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"
61 static pthread_t helper_thread;
62 static pthread_mutex_t helper_lock;
63 pthread_cond_t helper_cond;
64 int helper_do_stat = 0;
66 static struct fio_mutex *startup_mutex;
67 static struct flist_head *cgroup_list;
68 static char *cgroup_mnt;
69 static int exit_value;
70 static volatile int fio_abort;
71 static unsigned int nr_process = 0;
72 static unsigned int nr_thread = 0;
74 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
77 unsigned int thread_number = 0;
78 unsigned int stat_number = 0;
81 unsigned long done_secs = 0;
82 volatile int helper_exit = 0;
84 #define PAGE_ALIGN(buf) \
85 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
87 #define JOB_START_TIMEOUT (5 * 1000)
89 static void sig_int(int sig)
93 fio_server_got_signal(sig);
95 log_info("\nfio: terminating on signal %d\n", sig);
100 fio_terminate_threads(TERMINATE_ALL);
104 void sig_show_status(int sig)
106 show_running_run_stats();
109 static void set_sig_handlers(void)
111 struct sigaction act;
113 memset(&act, 0, sizeof(act));
114 act.sa_handler = sig_int;
115 act.sa_flags = SA_RESTART;
116 sigaction(SIGINT, &act, NULL);
118 memset(&act, 0, sizeof(act));
119 act.sa_handler = sig_int;
120 act.sa_flags = SA_RESTART;
121 sigaction(SIGTERM, &act, NULL);
123 /* Windows uses SIGBREAK as a quit signal from other applications */
125 memset(&act, 0, sizeof(act));
126 act.sa_handler = sig_int;
127 act.sa_flags = SA_RESTART;
128 sigaction(SIGBREAK, &act, NULL);
131 memset(&act, 0, sizeof(act));
132 act.sa_handler = sig_show_status;
133 act.sa_flags = SA_RESTART;
134 sigaction(SIGUSR1, &act, NULL);
137 memset(&act, 0, sizeof(act));
138 act.sa_handler = sig_int;
139 act.sa_flags = SA_RESTART;
140 sigaction(SIGPIPE, &act, NULL);
145 * Check if we are above the minimum rate given.
147 static bool __check_min_rate(struct thread_data *td, struct timeval *now,
150 unsigned long long bytes = 0;
151 unsigned long iops = 0;
154 unsigned int ratemin = 0;
155 unsigned int rate_iops = 0;
156 unsigned int rate_iops_min = 0;
158 assert(ddir_rw(ddir));
160 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
164 * allow a 2 second settle period in the beginning
166 if (mtime_since(&td->start, now) < 2000)
169 iops += td->this_io_blocks[ddir];
170 bytes += td->this_io_bytes[ddir];
171 ratemin += td->o.ratemin[ddir];
172 rate_iops += td->o.rate_iops[ddir];
173 rate_iops_min += td->o.rate_iops_min[ddir];
176 * if rate blocks is set, sample is running
178 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
179 spent = mtime_since(&td->lastrate[ddir], now);
180 if (spent < td->o.ratecycle)
183 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
185 * check bandwidth specified rate
187 if (bytes < td->rate_bytes[ddir]) {
188 log_err("%s: min rate %u not met\n", td->o.name,
193 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
197 if (rate < ratemin ||
198 bytes < td->rate_bytes[ddir]) {
199 log_err("%s: min rate %u not met, got"
200 " %luKB/sec\n", td->o.name,
207 * checks iops specified rate
209 if (iops < rate_iops) {
210 log_err("%s: min iops rate %u not met\n",
211 td->o.name, rate_iops);
215 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
219 if (rate < rate_iops_min ||
220 iops < td->rate_blocks[ddir]) {
221 log_err("%s: min iops rate %u not met,"
222 " got %lu\n", td->o.name,
223 rate_iops_min, rate);
230 td->rate_bytes[ddir] = bytes;
231 td->rate_blocks[ddir] = iops;
232 memcpy(&td->lastrate[ddir], now, sizeof(*now));
236 static bool check_min_rate(struct thread_data *td, struct timeval *now)
240 if (td->bytes_done[DDIR_READ])
241 ret |= __check_min_rate(td, now, DDIR_READ);
242 if (td->bytes_done[DDIR_WRITE])
243 ret |= __check_min_rate(td, now, DDIR_WRITE);
244 if (td->bytes_done[DDIR_TRIM])
245 ret |= __check_min_rate(td, now, DDIR_TRIM);
251 * When job exits, we can cancel the in-flight IO if we are using async
252 * io. Attempt to do so.
254 static void cleanup_pending_aio(struct thread_data *td)
259 * get immediately available events, if any
261 r = io_u_queued_complete(td, 0);
266 * now cancel remaining active events
268 if (td->io_ops->cancel) {
272 io_u_qiter(&td->io_u_all, io_u, i) {
273 if (io_u->flags & IO_U_F_FLIGHT) {
274 r = td->io_ops->cancel(td, io_u);
282 r = io_u_queued_complete(td, td->cur_depth);
286 * Helper to handle the final sync of a file. Works just like the normal
287 * io path, just does everything sync.
289 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
291 struct io_u *io_u = __get_io_u(td);
297 io_u->ddir = DDIR_SYNC;
300 if (td_io_prep(td, io_u)) {
306 ret = td_io_queue(td, io_u);
308 td_verror(td, io_u->error, "td_io_queue");
311 } else if (ret == FIO_Q_QUEUED) {
312 if (td_io_commit(td))
314 if (io_u_queued_complete(td, 1) < 0)
316 } else if (ret == FIO_Q_COMPLETED) {
318 td_verror(td, io_u->error, "td_io_queue");
322 if (io_u_sync_complete(td, io_u) < 0)
324 } else if (ret == FIO_Q_BUSY) {
325 if (td_io_commit(td))
333 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
337 if (fio_file_open(f))
338 return fio_io_sync(td, f);
340 if (td_io_open_file(td, f))
343 ret = fio_io_sync(td, f);
344 td_io_close_file(td, f);
348 static inline void __update_tv_cache(struct thread_data *td)
350 fio_gettime(&td->tv_cache, NULL);
353 static inline void update_tv_cache(struct thread_data *td)
355 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
356 __update_tv_cache(td);
359 static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
361 if (in_ramp_time(td))
365 if (utime_since(&td->epoch, t) >= td->o.timeout)
372 * We need to update the runtime consistently in ms, but keep a running
373 * tally of the current elapsed time in microseconds for sub millisecond
376 static inline void update_runtime(struct thread_data *td,
377 unsigned long long *elapsed_us,
378 const enum fio_ddir ddir)
380 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
383 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
384 elapsed_us[ddir] += utime_since_now(&td->start);
385 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
388 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
393 if (ret < 0 || td->error) {
395 enum error_type_bit eb;
400 eb = td_error_type(ddir, err);
401 if (!(td->o.continue_on_error & (1 << eb)))
404 if (td_non_fatal_error(td, eb, err)) {
406 * Continue with the I/Os in case of
409 update_error_count(td, err);
413 } else if (td->o.fill_device && err == ENOSPC) {
415 * We expect to hit this error if
416 * fill_device option is set.
419 fio_mark_td_terminate(td);
423 * Stop the I/O in case of a fatal
426 update_error_count(td, err);
434 static void check_update_rusage(struct thread_data *td)
436 if (td->update_rusage) {
437 td->update_rusage = 0;
438 update_rusage_stat(td);
439 fio_mutex_up(td->rusage_sem);
443 static int wait_for_completions(struct thread_data *td, struct timeval *time)
445 const int full = queue_full(td);
450 * if the queue is full, we MUST reap at least 1 event
452 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
453 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
456 if (time && (__should_check_rate(td, DDIR_READ) ||
457 __should_check_rate(td, DDIR_WRITE) ||
458 __should_check_rate(td, DDIR_TRIM)))
459 fio_gettime(time, NULL);
462 ret = io_u_queued_complete(td, min_evts);
465 } while (full && (td->cur_depth > td->o.iodepth_low));
470 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
471 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
472 struct timeval *comp_time)
477 case FIO_Q_COMPLETED:
480 clear_io_u(td, io_u);
481 } else if (io_u->resid) {
482 int bytes = io_u->xfer_buflen - io_u->resid;
483 struct fio_file *f = io_u->file;
486 *bytes_issued += bytes;
489 trim_io_piece(td, io_u);
496 unlog_io_piece(td, io_u);
497 td_verror(td, EIO, "full resid");
502 io_u->xfer_buflen = io_u->resid;
503 io_u->xfer_buf += bytes;
504 io_u->offset += bytes;
506 if (ddir_rw(io_u->ddir))
507 td->ts.short_io_u[io_u->ddir]++;
510 if (io_u->offset == f->real_file_size)
513 requeue_io_u(td, &io_u);
516 if (comp_time && (__should_check_rate(td, DDIR_READ) ||
517 __should_check_rate(td, DDIR_WRITE) ||
518 __should_check_rate(td, DDIR_TRIM)))
519 fio_gettime(comp_time, NULL);
521 *ret = io_u_sync_complete(td, io_u);
527 * when doing I/O (not when verifying),
528 * check for any errors that are to be ignored
536 * if the engine doesn't have a commit hook,
537 * the io_u is really queued. if it does have such
538 * a hook, it has to call io_u_queued() itself.
540 if (td->io_ops->commit == NULL)
541 io_u_queued(td, io_u);
543 *bytes_issued += io_u->xfer_buflen;
547 unlog_io_piece(td, io_u);
548 requeue_io_u(td, &io_u);
549 ret2 = td_io_commit(td);
555 td_verror(td, -(*ret), "td_io_queue");
559 if (break_on_this_error(td, ddir, ret))
565 static inline bool io_in_polling(struct thread_data *td)
567 return !td->o.iodepth_batch_complete_min &&
568 !td->o.iodepth_batch_complete_max;
572 * The main verify engine. Runs over the writes we previously submitted,
573 * reads the blocks back in, and checks the crc/md5 of the data.
575 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
582 dprint(FD_VERIFY, "starting loop\n");
585 * sync io first and invalidate cache, to make sure we really
588 for_each_file(td, f, i) {
589 if (!fio_file_open(f))
591 if (fio_io_sync(td, f))
593 if (file_invalidate_cache(td, f))
597 check_update_rusage(td);
602 td_set_runstate(td, TD_VERIFYING);
605 while (!td->terminate) {
610 check_update_rusage(td);
612 if (runtime_exceeded(td, &td->tv_cache)) {
613 __update_tv_cache(td);
614 if (runtime_exceeded(td, &td->tv_cache)) {
615 fio_mark_td_terminate(td);
620 if (flow_threshold_exceeded(td))
623 if (!td->o.experimental_verify) {
624 io_u = __get_io_u(td);
628 if (get_next_verify(td, io_u)) {
633 if (td_io_prep(td, io_u)) {
638 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
641 while ((io_u = get_io_u(td)) != NULL) {
649 * We are only interested in the places where
650 * we wrote or trimmed IOs. Turn those into
651 * reads for verification purposes.
653 if (io_u->ddir == DDIR_READ) {
655 * Pretend we issued it for rwmix
658 td->io_issues[DDIR_READ]++;
661 } else if (io_u->ddir == DDIR_TRIM) {
662 io_u->ddir = DDIR_READ;
663 io_u_set(io_u, IO_U_F_TRIMMED);
665 } else if (io_u->ddir == DDIR_WRITE) {
666 io_u->ddir = DDIR_READ;
678 if (verify_state_should_stop(td, io_u)) {
683 if (td->o.verify_async)
684 io_u->end_io = verify_io_u_async;
686 io_u->end_io = verify_io_u;
689 if (!td->o.disable_slat)
690 fio_gettime(&io_u->start_time, NULL);
692 ret = td_io_queue(td, io_u);
694 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
698 * if we can queue more, do so. but check if there are
699 * completed io_u's first. Note that we can get BUSY even
700 * without IO queued, if the system is resource starved.
703 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
704 if (full || io_in_polling(td))
705 ret = wait_for_completions(td, NULL);
711 check_update_rusage(td);
714 min_events = td->cur_depth;
717 ret = io_u_queued_complete(td, min_events);
719 cleanup_pending_aio(td);
721 td_set_runstate(td, TD_RUNNING);
723 dprint(FD_VERIFY, "exiting loop\n");
726 static bool exceeds_number_ios(struct thread_data *td)
728 unsigned long long number_ios;
730 if (!td->o.number_ios)
733 number_ios = ddir_rw_sum(td->io_blocks);
734 number_ios += td->io_u_queued + td->io_u_in_flight;
736 return number_ios >= (td->o.number_ios * td->loops);
739 static bool io_issue_bytes_exceeded(struct thread_data *td)
741 unsigned long long bytes, limit;
744 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
745 else if (td_write(td))
746 bytes = td->io_issue_bytes[DDIR_WRITE];
747 else if (td_read(td))
748 bytes = td->io_issue_bytes[DDIR_READ];
750 bytes = td->io_issue_bytes[DDIR_TRIM];
753 limit = td->o.io_limit;
758 return bytes >= limit || exceeds_number_ios(td);
761 static bool io_complete_bytes_exceeded(struct thread_data *td)
763 unsigned long long bytes, limit;
766 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
767 else if (td_write(td))
768 bytes = td->this_io_bytes[DDIR_WRITE];
769 else if (td_read(td))
770 bytes = td->this_io_bytes[DDIR_READ];
772 bytes = td->this_io_bytes[DDIR_TRIM];
775 limit = td->o.io_limit;
780 return bytes >= limit || exceeds_number_ios(td);
784 * used to calculate the next io time for rate control
787 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
789 uint64_t secs, remainder, bps, bytes, iops;
791 assert(!(td->flags & TD_F_CHILD));
792 bytes = td->rate_io_issue_bytes[ddir];
793 bps = td->rate_bps[ddir];
795 if (td->o.rate_process == RATE_PROCESS_POISSON) {
797 iops = bps / td->o.bs[ddir];
798 val = (int64_t) (1000000 / iops) *
799 -logf(__rand_0_1(&td->poisson_state));
801 dprint(FD_RATE, "poisson rate iops=%llu\n",
802 (unsigned long long) 1000000 / val);
804 td->last_usec += val;
805 return td->last_usec;
808 remainder = bytes % bps;
809 return remainder * 1000000 / bps + secs * 1000000;
816 * Main IO worker function. It retrieves io_u's to process and queues
817 * and reaps them, checking for rate and errors along the way.
819 * Returns number of bytes written and trimmed.
821 static void do_io(struct thread_data *td, uint64_t *bytes_done)
825 uint64_t total_bytes, bytes_issued = 0;
827 for (i = 0; i < DDIR_RWDIR_CNT; i++)
828 bytes_done[i] = td->bytes_done[i];
830 if (in_ramp_time(td))
831 td_set_runstate(td, TD_RAMP);
833 td_set_runstate(td, TD_RUNNING);
837 total_bytes = td->o.size;
839 * Allow random overwrite workloads to write up to io_limit
840 * before starting verification phase as 'size' doesn't apply.
842 if (td_write(td) && td_random(td) && td->o.norandommap)
843 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
845 * If verify_backlog is enabled, we'll run the verify in this
846 * handler as well. For that case, we may need up to twice the
849 if (td->o.verify != VERIFY_NONE &&
850 (td_write(td) && td->o.verify_backlog))
851 total_bytes += td->o.size;
853 /* In trimwrite mode, each byte is trimmed and then written, so
854 * allow total_bytes to be twice as big */
855 if (td_trimwrite(td))
856 total_bytes += td->total_io_size;
858 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
859 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
861 struct timeval comp_time;
866 check_update_rusage(td);
868 if (td->terminate || td->done)
873 if (runtime_exceeded(td, &td->tv_cache)) {
874 __update_tv_cache(td);
875 if (runtime_exceeded(td, &td->tv_cache)) {
876 fio_mark_td_terminate(td);
881 if (flow_threshold_exceeded(td))
885 * Break if we exceeded the bytes. The exception is time
886 * based runs, but we still need to break out of the loop
887 * for those to run verification, if enabled.
889 if (bytes_issued >= total_bytes &&
890 (!td->o.time_based ||
891 (td->o.time_based && td->o.verify != VERIFY_NONE)))
895 if (IS_ERR_OR_NULL(io_u)) {
896 int err = PTR_ERR(io_u);
903 if (td->o.latency_target)
911 * Add verification end_io handler if:
912 * - Asked to verify (!td_rw(td))
913 * - Or the io_u is from our verify list (mixed write/ver)
915 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
916 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
918 if (!td->o.verify_pattern_bytes) {
919 io_u->rand_seed = __rand(&td->verify_state);
920 if (sizeof(int) != sizeof(long *))
921 io_u->rand_seed *= __rand(&td->verify_state);
924 if (verify_state_should_stop(td, io_u)) {
929 if (td->o.verify_async)
930 io_u->end_io = verify_io_u_async;
932 io_u->end_io = verify_io_u;
933 td_set_runstate(td, TD_VERIFYING);
934 } else if (in_ramp_time(td))
935 td_set_runstate(td, TD_RAMP);
937 td_set_runstate(td, TD_RUNNING);
940 * Always log IO before it's issued, so we know the specific
941 * order of it. The logged unit will track when the IO has
944 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
946 td->o.verify != VERIFY_NONE &&
947 !td->o.experimental_verify)
948 log_io_piece(td, io_u);
950 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
951 const unsigned long blen = io_u->xfer_buflen;
952 const enum fio_ddir ddir = acct_ddir(io_u);
957 workqueue_enqueue(&td->io_wq, &io_u->work);
961 td->io_issues[ddir]++;
962 td->io_issue_bytes[ddir] += blen;
963 td->rate_io_issue_bytes[ddir] += blen;
966 if (should_check_rate(td))
967 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
970 ret = td_io_queue(td, io_u);
972 if (should_check_rate(td))
973 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
975 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
979 * See if we need to complete some commands. Note that
980 * we can get BUSY even without IO queued, if the
981 * system is resource starved.
984 full = queue_full(td) ||
985 (ret == FIO_Q_BUSY && td->cur_depth);
986 if (full || io_in_polling(td))
987 ret = wait_for_completions(td, &comp_time);
991 if (!ddir_rw_sum(td->bytes_done) &&
992 !(td->io_ops->flags & FIO_NOIO))
995 if (!in_ramp_time(td) && should_check_rate(td)) {
996 if (check_min_rate(td, &comp_time)) {
997 if (exitall_on_terminate || td->o.exitall_error)
998 fio_terminate_threads(td->groupid);
999 td_verror(td, EIO, "check_min_rate");
1003 if (!in_ramp_time(td) && td->o.latency_target)
1004 lat_target_check(td);
1006 if (td->o.thinktime) {
1007 unsigned long long b;
1009 b = ddir_rw_sum(td->io_blocks);
1010 if (!(b % td->o.thinktime_blocks)) {
1015 if (td->o.thinktime_spin)
1016 usec_spin(td->o.thinktime_spin);
1018 left = td->o.thinktime - td->o.thinktime_spin;
1020 usec_sleep(td, left);
1025 check_update_rusage(td);
1027 if (td->trim_entries)
1028 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1030 if (td->o.fill_device && td->error == ENOSPC) {
1032 fio_mark_td_terminate(td);
1037 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1038 workqueue_flush(&td->io_wq);
1044 ret = io_u_queued_complete(td, i);
1045 if (td->o.fill_device && td->error == ENOSPC)
1049 if (should_fsync(td) && td->o.end_fsync) {
1050 td_set_runstate(td, TD_FSYNCING);
1052 for_each_file(td, f, i) {
1053 if (!fio_file_fsync(td, f))
1056 log_err("fio: end_fsync failed for file %s\n",
1061 cleanup_pending_aio(td);
1064 * stop job if we failed doing any IO
1066 if (!ddir_rw_sum(td->this_io_bytes))
1069 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1070 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1073 static void cleanup_io_u(struct thread_data *td)
1077 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1079 if (td->io_ops->io_u_free)
1080 td->io_ops->io_u_free(td, io_u);
1082 fio_memfree(io_u, sizeof(*io_u));
1087 io_u_rexit(&td->io_u_requeues);
1088 io_u_qexit(&td->io_u_freelist);
1089 io_u_qexit(&td->io_u_all);
1091 if (td->last_write_comp)
1092 sfree(td->last_write_comp);
1095 static int init_io_u(struct thread_data *td)
1098 unsigned int max_bs, min_write;
1099 int cl_align, i, max_units;
1100 int data_xfer = 1, err;
1103 max_units = td->o.iodepth;
1104 max_bs = td_max_bs(td);
1105 min_write = td->o.min_bs[DDIR_WRITE];
1106 td->orig_buffer_size = (unsigned long long) max_bs
1107 * (unsigned long long) max_units;
1109 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1113 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1114 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1115 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1118 log_err("fio: failed setting up IO queues\n");
1123 * if we may later need to do address alignment, then add any
1124 * possible adjustment here so that we don't cause a buffer
1125 * overflow later. this adjustment may be too much if we get
1126 * lucky and the allocator gives us an aligned address.
1128 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1129 (td->io_ops->flags & FIO_RAWIO))
1130 td->orig_buffer_size += page_mask + td->o.mem_align;
1132 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1135 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1136 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1139 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1140 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1144 if (data_xfer && allocate_io_mem(td))
1147 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1148 (td->io_ops->flags & FIO_RAWIO))
1149 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1151 p = td->orig_buffer;
1153 cl_align = os_cache_line_size();
1155 for (i = 0; i < max_units; i++) {
1161 ptr = fio_memalign(cl_align, sizeof(*io_u));
1163 log_err("fio: unable to allocate aligned memory\n");
1168 memset(io_u, 0, sizeof(*io_u));
1169 INIT_FLIST_HEAD(&io_u->verify_list);
1170 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1174 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1177 io_u_fill_buffer(td, io_u, min_write, max_bs);
1178 if (td_write(td) && td->o.verify_pattern_bytes) {
1180 * Fill the buffer with the pattern if we are
1181 * going to be doing writes.
1183 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1188 io_u->flags = IO_U_F_FREE;
1189 io_u_qpush(&td->io_u_freelist, io_u);
1192 * io_u never leaves this stack, used for iteration of all
1195 io_u_qpush(&td->io_u_all, io_u);
1197 if (td->io_ops->io_u_init) {
1198 int ret = td->io_ops->io_u_init(td, io_u);
1201 log_err("fio: failed to init engine data: %d\n", ret);
1209 if (td->o.verify != VERIFY_NONE) {
1210 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1211 if (!td->last_write_comp) {
1212 log_err("fio: failed to alloc write comp data\n");
1220 static int switch_ioscheduler(struct thread_data *td)
1222 char tmp[256], tmp2[128];
1226 if (td->io_ops->flags & FIO_DISKLESSIO)
1229 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1231 f = fopen(tmp, "r+");
1233 if (errno == ENOENT) {
1234 log_err("fio: os or kernel doesn't support IO scheduler"
1238 td_verror(td, errno, "fopen iosched");
1245 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1246 if (ferror(f) || ret != 1) {
1247 td_verror(td, errno, "fwrite");
1255 * Read back and check that the selected scheduler is now the default.
1257 memset(tmp, 0, sizeof(tmp));
1258 ret = fread(tmp, sizeof(tmp), 1, f);
1259 if (ferror(f) || ret < 0) {
1260 td_verror(td, errno, "fread");
1265 * either a list of io schedulers or "none\n" is expected.
1267 tmp[strlen(tmp) - 1] = '\0';
1270 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1271 if (!strstr(tmp, tmp2)) {
1272 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1273 td_verror(td, EINVAL, "iosched_switch");
1282 static bool keep_running(struct thread_data *td)
1284 unsigned long long limit;
1288 if (td->o.time_based)
1294 if (exceeds_number_ios(td))
1298 limit = td->o.io_limit;
1302 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1306 * If the difference is less than the minimum IO size, we
1309 diff = limit - ddir_rw_sum(td->io_bytes);
1310 if (diff < td_max_bs(td))
1313 if (fio_files_done(td) && !td->o.io_limit)
1322 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1324 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1328 str = malloc(newlen);
1329 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1331 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1334 log_err("fio: exec of cmd <%s> failed\n", str);
1341 * Dry run to compute correct state of numberio for verification.
1343 static uint64_t do_dry_run(struct thread_data *td)
1345 td_set_runstate(td, TD_RUNNING);
1347 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1348 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1352 if (td->terminate || td->done)
1355 io_u = get_io_u(td);
1359 io_u_set(io_u, IO_U_F_FLIGHT);
1362 if (ddir_rw(acct_ddir(io_u)))
1363 td->io_issues[acct_ddir(io_u)]++;
1364 if (ddir_rw(io_u->ddir)) {
1365 io_u_mark_depth(td, 1);
1366 td->ts.total_io_u[io_u->ddir]++;
1369 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1371 td->o.verify != VERIFY_NONE &&
1372 !td->o.experimental_verify)
1373 log_io_piece(td, io_u);
1375 ret = io_u_sync_complete(td, io_u);
1379 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1383 struct thread_data *td;
1384 struct sk_out *sk_out;
1388 * Entry point for the thread based jobs. The process based jobs end up
1389 * here as well, after a little setup.
1391 static void *thread_main(void *data)
1393 struct fork_data *fd = data;
1394 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1395 struct thread_data *td = fd->td;
1396 struct thread_options *o = &td->o;
1397 struct sk_out *sk_out = fd->sk_out;
1398 pthread_condattr_t attr;
1402 sk_out_assign(sk_out);
1405 if (!o->use_thread) {
1411 fio_local_clock_init(o->use_thread);
1413 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1416 fio_server_send_start(td);
1418 INIT_FLIST_HEAD(&td->io_log_list);
1419 INIT_FLIST_HEAD(&td->io_hist_list);
1420 INIT_FLIST_HEAD(&td->verify_list);
1421 INIT_FLIST_HEAD(&td->trim_list);
1422 INIT_FLIST_HEAD(&td->next_rand_list);
1423 pthread_mutex_init(&td->io_u_lock, NULL);
1424 td->io_hist_tree = RB_ROOT;
1426 pthread_condattr_init(&attr);
1427 pthread_cond_init(&td->verify_cond, &attr);
1428 pthread_cond_init(&td->free_cond, &attr);
1430 td_set_runstate(td, TD_INITIALIZED);
1431 dprint(FD_MUTEX, "up startup_mutex\n");
1432 fio_mutex_up(startup_mutex);
1433 dprint(FD_MUTEX, "wait on td->mutex\n");
1434 fio_mutex_down(td->mutex);
1435 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1438 * A new gid requires privilege, so we need to do this before setting
1441 if (o->gid != -1U && setgid(o->gid)) {
1442 td_verror(td, errno, "setgid");
1445 if (o->uid != -1U && setuid(o->uid)) {
1446 td_verror(td, errno, "setuid");
1451 * If we have a gettimeofday() thread, make sure we exclude that
1452 * thread from this job
1455 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1458 * Set affinity first, in case it has an impact on the memory
1461 if (fio_option_is_set(o, cpumask)) {
1462 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1463 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1465 log_err("fio: no CPUs set\n");
1466 log_err("fio: Try increasing number of available CPUs\n");
1467 td_verror(td, EINVAL, "cpus_split");
1471 ret = fio_setaffinity(td->pid, o->cpumask);
1473 td_verror(td, errno, "cpu_set_affinity");
1478 #ifdef CONFIG_LIBNUMA
1479 /* numa node setup */
1480 if (fio_option_is_set(o, numa_cpunodes) ||
1481 fio_option_is_set(o, numa_memnodes)) {
1482 struct bitmask *mask;
1484 if (numa_available() < 0) {
1485 td_verror(td, errno, "Does not support NUMA API\n");
1489 if (fio_option_is_set(o, numa_cpunodes)) {
1490 mask = numa_parse_nodestring(o->numa_cpunodes);
1491 ret = numa_run_on_node_mask(mask);
1492 numa_free_nodemask(mask);
1494 td_verror(td, errno, \
1495 "numa_run_on_node_mask failed\n");
1500 if (fio_option_is_set(o, numa_memnodes)) {
1502 if (o->numa_memnodes)
1503 mask = numa_parse_nodestring(o->numa_memnodes);
1505 switch (o->numa_mem_mode) {
1506 case MPOL_INTERLEAVE:
1507 numa_set_interleave_mask(mask);
1510 numa_set_membind(mask);
1513 numa_set_localalloc();
1515 case MPOL_PREFERRED:
1516 numa_set_preferred(o->numa_mem_prefer_node);
1524 numa_free_nodemask(mask);
1530 if (fio_pin_memory(td))
1534 * May alter parameters that init_io_u() will use, so we need to
1543 if (o->verify_async && verify_async_init(td))
1546 if (fio_option_is_set(o, ioprio) ||
1547 fio_option_is_set(o, ioprio_class)) {
1548 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1550 td_verror(td, errno, "ioprio_set");
1555 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1559 if (nice(o->nice) == -1 && errno != 0) {
1560 td_verror(td, errno, "nice");
1564 if (o->ioscheduler && switch_ioscheduler(td))
1567 if (!o->create_serialize && setup_files(td))
1573 if (init_random_map(td))
1576 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1580 if (pre_read_files(td) < 0)
1584 if (iolog_compress_init(td, sk_out))
1587 fio_verify_init(td);
1589 if (rate_submit_init(td, sk_out))
1592 fio_gettime(&td->epoch, NULL);
1593 fio_getrusage(&td->ru_start);
1594 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1595 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1597 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1598 o->ratemin[DDIR_TRIM]) {
1599 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1600 sizeof(td->bw_sample_time));
1601 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1602 sizeof(td->bw_sample_time));
1603 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1604 sizeof(td->bw_sample_time));
1608 while (keep_running(td)) {
1609 uint64_t verify_bytes;
1611 fio_gettime(&td->start, NULL);
1612 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1615 clear_io_state(td, 0);
1617 prune_io_piece_log(td);
1619 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1620 verify_bytes = do_dry_run(td);
1622 uint64_t bytes_done[DDIR_RWDIR_CNT];
1624 do_io(td, bytes_done);
1626 if (!ddir_rw_sum(bytes_done)) {
1627 fio_mark_td_terminate(td);
1630 verify_bytes = bytes_done[DDIR_WRITE] +
1631 bytes_done[DDIR_TRIM];
1638 * Make sure we've successfully updated the rusage stats
1639 * before waiting on the stat mutex. Otherwise we could have
1640 * the stat thread holding stat mutex and waiting for
1641 * the rusage_sem, which would never get upped because
1642 * this thread is waiting for the stat mutex.
1644 check_update_rusage(td);
1646 fio_mutex_down(stat_mutex);
1647 if (td_read(td) && td->io_bytes[DDIR_READ])
1648 update_runtime(td, elapsed_us, DDIR_READ);
1649 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1650 update_runtime(td, elapsed_us, DDIR_WRITE);
1651 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1652 update_runtime(td, elapsed_us, DDIR_TRIM);
1653 fio_gettime(&td->start, NULL);
1654 fio_mutex_up(stat_mutex);
1656 if (td->error || td->terminate)
1659 if (!o->do_verify ||
1660 o->verify == VERIFY_NONE ||
1661 (td->io_ops->flags & FIO_UNIDIR))
1664 clear_io_state(td, 0);
1666 fio_gettime(&td->start, NULL);
1668 do_verify(td, verify_bytes);
1671 * See comment further up for why this is done here.
1673 check_update_rusage(td);
1675 fio_mutex_down(stat_mutex);
1676 update_runtime(td, elapsed_us, DDIR_READ);
1677 fio_gettime(&td->start, NULL);
1678 fio_mutex_up(stat_mutex);
1680 if (td->error || td->terminate)
1684 update_rusage_stat(td);
1685 td->ts.total_run_time = mtime_since_now(&td->epoch);
1686 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1687 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1688 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1690 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1691 (td->o.verify != VERIFY_NONE && td_write(td)))
1692 verify_save_state(td->thread_number);
1694 fio_unpin_memory(td);
1696 fio_writeout_logs(td);
1698 iolog_compress_exit(td);
1699 rate_submit_exit(td);
1701 if (o->exec_postrun)
1702 exec_string(o, o->exec_postrun, (const char *)"postrun");
1704 if (exitall_on_terminate || (o->exitall_error && td->error))
1705 fio_terminate_threads(td->groupid);
1709 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1712 if (o->verify_async)
1713 verify_async_exit(td);
1715 close_and_free_files(td);
1718 cgroup_shutdown(td, &cgroup_mnt);
1719 verify_free_state(td);
1721 if (td->zone_state_index) {
1724 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1725 free(td->zone_state_index[i]);
1726 free(td->zone_state_index);
1727 td->zone_state_index = NULL;
1730 if (fio_option_is_set(o, cpumask)) {
1731 ret = fio_cpuset_exit(&o->cpumask);
1733 td_verror(td, ret, "fio_cpuset_exit");
1737 * do this very late, it will log file closing as well
1739 if (o->write_iolog_file)
1740 write_iolog_close(td);
1742 fio_mutex_remove(td->mutex);
1745 td_set_runstate(td, TD_EXITED);
1748 * Do this last after setting our runstate to exited, so we
1749 * know that the stat thread is signaled.
1751 check_update_rusage(td);
1754 return (void *) (uintptr_t) td->error;
1759 * We cannot pass the td data into a forked process, so attach the td and
1760 * pass it to the thread worker.
1762 static int fork_main(struct sk_out *sk_out, int shmid, int offset)
1764 struct fork_data *fd;
1767 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1768 data = shmat(shmid, NULL, 0);
1769 if (data == (void *) -1) {
1777 * HP-UX inherits shm mappings?
1782 fd = calloc(1, sizeof(*fd));
1783 fd->td = data + offset * sizeof(struct thread_data);
1784 fd->sk_out = sk_out;
1785 ret = thread_main(fd);
1787 return (int) (uintptr_t) ret;
1790 static void dump_td_info(struct thread_data *td)
1792 log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1793 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1794 (unsigned long) time_since_now(&td->terminate_time));
1798 * Run over the job map and reap the threads that have exited, if any.
1800 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1801 unsigned int *m_rate)
1803 struct thread_data *td;
1804 unsigned int cputhreads, realthreads, pending;
1808 * reap exited threads (TD_EXITED -> TD_REAPED)
1810 realthreads = pending = cputhreads = 0;
1811 for_each_td(td, i) {
1815 * ->io_ops is NULL for a thread that has closed its
1818 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1827 if (td->runstate == TD_REAPED)
1829 if (td->o.use_thread) {
1830 if (td->runstate == TD_EXITED) {
1831 td_set_runstate(td, TD_REAPED);
1838 if (td->runstate == TD_EXITED)
1842 * check if someone quit or got killed in an unusual way
1844 ret = waitpid(td->pid, &status, flags);
1846 if (errno == ECHILD) {
1847 log_err("fio: pid=%d disappeared %d\n",
1848 (int) td->pid, td->runstate);
1850 td_set_runstate(td, TD_REAPED);
1854 } else if (ret == td->pid) {
1855 if (WIFSIGNALED(status)) {
1856 int sig = WTERMSIG(status);
1858 if (sig != SIGTERM && sig != SIGUSR2)
1859 log_err("fio: pid=%d, got signal=%d\n",
1860 (int) td->pid, sig);
1862 td_set_runstate(td, TD_REAPED);
1865 if (WIFEXITED(status)) {
1866 if (WEXITSTATUS(status) && !td->error)
1867 td->error = WEXITSTATUS(status);
1869 td_set_runstate(td, TD_REAPED);
1875 * If the job is stuck, do a forceful timeout of it and
1878 if (td->terminate &&
1879 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1881 td_set_runstate(td, TD_REAPED);
1886 * thread is not dead, continue
1892 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1893 (*t_rate) -= ddir_rw_sum(td->o.rate);
1900 done_secs += mtime_since_now(&td->epoch) / 1000;
1901 profile_td_exit(td);
1904 if (*nr_running == cputhreads && !pending && realthreads)
1905 fio_terminate_threads(TERMINATE_ALL);
1908 static bool __check_trigger_file(void)
1915 if (stat(trigger_file, &sb))
1918 if (unlink(trigger_file) < 0)
1919 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1925 static bool trigger_timedout(void)
1927 if (trigger_timeout)
1928 return time_since_genesis() >= trigger_timeout;
1933 void exec_trigger(const char *cmd)
1942 log_err("fio: failed executing %s trigger\n", cmd);
1945 void check_trigger_file(void)
1947 if (__check_trigger_file() || trigger_timedout()) {
1949 fio_clients_send_trigger(trigger_remote_cmd);
1951 verify_save_state(IO_LIST_ALL);
1952 fio_terminate_threads(TERMINATE_ALL);
1953 exec_trigger(trigger_cmd);
1958 static int fio_verify_load_state(struct thread_data *td)
1962 if (!td->o.verify_state)
1969 ret = fio_server_get_verify_state(td->o.name,
1970 td->thread_number - 1, &data, &ver);
1972 verify_convert_assign_state(td, data, ver);
1974 ret = verify_load_state(td, "local");
1979 static void do_usleep(unsigned int usecs)
1981 check_for_running_stats();
1982 check_trigger_file();
1986 static bool check_mount_writes(struct thread_data *td)
1991 if (!td_write(td) || td->o.allow_mounted_write)
1994 for_each_file(td, f, i) {
1995 if (f->filetype != FIO_TYPE_BD)
1997 if (device_is_mounted(f->file_name))
2003 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
2007 static bool waitee_running(struct thread_data *me)
2009 const char *waitee = me->o.wait_for;
2010 const char *self = me->o.name;
2011 struct thread_data *td;
2017 for_each_td(td, i) {
2018 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2021 if (td->runstate < TD_EXITED) {
2022 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2024 runstate_to_name(td->runstate));
2029 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2034 * Main function for kicking off and reaping jobs, as needed.
2036 static void run_threads(struct sk_out *sk_out)
2038 struct thread_data *td;
2039 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
2042 if (fio_gtod_offload && fio_start_gtod_thread())
2045 fio_idle_prof_init();
2049 nr_thread = nr_process = 0;
2050 for_each_td(td, i) {
2051 if (check_mount_writes(td))
2053 if (td->o.use_thread)
2059 if (output_format & FIO_OUTPUT_NORMAL) {
2060 log_info("Starting ");
2062 log_info("%d thread%s", nr_thread,
2063 nr_thread > 1 ? "s" : "");
2067 log_info("%d process%s", nr_process,
2068 nr_process > 1 ? "es" : "");
2074 todo = thread_number;
2077 m_rate = t_rate = 0;
2079 for_each_td(td, i) {
2080 print_status_init(td->thread_number - 1);
2082 if (!td->o.create_serialize)
2085 if (fio_verify_load_state(td))
2089 * do file setup here so it happens sequentially,
2090 * we don't want X number of threads getting their
2091 * client data interspersed on disk
2093 if (setup_files(td)) {
2097 log_err("fio: pid=%d, err=%d/%s\n",
2098 (int) td->pid, td->error, td->verror);
2099 td_set_runstate(td, TD_REAPED);
2106 * for sharing to work, each job must always open
2107 * its own files. so close them, if we opened them
2110 for_each_file(td, f, j) {
2111 if (fio_file_open(f))
2112 td_io_close_file(td, f);
2117 /* start idle threads before io threads start to run */
2118 fio_idle_prof_start();
2123 struct thread_data *map[REAL_MAX_JOBS];
2124 struct timeval this_start;
2125 int this_jobs = 0, left;
2128 * create threads (TD_NOT_CREATED -> TD_CREATED)
2130 for_each_td(td, i) {
2131 if (td->runstate != TD_NOT_CREATED)
2135 * never got a chance to start, killed by other
2136 * thread for some reason
2138 if (td->terminate) {
2143 if (td->o.start_delay) {
2144 spent = utime_since_genesis();
2146 if (td->o.start_delay > spent)
2150 if (td->o.stonewall && (nr_started || nr_running)) {
2151 dprint(FD_PROCESS, "%s: stonewall wait\n",
2156 if (waitee_running(td)) {
2157 dprint(FD_PROCESS, "%s: waiting for %s\n",
2158 td->o.name, td->o.wait_for);
2164 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2165 td->update_rusage = 0;
2168 * Set state to created. Thread will transition
2169 * to TD_INITIALIZED when it's done setting up.
2171 td_set_runstate(td, TD_CREATED);
2172 map[this_jobs++] = td;
2175 if (td->o.use_thread) {
2176 struct fork_data *fd;
2179 fd = calloc(1, sizeof(*fd));
2181 fd->sk_out = sk_out;
2183 dprint(FD_PROCESS, "will pthread_create\n");
2184 ret = pthread_create(&td->thread, NULL,
2187 log_err("pthread_create: %s\n",
2193 ret = pthread_detach(td->thread);
2195 log_err("pthread_detach: %s",
2199 dprint(FD_PROCESS, "will fork\n");
2202 int ret = fork_main(sk_out, shm_id, i);
2205 } else if (i == fio_debug_jobno)
2206 *fio_debug_jobp = pid;
2208 dprint(FD_MUTEX, "wait on startup_mutex\n");
2209 if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2210 log_err("fio: job startup hung? exiting.\n");
2211 fio_terminate_threads(TERMINATE_ALL);
2216 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2220 * Wait for the started threads to transition to
2223 fio_gettime(&this_start, NULL);
2225 while (left && !fio_abort) {
2226 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2231 for (i = 0; i < this_jobs; i++) {
2235 if (td->runstate == TD_INITIALIZED) {
2238 } else if (td->runstate >= TD_EXITED) {
2242 nr_running++; /* work-around... */
2248 log_err("fio: %d job%s failed to start\n", left,
2249 left > 1 ? "s" : "");
2250 for (i = 0; i < this_jobs; i++) {
2254 kill(td->pid, SIGTERM);
2260 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2262 for_each_td(td, i) {
2263 if (td->runstate != TD_INITIALIZED)
2266 if (in_ramp_time(td))
2267 td_set_runstate(td, TD_RAMP);
2269 td_set_runstate(td, TD_RUNNING);
2272 m_rate += ddir_rw_sum(td->o.ratemin);
2273 t_rate += ddir_rw_sum(td->o.rate);
2275 fio_mutex_up(td->mutex);
2278 reap_threads(&nr_running, &t_rate, &m_rate);
2284 while (nr_running) {
2285 reap_threads(&nr_running, &t_rate, &m_rate);
2289 fio_idle_prof_stop();
2294 static void wait_for_helper_thread_exit(void)
2299 pthread_cond_signal(&helper_cond);
2300 pthread_join(helper_thread, &ret);
2303 static void free_disk_util(void)
2305 disk_util_prune_entries();
2307 pthread_cond_destroy(&helper_cond);
2310 static void *helper_thread_main(void *data)
2312 struct sk_out *sk_out = data;
2315 sk_out_assign(sk_out);
2317 fio_mutex_up(startup_mutex);
2320 uint64_t sec = DISK_UTIL_MSEC / 1000;
2321 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2325 gettimeofday(&tv, NULL);
2326 ts.tv_sec = tv.tv_sec + sec;
2327 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2329 if (ts.tv_nsec >= 1000000000ULL) {
2330 ts.tv_nsec -= 1000000000ULL;
2334 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2336 ret = update_io_ticks();
2338 if (helper_do_stat) {
2340 __show_running_run_stats();
2344 print_thread_status();
2351 static int create_helper_thread(struct sk_out *sk_out)
2357 pthread_cond_init(&helper_cond, NULL);
2358 pthread_mutex_init(&helper_lock, NULL);
2360 ret = pthread_create(&helper_thread, NULL, helper_thread_main, sk_out);
2362 log_err("Can't create helper thread: %s\n", strerror(ret));
2366 dprint(FD_MUTEX, "wait on startup_mutex\n");
2367 fio_mutex_down(startup_mutex);
2368 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2372 int fio_backend(struct sk_out *sk_out)
2374 struct thread_data *td;
2378 if (load_profile(exec_profile))
2381 exec_profile = NULL;
2387 struct log_params p = {
2388 .log_type = IO_LOG_TYPE_BW,
2391 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2392 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2393 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2396 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2397 if (startup_mutex == NULL)
2402 create_helper_thread(sk_out);
2404 cgroup_list = smalloc(sizeof(*cgroup_list));
2405 INIT_FLIST_HEAD(cgroup_list);
2407 run_threads(sk_out);
2409 wait_for_helper_thread_exit();
2414 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2415 struct io_log *log = agg_io_log[i];
2423 for_each_td(td, i) {
2424 fio_options_free(td);
2425 if (td->rusage_sem) {
2426 fio_mutex_remove(td->rusage_sem);
2427 td->rusage_sem = NULL;
2432 cgroup_kill(cgroup_list);
2436 fio_mutex_remove(startup_mutex);