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 (ddir_rw_sum(td->bytes_done) + 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 to be twice as big */
975 if (td_trimwrite(td))
976 total_bytes += td->total_io_size;
978 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
979 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
981 struct timespec comp_time;
986 check_update_rusage(td);
988 if (td->terminate || td->done)
993 if (runtime_exceeded(td, &td->ts_cache)) {
994 __update_ts_cache(td);
995 if (runtime_exceeded(td, &td->ts_cache)) {
996 fio_mark_td_terminate(td);
1001 if (flow_threshold_exceeded(td))
1005 * Break if we exceeded the bytes. The exception is time
1006 * based runs, but we still need to break out of the loop
1007 * for those to run verification, if enabled.
1008 * Jobs read from iolog do not use this stop condition.
1010 if (bytes_issued >= total_bytes &&
1011 !td->o.read_iolog_file &&
1012 (!td->o.time_based ||
1013 (td->o.time_based && td->o.verify != VERIFY_NONE)))
1016 io_u = get_io_u(td);
1017 if (IS_ERR_OR_NULL(io_u)) {
1018 int err = PTR_ERR(io_u);
1022 if (err == -EBUSY) {
1026 if (td->o.latency_target)
1031 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
1032 populate_verify_io_u(td, io_u);
1037 * Add verification end_io handler if:
1038 * - Asked to verify (!td_rw(td))
1039 * - Or the io_u is from our verify list (mixed write/ver)
1041 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1042 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1044 if (verify_state_should_stop(td, io_u)) {
1049 if (td->o.verify_async)
1050 io_u->end_io = verify_io_u_async;
1052 io_u->end_io = verify_io_u;
1053 td_set_runstate(td, TD_VERIFYING);
1054 } else if (in_ramp_time(td))
1055 td_set_runstate(td, TD_RAMP);
1057 td_set_runstate(td, TD_RUNNING);
1060 * Always log IO before it's issued, so we know the specific
1061 * order of it. The logged unit will track when the IO has
1064 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1066 td->o.verify != VERIFY_NONE &&
1067 !td->o.experimental_verify)
1068 log_io_piece(td, io_u);
1070 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1071 const unsigned long long blen = io_u->xfer_buflen;
1072 const enum fio_ddir __ddir = acct_ddir(io_u);
1077 workqueue_enqueue(&td->io_wq, &io_u->work);
1080 if (ddir_rw(__ddir)) {
1081 td->io_issues[__ddir]++;
1082 td->io_issue_bytes[__ddir] += blen;
1083 td->rate_io_issue_bytes[__ddir] += blen;
1086 if (should_check_rate(td)) {
1087 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1088 fio_gettime(&comp_time, NULL);
1092 ret = io_u_submit(td, io_u);
1094 if (should_check_rate(td))
1095 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1097 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1101 * See if we need to complete some commands. Note that
1102 * we can get BUSY even without IO queued, if the
1103 * system is resource starved.
1106 full = queue_full(td) ||
1107 (ret == FIO_Q_BUSY && td->cur_depth);
1108 if (full || io_in_polling(td))
1109 ret = wait_for_completions(td, &comp_time);
1114 if (ddir_rw(ddir) && td->o.thinktime)
1115 handle_thinktime(td, ddir, &comp_time);
1117 if (!ddir_rw_sum(td->bytes_done) &&
1118 !td_ioengine_flagged(td, FIO_NOIO))
1121 if (!in_ramp_time(td) && should_check_rate(td)) {
1122 if (check_min_rate(td, &comp_time)) {
1123 if (exitall_on_terminate || td->o.exitall_error)
1124 fio_terminate_threads(td->groupid, td->o.exit_what);
1125 td_verror(td, EIO, "check_min_rate");
1129 if (!in_ramp_time(td) && td->o.latency_target)
1130 lat_target_check(td);
1133 check_update_rusage(td);
1135 if (td->trim_entries)
1136 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1138 if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
1140 fio_mark_td_terminate(td);
1145 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1146 workqueue_flush(&td->io_wq);
1152 ret = io_u_queued_complete(td, i);
1153 if (td->o.fill_device &&
1154 (td->error == ENOSPC || td->error == EDQUOT))
1158 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1159 td_set_runstate(td, TD_FSYNCING);
1161 for_each_file(td, f, i) {
1162 if (!fio_file_fsync(td, f))
1165 log_err("fio: end_fsync failed for file %s\n",
1170 if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
1171 workqueue_flush(&td->io_wq);
1172 cleanup_pending_aio(td);
1176 * stop job if we failed doing any IO
1178 if (!ddir_rw_sum(td->this_io_bytes))
1181 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1182 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1185 static void free_file_completion_logging(struct thread_data *td)
1190 for_each_file(td, f, i) {
1191 if (!f->last_write_comp)
1193 sfree(f->last_write_comp);
1197 static int init_file_completion_logging(struct thread_data *td,
1203 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1206 for_each_file(td, f, i) {
1207 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1208 if (!f->last_write_comp)
1215 free_file_completion_logging(td);
1216 log_err("fio: failed to alloc write comp data\n");
1220 static void cleanup_io_u(struct thread_data *td)
1224 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1226 if (td->io_ops->io_u_free)
1227 td->io_ops->io_u_free(td, io_u);
1229 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1234 io_u_rexit(&td->io_u_requeues);
1235 io_u_qexit(&td->io_u_freelist, false);
1236 io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1238 free_file_completion_logging(td);
1241 static int init_io_u(struct thread_data *td)
1244 int cl_align, i, max_units;
1247 max_units = td->o.iodepth;
1250 err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1251 err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1252 err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1255 log_err("fio: failed setting up IO queues\n");
1259 cl_align = os_cache_line_size();
1261 for (i = 0; i < max_units; i++) {
1267 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1269 log_err("fio: unable to allocate aligned memory\n");
1274 memset(io_u, 0, sizeof(*io_u));
1275 INIT_FLIST_HEAD(&io_u->verify_list);
1276 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1279 io_u->flags = IO_U_F_FREE;
1280 io_u_qpush(&td->io_u_freelist, io_u);
1283 * io_u never leaves this stack, used for iteration of all
1286 io_u_qpush(&td->io_u_all, io_u);
1288 if (td->io_ops->io_u_init) {
1289 int ret = td->io_ops->io_u_init(td, io_u);
1292 log_err("fio: failed to init engine data: %d\n", ret);
1298 init_io_u_buffers(td);
1300 if (init_file_completion_logging(td, max_units))
1306 int init_io_u_buffers(struct thread_data *td)
1309 unsigned long long max_bs, min_write;
1314 max_units = td->o.iodepth;
1315 max_bs = td_max_bs(td);
1316 min_write = td->o.min_bs[DDIR_WRITE];
1317 td->orig_buffer_size = (unsigned long long) max_bs
1318 * (unsigned long long) max_units;
1320 if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1324 * if we may later need to do address alignment, then add any
1325 * possible adjustment here so that we don't cause a buffer
1326 * overflow later. this adjustment may be too much if we get
1327 * lucky and the allocator gives us an aligned address.
1329 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1330 td_ioengine_flagged(td, FIO_RAWIO))
1331 td->orig_buffer_size += page_mask + td->o.mem_align;
1333 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1334 unsigned long long bs;
1336 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1337 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1340 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1341 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1345 if (data_xfer && allocate_io_mem(td))
1348 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1349 td_ioengine_flagged(td, FIO_RAWIO))
1350 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1352 p = td->orig_buffer;
1354 for (i = 0; i < max_units; i++) {
1355 io_u = td->io_u_all.io_us[i];
1356 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1360 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1363 io_u_fill_buffer(td, io_u, min_write, max_bs);
1364 if (td_write(td) && td->o.verify_pattern_bytes) {
1366 * Fill the buffer with the pattern if we are
1367 * going to be doing writes.
1369 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1378 #ifdef FIO_HAVE_IOSCHED_SWITCH
1380 * These functions are Linux specific.
1381 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1383 static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
1385 char tmp[256], tmp2[128], *p;
1389 assert(file->du && file->du->sysfs_root);
1390 sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);
1392 f = fopen(tmp, "r+");
1394 if (errno == ENOENT) {
1395 log_err("fio: os or kernel doesn't support IO scheduler"
1399 td_verror(td, errno, "fopen iosched");
1406 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1407 if (ferror(f) || ret != 1) {
1408 td_verror(td, errno, "fwrite");
1416 * Read back and check that the selected scheduler is now the default.
1418 ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1419 if (ferror(f) || ret < 0) {
1420 td_verror(td, errno, "fread");
1426 * either a list of io schedulers or "none\n" is expected. Strip the
1433 * Write to "none" entry doesn't fail, so check the result here.
1435 if (!strcmp(tmp, "none")) {
1436 log_err("fio: io scheduler is not tunable\n");
1441 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1442 if (!strstr(tmp, tmp2)) {
1443 log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
1444 td_verror(td, EINVAL, "iosched_switch");
1453 static int switch_ioscheduler(struct thread_data *td)
1459 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1462 assert(td->files && td->files[0]);
1464 for_each_file(td, f, i) {
1466 /* Only consider regular files and block device files */
1467 switch (f->filetype) {
1469 case FIO_TYPE_BLOCK:
1471 * Make sure that the device hosting the file could
1483 ret = set_ioscheduler(td, f);
1493 static int switch_ioscheduler(struct thread_data *td)
1498 #endif /* FIO_HAVE_IOSCHED_SWITCH */
1500 static bool keep_running(struct thread_data *td)
1502 unsigned long long limit;
1508 if (td->o.time_based)
1514 if (exceeds_number_ios(td))
1518 limit = td->o.io_size;
1522 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1526 * If the difference is less than the maximum IO size, we
1529 diff = limit - ddir_rw_sum(td->io_bytes);
1530 if (diff < td_max_bs(td))
1533 if (fio_files_done(td) && !td->o.io_size)
1542 static int exec_string(struct thread_options *o, const char *string,
1548 if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1551 log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1555 log_err("fio: exec of cmd <%s> failed\n", str);
1562 * Dry run to compute correct state of numberio for verification.
1564 static uint64_t do_dry_run(struct thread_data *td)
1566 td_set_runstate(td, TD_RUNNING);
1568 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1569 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1573 if (td->terminate || td->done)
1576 io_u = get_io_u(td);
1577 if (IS_ERR_OR_NULL(io_u))
1580 io_u_set(td, io_u, IO_U_F_FLIGHT);
1583 if (ddir_rw(acct_ddir(io_u)))
1584 td->io_issues[acct_ddir(io_u)]++;
1585 if (ddir_rw(io_u->ddir)) {
1586 io_u_mark_depth(td, 1);
1587 td->ts.total_io_u[io_u->ddir]++;
1590 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1592 td->o.verify != VERIFY_NONE &&
1593 !td->o.experimental_verify)
1594 log_io_piece(td, io_u);
1596 ret = io_u_sync_complete(td, io_u);
1600 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1604 struct thread_data *td;
1605 struct sk_out *sk_out;
1609 * Entry point for the thread based jobs. The process based jobs end up
1610 * here as well, after a little setup.
1612 static void *thread_main(void *data)
1614 struct fork_data *fd = data;
1615 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1616 struct thread_data *td = fd->td;
1617 struct thread_options *o = &td->o;
1618 struct sk_out *sk_out = fd->sk_out;
1619 uint64_t bytes_done[DDIR_RWDIR_CNT];
1620 int deadlock_loop_cnt;
1624 sk_out_assign(sk_out);
1627 if (!o->use_thread) {
1633 fio_local_clock_init();
1635 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1638 fio_server_send_start(td);
1640 INIT_FLIST_HEAD(&td->io_log_list);
1641 INIT_FLIST_HEAD(&td->io_hist_list);
1642 INIT_FLIST_HEAD(&td->verify_list);
1643 INIT_FLIST_HEAD(&td->trim_list);
1644 td->io_hist_tree = RB_ROOT;
1646 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1648 td_verror(td, ret, "mutex_cond_init_pshared");
1651 ret = cond_init_pshared(&td->verify_cond);
1653 td_verror(td, ret, "mutex_cond_pshared");
1657 td_set_runstate(td, TD_INITIALIZED);
1658 dprint(FD_MUTEX, "up startup_sem\n");
1659 fio_sem_up(startup_sem);
1660 dprint(FD_MUTEX, "wait on td->sem\n");
1661 fio_sem_down(td->sem);
1662 dprint(FD_MUTEX, "done waiting on td->sem\n");
1665 * A new gid requires privilege, so we need to do this before setting
1668 if (o->gid != -1U && setgid(o->gid)) {
1669 td_verror(td, errno, "setgid");
1672 if (o->uid != -1U && setuid(o->uid)) {
1673 td_verror(td, errno, "setuid");
1677 td_zone_gen_index(td);
1680 * Do this early, we don't want the compress threads to be limited
1681 * to the same CPUs as the IO workers. So do this before we set
1682 * any potential CPU affinity
1684 if (iolog_compress_init(td, sk_out))
1688 * If we have a gettimeofday() thread, make sure we exclude that
1689 * thread from this job
1692 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1695 * Set affinity first, in case it has an impact on the memory
1698 if (fio_option_is_set(o, cpumask)) {
1699 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1700 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1702 log_err("fio: no CPUs set\n");
1703 log_err("fio: Try increasing number of available CPUs\n");
1704 td_verror(td, EINVAL, "cpus_split");
1708 ret = fio_setaffinity(td->pid, o->cpumask);
1710 td_verror(td, errno, "cpu_set_affinity");
1715 #ifdef CONFIG_LIBNUMA
1716 /* numa node setup */
1717 if (fio_option_is_set(o, numa_cpunodes) ||
1718 fio_option_is_set(o, numa_memnodes)) {
1719 struct bitmask *mask;
1721 if (numa_available() < 0) {
1722 td_verror(td, errno, "Does not support NUMA API\n");
1726 if (fio_option_is_set(o, numa_cpunodes)) {
1727 mask = numa_parse_nodestring(o->numa_cpunodes);
1728 ret = numa_run_on_node_mask(mask);
1729 numa_free_nodemask(mask);
1731 td_verror(td, errno, \
1732 "numa_run_on_node_mask failed\n");
1737 if (fio_option_is_set(o, numa_memnodes)) {
1739 if (o->numa_memnodes)
1740 mask = numa_parse_nodestring(o->numa_memnodes);
1742 switch (o->numa_mem_mode) {
1743 case MPOL_INTERLEAVE:
1744 numa_set_interleave_mask(mask);
1747 numa_set_membind(mask);
1750 numa_set_localalloc();
1752 case MPOL_PREFERRED:
1753 numa_set_preferred(o->numa_mem_prefer_node);
1761 numa_free_nodemask(mask);
1767 if (fio_pin_memory(td))
1771 * May alter parameters that init_io_u() will use, so we need to
1774 if (!init_iolog(td))
1777 /* ioprio_set() has to be done before td_io_init() */
1778 if (fio_option_is_set(o, ioprio) ||
1779 fio_option_is_set(o, ioprio_class)) {
1780 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1782 td_verror(td, errno, "ioprio_set");
1785 td->ioprio = ioprio_value(o->ioprio_class, o->ioprio);
1786 td->ts.ioprio = td->ioprio;
1795 if (td->io_ops->post_init && td->io_ops->post_init(td))
1798 if (o->verify_async && verify_async_init(td))
1801 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1805 if (nice(o->nice) == -1 && errno != 0) {
1806 td_verror(td, errno, "nice");
1810 if (o->ioscheduler && switch_ioscheduler(td))
1813 if (!o->create_serialize && setup_files(td))
1816 if (!init_random_map(td))
1819 if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1822 if (o->pre_read && !pre_read_files(td))
1825 fio_verify_init(td);
1827 if (rate_submit_init(td, sk_out))
1830 set_epoch_time(td, o->log_unix_epoch | o->log_alternate_epoch, o->log_alternate_epoch_clock_id);
1831 fio_getrusage(&td->ru_start);
1832 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1833 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1834 memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1838 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1839 o->ratemin[DDIR_TRIM]) {
1840 memcpy(&td->last_rate_check_time[DDIR_READ], &td->bw_sample_time,
1841 sizeof(td->bw_sample_time));
1842 memcpy(&td->last_rate_check_time[DDIR_WRITE], &td->bw_sample_time,
1843 sizeof(td->bw_sample_time));
1844 memcpy(&td->last_rate_check_time[DDIR_TRIM], &td->bw_sample_time,
1845 sizeof(td->bw_sample_time));
1848 memset(bytes_done, 0, sizeof(bytes_done));
1849 clear_state = false;
1851 while (keep_running(td)) {
1852 uint64_t verify_bytes;
1854 fio_gettime(&td->start, NULL);
1855 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1858 clear_io_state(td, 0);
1860 if (o->unlink_each_loop && unlink_all_files(td))
1864 prune_io_piece_log(td);
1866 if (td->o.verify_only && td_write(td))
1867 verify_bytes = do_dry_run(td);
1869 do_io(td, bytes_done);
1871 if (!ddir_rw_sum(bytes_done)) {
1872 fio_mark_td_terminate(td);
1875 verify_bytes = bytes_done[DDIR_WRITE] +
1876 bytes_done[DDIR_TRIM];
1881 * If we took too long to shut down, the main thread could
1882 * already consider us reaped/exited. If that happens, break
1885 if (td->runstate >= TD_EXITED)
1891 * Make sure we've successfully updated the rusage stats
1892 * before waiting on the stat mutex. Otherwise we could have
1893 * the stat thread holding stat mutex and waiting for
1894 * the rusage_sem, which would never get upped because
1895 * this thread is waiting for the stat mutex.
1897 deadlock_loop_cnt = 0;
1899 check_update_rusage(td);
1900 if (!fio_sem_down_trylock(stat_sem))
1903 if (deadlock_loop_cnt++ > 5000) {
1904 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1905 td->error = EDEADLK;
1910 if (td_read(td) && td->io_bytes[DDIR_READ])
1911 update_runtime(td, elapsed_us, DDIR_READ);
1912 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1913 update_runtime(td, elapsed_us, DDIR_WRITE);
1914 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1915 update_runtime(td, elapsed_us, DDIR_TRIM);
1916 fio_gettime(&td->start, NULL);
1917 fio_sem_up(stat_sem);
1919 if (td->error || td->terminate)
1922 if (!o->do_verify ||
1923 o->verify == VERIFY_NONE ||
1924 td_ioengine_flagged(td, FIO_UNIDIR))
1927 clear_io_state(td, 0);
1929 fio_gettime(&td->start, NULL);
1931 do_verify(td, verify_bytes);
1934 * See comment further up for why this is done here.
1936 check_update_rusage(td);
1938 fio_sem_down(stat_sem);
1939 update_runtime(td, elapsed_us, DDIR_READ);
1940 fio_gettime(&td->start, NULL);
1941 fio_sem_up(stat_sem);
1943 if (td->error || td->terminate)
1948 * Acquire this lock if we were doing overlap checking in
1949 * offload mode so that we don't clean up this job while
1950 * another thread is checking its io_u's for overlap
1952 if (td_offload_overlap(td)) {
1953 int res = pthread_mutex_lock(&overlap_check);
1956 td_set_runstate(td, TD_FINISHING);
1957 if (td_offload_overlap(td)) {
1958 res = pthread_mutex_unlock(&overlap_check);
1962 update_rusage_stat(td);
1963 td->ts.total_run_time = mtime_since_now(&td->epoch);
1964 for_each_rw_ddir(ddir) {
1965 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
1968 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1969 (td->o.verify != VERIFY_NONE && td_write(td)))
1970 verify_save_state(td->thread_number);
1972 fio_unpin_memory(td);
1974 td_writeout_logs(td, true);
1976 iolog_compress_exit(td);
1977 rate_submit_exit(td);
1979 if (o->exec_postrun)
1980 exec_string(o, o->exec_postrun, "postrun");
1982 if (exitall_on_terminate || (o->exitall_error && td->error))
1983 fio_terminate_threads(td->groupid, td->o.exit_what);
1987 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1990 if (o->verify_async)
1991 verify_async_exit(td);
1993 close_and_free_files(td);
1996 cgroup_shutdown(td, cgroup_mnt);
1997 verify_free_state(td);
1998 td_zone_free_index(td);
2000 if (fio_option_is_set(o, cpumask)) {
2001 ret = fio_cpuset_exit(&o->cpumask);
2003 td_verror(td, ret, "fio_cpuset_exit");
2007 * do this very late, it will log file closing as well
2009 if (o->write_iolog_file)
2010 write_iolog_close(td);
2011 if (td->io_log_rfile)
2012 fclose(td->io_log_rfile);
2014 td_set_runstate(td, TD_EXITED);
2017 * Do this last after setting our runstate to exited, so we
2018 * know that the stat thread is signaled.
2020 check_update_rusage(td);
2023 return (void *) (uintptr_t) td->error;
2027 * Run over the job map and reap the threads that have exited, if any.
2029 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
2032 struct thread_data *td;
2033 unsigned int cputhreads, realthreads, pending;
2037 * reap exited threads (TD_EXITED -> TD_REAPED)
2039 realthreads = pending = cputhreads = 0;
2040 for_each_td(td, i) {
2043 if (!strcmp(td->o.ioengine, "cpuio"))
2052 if (td->runstate == TD_REAPED)
2054 if (td->o.use_thread) {
2055 if (td->runstate == TD_EXITED) {
2056 td_set_runstate(td, TD_REAPED);
2063 if (td->runstate == TD_EXITED)
2067 * check if someone quit or got killed in an unusual way
2069 ret = waitpid(td->pid, &status, flags);
2071 if (errno == ECHILD) {
2072 log_err("fio: pid=%d disappeared %d\n",
2073 (int) td->pid, td->runstate);
2075 td_set_runstate(td, TD_REAPED);
2079 } else if (ret == td->pid) {
2080 if (WIFSIGNALED(status)) {
2081 int sig = WTERMSIG(status);
2083 if (sig != SIGTERM && sig != SIGUSR2)
2084 log_err("fio: pid=%d, got signal=%d\n",
2085 (int) td->pid, sig);
2087 td_set_runstate(td, TD_REAPED);
2090 if (WIFEXITED(status)) {
2091 if (WEXITSTATUS(status) && !td->error)
2092 td->error = WEXITSTATUS(status);
2094 td_set_runstate(td, TD_REAPED);
2100 * If the job is stuck, do a forceful timeout of it and
2103 if (td->terminate &&
2104 td->runstate < TD_FSYNCING &&
2105 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2106 log_err("fio: job '%s' (state=%d) hasn't exited in "
2107 "%lu seconds, it appears to be stuck. Doing "
2108 "forceful exit of this job.\n",
2109 td->o.name, td->runstate,
2110 (unsigned long) time_since_now(&td->terminate_time));
2111 td_set_runstate(td, TD_REAPED);
2116 * thread is not dead, continue
2122 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2123 (*t_rate) -= ddir_rw_sum(td->o.rate);
2130 done_secs += mtime_since_now(&td->epoch) / 1000;
2131 profile_td_exit(td);
2135 if (*nr_running == cputhreads && !pending && realthreads)
2136 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2139 static bool __check_trigger_file(void)
2146 if (stat(trigger_file, &sb))
2149 if (unlink(trigger_file) < 0)
2150 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2156 static bool trigger_timedout(void)
2158 if (trigger_timeout)
2159 if (time_since_genesis() >= trigger_timeout) {
2160 trigger_timeout = 0;
2167 void exec_trigger(const char *cmd)
2171 if (!cmd || cmd[0] == '\0')
2176 log_err("fio: failed executing %s trigger\n", cmd);
2179 void check_trigger_file(void)
2181 if (__check_trigger_file() || trigger_timedout()) {
2183 fio_clients_send_trigger(trigger_remote_cmd);
2185 verify_save_state(IO_LIST_ALL);
2186 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2187 exec_trigger(trigger_cmd);
2192 static int fio_verify_load_state(struct thread_data *td)
2196 if (!td->o.verify_state)
2202 ret = fio_server_get_verify_state(td->o.name,
2203 td->thread_number - 1, &data);
2205 verify_assign_state(td, data);
2207 char prefix[PATH_MAX];
2210 sprintf(prefix, "%s%clocal", aux_path,
2211 FIO_OS_PATH_SEPARATOR);
2213 strcpy(prefix, "local");
2214 ret = verify_load_state(td, prefix);
2220 static void do_usleep(unsigned int usecs)
2222 check_for_running_stats();
2223 check_trigger_file();
2227 static bool check_mount_writes(struct thread_data *td)
2232 if (!td_write(td) || td->o.allow_mounted_write)
2236 * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2237 * are mkfs'd and mounted.
2239 for_each_file(td, f, i) {
2240 #ifdef FIO_HAVE_CHARDEV_SIZE
2241 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2243 if (f->filetype != FIO_TYPE_BLOCK)
2246 if (device_is_mounted(f->file_name))
2252 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2256 static bool waitee_running(struct thread_data *me)
2258 const char *waitee = me->o.wait_for;
2259 const char *self = me->o.name;
2260 struct thread_data *td;
2266 for_each_td(td, i) {
2267 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2270 if (td->runstate < TD_EXITED) {
2271 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2273 runstate_to_name(td->runstate));
2278 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2283 * Main function for kicking off and reaping jobs, as needed.
2285 static void run_threads(struct sk_out *sk_out)
2287 struct thread_data *td;
2288 unsigned int i, todo, nr_running, nr_started;
2289 uint64_t m_rate, t_rate;
2292 if (fio_gtod_offload && fio_start_gtod_thread())
2295 fio_idle_prof_init();
2299 nr_thread = nr_process = 0;
2300 for_each_td(td, i) {
2301 if (check_mount_writes(td))
2303 if (td->o.use_thread)
2309 if (output_format & FIO_OUTPUT_NORMAL) {
2310 struct buf_output out;
2312 buf_output_init(&out);
2313 __log_buf(&out, "Starting ");
2315 __log_buf(&out, "%d thread%s", nr_thread,
2316 nr_thread > 1 ? "s" : "");
2319 __log_buf(&out, " and ");
2320 __log_buf(&out, "%d process%s", nr_process,
2321 nr_process > 1 ? "es" : "");
2323 __log_buf(&out, "\n");
2324 log_info_buf(out.buf, out.buflen);
2325 buf_output_free(&out);
2328 todo = thread_number;
2331 m_rate = t_rate = 0;
2333 for_each_td(td, i) {
2334 print_status_init(td->thread_number - 1);
2336 if (!td->o.create_serialize)
2339 if (fio_verify_load_state(td))
2343 * do file setup here so it happens sequentially,
2344 * we don't want X number of threads getting their
2345 * client data interspersed on disk
2347 if (setup_files(td)) {
2351 log_err("fio: pid=%d, err=%d/%s\n",
2352 (int) td->pid, td->error, td->verror);
2353 td_set_runstate(td, TD_REAPED);
2360 * for sharing to work, each job must always open
2361 * its own files. so close them, if we opened them
2364 for_each_file(td, f, j) {
2365 if (fio_file_open(f))
2366 td_io_close_file(td, f);
2371 /* start idle threads before io threads start to run */
2372 fio_idle_prof_start();
2377 struct thread_data *map[REAL_MAX_JOBS];
2378 struct timespec this_start;
2379 int this_jobs = 0, left;
2380 struct fork_data *fd;
2383 * create threads (TD_NOT_CREATED -> TD_CREATED)
2385 for_each_td(td, i) {
2386 if (td->runstate != TD_NOT_CREATED)
2390 * never got a chance to start, killed by other
2391 * thread for some reason
2393 if (td->terminate) {
2398 if (td->o.start_delay) {
2399 spent = utime_since_genesis();
2401 if (td->o.start_delay > spent)
2405 if (td->o.stonewall && (nr_started || nr_running)) {
2406 dprint(FD_PROCESS, "%s: stonewall wait\n",
2411 if (waitee_running(td)) {
2412 dprint(FD_PROCESS, "%s: waiting for %s\n",
2413 td->o.name, td->o.wait_for);
2419 td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2420 td->update_rusage = 0;
2423 * Set state to created. Thread will transition
2424 * to TD_INITIALIZED when it's done setting up.
2426 td_set_runstate(td, TD_CREATED);
2427 map[this_jobs++] = td;
2430 fd = calloc(1, sizeof(*fd));
2432 fd->sk_out = sk_out;
2434 if (td->o.use_thread) {
2437 dprint(FD_PROCESS, "will pthread_create\n");
2438 ret = pthread_create(&td->thread, NULL,
2441 log_err("pthread_create: %s\n",
2448 ret = pthread_detach(td->thread);
2450 log_err("pthread_detach: %s",
2454 struct fio_file **files;
2456 dprint(FD_PROCESS, "will fork\n");
2464 ret = (int)(uintptr_t)thread_main(fd);
2466 } else if (i == fio_debug_jobno)
2467 *fio_debug_jobp = pid;
2468 // freeing previously allocated memory for files
2469 // this memory freed MUST NOT be shared between processes, only the pointer itself may be shared within TD
2475 dprint(FD_MUTEX, "wait on startup_sem\n");
2476 if (fio_sem_down_timeout(startup_sem, 10000)) {
2477 log_err("fio: job startup hung? exiting.\n");
2478 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2484 dprint(FD_MUTEX, "done waiting on startup_sem\n");
2488 * Wait for the started threads to transition to
2491 fio_gettime(&this_start, NULL);
2493 while (left && !fio_abort) {
2494 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2499 for (i = 0; i < this_jobs; i++) {
2503 if (td->runstate == TD_INITIALIZED) {
2506 } else if (td->runstate >= TD_EXITED) {
2510 nr_running++; /* work-around... */
2516 log_err("fio: %d job%s failed to start\n", left,
2517 left > 1 ? "s" : "");
2518 for (i = 0; i < this_jobs; i++) {
2522 kill(td->pid, SIGTERM);
2528 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2530 for_each_td(td, i) {
2531 if (td->runstate != TD_INITIALIZED)
2534 if (in_ramp_time(td))
2535 td_set_runstate(td, TD_RAMP);
2537 td_set_runstate(td, TD_RUNNING);
2540 m_rate += ddir_rw_sum(td->o.ratemin);
2541 t_rate += ddir_rw_sum(td->o.rate);
2543 fio_sem_up(td->sem);
2546 reap_threads(&nr_running, &t_rate, &m_rate);
2552 while (nr_running) {
2553 reap_threads(&nr_running, &t_rate, &m_rate);
2557 fio_idle_prof_stop();
2562 static void free_disk_util(void)
2564 disk_util_prune_entries();
2565 helper_thread_destroy();
2568 int fio_backend(struct sk_out *sk_out)
2570 struct thread_data *td;
2574 if (load_profile(exec_profile))
2577 exec_profile = NULL;
2583 struct log_params p = {
2584 .log_type = IO_LOG_TYPE_BW,
2587 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2588 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2589 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2592 if (init_global_dedupe_working_set_seeds()) {
2593 log_err("fio: failed to initialize global dedupe working set\n");
2597 startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2599 is_local_backend = true;
2600 if (startup_sem == NULL)
2605 if (helper_thread_create(startup_sem, sk_out))
2606 log_err("fio: failed to create helper thread\n");
2608 cgroup_list = smalloc(sizeof(*cgroup_list));
2610 INIT_FLIST_HEAD(cgroup_list);
2612 run_threads(sk_out);
2614 helper_thread_exit();
2619 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2620 struct io_log *log = agg_io_log[i];
2622 flush_log(log, false);
2628 for_each_td(td, i) {
2629 struct thread_stat *ts = &td->ts;
2631 free_clat_prio_stats(ts);
2632 steadystate_free(td);
2633 fio_options_free(td);
2634 fio_dump_options_free(td);
2635 if (td->rusage_sem) {
2636 fio_sem_remove(td->rusage_sem);
2637 td->rusage_sem = NULL;
2639 fio_sem_remove(td->sem);
2645 cgroup_kill(cgroup_list);
2649 fio_sem_remove(startup_sem);