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);
641 * verify_state needs to be reset before verification
642 * proceeds so that expected random seeds match actual
643 * random seeds in headers. The main loop will reset
644 * all random number generators if randrepeat is set.
646 if (!td->o.rand_repeatable)
647 td_fill_verify_state_seed(td);
649 td_set_runstate(td, TD_VERIFYING);
652 while (!td->terminate) {
657 check_update_rusage(td);
659 if (runtime_exceeded(td, &td->ts_cache)) {
660 __update_ts_cache(td);
661 if (runtime_exceeded(td, &td->ts_cache)) {
662 fio_mark_td_terminate(td);
667 if (flow_threshold_exceeded(td))
670 if (!td->o.experimental_verify) {
671 io_u = __get_io_u(td);
675 if (get_next_verify(td, io_u)) {
680 if (td_io_prep(td, io_u)) {
685 if (td->bytes_verified + td->o.rw_min_bs > verify_bytes)
688 while ((io_u = get_io_u(td)) != NULL) {
689 if (IS_ERR_OR_NULL(io_u)) {
696 * We are only interested in the places where
697 * we wrote or trimmed IOs. Turn those into
698 * reads for verification purposes.
700 if (io_u->ddir == DDIR_READ) {
702 * Pretend we issued it for rwmix
705 td->io_issues[DDIR_READ]++;
708 } else if (io_u->ddir == DDIR_TRIM) {
709 io_u->ddir = DDIR_READ;
710 io_u_set(td, io_u, IO_U_F_TRIMMED);
712 } else if (io_u->ddir == DDIR_WRITE) {
713 io_u->ddir = DDIR_READ;
714 populate_verify_io_u(td, io_u);
726 if (verify_state_should_stop(td, io_u)) {
731 if (td->o.verify_async)
732 io_u->end_io = verify_io_u_async;
734 io_u->end_io = verify_io_u;
737 if (!td->o.disable_slat)
738 fio_gettime(&io_u->start_time, NULL);
740 ret = io_u_submit(td, io_u);
742 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
746 * if we can queue more, do so. but check if there are
747 * completed io_u's first. Note that we can get BUSY even
748 * without IO queued, if the system is resource starved.
751 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
752 if (full || io_in_polling(td))
753 ret = wait_for_completions(td, NULL);
759 check_update_rusage(td);
762 min_events = td->cur_depth;
765 ret = io_u_queued_complete(td, min_events);
767 cleanup_pending_aio(td);
769 td_set_runstate(td, TD_RUNNING);
771 dprint(FD_VERIFY, "exiting loop\n");
774 static bool exceeds_number_ios(struct thread_data *td)
776 unsigned long long number_ios;
778 if (!td->o.number_ios)
781 number_ios = ddir_rw_sum(td->io_blocks);
782 number_ios += td->io_u_queued + td->io_u_in_flight;
784 return number_ios >= (td->o.number_ios * td->loops);
787 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
789 unsigned long long bytes, limit;
792 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
793 else if (td_write(td))
794 bytes = this_bytes[DDIR_WRITE];
795 else if (td_read(td))
796 bytes = this_bytes[DDIR_READ];
798 bytes = this_bytes[DDIR_TRIM];
801 limit = td->o.io_size;
806 return bytes >= limit || exceeds_number_ios(td);
809 static bool io_issue_bytes_exceeded(struct thread_data *td)
811 return io_bytes_exceeded(td, td->io_issue_bytes);
814 static bool io_complete_bytes_exceeded(struct thread_data *td)
816 return io_bytes_exceeded(td, td->this_io_bytes);
820 * used to calculate the next io time for rate control
823 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
825 uint64_t bps = td->rate_bps[ddir];
827 assert(!(td->flags & TD_F_CHILD));
829 if (td->o.rate_process == RATE_PROCESS_POISSON) {
832 iops = bps / td->o.min_bs[ddir];
833 val = (int64_t) (1000000 / iops) *
834 -logf(__rand_0_1(&td->poisson_state[ddir]));
836 dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
837 (unsigned long long) 1000000 / val,
840 td->last_usec[ddir] += val;
841 return td->last_usec[ddir];
843 uint64_t bytes = td->rate_io_issue_bytes[ddir];
844 uint64_t secs = bytes / bps;
845 uint64_t remainder = bytes % bps;
847 return remainder * 1000000 / bps + secs * 1000000;
853 static void init_thinktime(struct thread_data *td)
855 if (td->o.thinktime_blocks_type == THINKTIME_BLOCKS_TYPE_COMPLETE)
856 td->thinktime_blocks_counter = td->io_blocks;
858 td->thinktime_blocks_counter = td->io_issues;
859 td->last_thinktime = td->epoch;
860 td->last_thinktime_blocks = 0;
863 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir,
864 struct timespec *time)
866 unsigned long long b;
872 if (td->o.thinktime_iotime) {
873 fio_gettime(&now, NULL);
874 if (utime_since(&td->last_thinktime, &now)
875 >= td->o.thinktime_iotime + td->o.thinktime) {
877 } else if (!fio_option_is_set(&td->o, thinktime_blocks)) {
879 * When thinktime_iotime is set and thinktime_blocks is
880 * not set, skip the thinktime_blocks check, since
881 * thinktime_blocks default value 1 does not work
882 * together with thinktime_iotime.
889 b = ddir_rw_sum(td->thinktime_blocks_counter);
890 if (b >= td->last_thinktime_blocks + td->o.thinktime_blocks)
899 if (td->o.thinktime_spin)
900 total = usec_spin(td->o.thinktime_spin);
902 left = td->o.thinktime - total;
904 total += usec_sleep(td, left);
907 * If we're ignoring thinktime for the rate, add the number of bytes
908 * we would have done while sleeping, minus one block to ensure we
909 * start issuing immediately after the sleep.
911 if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
912 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
913 uint64_t bs = td->o.min_bs[ddir];
914 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
917 if (usperop <= total)
920 over = (usperop - total) / usperop * -bs;
922 td->rate_io_issue_bytes[ddir] += (missed - over);
923 /* adjust for rate_process=poisson */
924 td->last_usec[ddir] += total;
927 if (time && should_check_rate(td))
928 fio_gettime(time, NULL);
930 td->last_thinktime_blocks = b;
931 if (td->o.thinktime_iotime)
932 td->last_thinktime = now;
936 * Main IO worker function. It retrieves io_u's to process and queues
937 * and reaps them, checking for rate and errors along the way.
939 * Returns number of bytes written and trimmed.
941 static void do_io(struct thread_data *td, uint64_t *bytes_done)
945 uint64_t total_bytes, bytes_issued = 0;
947 for (i = 0; i < DDIR_RWDIR_CNT; i++)
948 bytes_done[i] = td->bytes_done[i];
950 if (in_ramp_time(td))
951 td_set_runstate(td, TD_RAMP);
953 td_set_runstate(td, TD_RUNNING);
957 total_bytes = td->o.size;
959 * Allow random overwrite workloads to write up to io_size
960 * before starting verification phase as 'size' doesn't apply.
962 if (td_write(td) && td_random(td) && td->o.norandommap)
963 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
965 * If verify_backlog is enabled, we'll run the verify in this
966 * handler as well. For that case, we may need up to twice the
969 if (td->o.verify != VERIFY_NONE &&
970 (td_write(td) && td->o.verify_backlog))
971 total_bytes += td->o.size;
973 /* In trimwrite mode, each byte is trimmed and then written, so
974 * allow total_bytes or number of ios to be twice as big */
975 if (td_trimwrite(td)) {
976 total_bytes += td->total_io_size;
977 td->o.number_ios *= 2;
980 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
981 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
983 struct timespec comp_time;
988 check_update_rusage(td);
990 if (td->terminate || td->done)
995 if (runtime_exceeded(td, &td->ts_cache)) {
996 __update_ts_cache(td);
997 if (runtime_exceeded(td, &td->ts_cache)) {
998 fio_mark_td_terminate(td);
1003 if (flow_threshold_exceeded(td))
1007 * Break if we exceeded the bytes. The exception is time
1008 * based runs, but we still need to break out of the loop
1009 * for those to run verification, if enabled.
1010 * Jobs read from iolog do not use this stop condition.
1012 if (bytes_issued >= total_bytes &&
1013 !td->o.read_iolog_file &&
1014 (!td->o.time_based ||
1015 (td->o.time_based && td->o.verify != VERIFY_NONE)))
1018 io_u = get_io_u(td);
1019 if (IS_ERR_OR_NULL(io_u)) {
1020 int err = PTR_ERR(io_u);
1024 if (err == -EBUSY) {
1028 if (td->o.latency_target)
1033 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
1034 populate_verify_io_u(td, io_u);
1039 * Add verification end_io handler if:
1040 * - Asked to verify (!td_rw(td))
1041 * - Or the io_u is from our verify list (mixed write/ver)
1043 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1044 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1046 if (verify_state_should_stop(td, io_u)) {
1051 if (td->o.verify_async)
1052 io_u->end_io = verify_io_u_async;
1054 io_u->end_io = verify_io_u;
1055 td_set_runstate(td, TD_VERIFYING);
1056 } else if (in_ramp_time(td))
1057 td_set_runstate(td, TD_RAMP);
1059 td_set_runstate(td, TD_RUNNING);
1062 * Always log IO before it's issued, so we know the specific
1063 * order of it. The logged unit will track when the IO has
1066 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1068 td->o.verify != VERIFY_NONE &&
1069 !td->o.experimental_verify)
1070 log_io_piece(td, io_u);
1072 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1073 const unsigned long long blen = io_u->xfer_buflen;
1074 const enum fio_ddir __ddir = acct_ddir(io_u);
1079 workqueue_enqueue(&td->io_wq, &io_u->work);
1082 if (ddir_rw(__ddir)) {
1083 td->io_issues[__ddir]++;
1084 td->io_issue_bytes[__ddir] += blen;
1085 td->rate_io_issue_bytes[__ddir] += blen;
1088 if (should_check_rate(td)) {
1089 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1090 fio_gettime(&comp_time, NULL);
1094 ret = io_u_submit(td, io_u);
1096 if (should_check_rate(td))
1097 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1099 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1103 * See if we need to complete some commands. Note that
1104 * we can get BUSY even without IO queued, if the
1105 * system is resource starved.
1108 full = queue_full(td) ||
1109 (ret == FIO_Q_BUSY && td->cur_depth);
1110 if (full || io_in_polling(td))
1111 ret = wait_for_completions(td, &comp_time);
1116 if (ddir_rw(ddir) && td->o.thinktime)
1117 handle_thinktime(td, ddir, &comp_time);
1119 if (!ddir_rw_sum(td->bytes_done) &&
1120 !td_ioengine_flagged(td, FIO_NOIO))
1123 if (!in_ramp_time(td) && should_check_rate(td)) {
1124 if (check_min_rate(td, &comp_time)) {
1125 if (exitall_on_terminate || td->o.exitall_error)
1126 fio_terminate_threads(td->groupid, td->o.exit_what);
1127 td_verror(td, EIO, "check_min_rate");
1131 if (!in_ramp_time(td) && td->o.latency_target)
1132 lat_target_check(td);
1135 check_update_rusage(td);
1137 if (td->trim_entries)
1138 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1140 if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
1142 fio_mark_td_terminate(td);
1147 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1148 workqueue_flush(&td->io_wq);
1154 ret = io_u_queued_complete(td, i);
1155 if (td->o.fill_device &&
1156 (td->error == ENOSPC || td->error == EDQUOT))
1160 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1161 td_set_runstate(td, TD_FSYNCING);
1163 for_each_file(td, f, i) {
1164 if (!fio_file_fsync(td, f))
1167 log_err("fio: end_fsync failed for file %s\n",
1172 if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
1173 workqueue_flush(&td->io_wq);
1174 cleanup_pending_aio(td);
1178 * stop job if we failed doing any IO
1180 if (!ddir_rw_sum(td->this_io_bytes))
1183 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1184 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1187 static void free_file_completion_logging(struct thread_data *td)
1192 for_each_file(td, f, i) {
1193 if (!f->last_write_comp)
1195 sfree(f->last_write_comp);
1199 static int init_file_completion_logging(struct thread_data *td,
1205 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1208 for_each_file(td, f, i) {
1209 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1210 if (!f->last_write_comp)
1217 free_file_completion_logging(td);
1218 log_err("fio: failed to alloc write comp data\n");
1222 static void cleanup_io_u(struct thread_data *td)
1226 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1228 if (td->io_ops->io_u_free)
1229 td->io_ops->io_u_free(td, io_u);
1231 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1236 io_u_rexit(&td->io_u_requeues);
1237 io_u_qexit(&td->io_u_freelist, false);
1238 io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1240 free_file_completion_logging(td);
1243 static int init_io_u(struct thread_data *td)
1246 int cl_align, i, max_units;
1249 max_units = td->o.iodepth;
1252 err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1253 err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1254 err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1257 log_err("fio: failed setting up IO queues\n");
1261 cl_align = os_cache_line_size();
1263 for (i = 0; i < max_units; i++) {
1269 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1271 log_err("fio: unable to allocate aligned memory\n");
1276 memset(io_u, 0, sizeof(*io_u));
1277 INIT_FLIST_HEAD(&io_u->verify_list);
1278 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
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);
1300 init_io_u_buffers(td);
1302 if (init_file_completion_logging(td, max_units))
1308 int init_io_u_buffers(struct thread_data *td)
1311 unsigned long long max_bs, min_write;
1316 max_units = td->o.iodepth;
1317 max_bs = td_max_bs(td);
1318 min_write = td->o.min_bs[DDIR_WRITE];
1319 td->orig_buffer_size = (unsigned long long) max_bs
1320 * (unsigned long long) max_units;
1322 if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1326 * if we may later need to do address alignment, then add any
1327 * possible adjustment here so that we don't cause a buffer
1328 * overflow later. this adjustment may be too much if we get
1329 * lucky and the allocator gives us an aligned address.
1331 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1332 td_ioengine_flagged(td, FIO_RAWIO))
1333 td->orig_buffer_size += page_mask + td->o.mem_align;
1335 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1336 unsigned long long bs;
1338 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1339 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1342 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1343 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1347 if (data_xfer && allocate_io_mem(td))
1350 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1351 td_ioengine_flagged(td, FIO_RAWIO))
1352 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1354 p = td->orig_buffer;
1356 for (i = 0; i < max_units; i++) {
1357 io_u = td->io_u_all.io_us[i];
1358 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1362 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1365 io_u_fill_buffer(td, io_u, min_write, max_bs);
1366 if (td_write(td) && td->o.verify_pattern_bytes) {
1368 * Fill the buffer with the pattern if we are
1369 * going to be doing writes.
1371 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1380 #ifdef FIO_HAVE_IOSCHED_SWITCH
1382 * These functions are Linux specific.
1383 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1385 static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
1387 char tmp[256], tmp2[128], *p;
1391 assert(file->du && file->du->sysfs_root);
1392 sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);
1394 f = fopen(tmp, "r+");
1396 if (errno == ENOENT) {
1397 log_err("fio: os or kernel doesn't support IO scheduler"
1401 td_verror(td, errno, "fopen iosched");
1408 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1409 if (ferror(f) || ret != 1) {
1410 td_verror(td, errno, "fwrite");
1418 * Read back and check that the selected scheduler is now the default.
1420 ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1421 if (ferror(f) || ret < 0) {
1422 td_verror(td, errno, "fread");
1428 * either a list of io schedulers or "none\n" is expected. Strip the
1435 * Write to "none" entry doesn't fail, so check the result here.
1437 if (!strcmp(tmp, "none")) {
1438 log_err("fio: io scheduler is not tunable\n");
1443 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1444 if (!strstr(tmp, tmp2)) {
1445 log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
1446 td_verror(td, EINVAL, "iosched_switch");
1455 static int switch_ioscheduler(struct thread_data *td)
1461 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1464 assert(td->files && td->files[0]);
1466 for_each_file(td, f, i) {
1468 /* Only consider regular files and block device files */
1469 switch (f->filetype) {
1471 case FIO_TYPE_BLOCK:
1473 * Make sure that the device hosting the file could
1485 ret = set_ioscheduler(td, f);
1495 static int switch_ioscheduler(struct thread_data *td)
1500 #endif /* FIO_HAVE_IOSCHED_SWITCH */
1502 static bool keep_running(struct thread_data *td)
1504 unsigned long long limit;
1510 if (td->o.time_based)
1516 if (exceeds_number_ios(td))
1520 limit = td->o.io_size;
1524 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1528 * If the difference is less than the maximum IO size, we
1531 diff = limit - ddir_rw_sum(td->io_bytes);
1532 if (diff < td_max_bs(td))
1535 if (fio_files_done(td) && !td->o.io_size)
1544 static int exec_string(struct thread_options *o, const char *string,
1550 if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1553 log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1557 log_err("fio: exec of cmd <%s> failed\n", str);
1564 * Dry run to compute correct state of numberio for verification.
1566 static uint64_t do_dry_run(struct thread_data *td)
1568 td_set_runstate(td, TD_RUNNING);
1570 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1571 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1575 if (td->terminate || td->done)
1578 io_u = get_io_u(td);
1579 if (IS_ERR_OR_NULL(io_u))
1582 io_u_set(td, io_u, IO_U_F_FLIGHT);
1585 if (ddir_rw(acct_ddir(io_u)))
1586 td->io_issues[acct_ddir(io_u)]++;
1587 if (ddir_rw(io_u->ddir)) {
1588 io_u_mark_depth(td, 1);
1589 td->ts.total_io_u[io_u->ddir]++;
1592 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1594 td->o.verify != VERIFY_NONE &&
1595 !td->o.experimental_verify)
1596 log_io_piece(td, io_u);
1598 ret = io_u_sync_complete(td, io_u);
1602 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1606 struct thread_data *td;
1607 struct sk_out *sk_out;
1611 * Entry point for the thread based jobs. The process based jobs end up
1612 * here as well, after a little setup.
1614 static void *thread_main(void *data)
1616 struct fork_data *fd = data;
1617 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1618 struct thread_data *td = fd->td;
1619 struct thread_options *o = &td->o;
1620 struct sk_out *sk_out = fd->sk_out;
1621 uint64_t bytes_done[DDIR_RWDIR_CNT];
1622 int deadlock_loop_cnt;
1626 sk_out_assign(sk_out);
1629 if (!o->use_thread) {
1635 fio_local_clock_init();
1637 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1640 fio_server_send_start(td);
1642 INIT_FLIST_HEAD(&td->io_log_list);
1643 INIT_FLIST_HEAD(&td->io_hist_list);
1644 INIT_FLIST_HEAD(&td->verify_list);
1645 INIT_FLIST_HEAD(&td->trim_list);
1646 td->io_hist_tree = RB_ROOT;
1648 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1650 td_verror(td, ret, "mutex_cond_init_pshared");
1653 ret = cond_init_pshared(&td->verify_cond);
1655 td_verror(td, ret, "mutex_cond_pshared");
1659 td_set_runstate(td, TD_INITIALIZED);
1660 dprint(FD_MUTEX, "up startup_sem\n");
1661 fio_sem_up(startup_sem);
1662 dprint(FD_MUTEX, "wait on td->sem\n");
1663 fio_sem_down(td->sem);
1664 dprint(FD_MUTEX, "done waiting on td->sem\n");
1667 * A new gid requires privilege, so we need to do this before setting
1670 if (o->gid != -1U && setgid(o->gid)) {
1671 td_verror(td, errno, "setgid");
1674 if (o->uid != -1U && setuid(o->uid)) {
1675 td_verror(td, errno, "setuid");
1679 td_zone_gen_index(td);
1682 * Do this early, we don't want the compress threads to be limited
1683 * to the same CPUs as the IO workers. So do this before we set
1684 * any potential CPU affinity
1686 if (iolog_compress_init(td, sk_out))
1690 * If we have a gettimeofday() thread, make sure we exclude that
1691 * thread from this job
1694 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1697 * Set affinity first, in case it has an impact on the memory
1700 if (fio_option_is_set(o, cpumask)) {
1701 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1702 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1704 log_err("fio: no CPUs set\n");
1705 log_err("fio: Try increasing number of available CPUs\n");
1706 td_verror(td, EINVAL, "cpus_split");
1710 ret = fio_setaffinity(td->pid, o->cpumask);
1712 td_verror(td, errno, "cpu_set_affinity");
1717 #ifdef CONFIG_LIBNUMA
1718 /* numa node setup */
1719 if (fio_option_is_set(o, numa_cpunodes) ||
1720 fio_option_is_set(o, numa_memnodes)) {
1721 struct bitmask *mask;
1723 if (numa_available() < 0) {
1724 td_verror(td, errno, "Does not support NUMA API\n");
1728 if (fio_option_is_set(o, numa_cpunodes)) {
1729 mask = numa_parse_nodestring(o->numa_cpunodes);
1730 ret = numa_run_on_node_mask(mask);
1731 numa_free_nodemask(mask);
1733 td_verror(td, errno, \
1734 "numa_run_on_node_mask failed\n");
1739 if (fio_option_is_set(o, numa_memnodes)) {
1741 if (o->numa_memnodes)
1742 mask = numa_parse_nodestring(o->numa_memnodes);
1744 switch (o->numa_mem_mode) {
1745 case MPOL_INTERLEAVE:
1746 numa_set_interleave_mask(mask);
1749 numa_set_membind(mask);
1752 numa_set_localalloc();
1754 case MPOL_PREFERRED:
1755 numa_set_preferred(o->numa_mem_prefer_node);
1763 numa_free_nodemask(mask);
1769 if (fio_pin_memory(td))
1773 * May alter parameters that init_io_u() will use, so we need to
1776 if (!init_iolog(td))
1779 /* ioprio_set() has to be done before td_io_init() */
1780 if (fio_option_is_set(o, ioprio) ||
1781 fio_option_is_set(o, ioprio_class)) {
1782 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1784 td_verror(td, errno, "ioprio_set");
1787 td->ioprio = ioprio_value(o->ioprio_class, o->ioprio);
1788 td->ts.ioprio = td->ioprio;
1794 if (td_ioengine_flagged(td, FIO_SYNCIO) && td->o.iodepth > 1) {
1795 log_info("note: both iodepth >= 1 and synchronous I/O engine "
1796 "are selected, queue depth will be capped at 1\n");
1802 if (td->io_ops->post_init && td->io_ops->post_init(td))
1805 if (o->verify_async && verify_async_init(td))
1808 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1812 if (nice(o->nice) == -1 && errno != 0) {
1813 td_verror(td, errno, "nice");
1817 if (o->ioscheduler && switch_ioscheduler(td))
1820 if (!o->create_serialize && setup_files(td))
1823 if (!init_random_map(td))
1826 if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1829 if (o->pre_read && !pre_read_files(td))
1832 fio_verify_init(td);
1834 if (rate_submit_init(td, sk_out))
1837 set_epoch_time(td, o->log_unix_epoch | o->log_alternate_epoch, o->log_alternate_epoch_clock_id);
1838 fio_getrusage(&td->ru_start);
1839 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1840 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1841 memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1845 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1846 o->ratemin[DDIR_TRIM]) {
1847 memcpy(&td->last_rate_check_time[DDIR_READ], &td->bw_sample_time,
1848 sizeof(td->bw_sample_time));
1849 memcpy(&td->last_rate_check_time[DDIR_WRITE], &td->bw_sample_time,
1850 sizeof(td->bw_sample_time));
1851 memcpy(&td->last_rate_check_time[DDIR_TRIM], &td->bw_sample_time,
1852 sizeof(td->bw_sample_time));
1855 memset(bytes_done, 0, sizeof(bytes_done));
1856 clear_state = false;
1858 while (keep_running(td)) {
1859 uint64_t verify_bytes;
1861 fio_gettime(&td->start, NULL);
1862 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1865 clear_io_state(td, 0);
1867 if (o->unlink_each_loop && unlink_all_files(td))
1871 prune_io_piece_log(td);
1873 if (td->o.verify_only && td_write(td))
1874 verify_bytes = do_dry_run(td);
1876 do_io(td, bytes_done);
1878 if (!ddir_rw_sum(bytes_done)) {
1879 fio_mark_td_terminate(td);
1882 verify_bytes = bytes_done[DDIR_WRITE] +
1883 bytes_done[DDIR_TRIM];
1888 * If we took too long to shut down, the main thread could
1889 * already consider us reaped/exited. If that happens, break
1892 if (td->runstate >= TD_EXITED)
1898 * Make sure we've successfully updated the rusage stats
1899 * before waiting on the stat mutex. Otherwise we could have
1900 * the stat thread holding stat mutex and waiting for
1901 * the rusage_sem, which would never get upped because
1902 * this thread is waiting for the stat mutex.
1904 deadlock_loop_cnt = 0;
1906 check_update_rusage(td);
1907 if (!fio_sem_down_trylock(stat_sem))
1910 if (deadlock_loop_cnt++ > 5000) {
1911 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1912 td->error = EDEADLK;
1917 if (td_read(td) && td->io_bytes[DDIR_READ])
1918 update_runtime(td, elapsed_us, DDIR_READ);
1919 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1920 update_runtime(td, elapsed_us, DDIR_WRITE);
1921 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1922 update_runtime(td, elapsed_us, DDIR_TRIM);
1923 fio_gettime(&td->start, NULL);
1924 fio_sem_up(stat_sem);
1926 if (td->error || td->terminate)
1929 if (!o->do_verify ||
1930 o->verify == VERIFY_NONE ||
1931 td_ioengine_flagged(td, FIO_UNIDIR))
1934 clear_io_state(td, 0);
1936 fio_gettime(&td->start, NULL);
1938 do_verify(td, verify_bytes);
1941 * See comment further up for why this is done here.
1943 check_update_rusage(td);
1945 fio_sem_down(stat_sem);
1946 update_runtime(td, elapsed_us, DDIR_READ);
1947 fio_gettime(&td->start, NULL);
1948 fio_sem_up(stat_sem);
1950 if (td->error || td->terminate)
1955 * Acquire this lock if we were doing overlap checking in
1956 * offload mode so that we don't clean up this job while
1957 * another thread is checking its io_u's for overlap
1959 if (td_offload_overlap(td)) {
1960 int res = pthread_mutex_lock(&overlap_check);
1963 td_set_runstate(td, TD_FINISHING);
1964 if (td_offload_overlap(td)) {
1965 res = pthread_mutex_unlock(&overlap_check);
1969 update_rusage_stat(td);
1970 td->ts.total_run_time = mtime_since_now(&td->epoch);
1971 for_each_rw_ddir(ddir) {
1972 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
1975 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1976 (td->o.verify != VERIFY_NONE && td_write(td)))
1977 verify_save_state(td->thread_number);
1979 fio_unpin_memory(td);
1981 td_writeout_logs(td, true);
1983 iolog_compress_exit(td);
1984 rate_submit_exit(td);
1986 if (o->exec_postrun)
1987 exec_string(o, o->exec_postrun, "postrun");
1989 if (exitall_on_terminate || (o->exitall_error && td->error))
1990 fio_terminate_threads(td->groupid, td->o.exit_what);
1994 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1997 if (o->verify_async)
1998 verify_async_exit(td);
2000 close_and_free_files(td);
2003 cgroup_shutdown(td, cgroup_mnt);
2004 verify_free_state(td);
2005 td_zone_free_index(td);
2007 if (fio_option_is_set(o, cpumask)) {
2008 ret = fio_cpuset_exit(&o->cpumask);
2010 td_verror(td, ret, "fio_cpuset_exit");
2014 * do this very late, it will log file closing as well
2016 if (o->write_iolog_file)
2017 write_iolog_close(td);
2018 if (td->io_log_rfile)
2019 fclose(td->io_log_rfile);
2021 td_set_runstate(td, TD_EXITED);
2024 * Do this last after setting our runstate to exited, so we
2025 * know that the stat thread is signaled.
2027 check_update_rusage(td);
2030 return (void *) (uintptr_t) td->error;
2034 * Run over the job map and reap the threads that have exited, if any.
2036 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
2039 struct thread_data *td;
2040 unsigned int cputhreads, realthreads, pending;
2044 * reap exited threads (TD_EXITED -> TD_REAPED)
2046 realthreads = pending = cputhreads = 0;
2047 for_each_td(td, i) {
2050 if (!strcmp(td->o.ioengine, "cpuio"))
2059 if (td->runstate == TD_REAPED)
2061 if (td->o.use_thread) {
2062 if (td->runstate == TD_EXITED) {
2063 td_set_runstate(td, TD_REAPED);
2070 if (td->runstate == TD_EXITED)
2074 * check if someone quit or got killed in an unusual way
2076 ret = waitpid(td->pid, &status, flags);
2078 if (errno == ECHILD) {
2079 log_err("fio: pid=%d disappeared %d\n",
2080 (int) td->pid, td->runstate);
2082 td_set_runstate(td, TD_REAPED);
2086 } else if (ret == td->pid) {
2087 if (WIFSIGNALED(status)) {
2088 int sig = WTERMSIG(status);
2090 if (sig != SIGTERM && sig != SIGUSR2)
2091 log_err("fio: pid=%d, got signal=%d\n",
2092 (int) td->pid, sig);
2094 td_set_runstate(td, TD_REAPED);
2097 if (WIFEXITED(status)) {
2098 if (WEXITSTATUS(status) && !td->error)
2099 td->error = WEXITSTATUS(status);
2101 td_set_runstate(td, TD_REAPED);
2107 * If the job is stuck, do a forceful timeout of it and
2110 if (td->terminate &&
2111 td->runstate < TD_FSYNCING &&
2112 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2113 log_err("fio: job '%s' (state=%d) hasn't exited in "
2114 "%lu seconds, it appears to be stuck. Doing "
2115 "forceful exit of this job.\n",
2116 td->o.name, td->runstate,
2117 (unsigned long) time_since_now(&td->terminate_time));
2118 td_set_runstate(td, TD_REAPED);
2123 * thread is not dead, continue
2129 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2130 (*t_rate) -= ddir_rw_sum(td->o.rate);
2137 done_secs += mtime_since_now(&td->epoch) / 1000;
2138 profile_td_exit(td);
2142 if (*nr_running == cputhreads && !pending && realthreads)
2143 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2146 static bool __check_trigger_file(void)
2153 if (stat(trigger_file, &sb))
2156 if (unlink(trigger_file) < 0)
2157 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2163 static bool trigger_timedout(void)
2165 if (trigger_timeout)
2166 if (time_since_genesis() >= trigger_timeout) {
2167 trigger_timeout = 0;
2174 void exec_trigger(const char *cmd)
2178 if (!cmd || cmd[0] == '\0')
2183 log_err("fio: failed executing %s trigger\n", cmd);
2186 void check_trigger_file(void)
2188 if (__check_trigger_file() || trigger_timedout()) {
2190 fio_clients_send_trigger(trigger_remote_cmd);
2192 verify_save_state(IO_LIST_ALL);
2193 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2194 exec_trigger(trigger_cmd);
2199 static int fio_verify_load_state(struct thread_data *td)
2203 if (!td->o.verify_state)
2209 ret = fio_server_get_verify_state(td->o.name,
2210 td->thread_number - 1, &data);
2212 verify_assign_state(td, data);
2214 char prefix[PATH_MAX];
2217 sprintf(prefix, "%s%clocal", aux_path,
2218 FIO_OS_PATH_SEPARATOR);
2220 strcpy(prefix, "local");
2221 ret = verify_load_state(td, prefix);
2227 static void do_usleep(unsigned int usecs)
2229 check_for_running_stats();
2230 check_trigger_file();
2234 static bool check_mount_writes(struct thread_data *td)
2239 if (!td_write(td) || td->o.allow_mounted_write)
2243 * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2244 * are mkfs'd and mounted.
2246 for_each_file(td, f, i) {
2247 #ifdef FIO_HAVE_CHARDEV_SIZE
2248 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2250 if (f->filetype != FIO_TYPE_BLOCK)
2253 if (device_is_mounted(f->file_name))
2259 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2263 static bool waitee_running(struct thread_data *me)
2265 const char *waitee = me->o.wait_for;
2266 const char *self = me->o.name;
2267 struct thread_data *td;
2273 for_each_td(td, i) {
2274 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2277 if (td->runstate < TD_EXITED) {
2278 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2280 runstate_to_name(td->runstate));
2285 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2290 * Main function for kicking off and reaping jobs, as needed.
2292 static void run_threads(struct sk_out *sk_out)
2294 struct thread_data *td;
2295 unsigned int i, todo, nr_running, nr_started;
2296 uint64_t m_rate, t_rate;
2299 if (fio_gtod_offload && fio_start_gtod_thread())
2302 fio_idle_prof_init();
2306 nr_thread = nr_process = 0;
2307 for_each_td(td, i) {
2308 if (check_mount_writes(td))
2310 if (td->o.use_thread)
2316 if (output_format & FIO_OUTPUT_NORMAL) {
2317 struct buf_output out;
2319 buf_output_init(&out);
2320 __log_buf(&out, "Starting ");
2322 __log_buf(&out, "%d thread%s", nr_thread,
2323 nr_thread > 1 ? "s" : "");
2326 __log_buf(&out, " and ");
2327 __log_buf(&out, "%d process%s", nr_process,
2328 nr_process > 1 ? "es" : "");
2330 __log_buf(&out, "\n");
2331 log_info_buf(out.buf, out.buflen);
2332 buf_output_free(&out);
2335 todo = thread_number;
2338 m_rate = t_rate = 0;
2340 for_each_td(td, i) {
2341 print_status_init(td->thread_number - 1);
2343 if (!td->o.create_serialize)
2346 if (fio_verify_load_state(td))
2350 * do file setup here so it happens sequentially,
2351 * we don't want X number of threads getting their
2352 * client data interspersed on disk
2354 if (setup_files(td)) {
2358 log_err("fio: pid=%d, err=%d/%s\n",
2359 (int) td->pid, td->error, td->verror);
2360 td_set_runstate(td, TD_REAPED);
2367 * for sharing to work, each job must always open
2368 * its own files. so close them, if we opened them
2371 for_each_file(td, f, j) {
2372 if (fio_file_open(f))
2373 td_io_close_file(td, f);
2378 /* start idle threads before io threads start to run */
2379 fio_idle_prof_start();
2384 struct thread_data *map[REAL_MAX_JOBS];
2385 struct timespec this_start;
2386 int this_jobs = 0, left;
2387 struct fork_data *fd;
2390 * create threads (TD_NOT_CREATED -> TD_CREATED)
2392 for_each_td(td, i) {
2393 if (td->runstate != TD_NOT_CREATED)
2397 * never got a chance to start, killed by other
2398 * thread for some reason
2400 if (td->terminate) {
2405 if (td->o.start_delay) {
2406 spent = utime_since_genesis();
2408 if (td->o.start_delay > spent)
2412 if (td->o.stonewall && (nr_started || nr_running)) {
2413 dprint(FD_PROCESS, "%s: stonewall wait\n",
2418 if (waitee_running(td)) {
2419 dprint(FD_PROCESS, "%s: waiting for %s\n",
2420 td->o.name, td->o.wait_for);
2426 td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2427 td->update_rusage = 0;
2430 * Set state to created. Thread will transition
2431 * to TD_INITIALIZED when it's done setting up.
2433 td_set_runstate(td, TD_CREATED);
2434 map[this_jobs++] = td;
2437 fd = calloc(1, sizeof(*fd));
2439 fd->sk_out = sk_out;
2441 if (td->o.use_thread) {
2444 dprint(FD_PROCESS, "will pthread_create\n");
2445 ret = pthread_create(&td->thread, NULL,
2448 log_err("pthread_create: %s\n",
2455 ret = pthread_detach(td->thread);
2457 log_err("pthread_detach: %s",
2462 dprint(FD_PROCESS, "will fork\n");
2469 ret = (int)(uintptr_t)thread_main(fd);
2471 } else if (i == fio_debug_jobno)
2472 *fio_debug_jobp = pid;
2477 dprint(FD_MUTEX, "wait on startup_sem\n");
2478 if (fio_sem_down_timeout(startup_sem, 10000)) {
2479 log_err("fio: job startup hung? exiting.\n");
2480 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2486 dprint(FD_MUTEX, "done waiting on startup_sem\n");
2490 * Wait for the started threads to transition to
2493 fio_gettime(&this_start, NULL);
2495 while (left && !fio_abort) {
2496 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2501 for (i = 0; i < this_jobs; i++) {
2505 if (td->runstate == TD_INITIALIZED) {
2508 } else if (td->runstate >= TD_EXITED) {
2512 nr_running++; /* work-around... */
2518 log_err("fio: %d job%s failed to start\n", left,
2519 left > 1 ? "s" : "");
2520 for (i = 0; i < this_jobs; i++) {
2524 kill(td->pid, SIGTERM);
2530 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2532 for_each_td(td, i) {
2533 if (td->runstate != TD_INITIALIZED)
2536 if (in_ramp_time(td))
2537 td_set_runstate(td, TD_RAMP);
2539 td_set_runstate(td, TD_RUNNING);
2542 m_rate += ddir_rw_sum(td->o.ratemin);
2543 t_rate += ddir_rw_sum(td->o.rate);
2545 fio_sem_up(td->sem);
2548 reap_threads(&nr_running, &t_rate, &m_rate);
2554 while (nr_running) {
2555 reap_threads(&nr_running, &t_rate, &m_rate);
2559 fio_idle_prof_stop();
2564 static void free_disk_util(void)
2566 disk_util_prune_entries();
2567 helper_thread_destroy();
2570 int fio_backend(struct sk_out *sk_out)
2572 struct thread_data *td;
2576 if (load_profile(exec_profile))
2579 exec_profile = NULL;
2585 struct log_params p = {
2586 .log_type = IO_LOG_TYPE_BW,
2589 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2590 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2591 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2594 if (init_global_dedupe_working_set_seeds()) {
2595 log_err("fio: failed to initialize global dedupe working set\n");
2599 startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2601 is_local_backend = true;
2602 if (startup_sem == NULL)
2607 if (helper_thread_create(startup_sem, sk_out))
2608 log_err("fio: failed to create helper thread\n");
2610 cgroup_list = smalloc(sizeof(*cgroup_list));
2612 INIT_FLIST_HEAD(cgroup_list);
2614 run_threads(sk_out);
2616 helper_thread_exit();
2621 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2622 struct io_log *log = agg_io_log[i];
2624 flush_log(log, false);
2630 for_each_td(td, i) {
2631 struct thread_stat *ts = &td->ts;
2633 free_clat_prio_stats(ts);
2634 steadystate_free(td);
2635 fio_options_free(td);
2636 fio_dump_options_free(td);
2637 if (td->rusage_sem) {
2638 fio_sem_remove(td->rusage_sem);
2639 td->rusage_sem = NULL;
2641 fio_sem_remove(td->sem);
2647 cgroup_kill(cgroup_list);
2651 fio_sem_remove(startup_sem);