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);
445 * if the queue is full, we MUST reap at least 1 event
447 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
448 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
451 if (time && (__should_check_rate(td, DDIR_READ) ||
452 __should_check_rate(td, DDIR_WRITE) ||
453 __should_check_rate(td, DDIR_TRIM)))
454 fio_gettime(time, NULL);
457 ret = io_u_queued_complete(td, min_evts);
460 } while (full && (td->cur_depth > td->o.iodepth_low));
465 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
466 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
467 struct timeval *comp_time)
472 case FIO_Q_COMPLETED:
475 clear_io_u(td, io_u);
476 } else if (io_u->resid) {
477 int bytes = io_u->xfer_buflen - io_u->resid;
478 struct fio_file *f = io_u->file;
481 *bytes_issued += bytes;
484 trim_io_piece(td, io_u);
491 unlog_io_piece(td, io_u);
492 td_verror(td, EIO, "full resid");
497 io_u->xfer_buflen = io_u->resid;
498 io_u->xfer_buf += bytes;
499 io_u->offset += bytes;
501 if (ddir_rw(io_u->ddir))
502 td->ts.short_io_u[io_u->ddir]++;
505 if (io_u->offset == f->real_file_size)
508 requeue_io_u(td, &io_u);
511 if (comp_time && (__should_check_rate(td, DDIR_READ) ||
512 __should_check_rate(td, DDIR_WRITE) ||
513 __should_check_rate(td, DDIR_TRIM)))
514 fio_gettime(comp_time, NULL);
516 *ret = io_u_sync_complete(td, io_u);
522 * when doing I/O (not when verifying),
523 * check for any errors that are to be ignored
531 * if the engine doesn't have a commit hook,
532 * the io_u is really queued. if it does have such
533 * a hook, it has to call io_u_queued() itself.
535 if (td->io_ops->commit == NULL)
536 io_u_queued(td, io_u);
538 *bytes_issued += io_u->xfer_buflen;
542 unlog_io_piece(td, io_u);
543 requeue_io_u(td, &io_u);
544 ret2 = td_io_commit(td);
550 td_verror(td, -(*ret), "td_io_queue");
554 if (break_on_this_error(td, ddir, ret))
560 static inline bool io_in_polling(struct thread_data *td)
562 return !td->o.iodepth_batch_complete_min &&
563 !td->o.iodepth_batch_complete_max;
567 * The main verify engine. Runs over the writes we previously submitted,
568 * reads the blocks back in, and checks the crc/md5 of the data.
570 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
577 dprint(FD_VERIFY, "starting loop\n");
580 * sync io first and invalidate cache, to make sure we really
583 for_each_file(td, f, i) {
584 if (!fio_file_open(f))
586 if (fio_io_sync(td, f))
588 if (file_invalidate_cache(td, f))
592 check_update_rusage(td);
597 td_set_runstate(td, TD_VERIFYING);
600 while (!td->terminate) {
605 check_update_rusage(td);
607 if (runtime_exceeded(td, &td->tv_cache)) {
608 __update_tv_cache(td);
609 if (runtime_exceeded(td, &td->tv_cache)) {
610 fio_mark_td_terminate(td);
615 if (flow_threshold_exceeded(td))
618 if (!td->o.experimental_verify) {
619 io_u = __get_io_u(td);
623 if (get_next_verify(td, io_u)) {
628 if (td_io_prep(td, io_u)) {
633 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
636 while ((io_u = get_io_u(td)) != NULL) {
644 * We are only interested in the places where
645 * we wrote or trimmed IOs. Turn those into
646 * reads for verification purposes.
648 if (io_u->ddir == DDIR_READ) {
650 * Pretend we issued it for rwmix
653 td->io_issues[DDIR_READ]++;
656 } else if (io_u->ddir == DDIR_TRIM) {
657 io_u->ddir = DDIR_READ;
658 io_u_set(io_u, IO_U_F_TRIMMED);
660 } else if (io_u->ddir == DDIR_WRITE) {
661 io_u->ddir = DDIR_READ;
673 if (verify_state_should_stop(td, io_u)) {
678 if (td->o.verify_async)
679 io_u->end_io = verify_io_u_async;
681 io_u->end_io = verify_io_u;
684 if (!td->o.disable_slat)
685 fio_gettime(&io_u->start_time, NULL);
687 ret = td_io_queue(td, io_u);
689 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
693 * if we can queue more, do so. but check if there are
694 * completed io_u's first. Note that we can get BUSY even
695 * without IO queued, if the system is resource starved.
698 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
699 if (full || io_in_polling(td))
700 ret = wait_for_completions(td, NULL);
706 check_update_rusage(td);
709 min_events = td->cur_depth;
712 ret = io_u_queued_complete(td, min_events);
714 cleanup_pending_aio(td);
716 td_set_runstate(td, TD_RUNNING);
718 dprint(FD_VERIFY, "exiting loop\n");
721 static bool exceeds_number_ios(struct thread_data *td)
723 unsigned long long number_ios;
725 if (!td->o.number_ios)
728 number_ios = ddir_rw_sum(td->io_blocks);
729 number_ios += td->io_u_queued + td->io_u_in_flight;
731 return number_ios >= (td->o.number_ios * td->loops);
734 static bool io_issue_bytes_exceeded(struct thread_data *td)
736 unsigned long long bytes, limit;
739 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
740 else if (td_write(td))
741 bytes = td->io_issue_bytes[DDIR_WRITE];
742 else if (td_read(td))
743 bytes = td->io_issue_bytes[DDIR_READ];
745 bytes = td->io_issue_bytes[DDIR_TRIM];
748 limit = td->o.io_limit;
753 return bytes >= limit || exceeds_number_ios(td);
756 static bool io_complete_bytes_exceeded(struct thread_data *td)
758 unsigned long long bytes, limit;
761 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
762 else if (td_write(td))
763 bytes = td->this_io_bytes[DDIR_WRITE];
764 else if (td_read(td))
765 bytes = td->this_io_bytes[DDIR_READ];
767 bytes = td->this_io_bytes[DDIR_TRIM];
770 limit = td->o.io_limit;
775 return bytes >= limit || exceeds_number_ios(td);
779 * used to calculate the next io time for rate control
782 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
784 uint64_t secs, remainder, bps, bytes, iops;
786 assert(!(td->flags & TD_F_CHILD));
787 bytes = td->rate_io_issue_bytes[ddir];
788 bps = td->rate_bps[ddir];
790 if (td->o.rate_process == RATE_PROCESS_POISSON) {
792 iops = bps / td->o.bs[ddir];
793 val = (int64_t) (1000000 / iops) *
794 -logf(__rand_0_1(&td->poisson_state));
796 dprint(FD_RATE, "poisson rate iops=%llu\n",
797 (unsigned long long) 1000000 / val);
799 td->last_usec += val;
800 return td->last_usec;
803 remainder = bytes % bps;
804 return remainder * 1000000 / bps + secs * 1000000;
811 * Main IO worker function. It retrieves io_u's to process and queues
812 * and reaps them, checking for rate and errors along the way.
814 * Returns number of bytes written and trimmed.
816 static void do_io(struct thread_data *td, uint64_t *bytes_done)
820 uint64_t total_bytes, bytes_issued = 0;
822 for (i = 0; i < DDIR_RWDIR_CNT; i++)
823 bytes_done[i] = td->bytes_done[i];
825 if (in_ramp_time(td))
826 td_set_runstate(td, TD_RAMP);
828 td_set_runstate(td, TD_RUNNING);
832 total_bytes = td->o.size;
834 * Allow random overwrite workloads to write up to io_limit
835 * before starting verification phase as 'size' doesn't apply.
837 if (td_write(td) && td_random(td) && td->o.norandommap)
838 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
840 * If verify_backlog is enabled, we'll run the verify in this
841 * handler as well. For that case, we may need up to twice the
844 if (td->o.verify != VERIFY_NONE &&
845 (td_write(td) && td->o.verify_backlog))
846 total_bytes += td->o.size;
848 /* In trimwrite mode, each byte is trimmed and then written, so
849 * allow total_bytes to be twice as big */
850 if (td_trimwrite(td))
851 total_bytes += td->total_io_size;
853 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
854 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
856 struct timeval comp_time;
861 check_update_rusage(td);
863 if (td->terminate || td->done)
868 if (runtime_exceeded(td, &td->tv_cache)) {
869 __update_tv_cache(td);
870 if (runtime_exceeded(td, &td->tv_cache)) {
871 fio_mark_td_terminate(td);
876 if (flow_threshold_exceeded(td))
880 * Break if we exceeded the bytes. The exception is time
881 * based runs, but we still need to break out of the loop
882 * for those to run verification, if enabled.
884 if (bytes_issued >= total_bytes &&
885 (!td->o.time_based ||
886 (td->o.time_based && td->o.verify != VERIFY_NONE)))
890 if (IS_ERR_OR_NULL(io_u)) {
891 int err = PTR_ERR(io_u);
898 if (td->o.latency_target)
906 * Add verification end_io handler if:
907 * - Asked to verify (!td_rw(td))
908 * - Or the io_u is from our verify list (mixed write/ver)
910 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
911 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
913 if (!td->o.verify_pattern_bytes) {
914 io_u->rand_seed = __rand(&td->verify_state);
915 if (sizeof(int) != sizeof(long *))
916 io_u->rand_seed *= __rand(&td->verify_state);
919 if (verify_state_should_stop(td, io_u)) {
924 if (td->o.verify_async)
925 io_u->end_io = verify_io_u_async;
927 io_u->end_io = verify_io_u;
928 td_set_runstate(td, TD_VERIFYING);
929 } else if (in_ramp_time(td))
930 td_set_runstate(td, TD_RAMP);
932 td_set_runstate(td, TD_RUNNING);
935 * Always log IO before it's issued, so we know the specific
936 * order of it. The logged unit will track when the IO has
939 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
941 td->o.verify != VERIFY_NONE &&
942 !td->o.experimental_verify)
943 log_io_piece(td, io_u);
945 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
946 const unsigned long blen = io_u->xfer_buflen;
947 const enum fio_ddir ddir = acct_ddir(io_u);
952 workqueue_enqueue(&td->io_wq, &io_u->work);
956 td->io_issues[ddir]++;
957 td->io_issue_bytes[ddir] += blen;
958 td->rate_io_issue_bytes[ddir] += blen;
961 if (should_check_rate(td))
962 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
965 ret = td_io_queue(td, io_u);
967 if (should_check_rate(td))
968 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
970 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
974 * See if we need to complete some commands. Note that
975 * we can get BUSY even without IO queued, if the
976 * system is resource starved.
979 full = queue_full(td) ||
980 (ret == FIO_Q_BUSY && td->cur_depth);
981 if (full || io_in_polling(td))
982 ret = wait_for_completions(td, &comp_time);
986 if (!ddir_rw_sum(td->bytes_done) &&
987 !(td->io_ops->flags & FIO_NOIO))
990 if (!in_ramp_time(td) && should_check_rate(td)) {
991 if (check_min_rate(td, &comp_time)) {
992 if (exitall_on_terminate || td->o.exitall_error)
993 fio_terminate_threads(td->groupid);
994 td_verror(td, EIO, "check_min_rate");
998 if (!in_ramp_time(td) && td->o.latency_target)
999 lat_target_check(td);
1001 if (td->o.thinktime) {
1002 unsigned long long b;
1004 b = ddir_rw_sum(td->io_blocks);
1005 if (!(b % td->o.thinktime_blocks)) {
1010 if (td->o.thinktime_spin)
1011 usec_spin(td->o.thinktime_spin);
1013 left = td->o.thinktime - td->o.thinktime_spin;
1015 usec_sleep(td, left);
1020 check_update_rusage(td);
1022 if (td->trim_entries)
1023 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1025 if (td->o.fill_device && td->error == ENOSPC) {
1027 fio_mark_td_terminate(td);
1032 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1033 workqueue_flush(&td->io_wq);
1039 ret = io_u_queued_complete(td, i);
1040 if (td->o.fill_device && td->error == ENOSPC)
1044 if (should_fsync(td) && td->o.end_fsync) {
1045 td_set_runstate(td, TD_FSYNCING);
1047 for_each_file(td, f, i) {
1048 if (!fio_file_fsync(td, f))
1051 log_err("fio: end_fsync failed for file %s\n",
1056 cleanup_pending_aio(td);
1059 * stop job if we failed doing any IO
1061 if (!ddir_rw_sum(td->this_io_bytes))
1064 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1065 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1068 static void free_file_completion_logging(struct thread_data *td)
1073 for_each_file(td, f, i) {
1074 if (!f->last_write_comp)
1076 sfree(f->last_write_comp);
1080 static int init_file_completion_logging(struct thread_data *td,
1086 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1089 for_each_file(td, f, i) {
1090 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1091 if (!f->last_write_comp)
1098 free_file_completion_logging(td);
1099 log_err("fio: failed to alloc write comp data\n");
1103 static void cleanup_io_u(struct thread_data *td)
1107 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1109 if (td->io_ops->io_u_free)
1110 td->io_ops->io_u_free(td, io_u);
1112 fio_memfree(io_u, sizeof(*io_u));
1117 io_u_rexit(&td->io_u_requeues);
1118 io_u_qexit(&td->io_u_freelist);
1119 io_u_qexit(&td->io_u_all);
1121 free_file_completion_logging(td);
1124 static int init_io_u(struct thread_data *td)
1127 unsigned int max_bs, min_write;
1128 int cl_align, i, max_units;
1129 int data_xfer = 1, err;
1132 max_units = td->o.iodepth;
1133 max_bs = td_max_bs(td);
1134 min_write = td->o.min_bs[DDIR_WRITE];
1135 td->orig_buffer_size = (unsigned long long) max_bs
1136 * (unsigned long long) max_units;
1138 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1142 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1143 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1144 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1147 log_err("fio: failed setting up IO queues\n");
1152 * if we may later need to do address alignment, then add any
1153 * possible adjustment here so that we don't cause a buffer
1154 * overflow later. this adjustment may be too much if we get
1155 * lucky and the allocator gives us an aligned address.
1157 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1158 (td->io_ops->flags & FIO_RAWIO))
1159 td->orig_buffer_size += page_mask + td->o.mem_align;
1161 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1164 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1165 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1168 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1169 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1173 if (data_xfer && allocate_io_mem(td))
1176 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1177 (td->io_ops->flags & FIO_RAWIO))
1178 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1180 p = td->orig_buffer;
1182 cl_align = os_cache_line_size();
1184 for (i = 0; i < max_units; i++) {
1190 ptr = fio_memalign(cl_align, sizeof(*io_u));
1192 log_err("fio: unable to allocate aligned memory\n");
1197 memset(io_u, 0, sizeof(*io_u));
1198 INIT_FLIST_HEAD(&io_u->verify_list);
1199 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1203 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1206 io_u_fill_buffer(td, io_u, min_write, max_bs);
1207 if (td_write(td) && td->o.verify_pattern_bytes) {
1209 * Fill the buffer with the pattern if we are
1210 * going to be doing writes.
1212 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1217 io_u->flags = IO_U_F_FREE;
1218 io_u_qpush(&td->io_u_freelist, io_u);
1221 * io_u never leaves this stack, used for iteration of all
1224 io_u_qpush(&td->io_u_all, io_u);
1226 if (td->io_ops->io_u_init) {
1227 int ret = td->io_ops->io_u_init(td, io_u);
1230 log_err("fio: failed to init engine data: %d\n", ret);
1238 if (init_file_completion_logging(td, max_units))
1244 static int switch_ioscheduler(struct thread_data *td)
1246 char tmp[256], tmp2[128];
1250 if (td->io_ops->flags & FIO_DISKLESSIO)
1253 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1255 f = fopen(tmp, "r+");
1257 if (errno == ENOENT) {
1258 log_err("fio: os or kernel doesn't support IO scheduler"
1262 td_verror(td, errno, "fopen iosched");
1269 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1270 if (ferror(f) || ret != 1) {
1271 td_verror(td, errno, "fwrite");
1279 * Read back and check that the selected scheduler is now the default.
1281 memset(tmp, 0, sizeof(tmp));
1282 ret = fread(tmp, sizeof(tmp), 1, f);
1283 if (ferror(f) || ret < 0) {
1284 td_verror(td, errno, "fread");
1289 * either a list of io schedulers or "none\n" is expected.
1291 tmp[strlen(tmp) - 1] = '\0';
1294 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1295 if (!strstr(tmp, tmp2)) {
1296 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1297 td_verror(td, EINVAL, "iosched_switch");
1306 static bool keep_running(struct thread_data *td)
1308 unsigned long long limit;
1312 if (td->o.time_based)
1318 if (exceeds_number_ios(td))
1322 limit = td->o.io_limit;
1326 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1330 * If the difference is less than the minimum IO size, we
1333 diff = limit - ddir_rw_sum(td->io_bytes);
1334 if (diff < td_max_bs(td))
1337 if (fio_files_done(td) && !td->o.io_limit)
1346 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1348 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1352 str = malloc(newlen);
1353 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1355 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1358 log_err("fio: exec of cmd <%s> failed\n", str);
1365 * Dry run to compute correct state of numberio for verification.
1367 static uint64_t do_dry_run(struct thread_data *td)
1369 td_set_runstate(td, TD_RUNNING);
1371 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1372 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1376 if (td->terminate || td->done)
1379 io_u = get_io_u(td);
1383 io_u_set(io_u, IO_U_F_FLIGHT);
1386 if (ddir_rw(acct_ddir(io_u)))
1387 td->io_issues[acct_ddir(io_u)]++;
1388 if (ddir_rw(io_u->ddir)) {
1389 io_u_mark_depth(td, 1);
1390 td->ts.total_io_u[io_u->ddir]++;
1393 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1395 td->o.verify != VERIFY_NONE &&
1396 !td->o.experimental_verify)
1397 log_io_piece(td, io_u);
1399 ret = io_u_sync_complete(td, io_u);
1403 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1407 struct thread_data *td;
1408 struct sk_out *sk_out;
1412 * Entry point for the thread based jobs. The process based jobs end up
1413 * here as well, after a little setup.
1415 static void *thread_main(void *data)
1417 struct fork_data *fd = data;
1418 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1419 struct thread_data *td = fd->td;
1420 struct thread_options *o = &td->o;
1421 struct sk_out *sk_out = fd->sk_out;
1422 pthread_condattr_t attr;
1426 sk_out_assign(sk_out);
1429 if (!o->use_thread) {
1435 fio_local_clock_init(o->use_thread);
1437 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1440 fio_server_send_start(td);
1442 INIT_FLIST_HEAD(&td->io_log_list);
1443 INIT_FLIST_HEAD(&td->io_hist_list);
1444 INIT_FLIST_HEAD(&td->verify_list);
1445 INIT_FLIST_HEAD(&td->trim_list);
1446 INIT_FLIST_HEAD(&td->next_rand_list);
1447 pthread_mutex_init(&td->io_u_lock, NULL);
1448 td->io_hist_tree = RB_ROOT;
1450 pthread_condattr_init(&attr);
1451 pthread_cond_init(&td->verify_cond, &attr);
1452 pthread_cond_init(&td->free_cond, &attr);
1454 td_set_runstate(td, TD_INITIALIZED);
1455 dprint(FD_MUTEX, "up startup_mutex\n");
1456 fio_mutex_up(startup_mutex);
1457 dprint(FD_MUTEX, "wait on td->mutex\n");
1458 fio_mutex_down(td->mutex);
1459 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1462 * A new gid requires privilege, so we need to do this before setting
1465 if (o->gid != -1U && setgid(o->gid)) {
1466 td_verror(td, errno, "setgid");
1469 if (o->uid != -1U && setuid(o->uid)) {
1470 td_verror(td, errno, "setuid");
1475 * If we have a gettimeofday() thread, make sure we exclude that
1476 * thread from this job
1479 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1482 * Set affinity first, in case it has an impact on the memory
1485 if (fio_option_is_set(o, cpumask)) {
1486 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1487 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1489 log_err("fio: no CPUs set\n");
1490 log_err("fio: Try increasing number of available CPUs\n");
1491 td_verror(td, EINVAL, "cpus_split");
1495 ret = fio_setaffinity(td->pid, o->cpumask);
1497 td_verror(td, errno, "cpu_set_affinity");
1502 #ifdef CONFIG_LIBNUMA
1503 /* numa node setup */
1504 if (fio_option_is_set(o, numa_cpunodes) ||
1505 fio_option_is_set(o, numa_memnodes)) {
1506 struct bitmask *mask;
1508 if (numa_available() < 0) {
1509 td_verror(td, errno, "Does not support NUMA API\n");
1513 if (fio_option_is_set(o, numa_cpunodes)) {
1514 mask = numa_parse_nodestring(o->numa_cpunodes);
1515 ret = numa_run_on_node_mask(mask);
1516 numa_free_nodemask(mask);
1518 td_verror(td, errno, \
1519 "numa_run_on_node_mask failed\n");
1524 if (fio_option_is_set(o, numa_memnodes)) {
1526 if (o->numa_memnodes)
1527 mask = numa_parse_nodestring(o->numa_memnodes);
1529 switch (o->numa_mem_mode) {
1530 case MPOL_INTERLEAVE:
1531 numa_set_interleave_mask(mask);
1534 numa_set_membind(mask);
1537 numa_set_localalloc();
1539 case MPOL_PREFERRED:
1540 numa_set_preferred(o->numa_mem_prefer_node);
1548 numa_free_nodemask(mask);
1554 if (fio_pin_memory(td))
1558 * May alter parameters that init_io_u() will use, so we need to
1567 if (o->verify_async && verify_async_init(td))
1570 if (fio_option_is_set(o, ioprio) ||
1571 fio_option_is_set(o, ioprio_class)) {
1572 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1574 td_verror(td, errno, "ioprio_set");
1579 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1583 if (nice(o->nice) == -1 && errno != 0) {
1584 td_verror(td, errno, "nice");
1588 if (o->ioscheduler && switch_ioscheduler(td))
1591 if (!o->create_serialize && setup_files(td))
1597 if (init_random_map(td))
1600 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1604 if (pre_read_files(td) < 0)
1608 if (iolog_compress_init(td, sk_out))
1611 fio_verify_init(td);
1613 if (rate_submit_init(td, sk_out))
1616 fio_gettime(&td->epoch, NULL);
1617 fio_getrusage(&td->ru_start);
1618 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1619 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1621 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1622 o->ratemin[DDIR_TRIM]) {
1623 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1624 sizeof(td->bw_sample_time));
1625 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1626 sizeof(td->bw_sample_time));
1627 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1628 sizeof(td->bw_sample_time));
1632 while (keep_running(td)) {
1633 uint64_t verify_bytes;
1635 fio_gettime(&td->start, NULL);
1636 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1639 clear_io_state(td, 0);
1641 prune_io_piece_log(td);
1643 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1644 verify_bytes = do_dry_run(td);
1646 uint64_t bytes_done[DDIR_RWDIR_CNT];
1648 do_io(td, bytes_done);
1650 if (!ddir_rw_sum(bytes_done)) {
1651 fio_mark_td_terminate(td);
1654 verify_bytes = bytes_done[DDIR_WRITE] +
1655 bytes_done[DDIR_TRIM];
1662 * Make sure we've successfully updated the rusage stats
1663 * before waiting on the stat mutex. Otherwise we could have
1664 * the stat thread holding stat mutex and waiting for
1665 * the rusage_sem, which would never get upped because
1666 * this thread is waiting for the stat mutex.
1668 check_update_rusage(td);
1670 fio_mutex_down(stat_mutex);
1671 if (td_read(td) && td->io_bytes[DDIR_READ])
1672 update_runtime(td, elapsed_us, DDIR_READ);
1673 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1674 update_runtime(td, elapsed_us, DDIR_WRITE);
1675 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1676 update_runtime(td, elapsed_us, DDIR_TRIM);
1677 fio_gettime(&td->start, NULL);
1678 fio_mutex_up(stat_mutex);
1680 if (td->error || td->terminate)
1683 if (!o->do_verify ||
1684 o->verify == VERIFY_NONE ||
1685 (td->io_ops->flags & FIO_UNIDIR))
1688 clear_io_state(td, 0);
1690 fio_gettime(&td->start, NULL);
1692 do_verify(td, verify_bytes);
1695 * See comment further up for why this is done here.
1697 check_update_rusage(td);
1699 fio_mutex_down(stat_mutex);
1700 update_runtime(td, elapsed_us, DDIR_READ);
1701 fio_gettime(&td->start, NULL);
1702 fio_mutex_up(stat_mutex);
1704 if (td->error || td->terminate)
1708 update_rusage_stat(td);
1709 td->ts.total_run_time = mtime_since_now(&td->epoch);
1710 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1711 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1712 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1714 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1715 (td->o.verify != VERIFY_NONE && td_write(td)))
1716 verify_save_state(td->thread_number);
1718 fio_unpin_memory(td);
1720 td_writeout_logs(td, true);
1722 iolog_compress_exit(td);
1723 rate_submit_exit(td);
1725 if (o->exec_postrun)
1726 exec_string(o, o->exec_postrun, (const char *)"postrun");
1728 if (exitall_on_terminate || (o->exitall_error && td->error))
1729 fio_terminate_threads(td->groupid);
1733 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1736 if (o->verify_async)
1737 verify_async_exit(td);
1739 close_and_free_files(td);
1742 cgroup_shutdown(td, &cgroup_mnt);
1743 verify_free_state(td);
1745 if (td->zone_state_index) {
1748 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1749 free(td->zone_state_index[i]);
1750 free(td->zone_state_index);
1751 td->zone_state_index = NULL;
1754 if (fio_option_is_set(o, cpumask)) {
1755 ret = fio_cpuset_exit(&o->cpumask);
1757 td_verror(td, ret, "fio_cpuset_exit");
1761 * do this very late, it will log file closing as well
1763 if (o->write_iolog_file)
1764 write_iolog_close(td);
1766 fio_mutex_remove(td->mutex);
1769 td_set_runstate(td, TD_EXITED);
1772 * Do this last after setting our runstate to exited, so we
1773 * know that the stat thread is signaled.
1775 check_update_rusage(td);
1778 return (void *) (uintptr_t) td->error;
1783 * We cannot pass the td data into a forked process, so attach the td and
1784 * pass it to the thread worker.
1786 static int fork_main(struct sk_out *sk_out, int shmid, int offset)
1788 struct fork_data *fd;
1791 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1792 data = shmat(shmid, NULL, 0);
1793 if (data == (void *) -1) {
1801 * HP-UX inherits shm mappings?
1806 fd = calloc(1, sizeof(*fd));
1807 fd->td = data + offset * sizeof(struct thread_data);
1808 fd->sk_out = sk_out;
1809 ret = thread_main(fd);
1811 return (int) (uintptr_t) ret;
1814 static void dump_td_info(struct thread_data *td)
1816 log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1817 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1818 (unsigned long) time_since_now(&td->terminate_time));
1822 * Run over the job map and reap the threads that have exited, if any.
1824 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1825 unsigned int *m_rate)
1827 struct thread_data *td;
1828 unsigned int cputhreads, realthreads, pending;
1832 * reap exited threads (TD_EXITED -> TD_REAPED)
1834 realthreads = pending = cputhreads = 0;
1835 for_each_td(td, i) {
1839 * ->io_ops is NULL for a thread that has closed its
1842 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1851 if (td->runstate == TD_REAPED)
1853 if (td->o.use_thread) {
1854 if (td->runstate == TD_EXITED) {
1855 td_set_runstate(td, TD_REAPED);
1862 if (td->runstate == TD_EXITED)
1866 * check if someone quit or got killed in an unusual way
1868 ret = waitpid(td->pid, &status, flags);
1870 if (errno == ECHILD) {
1871 log_err("fio: pid=%d disappeared %d\n",
1872 (int) td->pid, td->runstate);
1874 td_set_runstate(td, TD_REAPED);
1878 } else if (ret == td->pid) {
1879 if (WIFSIGNALED(status)) {
1880 int sig = WTERMSIG(status);
1882 if (sig != SIGTERM && sig != SIGUSR2)
1883 log_err("fio: pid=%d, got signal=%d\n",
1884 (int) td->pid, sig);
1886 td_set_runstate(td, TD_REAPED);
1889 if (WIFEXITED(status)) {
1890 if (WEXITSTATUS(status) && !td->error)
1891 td->error = WEXITSTATUS(status);
1893 td_set_runstate(td, TD_REAPED);
1899 * If the job is stuck, do a forceful timeout of it and
1902 if (td->terminate &&
1903 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1905 td_set_runstate(td, TD_REAPED);
1910 * thread is not dead, continue
1916 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1917 (*t_rate) -= ddir_rw_sum(td->o.rate);
1924 done_secs += mtime_since_now(&td->epoch) / 1000;
1925 profile_td_exit(td);
1928 if (*nr_running == cputhreads && !pending && realthreads)
1929 fio_terminate_threads(TERMINATE_ALL);
1932 static bool __check_trigger_file(void)
1939 if (stat(trigger_file, &sb))
1942 if (unlink(trigger_file) < 0)
1943 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1949 static bool trigger_timedout(void)
1951 if (trigger_timeout)
1952 return time_since_genesis() >= trigger_timeout;
1957 void exec_trigger(const char *cmd)
1966 log_err("fio: failed executing %s trigger\n", cmd);
1969 void check_trigger_file(void)
1971 if (__check_trigger_file() || trigger_timedout()) {
1973 fio_clients_send_trigger(trigger_remote_cmd);
1975 verify_save_state(IO_LIST_ALL);
1976 fio_terminate_threads(TERMINATE_ALL);
1977 exec_trigger(trigger_cmd);
1982 static int fio_verify_load_state(struct thread_data *td)
1986 if (!td->o.verify_state)
1992 ret = fio_server_get_verify_state(td->o.name,
1993 td->thread_number - 1, &data);
1995 verify_assign_state(td, data);
1997 ret = verify_load_state(td, "local");
2002 static void do_usleep(unsigned int usecs)
2004 check_for_running_stats();
2005 check_trigger_file();
2009 static bool check_mount_writes(struct thread_data *td)
2014 if (!td_write(td) || td->o.allow_mounted_write)
2017 for_each_file(td, f, i) {
2018 if (f->filetype != FIO_TYPE_BD)
2020 if (device_is_mounted(f->file_name))
2026 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
2030 static bool waitee_running(struct thread_data *me)
2032 const char *waitee = me->o.wait_for;
2033 const char *self = me->o.name;
2034 struct thread_data *td;
2040 for_each_td(td, i) {
2041 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2044 if (td->runstate < TD_EXITED) {
2045 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2047 runstate_to_name(td->runstate));
2052 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2057 * Main function for kicking off and reaping jobs, as needed.
2059 static void run_threads(struct sk_out *sk_out)
2061 struct thread_data *td;
2062 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
2065 if (fio_gtod_offload && fio_start_gtod_thread())
2068 fio_idle_prof_init();
2072 nr_thread = nr_process = 0;
2073 for_each_td(td, i) {
2074 if (check_mount_writes(td))
2076 if (td->o.use_thread)
2082 if (output_format & FIO_OUTPUT_NORMAL) {
2083 log_info("Starting ");
2085 log_info("%d thread%s", nr_thread,
2086 nr_thread > 1 ? "s" : "");
2090 log_info("%d process%s", nr_process,
2091 nr_process > 1 ? "es" : "");
2097 todo = thread_number;
2100 m_rate = t_rate = 0;
2102 for_each_td(td, i) {
2103 print_status_init(td->thread_number - 1);
2105 if (!td->o.create_serialize)
2108 if (fio_verify_load_state(td))
2112 * do file setup here so it happens sequentially,
2113 * we don't want X number of threads getting their
2114 * client data interspersed on disk
2116 if (setup_files(td)) {
2120 log_err("fio: pid=%d, err=%d/%s\n",
2121 (int) td->pid, td->error, td->verror);
2122 td_set_runstate(td, TD_REAPED);
2129 * for sharing to work, each job must always open
2130 * its own files. so close them, if we opened them
2133 for_each_file(td, f, j) {
2134 if (fio_file_open(f))
2135 td_io_close_file(td, f);
2140 /* start idle threads before io threads start to run */
2141 fio_idle_prof_start();
2146 struct thread_data *map[REAL_MAX_JOBS];
2147 struct timeval this_start;
2148 int this_jobs = 0, left;
2151 * create threads (TD_NOT_CREATED -> TD_CREATED)
2153 for_each_td(td, i) {
2154 if (td->runstate != TD_NOT_CREATED)
2158 * never got a chance to start, killed by other
2159 * thread for some reason
2161 if (td->terminate) {
2166 if (td->o.start_delay) {
2167 spent = utime_since_genesis();
2169 if (td->o.start_delay > spent)
2173 if (td->o.stonewall && (nr_started || nr_running)) {
2174 dprint(FD_PROCESS, "%s: stonewall wait\n",
2179 if (waitee_running(td)) {
2180 dprint(FD_PROCESS, "%s: waiting for %s\n",
2181 td->o.name, td->o.wait_for);
2187 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2188 td->update_rusage = 0;
2191 * Set state to created. Thread will transition
2192 * to TD_INITIALIZED when it's done setting up.
2194 td_set_runstate(td, TD_CREATED);
2195 map[this_jobs++] = td;
2198 if (td->o.use_thread) {
2199 struct fork_data *fd;
2202 fd = calloc(1, sizeof(*fd));
2204 fd->sk_out = sk_out;
2206 dprint(FD_PROCESS, "will pthread_create\n");
2207 ret = pthread_create(&td->thread, NULL,
2210 log_err("pthread_create: %s\n",
2216 ret = pthread_detach(td->thread);
2218 log_err("pthread_detach: %s",
2222 dprint(FD_PROCESS, "will fork\n");
2225 int ret = fork_main(sk_out, shm_id, i);
2228 } else if (i == fio_debug_jobno)
2229 *fio_debug_jobp = pid;
2231 dprint(FD_MUTEX, "wait on startup_mutex\n");
2232 if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2233 log_err("fio: job startup hung? exiting.\n");
2234 fio_terminate_threads(TERMINATE_ALL);
2239 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2243 * Wait for the started threads to transition to
2246 fio_gettime(&this_start, NULL);
2248 while (left && !fio_abort) {
2249 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2254 for (i = 0; i < this_jobs; i++) {
2258 if (td->runstate == TD_INITIALIZED) {
2261 } else if (td->runstate >= TD_EXITED) {
2265 nr_running++; /* work-around... */
2271 log_err("fio: %d job%s failed to start\n", left,
2272 left > 1 ? "s" : "");
2273 for (i = 0; i < this_jobs; i++) {
2277 kill(td->pid, SIGTERM);
2283 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2285 for_each_td(td, i) {
2286 if (td->runstate != TD_INITIALIZED)
2289 if (in_ramp_time(td))
2290 td_set_runstate(td, TD_RAMP);
2292 td_set_runstate(td, TD_RUNNING);
2295 m_rate += ddir_rw_sum(td->o.ratemin);
2296 t_rate += ddir_rw_sum(td->o.rate);
2298 fio_mutex_up(td->mutex);
2301 reap_threads(&nr_running, &t_rate, &m_rate);
2307 while (nr_running) {
2308 reap_threads(&nr_running, &t_rate, &m_rate);
2312 fio_idle_prof_stop();
2317 static void free_disk_util(void)
2319 disk_util_prune_entries();
2320 helper_thread_destroy();
2323 int fio_backend(struct sk_out *sk_out)
2325 struct thread_data *td;
2329 if (load_profile(exec_profile))
2332 exec_profile = NULL;
2338 struct log_params p = {
2339 .log_type = IO_LOG_TYPE_BW,
2342 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2343 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2344 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2347 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2348 if (startup_mutex == NULL)
2353 helper_thread_create(startup_mutex, sk_out);
2355 cgroup_list = smalloc(sizeof(*cgroup_list));
2356 INIT_FLIST_HEAD(cgroup_list);
2358 run_threads(sk_out);
2360 helper_thread_exit();
2365 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2366 struct io_log *log = agg_io_log[i];
2374 for_each_td(td, i) {
2375 fio_options_free(td);
2376 if (td->rusage_sem) {
2377 fio_mutex_remove(td->rusage_sem);
2378 td->rusage_sem = NULL;
2383 cgroup_kill(cgroup_list);
2387 fio_mutex_remove(startup_mutex);