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
54 #include "lib/getrusage.h"
58 #include "workqueue.h"
59 #include "lib/mountcheck.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 int __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 int 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 int 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 (io_u_queued_complete(td, 1) < 0)
314 } else if (ret == FIO_Q_COMPLETED) {
316 td_verror(td, io_u->error, "td_io_queue");
320 if (io_u_sync_complete(td, io_u) < 0)
322 } else if (ret == FIO_Q_BUSY) {
323 if (td_io_commit(td))
331 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
335 if (fio_file_open(f))
336 return fio_io_sync(td, f);
338 if (td_io_open_file(td, f))
341 ret = fio_io_sync(td, f);
342 td_io_close_file(td, f);
346 static inline void __update_tv_cache(struct thread_data *td)
348 fio_gettime(&td->tv_cache, NULL);
351 static inline void update_tv_cache(struct thread_data *td)
353 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
354 __update_tv_cache(td);
357 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
359 if (in_ramp_time(td))
363 if (utime_since(&td->epoch, t) >= td->o.timeout)
370 * We need to update the runtime consistently in ms, but keep a running
371 * tally of the current elapsed time in microseconds for sub millisecond
374 static inline void update_runtime(struct thread_data *td,
375 unsigned long long *elapsed_us,
376 const enum fio_ddir ddir)
378 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
381 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
382 elapsed_us[ddir] += utime_since_now(&td->start);
383 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
386 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
391 if (ret < 0 || td->error) {
393 enum error_type_bit eb;
398 eb = td_error_type(ddir, err);
399 if (!(td->o.continue_on_error & (1 << eb)))
402 if (td_non_fatal_error(td, eb, err)) {
404 * Continue with the I/Os in case of
407 update_error_count(td, err);
411 } else if (td->o.fill_device && err == ENOSPC) {
413 * We expect to hit this error if
414 * fill_device option is set.
417 fio_mark_td_terminate(td);
421 * Stop the I/O in case of a fatal
424 update_error_count(td, err);
432 static void check_update_rusage(struct thread_data *td)
434 if (td->update_rusage) {
435 td->update_rusage = 0;
436 update_rusage_stat(td);
437 fio_mutex_up(td->rusage_sem);
441 static int wait_for_completions(struct thread_data *td, struct timeval *time)
443 const int full = queue_full(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);
526 * if the engine doesn't have a commit hook,
527 * the io_u is really queued. if it does have such
528 * a hook, it has to call io_u_queued() itself.
530 if (td->io_ops->commit == NULL)
531 io_u_queued(td, io_u);
533 *bytes_issued += io_u->xfer_buflen;
537 unlog_io_piece(td, io_u);
538 requeue_io_u(td, &io_u);
539 ret2 = td_io_commit(td);
545 td_verror(td, -(*ret), "td_io_queue");
549 if (break_on_this_error(td, ddir, ret))
555 static inline int io_in_polling(struct thread_data *td)
557 return !td->o.iodepth_batch_complete_min &&
558 !td->o.iodepth_batch_complete_max;
562 * The main verify engine. Runs over the writes we previously submitted,
563 * reads the blocks back in, and checks the crc/md5 of the data.
565 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
572 dprint(FD_VERIFY, "starting loop\n");
575 * sync io first and invalidate cache, to make sure we really
578 for_each_file(td, f, i) {
579 if (!fio_file_open(f))
581 if (fio_io_sync(td, f))
583 if (file_invalidate_cache(td, f))
587 check_update_rusage(td);
592 td_set_runstate(td, TD_VERIFYING);
595 while (!td->terminate) {
600 check_update_rusage(td);
602 if (runtime_exceeded(td, &td->tv_cache)) {
603 __update_tv_cache(td);
604 if (runtime_exceeded(td, &td->tv_cache)) {
605 fio_mark_td_terminate(td);
610 if (flow_threshold_exceeded(td))
613 if (!td->o.experimental_verify) {
614 io_u = __get_io_u(td);
618 if (get_next_verify(td, io_u)) {
623 if (td_io_prep(td, io_u)) {
628 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
631 while ((io_u = get_io_u(td)) != NULL) {
639 * We are only interested in the places where
640 * we wrote or trimmed IOs. Turn those into
641 * reads for verification purposes.
643 if (io_u->ddir == DDIR_READ) {
645 * Pretend we issued it for rwmix
648 td->io_issues[DDIR_READ]++;
651 } else if (io_u->ddir == DDIR_TRIM) {
652 io_u->ddir = DDIR_READ;
653 io_u_set(io_u, IO_U_F_TRIMMED);
655 } else if (io_u->ddir == DDIR_WRITE) {
656 io_u->ddir = DDIR_READ;
668 if (verify_state_should_stop(td, io_u)) {
673 if (td->o.verify_async)
674 io_u->end_io = verify_io_u_async;
676 io_u->end_io = verify_io_u;
679 if (!td->o.disable_slat)
680 fio_gettime(&io_u->start_time, NULL);
682 ret = td_io_queue(td, io_u);
684 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
688 * if we can queue more, do so. but check if there are
689 * completed io_u's first. Note that we can get BUSY even
690 * without IO queued, if the system is resource starved.
693 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
694 if (full || io_in_polling(td))
695 ret = wait_for_completions(td, NULL);
701 check_update_rusage(td);
704 min_events = td->cur_depth;
707 ret = io_u_queued_complete(td, min_events);
709 cleanup_pending_aio(td);
711 td_set_runstate(td, TD_RUNNING);
713 dprint(FD_VERIFY, "exiting loop\n");
716 static unsigned int exceeds_number_ios(struct thread_data *td)
718 unsigned long long number_ios;
720 if (!td->o.number_ios)
723 number_ios = ddir_rw_sum(td->io_blocks);
724 number_ios += td->io_u_queued + td->io_u_in_flight;
726 return number_ios >= (td->o.number_ios * td->loops);
729 static int io_issue_bytes_exceeded(struct thread_data *td)
731 unsigned long long bytes, limit;
734 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
735 else if (td_write(td))
736 bytes = td->io_issue_bytes[DDIR_WRITE];
737 else if (td_read(td))
738 bytes = td->io_issue_bytes[DDIR_READ];
740 bytes = td->io_issue_bytes[DDIR_TRIM];
743 limit = td->o.io_limit;
748 return bytes >= limit || exceeds_number_ios(td);
751 static int io_complete_bytes_exceeded(struct thread_data *td)
753 unsigned long long bytes, limit;
756 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
757 else if (td_write(td))
758 bytes = td->this_io_bytes[DDIR_WRITE];
759 else if (td_read(td))
760 bytes = td->this_io_bytes[DDIR_READ];
762 bytes = td->this_io_bytes[DDIR_TRIM];
765 limit = td->o.io_limit;
770 return bytes >= limit || exceeds_number_ios(td);
774 * used to calculate the next io time for rate control
777 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
779 uint64_t secs, remainder, bps, bytes, iops;
781 assert(!(td->flags & TD_F_CHILD));
782 bytes = td->rate_io_issue_bytes[ddir];
783 bps = td->rate_bps[ddir];
785 if (td->o.poisson_rate) {
786 iops = bps / td->o.bs[ddir];
787 td->last_usec += (int64_t) (1000000 / iops) *
788 -logf(__rand_0_1(&td->poisson_state));
789 return td->last_usec;
792 remainder = bytes % bps;
793 return remainder * 1000000 / bps + secs * 1000000;
800 * Main IO worker function. It retrieves io_u's to process and queues
801 * and reaps them, checking for rate and errors along the way.
803 * Returns number of bytes written and trimmed.
805 static uint64_t do_io(struct thread_data *td)
809 uint64_t total_bytes, bytes_issued = 0;
811 if (in_ramp_time(td))
812 td_set_runstate(td, TD_RAMP);
814 td_set_runstate(td, TD_RUNNING);
818 total_bytes = td->o.size;
820 * Allow random overwrite workloads to write up to io_limit
821 * before starting verification phase as 'size' doesn't apply.
823 if (td_write(td) && td_random(td) && td->o.norandommap)
824 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
826 * If verify_backlog is enabled, we'll run the verify in this
827 * handler as well. For that case, we may need up to twice the
830 if (td->o.verify != VERIFY_NONE &&
831 (td_write(td) && td->o.verify_backlog))
832 total_bytes += td->o.size;
834 /* In trimwrite mode, each byte is trimmed and then written, so
835 * allow total_bytes to be twice as big */
836 if (td_trimwrite(td))
837 total_bytes += td->total_io_size;
839 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
840 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
842 struct timeval comp_time;
847 check_update_rusage(td);
849 if (td->terminate || td->done)
854 if (runtime_exceeded(td, &td->tv_cache)) {
855 __update_tv_cache(td);
856 if (runtime_exceeded(td, &td->tv_cache)) {
857 fio_mark_td_terminate(td);
862 if (flow_threshold_exceeded(td))
865 if (bytes_issued >= total_bytes)
869 if (IS_ERR_OR_NULL(io_u)) {
870 int err = PTR_ERR(io_u);
877 if (td->o.latency_target)
885 * Add verification end_io handler if:
886 * - Asked to verify (!td_rw(td))
887 * - Or the io_u is from our verify list (mixed write/ver)
889 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
890 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
892 if (!td->o.verify_pattern_bytes) {
893 io_u->rand_seed = __rand(&td->verify_state);
894 if (sizeof(int) != sizeof(long *))
895 io_u->rand_seed *= __rand(&td->verify_state);
898 if (verify_state_should_stop(td, io_u)) {
903 if (td->o.verify_async)
904 io_u->end_io = verify_io_u_async;
906 io_u->end_io = verify_io_u;
907 td_set_runstate(td, TD_VERIFYING);
908 } else if (in_ramp_time(td))
909 td_set_runstate(td, TD_RAMP);
911 td_set_runstate(td, TD_RUNNING);
914 * Always log IO before it's issued, so we know the specific
915 * order of it. The logged unit will track when the IO has
918 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
920 td->o.verify != VERIFY_NONE &&
921 !td->o.experimental_verify)
922 log_io_piece(td, io_u);
924 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
927 ret = workqueue_enqueue(&td->io_wq, io_u);
929 if (should_check_rate(td))
930 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
933 ret = td_io_queue(td, io_u);
935 if (should_check_rate(td))
936 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
938 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
942 * See if we need to complete some commands. Note that
943 * we can get BUSY even without IO queued, if the
944 * system is resource starved.
947 full = queue_full(td) ||
948 (ret == FIO_Q_BUSY && td->cur_depth);
949 if (full || io_in_polling(td))
950 ret = wait_for_completions(td, &comp_time);
954 if (!ddir_rw_sum(td->bytes_done) &&
955 !(td->io_ops->flags & FIO_NOIO))
958 if (!in_ramp_time(td) && should_check_rate(td)) {
959 if (check_min_rate(td, &comp_time)) {
960 if (exitall_on_terminate)
961 fio_terminate_threads(td->groupid);
962 td_verror(td, EIO, "check_min_rate");
966 if (!in_ramp_time(td) && td->o.latency_target)
967 lat_target_check(td);
969 if (td->o.thinktime) {
970 unsigned long long b;
972 b = ddir_rw_sum(td->io_blocks);
973 if (!(b % td->o.thinktime_blocks)) {
978 if (td->o.thinktime_spin)
979 usec_spin(td->o.thinktime_spin);
981 left = td->o.thinktime - td->o.thinktime_spin;
983 usec_sleep(td, left);
988 check_update_rusage(td);
990 if (td->trim_entries)
991 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
993 if (td->o.fill_device && td->error == ENOSPC) {
995 fio_mark_td_terminate(td);
1000 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1001 workqueue_flush(&td->io_wq);
1007 ret = io_u_queued_complete(td, i);
1008 if (td->o.fill_device && td->error == ENOSPC)
1012 if (should_fsync(td) && td->o.end_fsync) {
1013 td_set_runstate(td, TD_FSYNCING);
1015 for_each_file(td, f, i) {
1016 if (!fio_file_fsync(td, f))
1019 log_err("fio: end_fsync failed for file %s\n",
1024 cleanup_pending_aio(td);
1027 * stop job if we failed doing any IO
1029 if (!ddir_rw_sum(td->this_io_bytes))
1032 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1035 static void cleanup_io_u(struct thread_data *td)
1039 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1041 if (td->io_ops->io_u_free)
1042 td->io_ops->io_u_free(td, io_u);
1044 fio_memfree(io_u, sizeof(*io_u));
1049 io_u_rexit(&td->io_u_requeues);
1050 io_u_qexit(&td->io_u_freelist);
1051 io_u_qexit(&td->io_u_all);
1053 if (td->last_write_comp)
1054 sfree(td->last_write_comp);
1057 static int init_io_u(struct thread_data *td)
1060 unsigned int max_bs, min_write;
1061 int cl_align, i, max_units;
1062 int data_xfer = 1, err;
1065 max_units = td->o.iodepth;
1066 max_bs = td_max_bs(td);
1067 min_write = td->o.min_bs[DDIR_WRITE];
1068 td->orig_buffer_size = (unsigned long long) max_bs
1069 * (unsigned long long) max_units;
1071 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1075 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1076 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1077 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1080 log_err("fio: failed setting up IO queues\n");
1085 * if we may later need to do address alignment, then add any
1086 * possible adjustment here so that we don't cause a buffer
1087 * overflow later. this adjustment may be too much if we get
1088 * lucky and the allocator gives us an aligned address.
1090 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1091 (td->io_ops->flags & FIO_RAWIO))
1092 td->orig_buffer_size += page_mask + td->o.mem_align;
1094 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1097 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1098 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1101 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1102 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1106 if (data_xfer && allocate_io_mem(td))
1109 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1110 (td->io_ops->flags & FIO_RAWIO))
1111 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1113 p = td->orig_buffer;
1115 cl_align = os_cache_line_size();
1117 for (i = 0; i < max_units; i++) {
1123 ptr = fio_memalign(cl_align, sizeof(*io_u));
1125 log_err("fio: unable to allocate aligned memory\n");
1130 memset(io_u, 0, sizeof(*io_u));
1131 INIT_FLIST_HEAD(&io_u->verify_list);
1132 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1136 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1139 io_u_fill_buffer(td, io_u, min_write, max_bs);
1140 if (td_write(td) && td->o.verify_pattern_bytes) {
1142 * Fill the buffer with the pattern if we are
1143 * going to be doing writes.
1145 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1150 io_u->flags = IO_U_F_FREE;
1151 io_u_qpush(&td->io_u_freelist, io_u);
1154 * io_u never leaves this stack, used for iteration of all
1157 io_u_qpush(&td->io_u_all, io_u);
1159 if (td->io_ops->io_u_init) {
1160 int ret = td->io_ops->io_u_init(td, io_u);
1163 log_err("fio: failed to init engine data: %d\n", ret);
1171 if (td->o.verify != VERIFY_NONE) {
1172 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1173 if (!td->last_write_comp) {
1174 log_err("fio: failed to alloc write comp data\n");
1182 static int switch_ioscheduler(struct thread_data *td)
1184 char tmp[256], tmp2[128];
1188 if (td->io_ops->flags & FIO_DISKLESSIO)
1191 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1193 f = fopen(tmp, "r+");
1195 if (errno == ENOENT) {
1196 log_err("fio: os or kernel doesn't support IO scheduler"
1200 td_verror(td, errno, "fopen iosched");
1207 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1208 if (ferror(f) || ret != 1) {
1209 td_verror(td, errno, "fwrite");
1217 * Read back and check that the selected scheduler is now the default.
1219 memset(tmp, 0, sizeof(tmp));
1220 ret = fread(tmp, sizeof(tmp), 1, f);
1221 if (ferror(f) || ret < 0) {
1222 td_verror(td, errno, "fread");
1227 * either a list of io schedulers or "none\n" is expected.
1229 tmp[strlen(tmp) - 1] = '\0';
1232 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1233 if (!strstr(tmp, tmp2)) {
1234 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1235 td_verror(td, EINVAL, "iosched_switch");
1244 static int keep_running(struct thread_data *td)
1246 unsigned long long limit;
1250 if (td->o.time_based)
1256 if (exceeds_number_ios(td))
1260 limit = td->o.io_limit;
1264 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1268 * If the difference is less than the minimum IO size, we
1271 diff = limit - ddir_rw_sum(td->io_bytes);
1272 if (diff < td_max_bs(td))
1275 if (fio_files_done(td))
1284 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1286 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1290 str = malloc(newlen);
1291 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1293 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1296 log_err("fio: exec of cmd <%s> failed\n", str);
1303 * Dry run to compute correct state of numberio for verification.
1305 static uint64_t do_dry_run(struct thread_data *td)
1307 td_set_runstate(td, TD_RUNNING);
1309 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1310 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1314 if (td->terminate || td->done)
1317 io_u = get_io_u(td);
1321 io_u_set(io_u, IO_U_F_FLIGHT);
1324 if (ddir_rw(acct_ddir(io_u)))
1325 td->io_issues[acct_ddir(io_u)]++;
1326 if (ddir_rw(io_u->ddir)) {
1327 io_u_mark_depth(td, 1);
1328 td->ts.total_io_u[io_u->ddir]++;
1331 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1333 td->o.verify != VERIFY_NONE &&
1334 !td->o.experimental_verify)
1335 log_io_piece(td, io_u);
1337 ret = io_u_sync_complete(td, io_u);
1341 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1344 static void io_workqueue_fn(struct thread_data *td, struct io_u *io_u)
1346 const enum fio_ddir ddir = io_u->ddir;
1349 dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
1351 io_u_set(io_u, IO_U_F_NO_FILE_PUT);
1355 ret = td_io_queue(td, io_u);
1357 dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
1359 io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
1361 if (ret == FIO_Q_QUEUED)
1362 ret = io_u_queued_complete(td, 1);
1368 * Entry point for the thread based jobs. The process based jobs end up
1369 * here as well, after a little setup.
1371 static void *thread_main(void *data)
1373 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1374 struct thread_data *td = data;
1375 struct thread_options *o = &td->o;
1376 pthread_condattr_t attr;
1380 if (!o->use_thread) {
1386 fio_local_clock_init(o->use_thread);
1388 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1391 fio_server_send_start(td);
1393 INIT_FLIST_HEAD(&td->io_log_list);
1394 INIT_FLIST_HEAD(&td->io_hist_list);
1395 INIT_FLIST_HEAD(&td->verify_list);
1396 INIT_FLIST_HEAD(&td->trim_list);
1397 INIT_FLIST_HEAD(&td->next_rand_list);
1398 pthread_mutex_init(&td->io_u_lock, NULL);
1399 td->io_hist_tree = RB_ROOT;
1401 pthread_condattr_init(&attr);
1402 pthread_cond_init(&td->verify_cond, &attr);
1403 pthread_cond_init(&td->free_cond, &attr);
1405 td_set_runstate(td, TD_INITIALIZED);
1406 dprint(FD_MUTEX, "up startup_mutex\n");
1407 fio_mutex_up(startup_mutex);
1408 dprint(FD_MUTEX, "wait on td->mutex\n");
1409 fio_mutex_down(td->mutex);
1410 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1413 * A new gid requires privilege, so we need to do this before setting
1416 if (o->gid != -1U && setgid(o->gid)) {
1417 td_verror(td, errno, "setgid");
1420 if (o->uid != -1U && setuid(o->uid)) {
1421 td_verror(td, errno, "setuid");
1426 * If we have a gettimeofday() thread, make sure we exclude that
1427 * thread from this job
1430 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1433 * Set affinity first, in case it has an impact on the memory
1436 if (fio_option_is_set(o, cpumask)) {
1437 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1438 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1440 log_err("fio: no CPUs set\n");
1441 log_err("fio: Try increasing number of available CPUs\n");
1442 td_verror(td, EINVAL, "cpus_split");
1446 ret = fio_setaffinity(td->pid, o->cpumask);
1448 td_verror(td, errno, "cpu_set_affinity");
1453 #ifdef CONFIG_LIBNUMA
1454 /* numa node setup */
1455 if (fio_option_is_set(o, numa_cpunodes) ||
1456 fio_option_is_set(o, numa_memnodes)) {
1457 struct bitmask *mask;
1459 if (numa_available() < 0) {
1460 td_verror(td, errno, "Does not support NUMA API\n");
1464 if (fio_option_is_set(o, numa_cpunodes)) {
1465 mask = numa_parse_nodestring(o->numa_cpunodes);
1466 ret = numa_run_on_node_mask(mask);
1467 numa_free_nodemask(mask);
1469 td_verror(td, errno, \
1470 "numa_run_on_node_mask failed\n");
1475 if (fio_option_is_set(o, numa_memnodes)) {
1477 if (o->numa_memnodes)
1478 mask = numa_parse_nodestring(o->numa_memnodes);
1480 switch (o->numa_mem_mode) {
1481 case MPOL_INTERLEAVE:
1482 numa_set_interleave_mask(mask);
1485 numa_set_membind(mask);
1488 numa_set_localalloc();
1490 case MPOL_PREFERRED:
1491 numa_set_preferred(o->numa_mem_prefer_node);
1499 numa_free_nodemask(mask);
1505 if (fio_pin_memory(td))
1509 * May alter parameters that init_io_u() will use, so we need to
1518 if (o->verify_async && verify_async_init(td))
1521 if (fio_option_is_set(o, ioprio) ||
1522 fio_option_is_set(o, ioprio_class)) {
1523 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1525 td_verror(td, errno, "ioprio_set");
1530 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1534 if (nice(o->nice) == -1 && errno != 0) {
1535 td_verror(td, errno, "nice");
1539 if (o->ioscheduler && switch_ioscheduler(td))
1542 if (!o->create_serialize && setup_files(td))
1548 if (init_random_map(td))
1551 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1555 if (pre_read_files(td) < 0)
1559 if (td->flags & TD_F_COMPRESS_LOG)
1560 tp_init(&td->tp_data);
1562 fio_verify_init(td);
1564 if ((o->io_submit_mode == IO_MODE_OFFLOAD) &&
1565 workqueue_init(td, &td->io_wq, io_workqueue_fn, td->o.iodepth))
1568 fio_gettime(&td->epoch, NULL);
1569 fio_getrusage(&td->ru_start);
1570 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1571 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1573 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1574 o->ratemin[DDIR_TRIM]) {
1575 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1576 sizeof(td->bw_sample_time));
1577 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1578 sizeof(td->bw_sample_time));
1579 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1580 sizeof(td->bw_sample_time));
1584 while (keep_running(td)) {
1585 uint64_t verify_bytes;
1587 fio_gettime(&td->start, NULL);
1588 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1591 clear_io_state(td, 0);
1593 prune_io_piece_log(td);
1595 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1596 verify_bytes = do_dry_run(td);
1598 verify_bytes = do_io(td);
1603 * Make sure we've successfully updated the rusage stats
1604 * before waiting on the stat mutex. Otherwise we could have
1605 * the stat thread holding stat mutex and waiting for
1606 * the rusage_sem, which would never get upped because
1607 * this thread is waiting for the stat mutex.
1609 check_update_rusage(td);
1611 fio_mutex_down(stat_mutex);
1612 if (td_read(td) && td->io_bytes[DDIR_READ])
1613 update_runtime(td, elapsed_us, DDIR_READ);
1614 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1615 update_runtime(td, elapsed_us, DDIR_WRITE);
1616 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1617 update_runtime(td, elapsed_us, DDIR_TRIM);
1618 fio_gettime(&td->start, NULL);
1619 fio_mutex_up(stat_mutex);
1621 if (td->error || td->terminate)
1624 if (!o->do_verify ||
1625 o->verify == VERIFY_NONE ||
1626 (td->io_ops->flags & FIO_UNIDIR))
1629 clear_io_state(td, 0);
1631 fio_gettime(&td->start, NULL);
1633 do_verify(td, verify_bytes);
1636 * See comment further up for why this is done here.
1638 check_update_rusage(td);
1640 fio_mutex_down(stat_mutex);
1641 update_runtime(td, elapsed_us, DDIR_READ);
1642 fio_gettime(&td->start, NULL);
1643 fio_mutex_up(stat_mutex);
1645 if (td->error || td->terminate)
1649 update_rusage_stat(td);
1650 td->ts.total_run_time = mtime_since_now(&td->epoch);
1651 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1652 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1653 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1655 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1656 (td->o.verify != VERIFY_NONE && td_write(td)))
1657 verify_save_state(td->thread_number);
1659 fio_unpin_memory(td);
1661 fio_writeout_logs(td);
1663 if (o->io_submit_mode == IO_MODE_OFFLOAD)
1664 workqueue_exit(&td->io_wq);
1666 if (td->flags & TD_F_COMPRESS_LOG)
1667 tp_exit(&td->tp_data);
1669 if (o->exec_postrun)
1670 exec_string(o, o->exec_postrun, (const char *)"postrun");
1672 if (exitall_on_terminate)
1673 fio_terminate_threads(td->groupid);
1677 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1680 if (o->verify_async)
1681 verify_async_exit(td);
1683 close_and_free_files(td);
1686 cgroup_shutdown(td, &cgroup_mnt);
1687 verify_free_state(td);
1689 if (fio_option_is_set(o, cpumask)) {
1690 ret = fio_cpuset_exit(&o->cpumask);
1692 td_verror(td, ret, "fio_cpuset_exit");
1696 * do this very late, it will log file closing as well
1698 if (o->write_iolog_file)
1699 write_iolog_close(td);
1701 fio_mutex_remove(td->mutex);
1704 td_set_runstate(td, TD_EXITED);
1707 * Do this last after setting our runstate to exited, so we
1708 * know that the stat thread is signaled.
1710 check_update_rusage(td);
1712 return (void *) (uintptr_t) td->error;
1717 * We cannot pass the td data into a forked process, so attach the td and
1718 * pass it to the thread worker.
1720 static int fork_main(int shmid, int offset)
1722 struct thread_data *td;
1725 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1726 data = shmat(shmid, NULL, 0);
1727 if (data == (void *) -1) {
1735 * HP-UX inherits shm mappings?
1740 td = data + offset * sizeof(struct thread_data);
1741 ret = thread_main(td);
1743 return (int) (uintptr_t) ret;
1746 static void dump_td_info(struct thread_data *td)
1748 log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1749 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1750 (unsigned long) time_since_now(&td->terminate_time));
1754 * Run over the job map and reap the threads that have exited, if any.
1756 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1757 unsigned int *m_rate)
1759 struct thread_data *td;
1760 unsigned int cputhreads, realthreads, pending;
1764 * reap exited threads (TD_EXITED -> TD_REAPED)
1766 realthreads = pending = cputhreads = 0;
1767 for_each_td(td, i) {
1771 * ->io_ops is NULL for a thread that has closed its
1774 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1783 if (td->runstate == TD_REAPED)
1785 if (td->o.use_thread) {
1786 if (td->runstate == TD_EXITED) {
1787 td_set_runstate(td, TD_REAPED);
1794 if (td->runstate == TD_EXITED)
1798 * check if someone quit or got killed in an unusual way
1800 ret = waitpid(td->pid, &status, flags);
1802 if (errno == ECHILD) {
1803 log_err("fio: pid=%d disappeared %d\n",
1804 (int) td->pid, td->runstate);
1806 td_set_runstate(td, TD_REAPED);
1810 } else if (ret == td->pid) {
1811 if (WIFSIGNALED(status)) {
1812 int sig = WTERMSIG(status);
1814 if (sig != SIGTERM && sig != SIGUSR2)
1815 log_err("fio: pid=%d, got signal=%d\n",
1816 (int) td->pid, sig);
1818 td_set_runstate(td, TD_REAPED);
1821 if (WIFEXITED(status)) {
1822 if (WEXITSTATUS(status) && !td->error)
1823 td->error = WEXITSTATUS(status);
1825 td_set_runstate(td, TD_REAPED);
1831 * If the job is stuck, do a forceful timeout of it and
1834 if (td->terminate &&
1835 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1837 td_set_runstate(td, TD_REAPED);
1842 * thread is not dead, continue
1848 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1849 (*t_rate) -= ddir_rw_sum(td->o.rate);
1856 done_secs += mtime_since_now(&td->epoch) / 1000;
1857 profile_td_exit(td);
1860 if (*nr_running == cputhreads && !pending && realthreads)
1861 fio_terminate_threads(TERMINATE_ALL);
1864 static int __check_trigger_file(void)
1871 if (stat(trigger_file, &sb))
1874 if (unlink(trigger_file) < 0)
1875 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1881 static int trigger_timedout(void)
1883 if (trigger_timeout)
1884 return time_since_genesis() >= trigger_timeout;
1889 void exec_trigger(const char *cmd)
1898 log_err("fio: failed executing %s trigger\n", cmd);
1901 void check_trigger_file(void)
1903 if (__check_trigger_file() || trigger_timedout()) {
1905 fio_clients_send_trigger(trigger_remote_cmd);
1907 verify_save_state(IO_LIST_ALL);
1908 fio_terminate_threads(TERMINATE_ALL);
1909 exec_trigger(trigger_cmd);
1914 static int fio_verify_load_state(struct thread_data *td)
1918 if (!td->o.verify_state)
1925 ret = fio_server_get_verify_state(td->o.name,
1926 td->thread_number - 1, &data, &ver);
1928 verify_convert_assign_state(td, data, ver);
1930 ret = verify_load_state(td, "local");
1935 static void do_usleep(unsigned int usecs)
1937 check_for_running_stats();
1938 check_trigger_file();
1942 static int check_mount_writes(struct thread_data *td)
1947 if (!td_write(td) || td->o.allow_mounted_write)
1950 for_each_file(td, f, i) {
1951 if (f->filetype != FIO_TYPE_BD)
1953 if (device_is_mounted(f->file_name))
1959 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
1964 * Main function for kicking off and reaping jobs, as needed.
1966 static void run_threads(void)
1968 struct thread_data *td;
1969 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1972 if (fio_gtod_offload && fio_start_gtod_thread())
1975 fio_idle_prof_init();
1979 nr_thread = nr_process = 0;
1980 for_each_td(td, i) {
1981 if (check_mount_writes(td))
1983 if (td->o.use_thread)
1989 if (output_format & FIO_OUTPUT_NORMAL) {
1990 log_info("Starting ");
1992 log_info("%d thread%s", nr_thread,
1993 nr_thread > 1 ? "s" : "");
1997 log_info("%d process%s", nr_process,
1998 nr_process > 1 ? "es" : "");
2004 todo = thread_number;
2007 m_rate = t_rate = 0;
2009 for_each_td(td, i) {
2010 print_status_init(td->thread_number - 1);
2012 if (!td->o.create_serialize)
2015 if (fio_verify_load_state(td))
2019 * do file setup here so it happens sequentially,
2020 * we don't want X number of threads getting their
2021 * client data interspersed on disk
2023 if (setup_files(td)) {
2027 log_err("fio: pid=%d, err=%d/%s\n",
2028 (int) td->pid, td->error, td->verror);
2029 td_set_runstate(td, TD_REAPED);
2036 * for sharing to work, each job must always open
2037 * its own files. so close them, if we opened them
2040 for_each_file(td, f, j) {
2041 if (fio_file_open(f))
2042 td_io_close_file(td, f);
2047 /* start idle threads before io threads start to run */
2048 fio_idle_prof_start();
2053 struct thread_data *map[REAL_MAX_JOBS];
2054 struct timeval this_start;
2055 int this_jobs = 0, left;
2058 * create threads (TD_NOT_CREATED -> TD_CREATED)
2060 for_each_td(td, i) {
2061 if (td->runstate != TD_NOT_CREATED)
2065 * never got a chance to start, killed by other
2066 * thread for some reason
2068 if (td->terminate) {
2073 if (td->o.start_delay) {
2074 spent = utime_since_genesis();
2076 if (td->o.start_delay > spent)
2080 if (td->o.stonewall && (nr_started || nr_running)) {
2081 dprint(FD_PROCESS, "%s: stonewall wait\n",
2088 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2089 td->update_rusage = 0;
2092 * Set state to created. Thread will transition
2093 * to TD_INITIALIZED when it's done setting up.
2095 td_set_runstate(td, TD_CREATED);
2096 map[this_jobs++] = td;
2099 if (td->o.use_thread) {
2102 dprint(FD_PROCESS, "will pthread_create\n");
2103 ret = pthread_create(&td->thread, NULL,
2106 log_err("pthread_create: %s\n",
2111 ret = pthread_detach(td->thread);
2113 log_err("pthread_detach: %s",
2117 dprint(FD_PROCESS, "will fork\n");
2120 int ret = fork_main(shm_id, i);
2123 } else if (i == fio_debug_jobno)
2124 *fio_debug_jobp = pid;
2126 dprint(FD_MUTEX, "wait on startup_mutex\n");
2127 if (fio_mutex_down_timeout(startup_mutex, 10)) {
2128 log_err("fio: job startup hung? exiting.\n");
2129 fio_terminate_threads(TERMINATE_ALL);
2134 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2138 * Wait for the started threads to transition to
2141 fio_gettime(&this_start, NULL);
2143 while (left && !fio_abort) {
2144 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2149 for (i = 0; i < this_jobs; i++) {
2153 if (td->runstate == TD_INITIALIZED) {
2156 } else if (td->runstate >= TD_EXITED) {
2160 nr_running++; /* work-around... */
2166 log_err("fio: %d job%s failed to start\n", left,
2167 left > 1 ? "s" : "");
2168 for (i = 0; i < this_jobs; i++) {
2172 kill(td->pid, SIGTERM);
2178 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2180 for_each_td(td, i) {
2181 if (td->runstate != TD_INITIALIZED)
2184 if (in_ramp_time(td))
2185 td_set_runstate(td, TD_RAMP);
2187 td_set_runstate(td, TD_RUNNING);
2190 m_rate += ddir_rw_sum(td->o.ratemin);
2191 t_rate += ddir_rw_sum(td->o.rate);
2193 fio_mutex_up(td->mutex);
2196 reap_threads(&nr_running, &t_rate, &m_rate);
2202 while (nr_running) {
2203 reap_threads(&nr_running, &t_rate, &m_rate);
2207 fio_idle_prof_stop();
2212 static void wait_for_helper_thread_exit(void)
2217 pthread_cond_signal(&helper_cond);
2218 pthread_join(helper_thread, &ret);
2221 static void free_disk_util(void)
2223 disk_util_prune_entries();
2225 pthread_cond_destroy(&helper_cond);
2228 static void *helper_thread_main(void *data)
2232 fio_mutex_up(startup_mutex);
2235 uint64_t sec = DISK_UTIL_MSEC / 1000;
2236 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2240 gettimeofday(&tv, NULL);
2241 ts.tv_sec = tv.tv_sec + sec;
2242 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2244 if (ts.tv_nsec >= 1000000000ULL) {
2245 ts.tv_nsec -= 1000000000ULL;
2249 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2251 ret = update_io_ticks();
2253 if (helper_do_stat) {
2255 __show_running_run_stats();
2259 print_thread_status();
2265 static int create_helper_thread(void)
2271 pthread_cond_init(&helper_cond, NULL);
2272 pthread_mutex_init(&helper_lock, NULL);
2274 ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2276 log_err("Can't create helper thread: %s\n", strerror(ret));
2280 dprint(FD_MUTEX, "wait on startup_mutex\n");
2281 fio_mutex_down(startup_mutex);
2282 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2286 int fio_backend(void)
2288 struct thread_data *td;
2292 if (load_profile(exec_profile))
2295 exec_profile = NULL;
2301 struct log_params p = {
2302 .log_type = IO_LOG_TYPE_BW,
2305 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2306 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2307 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2310 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2311 if (startup_mutex == NULL)
2316 create_helper_thread();
2318 cgroup_list = smalloc(sizeof(*cgroup_list));
2319 INIT_FLIST_HEAD(cgroup_list);
2323 wait_for_helper_thread_exit();
2328 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2329 struct io_log *log = agg_io_log[i];
2337 for_each_td(td, i) {
2338 fio_options_free(td);
2339 if (td->rusage_sem) {
2340 fio_mutex_remove(td->rusage_sem);
2341 td->rusage_sem = NULL;
2346 cgroup_kill(cgroup_list);
2350 fio_mutex_remove(startup_mutex);