {
2525 } else if (td->runstate >= TD_EXITED) {
2529 nr_running++; /* work-around... */
2535 log_err("fio: %d job%s failed to start\n", left,
2536 left > 1 ? "s" : "");
2537 for (i = 0; i < this_jobs; i++) {
2541 kill(td->pid, SIGTERM);
2547 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2549 for_each_td(td, i) {
2550 if (td->runstate != TD_INITIALIZED)
2553 if (in_ramp_time(td))
2554 td_set_runstate(td, TD_RAMP);
2556 td_set_runstate(td, TD_RUNNING);
2559 m_rate += ddir_rw_sum(td->o.ratemin);
2560 t_rate += ddir_rw_sum(td->o.rate);
2562 fio_sem_up(td->sem);
2565 reap_threads(&nr_running, &t_rate, &m_rate);
2571 while (nr_running) {
2572 reap_threads(&nr_running, &t_rate, &m_rate);
2576 fio_idle_prof_stop();
2581 static void free_disk_util(void)
2583 disk_util_prune_entries();
2584 helper_thread_destroy();
2587 int fio_backend(struct sk_out *sk_out)
2589 struct thread_data *td;
2593 if (load_profile(exec_profile))
2596 exec_profile = NULL;
2602 struct log_params p = {
2603 .log_type = IO_LOG_TYPE_BW,
2606 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2607 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2608 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2611 if (init_global_dedupe_working_set_seeds()) {
2612 log_err("fio: failed to initialize global dedupe working set\n");
2616 startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2618 is_local_backend = true;
2619 if (startup_sem == NULL)
2624 if (helper_thread_create(startup_sem, sk_out))
2625 log_err("fio: failed to create helper thread\n");
2627 cgroup_list = smalloc(sizeof(*cgroup_list));
2629 INIT_FLIST_HEAD(cgroup_list);
2631 run_threads(sk_out);
2633 helper_thread_exit();
2638 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2639 struct io_log *log = agg_io_log[i];
2641 flush_log(log, false);
2647 for_each_td(td, i) {
2648 struct thread_stat *ts = &td->ts;
2650 free_clat_prio_stats(ts);
2651 steadystate_free(td);
2652 fio_options_free(td);
2653 fio_dump_options_free(td);
2654 if (td->rusage_sem) {
2655 fio_sem_remove(td->rusage_sem);
2656 td->rusage_sem = NULL;
2658 fio_sem_remove(td->sem);
2664 cgroup_kill(cgroup_list);
2668 fio_sem_remove(startup_sem);