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., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
41 #ifndef FIO_NO_HAVE_SHM_H
52 #include "lib/memalign.h"
54 #include "lib/getrusage.h"
57 #include "workqueue.h"
58 #include "lib/mountcheck.h"
59 #include "rate-submit.h"
60 #include "helper_thread.h"
62 static struct fio_mutex *startup_mutex;
63 static struct flist_head *cgroup_list;
64 static char *cgroup_mnt;
65 static int exit_value;
66 static volatile int fio_abort;
67 static unsigned int nr_process = 0;
68 static unsigned int nr_thread = 0;
70 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
73 unsigned int thread_number = 0;
74 unsigned int stat_number = 0;
77 unsigned long done_secs = 0;
79 #define PAGE_ALIGN(buf) \
80 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
82 #define JOB_START_TIMEOUT (5 * 1000)
84 static void sig_int(int sig)
88 fio_server_got_signal(sig);
90 log_info("\nfio: terminating on signal %d\n", sig);
95 fio_terminate_threads(TERMINATE_ALL);
99 void sig_show_status(int sig)
101 show_running_run_stats();
104 static void set_sig_handlers(void)
106 struct sigaction act;
108 memset(&act, 0, sizeof(act));
109 act.sa_handler = sig_int;
110 act.sa_flags = SA_RESTART;
111 sigaction(SIGINT, &act, NULL);
113 memset(&act, 0, sizeof(act));
114 act.sa_handler = sig_int;
115 act.sa_flags = SA_RESTART;
116 sigaction(SIGTERM, &act, NULL);
118 /* Windows uses SIGBREAK as a quit signal from other applications */
120 memset(&act, 0, sizeof(act));
121 act.sa_handler = sig_int;
122 act.sa_flags = SA_RESTART;
123 sigaction(SIGBREAK, &act, NULL);
126 memset(&act, 0, sizeof(act));
127 act.sa_handler = sig_show_status;
128 act.sa_flags = SA_RESTART;
129 sigaction(SIGUSR1, &act, NULL);
132 memset(&act, 0, sizeof(act));
133 act.sa_handler = sig_int;
134 act.sa_flags = SA_RESTART;
135 sigaction(SIGPIPE, &act, NULL);
140 * Check if we are above the minimum rate given.
142 static bool __check_min_rate(struct thread_data *td, struct timeval *now,
145 unsigned long long bytes = 0;
146 unsigned long iops = 0;
149 unsigned int ratemin = 0;
150 unsigned int rate_iops = 0;
151 unsigned int rate_iops_min = 0;
153 assert(ddir_rw(ddir));
155 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
159 * allow a 2 second settle period in the beginning
161 if (mtime_since(&td->start, now) < 2000)
164 iops += td->this_io_blocks[ddir];
165 bytes += td->this_io_bytes[ddir];
166 ratemin += td->o.ratemin[ddir];
167 rate_iops += td->o.rate_iops[ddir];
168 rate_iops_min += td->o.rate_iops_min[ddir];
171 * if rate blocks is set, sample is running
173 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
174 spent = mtime_since(&td->lastrate[ddir], now);
175 if (spent < td->o.ratecycle)
178 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
180 * check bandwidth specified rate
182 if (bytes < td->rate_bytes[ddir]) {
183 log_err("%s: min rate %u not met\n", td->o.name,
188 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
192 if (rate < ratemin ||
193 bytes < td->rate_bytes[ddir]) {
194 log_err("%s: min rate %u not met, got"
195 " %luKB/sec\n", td->o.name,
202 * checks iops specified rate
204 if (iops < rate_iops) {
205 log_err("%s: min iops rate %u not met\n",
206 td->o.name, rate_iops);
210 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
214 if (rate < rate_iops_min ||
215 iops < td->rate_blocks[ddir]) {
216 log_err("%s: min iops rate %u not met,"
217 " got %lu\n", td->o.name,
218 rate_iops_min, rate);
225 td->rate_bytes[ddir] = bytes;
226 td->rate_blocks[ddir] = iops;
227 memcpy(&td->lastrate[ddir], now, sizeof(*now));
231 static bool check_min_rate(struct thread_data *td, struct timeval *now)
235 if (td->bytes_done[DDIR_READ])
236 ret |= __check_min_rate(td, now, DDIR_READ);
237 if (td->bytes_done[DDIR_WRITE])
238 ret |= __check_min_rate(td, now, DDIR_WRITE);
239 if (td->bytes_done[DDIR_TRIM])
240 ret |= __check_min_rate(td, now, DDIR_TRIM);
246 * When job exits, we can cancel the in-flight IO if we are using async
247 * io. Attempt to do so.
249 static void cleanup_pending_aio(struct thread_data *td)
254 * get immediately available events, if any
256 r = io_u_queued_complete(td, 0);
261 * now cancel remaining active events
263 if (td->io_ops->cancel) {
267 io_u_qiter(&td->io_u_all, io_u, i) {
268 if (io_u->flags & IO_U_F_FLIGHT) {
269 r = td->io_ops->cancel(td, io_u);
277 r = io_u_queued_complete(td, td->cur_depth);
281 * Helper to handle the final sync of a file. Works just like the normal
282 * io path, just does everything sync.
284 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
286 struct io_u *io_u = __get_io_u(td);
292 io_u->ddir = DDIR_SYNC;
295 if (td_io_prep(td, io_u)) {
301 ret = td_io_queue(td, io_u);
303 td_verror(td, io_u->error, "td_io_queue");
306 } else if (ret == FIO_Q_QUEUED) {
307 if (td_io_commit(td))
309 if (io_u_queued_complete(td, 1) < 0)
311 } else if (ret == FIO_Q_COMPLETED) {
313 td_verror(td, io_u->error, "td_io_queue");
317 if (io_u_sync_complete(td, io_u) < 0)
319 } else if (ret == FIO_Q_BUSY) {
320 if (td_io_commit(td))
328 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
332 if (fio_file_open(f))
333 return fio_io_sync(td, f);
335 if (td_io_open_file(td, f))
338 ret = fio_io_sync(td, f);
339 td_io_close_file(td, f);
343 static inline void __update_tv_cache(struct thread_data *td)
345 fio_gettime(&td->tv_cache, NULL);
348 static inline void update_tv_cache(struct thread_data *td)
350 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
351 __update_tv_cache(td);
354 static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
356 if (in_ramp_time(td))
360 if (utime_since(&td->epoch, t) >= td->o.timeout)
367 * We need to update the runtime consistently in ms, but keep a running
368 * tally of the current elapsed time in microseconds for sub millisecond
371 static inline void update_runtime(struct thread_data *td,
372 unsigned long long *elapsed_us,
373 const enum fio_ddir ddir)
375 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
378 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
379 elapsed_us[ddir] += utime_since_now(&td->start);
380 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
383 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
388 if (ret < 0 || td->error) {
390 enum error_type_bit eb;
395 eb = td_error_type(ddir, err);
396 if (!(td->o.continue_on_error & (1 << eb)))
399 if (td_non_fatal_error(td, eb, err)) {
401 * Continue with the I/Os in case of
404 update_error_count(td, err);
408 } else if (td->o.fill_device && err == ENOSPC) {
410 * We expect to hit this error if
411 * fill_device option is set.
414 fio_mark_td_terminate(td);
418 * Stop the I/O in case of a fatal
421 update_error_count(td, err);
429 static void check_update_rusage(struct thread_data *td)
431 if (td->update_rusage) {
432 td->update_rusage = 0;
433 update_rusage_stat(td);
434 fio_mutex_up(td->rusage_sem);
438 static int wait_for_completions(struct thread_data *td, struct timeval *time)
440 const int full = queue_full(td);
444 if (td->flags & TD_F_REGROW_LOGS)
445 return io_u_quiesce(td);
448 * if the queue is full, we MUST reap at least 1 event
450 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
451 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
454 if (time && (__should_check_rate(td, DDIR_READ) ||
455 __should_check_rate(td, DDIR_WRITE) ||
456 __should_check_rate(td, DDIR_TRIM)))
457 fio_gettime(time, NULL);
460 ret = io_u_queued_complete(td, min_evts);
463 } while (full && (td->cur_depth > td->o.iodepth_low));
468 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
469 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
470 struct timeval *comp_time)
475 case FIO_Q_COMPLETED:
478 clear_io_u(td, io_u);
479 } else if (io_u->resid) {
480 int bytes = io_u->xfer_buflen - io_u->resid;
481 struct fio_file *f = io_u->file;
484 *bytes_issued += bytes;
487 trim_io_piece(td, io_u);
494 unlog_io_piece(td, io_u);
495 td_verror(td, EIO, "full resid");
500 io_u->xfer_buflen = io_u->resid;
501 io_u->xfer_buf += bytes;
502 io_u->offset += bytes;
504 if (ddir_rw(io_u->ddir))
505 td->ts.short_io_u[io_u->ddir]++;
508 if (io_u->offset == f->real_file_size)
511 requeue_io_u(td, &io_u);
514 if (comp_time && (__should_check_rate(td, DDIR_READ) ||
515 __should_check_rate(td, DDIR_WRITE) ||
516 __should_check_rate(td, DDIR_TRIM)))
517 fio_gettime(comp_time, NULL);
519 *ret = io_u_sync_complete(td, io_u);
524 if (td->flags & TD_F_REGROW_LOGS)
528 * when doing I/O (not when verifying),
529 * check for any errors that are to be ignored
537 * if the engine doesn't have a commit hook,
538 * the io_u is really queued. if it does have such
539 * a hook, it has to call io_u_queued() itself.
541 if (td->io_ops->commit == NULL)
542 io_u_queued(td, io_u);
544 *bytes_issued += io_u->xfer_buflen;
548 unlog_io_piece(td, io_u);
549 requeue_io_u(td, &io_u);
550 ret2 = td_io_commit(td);
556 td_verror(td, -(*ret), "td_io_queue");
560 if (break_on_this_error(td, ddir, ret))
566 static inline bool io_in_polling(struct thread_data *td)
568 return !td->o.iodepth_batch_complete_min &&
569 !td->o.iodepth_batch_complete_max;
572 * Unlinks files from thread data fio_file structure
574 static int unlink_all_files(struct thread_data *td)
580 for_each_file(td, f, i) {
581 if (f->filetype != FIO_TYPE_FILE)
583 ret = td_io_unlink_file(td, f);
589 td_verror(td, ret, "unlink_all_files");
595 * The main verify engine. Runs over the writes we previously submitted,
596 * reads the blocks back in, and checks the crc/md5 of the data.
598 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
605 dprint(FD_VERIFY, "starting loop\n");
608 * sync io first and invalidate cache, to make sure we really
611 for_each_file(td, f, i) {
612 if (!fio_file_open(f))
614 if (fio_io_sync(td, f))
616 if (file_invalidate_cache(td, f))
620 check_update_rusage(td);
626 * verify_state needs to be reset before verification
627 * proceeds so that expected random seeds match actual
628 * random seeds in headers. The main loop will reset
629 * all random number generators if randrepeat is set.
631 if (!td->o.rand_repeatable)
632 td_fill_verify_state_seed(td);
634 td_set_runstate(td, TD_VERIFYING);
637 while (!td->terminate) {
642 check_update_rusage(td);
644 if (runtime_exceeded(td, &td->tv_cache)) {
645 __update_tv_cache(td);
646 if (runtime_exceeded(td, &td->tv_cache)) {
647 fio_mark_td_terminate(td);
652 if (flow_threshold_exceeded(td))
655 if (!td->o.experimental_verify) {
656 io_u = __get_io_u(td);
660 if (get_next_verify(td, io_u)) {
665 if (td_io_prep(td, io_u)) {
670 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
673 while ((io_u = get_io_u(td)) != NULL) {
674 if (IS_ERR_OR_NULL(io_u)) {
681 * We are only interested in the places where
682 * we wrote or trimmed IOs. Turn those into
683 * reads for verification purposes.
685 if (io_u->ddir == DDIR_READ) {
687 * Pretend we issued it for rwmix
690 td->io_issues[DDIR_READ]++;
693 } else if (io_u->ddir == DDIR_TRIM) {
694 io_u->ddir = DDIR_READ;
695 io_u_set(td, io_u, IO_U_F_TRIMMED);
697 } else if (io_u->ddir == DDIR_WRITE) {
698 io_u->ddir = DDIR_READ;
710 if (verify_state_should_stop(td, io_u)) {
715 if (td->o.verify_async)
716 io_u->end_io = verify_io_u_async;
718 io_u->end_io = verify_io_u;
721 if (!td->o.disable_slat)
722 fio_gettime(&io_u->start_time, NULL);
724 ret = td_io_queue(td, io_u);
726 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
730 * if we can queue more, do so. but check if there are
731 * completed io_u's first. Note that we can get BUSY even
732 * without IO queued, if the system is resource starved.
735 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
736 if (full || io_in_polling(td))
737 ret = wait_for_completions(td, NULL);
743 check_update_rusage(td);
746 min_events = td->cur_depth;
749 ret = io_u_queued_complete(td, min_events);
751 cleanup_pending_aio(td);
753 td_set_runstate(td, TD_RUNNING);
755 dprint(FD_VERIFY, "exiting loop\n");
758 static bool exceeds_number_ios(struct thread_data *td)
760 unsigned long long number_ios;
762 if (!td->o.number_ios)
765 number_ios = ddir_rw_sum(td->io_blocks);
766 number_ios += td->io_u_queued + td->io_u_in_flight;
768 return number_ios >= (td->o.number_ios * td->loops);
771 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
773 unsigned long long bytes, limit;
776 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
777 else if (td_write(td))
778 bytes = this_bytes[DDIR_WRITE];
779 else if (td_read(td))
780 bytes = this_bytes[DDIR_READ];
782 bytes = this_bytes[DDIR_TRIM];
785 limit = td->o.io_limit;
790 return bytes >= limit || exceeds_number_ios(td);
793 static bool io_issue_bytes_exceeded(struct thread_data *td)
795 return io_bytes_exceeded(td, td->io_issue_bytes);
798 static bool io_complete_bytes_exceeded(struct thread_data *td)
800 return io_bytes_exceeded(td, td->this_io_bytes);
804 * used to calculate the next io time for rate control
807 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
809 uint64_t secs, remainder, bps, bytes, iops;
811 assert(!(td->flags & TD_F_CHILD));
812 bytes = td->rate_io_issue_bytes[ddir];
813 bps = td->rate_bps[ddir];
815 if (td->o.rate_process == RATE_PROCESS_POISSON) {
817 iops = bps / td->o.bs[ddir];
818 val = (int64_t) (1000000 / iops) *
819 -logf(__rand_0_1(&td->poisson_state));
821 dprint(FD_RATE, "poisson rate iops=%llu\n",
822 (unsigned long long) 1000000 / val);
824 td->last_usec += val;
825 return td->last_usec;
828 remainder = bytes % bps;
829 return remainder * 1000000 / bps + secs * 1000000;
836 * Main IO worker function. It retrieves io_u's to process and queues
837 * and reaps them, checking for rate and errors along the way.
839 * Returns number of bytes written and trimmed.
841 static void do_io(struct thread_data *td, uint64_t *bytes_done)
845 uint64_t total_bytes, bytes_issued = 0;
847 for (i = 0; i < DDIR_RWDIR_CNT; i++)
848 bytes_done[i] = td->bytes_done[i];
850 if (in_ramp_time(td))
851 td_set_runstate(td, TD_RAMP);
853 td_set_runstate(td, TD_RUNNING);
857 total_bytes = td->o.size;
859 * Allow random overwrite workloads to write up to io_limit
860 * before starting verification phase as 'size' doesn't apply.
862 if (td_write(td) && td_random(td) && td->o.norandommap)
863 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
865 * If verify_backlog is enabled, we'll run the verify in this
866 * handler as well. For that case, we may need up to twice the
869 if (td->o.verify != VERIFY_NONE &&
870 (td_write(td) && td->o.verify_backlog))
871 total_bytes += td->o.size;
873 /* In trimwrite mode, each byte is trimmed and then written, so
874 * allow total_bytes to be twice as big */
875 if (td_trimwrite(td))
876 total_bytes += td->total_io_size;
878 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
879 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
881 struct timeval comp_time;
886 check_update_rusage(td);
888 if (td->terminate || td->done)
893 if (runtime_exceeded(td, &td->tv_cache)) {
894 __update_tv_cache(td);
895 if (runtime_exceeded(td, &td->tv_cache)) {
896 fio_mark_td_terminate(td);
901 if (flow_threshold_exceeded(td))
905 * Break if we exceeded the bytes. The exception is time
906 * based runs, but we still need to break out of the loop
907 * for those to run verification, if enabled.
909 if (bytes_issued >= total_bytes &&
910 (!td->o.time_based ||
911 (td->o.time_based && td->o.verify != VERIFY_NONE)))
915 if (IS_ERR_OR_NULL(io_u)) {
916 int err = PTR_ERR(io_u);
923 if (td->o.latency_target)
931 * Add verification end_io handler if:
932 * - Asked to verify (!td_rw(td))
933 * - Or the io_u is from our verify list (mixed write/ver)
935 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
936 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
938 if (!td->o.verify_pattern_bytes) {
939 io_u->rand_seed = __rand(&td->verify_state);
940 if (sizeof(int) != sizeof(long *))
941 io_u->rand_seed *= __rand(&td->verify_state);
944 if (verify_state_should_stop(td, io_u)) {
949 if (td->o.verify_async)
950 io_u->end_io = verify_io_u_async;
952 io_u->end_io = verify_io_u;
953 td_set_runstate(td, TD_VERIFYING);
954 } else if (in_ramp_time(td))
955 td_set_runstate(td, TD_RAMP);
957 td_set_runstate(td, TD_RUNNING);
960 * Always log IO before it's issued, so we know the specific
961 * order of it. The logged unit will track when the IO has
964 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
966 td->o.verify != VERIFY_NONE &&
967 !td->o.experimental_verify)
968 log_io_piece(td, io_u);
970 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
971 const unsigned long blen = io_u->xfer_buflen;
972 const enum fio_ddir ddir = acct_ddir(io_u);
977 workqueue_enqueue(&td->io_wq, &io_u->work);
981 td->io_issues[ddir]++;
982 td->io_issue_bytes[ddir] += blen;
983 td->rate_io_issue_bytes[ddir] += blen;
986 if (should_check_rate(td))
987 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
990 ret = td_io_queue(td, io_u);
992 if (should_check_rate(td))
993 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
995 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
999 * See if we need to complete some commands. Note that
1000 * we can get BUSY even without IO queued, if the
1001 * system is resource starved.
1004 full = queue_full(td) ||
1005 (ret == FIO_Q_BUSY && td->cur_depth);
1006 if (full || io_in_polling(td))
1007 ret = wait_for_completions(td, &comp_time);
1011 if (!ddir_rw_sum(td->bytes_done) &&
1012 !td_ioengine_flagged(td, FIO_NOIO))
1015 if (!in_ramp_time(td) && should_check_rate(td)) {
1016 if (check_min_rate(td, &comp_time)) {
1017 if (exitall_on_terminate || td->o.exitall_error)
1018 fio_terminate_threads(td->groupid);
1019 td_verror(td, EIO, "check_min_rate");
1023 if (!in_ramp_time(td) && td->o.latency_target)
1024 lat_target_check(td);
1026 if (td->o.thinktime) {
1027 unsigned long long b;
1029 b = ddir_rw_sum(td->io_blocks);
1030 if (!(b % td->o.thinktime_blocks)) {
1035 if (td->o.thinktime_spin)
1036 usec_spin(td->o.thinktime_spin);
1038 left = td->o.thinktime - td->o.thinktime_spin;
1040 usec_sleep(td, left);
1045 check_update_rusage(td);
1047 if (td->trim_entries)
1048 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1050 if (td->o.fill_device && td->error == ENOSPC) {
1052 fio_mark_td_terminate(td);
1057 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1058 workqueue_flush(&td->io_wq);
1064 ret = io_u_queued_complete(td, i);
1065 if (td->o.fill_device && td->error == ENOSPC)
1069 if (should_fsync(td) && td->o.end_fsync) {
1070 td_set_runstate(td, TD_FSYNCING);
1072 for_each_file(td, f, i) {
1073 if (!fio_file_fsync(td, f))
1076 log_err("fio: end_fsync failed for file %s\n",
1081 cleanup_pending_aio(td);
1084 * stop job if we failed doing any IO
1086 if (!ddir_rw_sum(td->this_io_bytes))
1089 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1090 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1093 static void free_file_completion_logging(struct thread_data *td)
1098 for_each_file(td, f, i) {
1099 if (!f->last_write_comp)
1101 sfree(f->last_write_comp);
1105 static int init_file_completion_logging(struct thread_data *td,
1111 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1114 for_each_file(td, f, i) {
1115 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1116 if (!f->last_write_comp)
1123 free_file_completion_logging(td);
1124 log_err("fio: failed to alloc write comp data\n");
1128 static void cleanup_io_u(struct thread_data *td)
1132 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1134 if (td->io_ops->io_u_free)
1135 td->io_ops->io_u_free(td, io_u);
1137 fio_memfree(io_u, sizeof(*io_u));
1142 io_u_rexit(&td->io_u_requeues);
1143 io_u_qexit(&td->io_u_freelist);
1144 io_u_qexit(&td->io_u_all);
1146 free_file_completion_logging(td);
1149 static int init_io_u(struct thread_data *td)
1152 unsigned int max_bs, min_write;
1153 int cl_align, i, max_units;
1154 int data_xfer = 1, err;
1157 max_units = td->o.iodepth;
1158 max_bs = td_max_bs(td);
1159 min_write = td->o.min_bs[DDIR_WRITE];
1160 td->orig_buffer_size = (unsigned long long) max_bs
1161 * (unsigned long long) max_units;
1163 if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1167 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1168 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1169 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1172 log_err("fio: failed setting up IO queues\n");
1177 * if we may later need to do address alignment, then add any
1178 * possible adjustment here so that we don't cause a buffer
1179 * overflow later. this adjustment may be too much if we get
1180 * lucky and the allocator gives us an aligned address.
1182 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1183 td_ioengine_flagged(td, FIO_RAWIO))
1184 td->orig_buffer_size += page_mask + td->o.mem_align;
1186 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1189 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1190 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1193 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1194 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1198 if (data_xfer && allocate_io_mem(td))
1201 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1202 td_ioengine_flagged(td, FIO_RAWIO))
1203 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1205 p = td->orig_buffer;
1207 cl_align = os_cache_line_size();
1209 for (i = 0; i < max_units; i++) {
1215 ptr = fio_memalign(cl_align, sizeof(*io_u));
1217 log_err("fio: unable to allocate aligned memory\n");
1222 memset(io_u, 0, sizeof(*io_u));
1223 INIT_FLIST_HEAD(&io_u->verify_list);
1224 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1228 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1231 io_u_fill_buffer(td, io_u, min_write, max_bs);
1232 if (td_write(td) && td->o.verify_pattern_bytes) {
1234 * Fill the buffer with the pattern if we are
1235 * going to be doing writes.
1237 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1242 io_u->flags = IO_U_F_FREE;
1243 io_u_qpush(&td->io_u_freelist, io_u);
1246 * io_u never leaves this stack, used for iteration of all
1249 io_u_qpush(&td->io_u_all, io_u);
1251 if (td->io_ops->io_u_init) {
1252 int ret = td->io_ops->io_u_init(td, io_u);
1255 log_err("fio: failed to init engine data: %d\n", ret);
1263 if (init_file_completion_logging(td, max_units))
1269 static int switch_ioscheduler(struct thread_data *td)
1271 #ifdef FIO_HAVE_IOSCHED_SWITCH
1272 char tmp[256], tmp2[128];
1276 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1279 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1281 f = fopen(tmp, "r+");
1283 if (errno == ENOENT) {
1284 log_err("fio: os or kernel doesn't support IO scheduler"
1288 td_verror(td, errno, "fopen iosched");
1295 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1296 if (ferror(f) || ret != 1) {
1297 td_verror(td, errno, "fwrite");
1305 * Read back and check that the selected scheduler is now the default.
1307 memset(tmp, 0, sizeof(tmp));
1308 ret = fread(tmp, sizeof(tmp), 1, f);
1309 if (ferror(f) || ret < 0) {
1310 td_verror(td, errno, "fread");
1315 * either a list of io schedulers or "none\n" is expected.
1317 tmp[strlen(tmp) - 1] = '\0';
1320 * Write to "none" entry doesn't fail, so check the result here.
1322 if (!strcmp(tmp, "none")) {
1323 log_err("fio: io scheduler is not tunable\n");
1328 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1329 if (!strstr(tmp, tmp2)) {
1330 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1331 td_verror(td, EINVAL, "iosched_switch");
1343 static bool keep_running(struct thread_data *td)
1345 unsigned long long limit;
1349 if (td->o.time_based)
1355 if (exceeds_number_ios(td))
1359 limit = td->o.io_limit;
1363 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1367 * If the difference is less than the minimum IO size, we
1370 diff = limit - ddir_rw_sum(td->io_bytes);
1371 if (diff < td_max_bs(td))
1374 if (fio_files_done(td) && !td->o.io_limit)
1383 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1385 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1389 str = malloc(newlen);
1390 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1392 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1395 log_err("fio: exec of cmd <%s> failed\n", str);
1402 * Dry run to compute correct state of numberio for verification.
1404 static uint64_t do_dry_run(struct thread_data *td)
1406 td_set_runstate(td, TD_RUNNING);
1408 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1409 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1413 if (td->terminate || td->done)
1416 io_u = get_io_u(td);
1417 if (IS_ERR_OR_NULL(io_u))
1420 io_u_set(td, io_u, IO_U_F_FLIGHT);
1423 if (ddir_rw(acct_ddir(io_u)))
1424 td->io_issues[acct_ddir(io_u)]++;
1425 if (ddir_rw(io_u->ddir)) {
1426 io_u_mark_depth(td, 1);
1427 td->ts.total_io_u[io_u->ddir]++;
1430 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1432 td->o.verify != VERIFY_NONE &&
1433 !td->o.experimental_verify)
1434 log_io_piece(td, io_u);
1436 ret = io_u_sync_complete(td, io_u);
1440 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1444 struct thread_data *td;
1445 struct sk_out *sk_out;
1449 * Entry point for the thread based jobs. The process based jobs end up
1450 * here as well, after a little setup.
1452 static void *thread_main(void *data)
1454 struct fork_data *fd = data;
1455 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1456 struct thread_data *td = fd->td;
1457 struct thread_options *o = &td->o;
1458 struct sk_out *sk_out = fd->sk_out;
1459 int deadlock_loop_cnt;
1463 sk_out_assign(sk_out);
1466 if (!o->use_thread) {
1472 fio_local_clock_init(o->use_thread);
1474 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1477 fio_server_send_start(td);
1479 INIT_FLIST_HEAD(&td->io_log_list);
1480 INIT_FLIST_HEAD(&td->io_hist_list);
1481 INIT_FLIST_HEAD(&td->verify_list);
1482 INIT_FLIST_HEAD(&td->trim_list);
1483 INIT_FLIST_HEAD(&td->next_rand_list);
1484 td->io_hist_tree = RB_ROOT;
1486 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1488 td_verror(td, ret, "mutex_cond_init_pshared");
1491 ret = cond_init_pshared(&td->verify_cond);
1493 td_verror(td, ret, "mutex_cond_pshared");
1497 td_set_runstate(td, TD_INITIALIZED);
1498 dprint(FD_MUTEX, "up startup_mutex\n");
1499 fio_mutex_up(startup_mutex);
1500 dprint(FD_MUTEX, "wait on td->mutex\n");
1501 fio_mutex_down(td->mutex);
1502 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1505 * A new gid requires privilege, so we need to do this before setting
1508 if (o->gid != -1U && setgid(o->gid)) {
1509 td_verror(td, errno, "setgid");
1512 if (o->uid != -1U && setuid(o->uid)) {
1513 td_verror(td, errno, "setuid");
1518 * Do this early, we don't want the compress threads to be limited
1519 * to the same CPUs as the IO workers. So do this before we set
1520 * any potential CPU affinity
1522 if (iolog_compress_init(td, sk_out))
1526 * If we have a gettimeofday() thread, make sure we exclude that
1527 * thread from this job
1530 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1533 * Set affinity first, in case it has an impact on the memory
1536 if (fio_option_is_set(o, cpumask)) {
1537 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1538 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1540 log_err("fio: no CPUs set\n");
1541 log_err("fio: Try increasing number of available CPUs\n");
1542 td_verror(td, EINVAL, "cpus_split");
1546 ret = fio_setaffinity(td->pid, o->cpumask);
1548 td_verror(td, errno, "cpu_set_affinity");
1553 #ifdef CONFIG_LIBNUMA
1554 /* numa node setup */
1555 if (fio_option_is_set(o, numa_cpunodes) ||
1556 fio_option_is_set(o, numa_memnodes)) {
1557 struct bitmask *mask;
1559 if (numa_available() < 0) {
1560 td_verror(td, errno, "Does not support NUMA API\n");
1564 if (fio_option_is_set(o, numa_cpunodes)) {
1565 mask = numa_parse_nodestring(o->numa_cpunodes);
1566 ret = numa_run_on_node_mask(mask);
1567 numa_free_nodemask(mask);
1569 td_verror(td, errno, \
1570 "numa_run_on_node_mask failed\n");
1575 if (fio_option_is_set(o, numa_memnodes)) {
1577 if (o->numa_memnodes)
1578 mask = numa_parse_nodestring(o->numa_memnodes);
1580 switch (o->numa_mem_mode) {
1581 case MPOL_INTERLEAVE:
1582 numa_set_interleave_mask(mask);
1585 numa_set_membind(mask);
1588 numa_set_localalloc();
1590 case MPOL_PREFERRED:
1591 numa_set_preferred(o->numa_mem_prefer_node);
1599 numa_free_nodemask(mask);
1605 if (fio_pin_memory(td))
1609 * May alter parameters that init_io_u() will use, so we need to
1618 if (o->verify_async && verify_async_init(td))
1621 if (fio_option_is_set(o, ioprio) ||
1622 fio_option_is_set(o, ioprio_class)) {
1623 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1625 td_verror(td, errno, "ioprio_set");
1630 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1634 if (nice(o->nice) == -1 && errno != 0) {
1635 td_verror(td, errno, "nice");
1639 if (o->ioscheduler && switch_ioscheduler(td))
1642 if (!o->create_serialize && setup_files(td))
1648 if (init_random_map(td))
1651 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1655 if (pre_read_files(td) < 0)
1659 fio_verify_init(td);
1661 if (rate_submit_init(td, sk_out))
1664 set_epoch_time(td, o->log_unix_epoch);
1665 fio_getrusage(&td->ru_start);
1666 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1667 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1668 memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1670 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1671 o->ratemin[DDIR_TRIM]) {
1672 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1673 sizeof(td->bw_sample_time));
1674 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1675 sizeof(td->bw_sample_time));
1676 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1677 sizeof(td->bw_sample_time));
1681 while (keep_running(td)) {
1682 uint64_t verify_bytes;
1684 fio_gettime(&td->start, NULL);
1685 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1688 clear_io_state(td, 0);
1690 if (o->unlink_each_loop && unlink_all_files(td))
1694 prune_io_piece_log(td);
1696 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1697 verify_bytes = do_dry_run(td);
1699 uint64_t bytes_done[DDIR_RWDIR_CNT];
1701 do_io(td, bytes_done);
1703 if (!ddir_rw_sum(bytes_done)) {
1704 fio_mark_td_terminate(td);
1707 verify_bytes = bytes_done[DDIR_WRITE] +
1708 bytes_done[DDIR_TRIM];
1713 * If we took too long to shut down, the main thread could
1714 * already consider us reaped/exited. If that happens, break
1717 if (td->runstate >= TD_EXITED)
1723 * Make sure we've successfully updated the rusage stats
1724 * before waiting on the stat mutex. Otherwise we could have
1725 * the stat thread holding stat mutex and waiting for
1726 * the rusage_sem, which would never get upped because
1727 * this thread is waiting for the stat mutex.
1729 deadlock_loop_cnt = 0;
1731 check_update_rusage(td);
1732 if (!fio_mutex_down_trylock(stat_mutex))
1735 if (deadlock_loop_cnt++ > 5000) {
1736 log_err("fio seems to be stuck grabbing stat_mutex, forcibly exiting\n");
1737 td->error = EDEADLK;
1742 if (td_read(td) && td->io_bytes[DDIR_READ])
1743 update_runtime(td, elapsed_us, DDIR_READ);
1744 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1745 update_runtime(td, elapsed_us, DDIR_WRITE);
1746 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1747 update_runtime(td, elapsed_us, DDIR_TRIM);
1748 fio_gettime(&td->start, NULL);
1749 fio_mutex_up(stat_mutex);
1751 if (td->error || td->terminate)
1754 if (!o->do_verify ||
1755 o->verify == VERIFY_NONE ||
1756 td_ioengine_flagged(td, FIO_UNIDIR))
1759 clear_io_state(td, 0);
1761 fio_gettime(&td->start, NULL);
1763 do_verify(td, verify_bytes);
1766 * See comment further up for why this is done here.
1768 check_update_rusage(td);
1770 fio_mutex_down(stat_mutex);
1771 update_runtime(td, elapsed_us, DDIR_READ);
1772 fio_gettime(&td->start, NULL);
1773 fio_mutex_up(stat_mutex);
1775 if (td->error || td->terminate)
1779 td_set_runstate(td, TD_FINISHING);
1781 update_rusage_stat(td);
1782 td->ts.total_run_time = mtime_since_now(&td->epoch);
1783 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1784 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1785 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1787 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1788 (td->o.verify != VERIFY_NONE && td_write(td)))
1789 verify_save_state(td->thread_number);
1791 fio_unpin_memory(td);
1793 td_writeout_logs(td, true);
1795 iolog_compress_exit(td);
1796 rate_submit_exit(td);
1798 if (o->exec_postrun)
1799 exec_string(o, o->exec_postrun, (const char *)"postrun");
1801 if (exitall_on_terminate || (o->exitall_error && td->error))
1802 fio_terminate_threads(td->groupid);
1806 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1809 if (o->verify_async)
1810 verify_async_exit(td);
1812 close_and_free_files(td);
1815 cgroup_shutdown(td, &cgroup_mnt);
1816 verify_free_state(td);
1818 if (td->zone_state_index) {
1821 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1822 free(td->zone_state_index[i]);
1823 free(td->zone_state_index);
1824 td->zone_state_index = NULL;
1827 if (fio_option_is_set(o, cpumask)) {
1828 ret = fio_cpuset_exit(&o->cpumask);
1830 td_verror(td, ret, "fio_cpuset_exit");
1834 * do this very late, it will log file closing as well
1836 if (o->write_iolog_file)
1837 write_iolog_close(td);
1839 fio_mutex_remove(td->mutex);
1842 td_set_runstate(td, TD_EXITED);
1845 * Do this last after setting our runstate to exited, so we
1846 * know that the stat thread is signaled.
1848 check_update_rusage(td);
1851 return (void *) (uintptr_t) td->error;
1854 static void dump_td_info(struct thread_data *td)
1856 log_err("fio: job '%s' (state=%d) hasn't exited in %lu seconds, it "
1857 "appears to be stuck. Doing forceful exit of this job.\n",
1858 td->o.name, td->runstate,
1859 (unsigned long) time_since_now(&td->terminate_time));
1863 * Run over the job map and reap the threads that have exited, if any.
1865 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1868 struct thread_data *td;
1869 unsigned int cputhreads, realthreads, pending;
1873 * reap exited threads (TD_EXITED -> TD_REAPED)
1875 realthreads = pending = cputhreads = 0;
1876 for_each_td(td, i) {
1880 * ->io_ops is NULL for a thread that has closed its
1883 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1892 if (td->runstate == TD_REAPED)
1894 if (td->o.use_thread) {
1895 if (td->runstate == TD_EXITED) {
1896 td_set_runstate(td, TD_REAPED);
1903 if (td->runstate == TD_EXITED)
1907 * check if someone quit or got killed in an unusual way
1909 ret = waitpid(td->pid, &status, flags);
1911 if (errno == ECHILD) {
1912 log_err("fio: pid=%d disappeared %d\n",
1913 (int) td->pid, td->runstate);
1915 td_set_runstate(td, TD_REAPED);
1919 } else if (ret == td->pid) {
1920 if (WIFSIGNALED(status)) {
1921 int sig = WTERMSIG(status);
1923 if (sig != SIGTERM && sig != SIGUSR2)
1924 log_err("fio: pid=%d, got signal=%d\n",
1925 (int) td->pid, sig);
1927 td_set_runstate(td, TD_REAPED);
1930 if (WIFEXITED(status)) {
1931 if (WEXITSTATUS(status) && !td->error)
1932 td->error = WEXITSTATUS(status);
1934 td_set_runstate(td, TD_REAPED);
1940 * If the job is stuck, do a forceful timeout of it and
1943 if (td->terminate &&
1944 td->runstate < TD_FSYNCING &&
1945 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1947 td_set_runstate(td, TD_REAPED);
1952 * thread is not dead, continue
1958 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1959 (*t_rate) -= ddir_rw_sum(td->o.rate);
1966 done_secs += mtime_since_now(&td->epoch) / 1000;
1967 profile_td_exit(td);
1970 if (*nr_running == cputhreads && !pending && realthreads)
1971 fio_terminate_threads(TERMINATE_ALL);
1974 static bool __check_trigger_file(void)
1981 if (stat(trigger_file, &sb))
1984 if (unlink(trigger_file) < 0)
1985 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1991 static bool trigger_timedout(void)
1993 if (trigger_timeout)
1994 return time_since_genesis() >= trigger_timeout;
1999 void exec_trigger(const char *cmd)
2008 log_err("fio: failed executing %s trigger\n", cmd);
2011 void check_trigger_file(void)
2013 if (__check_trigger_file() || trigger_timedout()) {
2015 fio_clients_send_trigger(trigger_remote_cmd);
2017 verify_save_state(IO_LIST_ALL);
2018 fio_terminate_threads(TERMINATE_ALL);
2019 exec_trigger(trigger_cmd);
2024 static int fio_verify_load_state(struct thread_data *td)
2028 if (!td->o.verify_state)
2034 ret = fio_server_get_verify_state(td->o.name,
2035 td->thread_number - 1, &data);
2037 verify_assign_state(td, data);
2039 ret = verify_load_state(td, "local");
2044 static void do_usleep(unsigned int usecs)
2046 check_for_running_stats();
2047 check_trigger_file();
2051 static bool check_mount_writes(struct thread_data *td)
2056 if (!td_write(td) || td->o.allow_mounted_write)
2059 for_each_file(td, f, i) {
2060 if (f->filetype != FIO_TYPE_BD)
2062 if (device_is_mounted(f->file_name))
2068 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
2072 static bool waitee_running(struct thread_data *me)
2074 const char *waitee = me->o.wait_for;
2075 const char *self = me->o.name;
2076 struct thread_data *td;
2082 for_each_td(td, i) {
2083 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2086 if (td->runstate < TD_EXITED) {
2087 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2089 runstate_to_name(td->runstate));
2094 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2099 * Main function for kicking off and reaping jobs, as needed.
2101 static void run_threads(struct sk_out *sk_out)
2103 struct thread_data *td;
2104 unsigned int i, todo, nr_running, nr_started;
2105 uint64_t m_rate, t_rate;
2108 if (fio_gtod_offload && fio_start_gtod_thread())
2111 fio_idle_prof_init();
2115 nr_thread = nr_process = 0;
2116 for_each_td(td, i) {
2117 if (check_mount_writes(td))
2119 if (td->o.use_thread)
2125 if (output_format & FIO_OUTPUT_NORMAL) {
2126 log_info("Starting ");
2128 log_info("%d thread%s", nr_thread,
2129 nr_thread > 1 ? "s" : "");
2133 log_info("%d process%s", nr_process,
2134 nr_process > 1 ? "es" : "");
2140 todo = thread_number;
2143 m_rate = t_rate = 0;
2145 for_each_td(td, i) {
2146 print_status_init(td->thread_number - 1);
2148 if (!td->o.create_serialize)
2151 if (fio_verify_load_state(td))
2155 * do file setup here so it happens sequentially,
2156 * we don't want X number of threads getting their
2157 * client data interspersed on disk
2159 if (setup_files(td)) {
2163 log_err("fio: pid=%d, err=%d/%s\n",
2164 (int) td->pid, td->error, td->verror);
2165 td_set_runstate(td, TD_REAPED);
2172 * for sharing to work, each job must always open
2173 * its own files. so close them, if we opened them
2176 for_each_file(td, f, j) {
2177 if (fio_file_open(f))
2178 td_io_close_file(td, f);
2183 /* start idle threads before io threads start to run */
2184 fio_idle_prof_start();
2189 struct thread_data *map[REAL_MAX_JOBS];
2190 struct timeval this_start;
2191 int this_jobs = 0, left;
2192 struct fork_data *fd;
2195 * create threads (TD_NOT_CREATED -> TD_CREATED)
2197 for_each_td(td, i) {
2198 if (td->runstate != TD_NOT_CREATED)
2202 * never got a chance to start, killed by other
2203 * thread for some reason
2205 if (td->terminate) {
2210 if (td->o.start_delay) {
2211 spent = utime_since_genesis();
2213 if (td->o.start_delay > spent)
2217 if (td->o.stonewall && (nr_started || nr_running)) {
2218 dprint(FD_PROCESS, "%s: stonewall wait\n",
2223 if (waitee_running(td)) {
2224 dprint(FD_PROCESS, "%s: waiting for %s\n",
2225 td->o.name, td->o.wait_for);
2231 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2232 td->update_rusage = 0;
2235 * Set state to created. Thread will transition
2236 * to TD_INITIALIZED when it's done setting up.
2238 td_set_runstate(td, TD_CREATED);
2239 map[this_jobs++] = td;
2242 fd = calloc(1, sizeof(*fd));
2244 fd->sk_out = sk_out;
2246 if (td->o.use_thread) {
2249 dprint(FD_PROCESS, "will pthread_create\n");
2250 ret = pthread_create(&td->thread, NULL,
2253 log_err("pthread_create: %s\n",
2259 ret = pthread_detach(td->thread);
2261 log_err("pthread_detach: %s",
2265 dprint(FD_PROCESS, "will fork\n");
2270 ret = (int)(uintptr_t)thread_main(fd);
2272 } else if (i == fio_debug_jobno)
2273 *fio_debug_jobp = pid;
2275 dprint(FD_MUTEX, "wait on startup_mutex\n");
2276 if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2277 log_err("fio: job startup hung? exiting.\n");
2278 fio_terminate_threads(TERMINATE_ALL);
2283 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2287 * Wait for the started threads to transition to
2290 fio_gettime(&this_start, NULL);
2292 while (left && !fio_abort) {
2293 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2298 for (i = 0; i < this_jobs; i++) {
2302 if (td->runstate == TD_INITIALIZED) {
2305 } else if (td->runstate >= TD_EXITED) {
2309 nr_running++; /* work-around... */
2315 log_err("fio: %d job%s failed to start\n", left,
2316 left > 1 ? "s" : "");
2317 for (i = 0; i < this_jobs; i++) {
2321 kill(td->pid, SIGTERM);
2327 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2329 for_each_td(td, i) {
2330 if (td->runstate != TD_INITIALIZED)
2333 if (in_ramp_time(td))
2334 td_set_runstate(td, TD_RAMP);
2336 td_set_runstate(td, TD_RUNNING);
2339 m_rate += ddir_rw_sum(td->o.ratemin);
2340 t_rate += ddir_rw_sum(td->o.rate);
2342 fio_mutex_up(td->mutex);
2345 reap_threads(&nr_running, &t_rate, &m_rate);
2351 while (nr_running) {
2352 reap_threads(&nr_running, &t_rate, &m_rate);
2356 fio_idle_prof_stop();
2361 static void free_disk_util(void)
2363 disk_util_prune_entries();
2364 helper_thread_destroy();
2367 int fio_backend(struct sk_out *sk_out)
2369 struct thread_data *td;
2373 if (load_profile(exec_profile))
2376 exec_profile = NULL;
2382 struct log_params p = {
2383 .log_type = IO_LOG_TYPE_BW,
2386 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2387 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2388 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2391 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2392 if (startup_mutex == NULL)
2397 helper_thread_create(startup_mutex, sk_out);
2399 cgroup_list = smalloc(sizeof(*cgroup_list));
2400 INIT_FLIST_HEAD(cgroup_list);
2402 run_threads(sk_out);
2404 helper_thread_exit();
2409 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2410 struct io_log *log = agg_io_log[i];
2412 flush_log(log, false);
2418 for_each_td(td, i) {
2420 if (td->ss.iops_data != NULL) {
2421 free(td->ss.iops_data);
2422 free(td->ss.bw_data);
2425 fio_options_free(td);
2426 if (td->rusage_sem) {
2427 fio_mutex_remove(td->rusage_sem);
2428 td->rusage_sem = NULL;
2433 cgroup_kill(cgroup_list);
2437 fio_mutex_remove(startup_mutex);