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
40 #ifndef FIO_NO_HAVE_SHM_H
53 #include "lib/getrusage.h"
57 #include "workqueue.h"
58 #include "lib/mountcheck.h"
60 static pthread_t helper_thread;
61 static pthread_mutex_t helper_lock;
62 pthread_cond_t helper_cond;
63 int helper_do_stat = 0;
65 static struct fio_mutex *startup_mutex;
66 static struct flist_head *cgroup_list;
67 static char *cgroup_mnt;
68 static int exit_value;
69 static volatile int fio_abort;
70 static unsigned int nr_process = 0;
71 static unsigned int nr_thread = 0;
73 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
76 unsigned int thread_number = 0;
77 unsigned int stat_number = 0;
80 unsigned long done_secs = 0;
81 volatile int helper_exit = 0;
83 #define PAGE_ALIGN(buf) \
84 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
86 #define JOB_START_TIMEOUT (5 * 1000)
88 static void sig_int(int sig)
92 fio_server_got_signal(sig);
94 log_info("\nfio: terminating on signal %d\n", sig);
99 fio_terminate_threads(TERMINATE_ALL);
103 static void sig_show_status(int sig)
105 show_running_run_stats();
108 static void set_sig_handlers(void)
110 struct sigaction act;
112 memset(&act, 0, sizeof(act));
113 act.sa_handler = sig_int;
114 act.sa_flags = SA_RESTART;
115 sigaction(SIGINT, &act, NULL);
117 memset(&act, 0, sizeof(act));
118 act.sa_handler = sig_int;
119 act.sa_flags = SA_RESTART;
120 sigaction(SIGTERM, &act, NULL);
122 /* Windows uses SIGBREAK as a quit signal from other applications */
124 memset(&act, 0, sizeof(act));
125 act.sa_handler = sig_int;
126 act.sa_flags = SA_RESTART;
127 sigaction(SIGBREAK, &act, NULL);
130 memset(&act, 0, sizeof(act));
131 act.sa_handler = sig_show_status;
132 act.sa_flags = SA_RESTART;
133 sigaction(SIGUSR1, &act, NULL);
136 memset(&act, 0, sizeof(act));
137 act.sa_handler = sig_int;
138 act.sa_flags = SA_RESTART;
139 sigaction(SIGPIPE, &act, NULL);
144 * Check if we are above the minimum rate given.
146 static int __check_min_rate(struct thread_data *td, struct timeval *now,
149 unsigned long long bytes = 0;
150 unsigned long iops = 0;
153 unsigned int ratemin = 0;
154 unsigned int rate_iops = 0;
155 unsigned int rate_iops_min = 0;
157 assert(ddir_rw(ddir));
159 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
163 * allow a 2 second settle period in the beginning
165 if (mtime_since(&td->start, now) < 2000)
168 iops += td->this_io_blocks[ddir];
169 bytes += td->this_io_bytes[ddir];
170 ratemin += td->o.ratemin[ddir];
171 rate_iops += td->o.rate_iops[ddir];
172 rate_iops_min += td->o.rate_iops_min[ddir];
175 * if rate blocks is set, sample is running
177 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
178 spent = mtime_since(&td->lastrate[ddir], now);
179 if (spent < td->o.ratecycle)
182 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
184 * check bandwidth specified rate
186 if (bytes < td->rate_bytes[ddir]) {
187 log_err("%s: min rate %u not met\n", td->o.name,
192 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
196 if (rate < ratemin ||
197 bytes < td->rate_bytes[ddir]) {
198 log_err("%s: min rate %u not met, got"
199 " %luKB/sec\n", td->o.name,
206 * checks iops specified rate
208 if (iops < rate_iops) {
209 log_err("%s: min iops rate %u not met\n",
210 td->o.name, rate_iops);
214 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
218 if (rate < rate_iops_min ||
219 iops < td->rate_blocks[ddir]) {
220 log_err("%s: min iops rate %u not met,"
221 " got %lu\n", td->o.name,
222 rate_iops_min, rate);
229 td->rate_bytes[ddir] = bytes;
230 td->rate_blocks[ddir] = iops;
231 memcpy(&td->lastrate[ddir], now, sizeof(*now));
235 static int check_min_rate(struct thread_data *td, struct timeval *now)
239 if (td->bytes_done[DDIR_READ])
240 ret |= __check_min_rate(td, now, DDIR_READ);
241 if (td->bytes_done[DDIR_WRITE])
242 ret |= __check_min_rate(td, now, DDIR_WRITE);
243 if (td->bytes_done[DDIR_TRIM])
244 ret |= __check_min_rate(td, now, DDIR_TRIM);
250 * When job exits, we can cancel the in-flight IO if we are using async
251 * io. Attempt to do so.
253 static void cleanup_pending_aio(struct thread_data *td)
258 * get immediately available events, if any
260 r = io_u_queued_complete(td, 0);
265 * now cancel remaining active events
267 if (td->io_ops->cancel) {
271 io_u_qiter(&td->io_u_all, io_u, i) {
272 if (io_u->flags & IO_U_F_FLIGHT) {
273 r = td->io_ops->cancel(td, io_u);
281 r = io_u_queued_complete(td, td->cur_depth);
285 * Helper to handle the final sync of a file. Works just like the normal
286 * io path, just does everything sync.
288 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
290 struct io_u *io_u = __get_io_u(td);
296 io_u->ddir = DDIR_SYNC;
299 if (td_io_prep(td, io_u)) {
305 ret = td_io_queue(td, io_u);
307 td_verror(td, io_u->error, "td_io_queue");
310 } else if (ret == FIO_Q_QUEUED) {
311 if (io_u_queued_complete(td, 1) < 0)
313 } else if (ret == FIO_Q_COMPLETED) {
315 td_verror(td, io_u->error, "td_io_queue");
319 if (io_u_sync_complete(td, io_u) < 0)
321 } else if (ret == FIO_Q_BUSY) {
322 if (td_io_commit(td))
330 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
334 if (fio_file_open(f))
335 return fio_io_sync(td, f);
337 if (td_io_open_file(td, f))
340 ret = fio_io_sync(td, f);
341 td_io_close_file(td, f);
345 static inline void __update_tv_cache(struct thread_data *td)
347 fio_gettime(&td->tv_cache, NULL);
350 static inline void update_tv_cache(struct thread_data *td)
352 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
353 __update_tv_cache(td);
356 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
358 if (in_ramp_time(td))
362 if (utime_since(&td->epoch, t) >= td->o.timeout)
369 * We need to update the runtime consistently in ms, but keep a running
370 * tally of the current elapsed time in microseconds for sub millisecond
373 static inline void update_runtime(struct thread_data *td,
374 unsigned long long *elapsed_us,
375 const enum fio_ddir ddir)
377 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
380 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
381 elapsed_us[ddir] += utime_since_now(&td->start);
382 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
385 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
390 if (ret < 0 || td->error) {
392 enum error_type_bit eb;
397 eb = td_error_type(ddir, err);
398 if (!(td->o.continue_on_error & (1 << eb)))
401 if (td_non_fatal_error(td, eb, err)) {
403 * Continue with the I/Os in case of
406 update_error_count(td, err);
410 } else if (td->o.fill_device && err == ENOSPC) {
412 * We expect to hit this error if
413 * fill_device option is set.
416 fio_mark_td_terminate(td);
420 * Stop the I/O in case of a fatal
423 update_error_count(td, err);
431 static void check_update_rusage(struct thread_data *td)
433 if (td->update_rusage) {
434 td->update_rusage = 0;
435 update_rusage_stat(td);
436 fio_mutex_up(td->rusage_sem);
440 static int wait_for_completions(struct thread_data *td, struct timeval *time)
442 const int full = queue_full(td);
447 * if the queue is full, we MUST reap at least 1 event
449 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
450 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
453 if (time && (__should_check_rate(td, DDIR_READ) ||
454 __should_check_rate(td, DDIR_WRITE) ||
455 __should_check_rate(td, DDIR_TRIM)))
456 fio_gettime(time, NULL);
459 ret = io_u_queued_complete(td, min_evts);
462 } while (full && (td->cur_depth > td->o.iodepth_low));
467 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
468 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
469 struct timeval *comp_time)
474 case FIO_Q_COMPLETED:
477 clear_io_u(td, io_u);
478 } else if (io_u->resid) {
479 int bytes = io_u->xfer_buflen - io_u->resid;
480 struct fio_file *f = io_u->file;
483 *bytes_issued += bytes;
486 trim_io_piece(td, io_u);
493 unlog_io_piece(td, io_u);
494 td_verror(td, EIO, "full resid");
499 io_u->xfer_buflen = io_u->resid;
500 io_u->xfer_buf += bytes;
501 io_u->offset += bytes;
503 if (ddir_rw(io_u->ddir))
504 td->ts.short_io_u[io_u->ddir]++;
507 if (io_u->offset == f->real_file_size)
510 requeue_io_u(td, &io_u);
513 if (comp_time && (__should_check_rate(td, DDIR_READ) ||
514 __should_check_rate(td, DDIR_WRITE) ||
515 __should_check_rate(td, DDIR_TRIM)))
516 fio_gettime(comp_time, NULL);
518 *ret = io_u_sync_complete(td, io_u);
525 * if the engine doesn't have a commit hook,
526 * the io_u is really queued. if it does have such
527 * a hook, it has to call io_u_queued() itself.
529 if (td->io_ops->commit == NULL)
530 io_u_queued(td, io_u);
532 *bytes_issued += io_u->xfer_buflen;
536 unlog_io_piece(td, io_u);
537 requeue_io_u(td, &io_u);
538 ret2 = td_io_commit(td);
544 td_verror(td, -(*ret), "td_io_queue");
548 if (break_on_this_error(td, ddir, ret))
554 static inline int io_in_polling(struct thread_data *td)
556 return !td->o.iodepth_batch_complete_min &&
557 !td->o.iodepth_batch_complete_max;
561 * The main verify engine. Runs over the writes we previously submitted,
562 * reads the blocks back in, and checks the crc/md5 of the data.
564 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
571 dprint(FD_VERIFY, "starting loop\n");
574 * sync io first and invalidate cache, to make sure we really
577 for_each_file(td, f, i) {
578 if (!fio_file_open(f))
580 if (fio_io_sync(td, f))
582 if (file_invalidate_cache(td, f))
586 check_update_rusage(td);
591 td_set_runstate(td, TD_VERIFYING);
594 while (!td->terminate) {
599 check_update_rusage(td);
601 if (runtime_exceeded(td, &td->tv_cache)) {
602 __update_tv_cache(td);
603 if (runtime_exceeded(td, &td->tv_cache)) {
604 fio_mark_td_terminate(td);
609 if (flow_threshold_exceeded(td))
612 if (!td->o.experimental_verify) {
613 io_u = __get_io_u(td);
617 if (get_next_verify(td, io_u)) {
622 if (td_io_prep(td, io_u)) {
627 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
630 while ((io_u = get_io_u(td)) != NULL) {
638 * We are only interested in the places where
639 * we wrote or trimmed IOs. Turn those into
640 * reads for verification purposes.
642 if (io_u->ddir == DDIR_READ) {
644 * Pretend we issued it for rwmix
647 td->io_issues[DDIR_READ]++;
650 } else if (io_u->ddir == DDIR_TRIM) {
651 io_u->ddir = DDIR_READ;
652 io_u_set(io_u, IO_U_F_TRIMMED);
654 } else if (io_u->ddir == DDIR_WRITE) {
655 io_u->ddir = DDIR_READ;
667 if (verify_state_should_stop(td, io_u)) {
672 if (td->o.verify_async)
673 io_u->end_io = verify_io_u_async;
675 io_u->end_io = verify_io_u;
678 if (!td->o.disable_slat)
679 fio_gettime(&io_u->start_time, NULL);
681 ret = td_io_queue(td, io_u);
683 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
687 * if we can queue more, do so. but check if there are
688 * completed io_u's first. Note that we can get BUSY even
689 * without IO queued, if the system is resource starved.
692 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
693 if (full || io_in_polling(td))
694 ret = wait_for_completions(td, NULL);
700 check_update_rusage(td);
703 min_events = td->cur_depth;
706 ret = io_u_queued_complete(td, min_events);
708 cleanup_pending_aio(td);
710 td_set_runstate(td, TD_RUNNING);
712 dprint(FD_VERIFY, "exiting loop\n");
715 static unsigned int exceeds_number_ios(struct thread_data *td)
717 unsigned long long number_ios;
719 if (!td->o.number_ios)
722 number_ios = ddir_rw_sum(td->io_blocks);
723 number_ios += td->io_u_queued + td->io_u_in_flight;
725 return number_ios >= (td->o.number_ios * td->loops);
728 static int io_issue_bytes_exceeded(struct thread_data *td)
730 unsigned long long bytes, limit;
733 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
734 else if (td_write(td))
735 bytes = td->io_issue_bytes[DDIR_WRITE];
736 else if (td_read(td))
737 bytes = td->io_issue_bytes[DDIR_READ];
739 bytes = td->io_issue_bytes[DDIR_TRIM];
742 limit = td->o.io_limit;
747 return bytes >= limit || exceeds_number_ios(td);
750 static int io_complete_bytes_exceeded(struct thread_data *td)
752 unsigned long long bytes, limit;
755 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
756 else if (td_write(td))
757 bytes = td->this_io_bytes[DDIR_WRITE];
758 else if (td_read(td))
759 bytes = td->this_io_bytes[DDIR_READ];
761 bytes = td->this_io_bytes[DDIR_TRIM];
764 limit = td->o.io_limit;
769 return bytes >= limit || exceeds_number_ios(td);
773 * used to calculate the next io time for rate control
776 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
778 uint64_t secs, remainder, bps, bytes;
780 assert(!(td->flags & TD_F_CHILD));
781 bytes = td->rate_io_issue_bytes[ddir];
782 bps = td->rate_bps[ddir];
785 remainder = bytes % bps;
786 return remainder * 1000000 / bps + secs * 1000000;
792 * Main IO worker function. It retrieves io_u's to process and queues
793 * and reaps them, checking for rate and errors along the way.
795 * Returns number of bytes written and trimmed.
797 static uint64_t do_io(struct thread_data *td)
801 uint64_t total_bytes, bytes_issued = 0;
803 if (in_ramp_time(td))
804 td_set_runstate(td, TD_RAMP);
806 td_set_runstate(td, TD_RUNNING);
810 total_bytes = td->o.size;
812 * Allow random overwrite workloads to write up to io_limit
813 * before starting verification phase as 'size' doesn't apply.
815 if (td_write(td) && td_random(td) && td->o.norandommap)
816 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
818 * If verify_backlog is enabled, we'll run the verify in this
819 * handler as well. For that case, we may need up to twice the
822 if (td->o.verify != VERIFY_NONE &&
823 (td_write(td) && td->o.verify_backlog))
824 total_bytes += td->o.size;
826 /* In trimwrite mode, each byte is trimmed and then written, so
827 * allow total_bytes to be twice as big */
828 if (td_trimwrite(td))
829 total_bytes += td->total_io_size;
831 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
832 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
834 struct timeval comp_time;
839 check_update_rusage(td);
841 if (td->terminate || td->done)
846 if (runtime_exceeded(td, &td->tv_cache)) {
847 __update_tv_cache(td);
848 if (runtime_exceeded(td, &td->tv_cache)) {
849 fio_mark_td_terminate(td);
854 if (flow_threshold_exceeded(td))
857 if (bytes_issued >= total_bytes)
861 if (IS_ERR_OR_NULL(io_u)) {
862 int err = PTR_ERR(io_u);
869 if (td->o.latency_target)
877 * Add verification end_io handler if:
878 * - Asked to verify (!td_rw(td))
879 * - Or the io_u is from our verify list (mixed write/ver)
881 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
882 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
884 if (!td->o.verify_pattern_bytes) {
885 io_u->rand_seed = __rand(&td->verify_state);
886 if (sizeof(int) != sizeof(long *))
887 io_u->rand_seed *= __rand(&td->verify_state);
890 if (verify_state_should_stop(td, io_u)) {
895 if (td->o.verify_async)
896 io_u->end_io = verify_io_u_async;
898 io_u->end_io = verify_io_u;
899 td_set_runstate(td, TD_VERIFYING);
900 } else if (in_ramp_time(td))
901 td_set_runstate(td, TD_RAMP);
903 td_set_runstate(td, TD_RUNNING);
906 * Always log IO before it's issued, so we know the specific
907 * order of it. The logged unit will track when the IO has
910 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
912 td->o.verify != VERIFY_NONE &&
913 !td->o.experimental_verify)
914 log_io_piece(td, io_u);
916 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
919 ret = workqueue_enqueue(&td->io_wq, io_u);
921 if (should_check_rate(td))
922 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
925 ret = td_io_queue(td, io_u);
927 if (should_check_rate(td))
928 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
930 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
934 * See if we need to complete some commands. Note that
935 * we can get BUSY even without IO queued, if the
936 * system is resource starved.
939 full = queue_full(td) ||
940 (ret == FIO_Q_BUSY && td->cur_depth);
941 if (full || io_in_polling(td))
942 ret = wait_for_completions(td, &comp_time);
946 if (!ddir_rw_sum(td->bytes_done) &&
947 !(td->io_ops->flags & FIO_NOIO))
950 if (!in_ramp_time(td) && should_check_rate(td)) {
951 if (check_min_rate(td, &comp_time)) {
952 if (exitall_on_terminate)
953 fio_terminate_threads(td->groupid);
954 td_verror(td, EIO, "check_min_rate");
958 if (!in_ramp_time(td) && td->o.latency_target)
959 lat_target_check(td);
961 if (td->o.thinktime) {
962 unsigned long long b;
964 b = ddir_rw_sum(td->io_blocks);
965 if (!(b % td->o.thinktime_blocks)) {
970 if (td->o.thinktime_spin)
971 usec_spin(td->o.thinktime_spin);
973 left = td->o.thinktime - td->o.thinktime_spin;
975 usec_sleep(td, left);
980 check_update_rusage(td);
982 if (td->trim_entries)
983 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
985 if (td->o.fill_device && td->error == ENOSPC) {
987 fio_mark_td_terminate(td);
992 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
993 workqueue_flush(&td->io_wq);
999 ret = io_u_queued_complete(td, i);
1000 if (td->o.fill_device && td->error == ENOSPC)
1004 if (should_fsync(td) && td->o.end_fsync) {
1005 td_set_runstate(td, TD_FSYNCING);
1007 for_each_file(td, f, i) {
1008 if (!fio_file_fsync(td, f))
1011 log_err("fio: end_fsync failed for file %s\n",
1016 cleanup_pending_aio(td);
1019 * stop job if we failed doing any IO
1021 if (!ddir_rw_sum(td->this_io_bytes))
1024 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1027 static void cleanup_io_u(struct thread_data *td)
1031 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1033 if (td->io_ops->io_u_free)
1034 td->io_ops->io_u_free(td, io_u);
1036 fio_memfree(io_u, sizeof(*io_u));
1041 io_u_rexit(&td->io_u_requeues);
1042 io_u_qexit(&td->io_u_freelist);
1043 io_u_qexit(&td->io_u_all);
1045 if (td->last_write_comp)
1046 sfree(td->last_write_comp);
1049 static int init_io_u(struct thread_data *td)
1052 unsigned int max_bs, min_write;
1053 int cl_align, i, max_units;
1054 int data_xfer = 1, err;
1057 max_units = td->o.iodepth;
1058 max_bs = td_max_bs(td);
1059 min_write = td->o.min_bs[DDIR_WRITE];
1060 td->orig_buffer_size = (unsigned long long) max_bs
1061 * (unsigned long long) max_units;
1063 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1067 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1068 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1069 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1072 log_err("fio: failed setting up IO queues\n");
1077 * if we may later need to do address alignment, then add any
1078 * possible adjustment here so that we don't cause a buffer
1079 * overflow later. this adjustment may be too much if we get
1080 * lucky and the allocator gives us an aligned address.
1082 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1083 (td->io_ops->flags & FIO_RAWIO))
1084 td->orig_buffer_size += page_mask + td->o.mem_align;
1086 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1089 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1090 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1093 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1094 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1098 if (data_xfer && allocate_io_mem(td))
1101 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1102 (td->io_ops->flags & FIO_RAWIO))
1103 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1105 p = td->orig_buffer;
1107 cl_align = os_cache_line_size();
1109 for (i = 0; i < max_units; i++) {
1115 ptr = fio_memalign(cl_align, sizeof(*io_u));
1117 log_err("fio: unable to allocate aligned memory\n");
1122 memset(io_u, 0, sizeof(*io_u));
1123 INIT_FLIST_HEAD(&io_u->verify_list);
1124 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1128 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1131 io_u_fill_buffer(td, io_u, min_write, max_bs);
1132 if (td_write(td) && td->o.verify_pattern_bytes) {
1134 * Fill the buffer with the pattern if we are
1135 * going to be doing writes.
1137 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1142 io_u->flags = IO_U_F_FREE;
1143 io_u_qpush(&td->io_u_freelist, io_u);
1146 * io_u never leaves this stack, used for iteration of all
1149 io_u_qpush(&td->io_u_all, io_u);
1151 if (td->io_ops->io_u_init) {
1152 int ret = td->io_ops->io_u_init(td, io_u);
1155 log_err("fio: failed to init engine data: %d\n", ret);
1163 if (td->o.verify != VERIFY_NONE) {
1164 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1165 if (!td->last_write_comp) {
1166 log_err("fio: failed to alloc write comp data\n");
1174 static int switch_ioscheduler(struct thread_data *td)
1176 char tmp[256], tmp2[128];
1180 if (td->io_ops->flags & FIO_DISKLESSIO)
1183 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1185 f = fopen(tmp, "r+");
1187 if (errno == ENOENT) {
1188 log_err("fio: os or kernel doesn't support IO scheduler"
1192 td_verror(td, errno, "fopen iosched");
1199 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1200 if (ferror(f) || ret != 1) {
1201 td_verror(td, errno, "fwrite");
1209 * Read back and check that the selected scheduler is now the default.
1211 memset(tmp, 0, sizeof(tmp));
1212 ret = fread(tmp, sizeof(tmp), 1, f);
1213 if (ferror(f) || ret < 0) {
1214 td_verror(td, errno, "fread");
1219 * either a list of io schedulers or "none\n" is expected.
1221 tmp[strlen(tmp) - 1] = '\0';
1224 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1225 if (!strstr(tmp, tmp2)) {
1226 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1227 td_verror(td, EINVAL, "iosched_switch");
1236 static int keep_running(struct thread_data *td)
1238 unsigned long long limit;
1242 if (td->o.time_based)
1248 if (exceeds_number_ios(td))
1252 limit = td->o.io_limit;
1256 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1260 * If the difference is less than the minimum IO size, we
1263 diff = limit - ddir_rw_sum(td->io_bytes);
1264 if (diff < td_max_bs(td))
1267 if (fio_files_done(td))
1276 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1278 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1282 str = malloc(newlen);
1283 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1285 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1288 log_err("fio: exec of cmd <%s> failed\n", str);
1295 * Dry run to compute correct state of numberio for verification.
1297 static uint64_t do_dry_run(struct thread_data *td)
1299 td_set_runstate(td, TD_RUNNING);
1301 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1302 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1306 if (td->terminate || td->done)
1309 io_u = get_io_u(td);
1313 io_u_set(io_u, IO_U_F_FLIGHT);
1316 if (ddir_rw(acct_ddir(io_u)))
1317 td->io_issues[acct_ddir(io_u)]++;
1318 if (ddir_rw(io_u->ddir)) {
1319 io_u_mark_depth(td, 1);
1320 td->ts.total_io_u[io_u->ddir]++;
1323 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1325 td->o.verify != VERIFY_NONE &&
1326 !td->o.experimental_verify)
1327 log_io_piece(td, io_u);
1329 ret = io_u_sync_complete(td, io_u);
1333 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1336 static void io_workqueue_fn(struct thread_data *td, struct io_u *io_u)
1338 const enum fio_ddir ddir = io_u->ddir;
1341 dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
1343 io_u_set(io_u, IO_U_F_NO_FILE_PUT);
1347 ret = td_io_queue(td, io_u);
1349 dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
1351 io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
1353 if (ret == FIO_Q_QUEUED)
1354 ret = io_u_queued_complete(td, 1);
1360 * Entry point for the thread based jobs. The process based jobs end up
1361 * here as well, after a little setup.
1363 static void *thread_main(void *data)
1365 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1366 struct thread_data *td = data;
1367 struct thread_options *o = &td->o;
1368 pthread_condattr_t attr;
1372 if (!o->use_thread) {
1378 fio_local_clock_init(o->use_thread);
1380 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1383 fio_server_send_start(td);
1385 INIT_FLIST_HEAD(&td->io_log_list);
1386 INIT_FLIST_HEAD(&td->io_hist_list);
1387 INIT_FLIST_HEAD(&td->verify_list);
1388 INIT_FLIST_HEAD(&td->trim_list);
1389 INIT_FLIST_HEAD(&td->next_rand_list);
1390 pthread_mutex_init(&td->io_u_lock, NULL);
1391 td->io_hist_tree = RB_ROOT;
1393 pthread_condattr_init(&attr);
1394 pthread_cond_init(&td->verify_cond, &attr);
1395 pthread_cond_init(&td->free_cond, &attr);
1397 td_set_runstate(td, TD_INITIALIZED);
1398 dprint(FD_MUTEX, "up startup_mutex\n");
1399 fio_mutex_up(startup_mutex);
1400 dprint(FD_MUTEX, "wait on td->mutex\n");
1401 fio_mutex_down(td->mutex);
1402 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1405 * A new gid requires privilege, so we need to do this before setting
1408 if (o->gid != -1U && setgid(o->gid)) {
1409 td_verror(td, errno, "setgid");
1412 if (o->uid != -1U && setuid(o->uid)) {
1413 td_verror(td, errno, "setuid");
1418 * If we have a gettimeofday() thread, make sure we exclude that
1419 * thread from this job
1422 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1425 * Set affinity first, in case it has an impact on the memory
1428 if (fio_option_is_set(o, cpumask)) {
1429 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1430 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1432 log_err("fio: no CPUs set\n");
1433 log_err("fio: Try increasing number of available CPUs\n");
1434 td_verror(td, EINVAL, "cpus_split");
1438 ret = fio_setaffinity(td->pid, o->cpumask);
1440 td_verror(td, errno, "cpu_set_affinity");
1445 #ifdef CONFIG_LIBNUMA
1446 /* numa node setup */
1447 if (fio_option_is_set(o, numa_cpunodes) ||
1448 fio_option_is_set(o, numa_memnodes)) {
1449 struct bitmask *mask;
1451 if (numa_available() < 0) {
1452 td_verror(td, errno, "Does not support NUMA API\n");
1456 if (fio_option_is_set(o, numa_cpunodes)) {
1457 mask = numa_parse_nodestring(o->numa_cpunodes);
1458 ret = numa_run_on_node_mask(mask);
1459 numa_free_nodemask(mask);
1461 td_verror(td, errno, \
1462 "numa_run_on_node_mask failed\n");
1467 if (fio_option_is_set(o, numa_memnodes)) {
1469 if (o->numa_memnodes)
1470 mask = numa_parse_nodestring(o->numa_memnodes);
1472 switch (o->numa_mem_mode) {
1473 case MPOL_INTERLEAVE:
1474 numa_set_interleave_mask(mask);
1477 numa_set_membind(mask);
1480 numa_set_localalloc();
1482 case MPOL_PREFERRED:
1483 numa_set_preferred(o->numa_mem_prefer_node);
1491 numa_free_nodemask(mask);
1497 if (fio_pin_memory(td))
1501 * May alter parameters that init_io_u() will use, so we need to
1510 if (o->verify_async && verify_async_init(td))
1513 if (fio_option_is_set(o, ioprio) ||
1514 fio_option_is_set(o, ioprio_class)) {
1515 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1517 td_verror(td, errno, "ioprio_set");
1522 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1526 if (nice(o->nice) == -1 && errno != 0) {
1527 td_verror(td, errno, "nice");
1531 if (o->ioscheduler && switch_ioscheduler(td))
1534 if (!o->create_serialize && setup_files(td))
1540 if (init_random_map(td))
1543 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1547 if (pre_read_files(td) < 0)
1551 if (td->flags & TD_F_COMPRESS_LOG)
1552 tp_init(&td->tp_data);
1554 fio_verify_init(td);
1556 if ((o->io_submit_mode == IO_MODE_OFFLOAD) &&
1557 workqueue_init(td, &td->io_wq, io_workqueue_fn, td->o.iodepth))
1560 fio_gettime(&td->epoch, NULL);
1561 fio_getrusage(&td->ru_start);
1563 while (keep_running(td)) {
1564 uint64_t verify_bytes;
1566 fio_gettime(&td->start, NULL);
1567 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1568 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1569 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1571 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1572 o->ratemin[DDIR_TRIM]) {
1573 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1574 sizeof(td->bw_sample_time));
1575 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1576 sizeof(td->bw_sample_time));
1577 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1578 sizeof(td->bw_sample_time));
1584 prune_io_piece_log(td);
1586 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1587 verify_bytes = do_dry_run(td);
1589 verify_bytes = do_io(td);
1594 * Make sure we've successfully updated the rusage stats
1595 * before waiting on the stat mutex. Otherwise we could have
1596 * the stat thread holding stat mutex and waiting for
1597 * the rusage_sem, which would never get upped because
1598 * this thread is waiting for the stat mutex.
1600 check_update_rusage(td);
1602 fio_mutex_down(stat_mutex);
1603 if (td_read(td) && td->io_bytes[DDIR_READ])
1604 update_runtime(td, elapsed_us, DDIR_READ);
1605 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1606 update_runtime(td, elapsed_us, DDIR_WRITE);
1607 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1608 update_runtime(td, elapsed_us, DDIR_TRIM);
1609 fio_gettime(&td->start, NULL);
1610 fio_mutex_up(stat_mutex);
1612 if (td->error || td->terminate)
1615 if (!o->do_verify ||
1616 o->verify == VERIFY_NONE ||
1617 (td->io_ops->flags & FIO_UNIDIR))
1622 fio_gettime(&td->start, NULL);
1624 do_verify(td, verify_bytes);
1627 * See comment further up for why this is done here.
1629 check_update_rusage(td);
1631 fio_mutex_down(stat_mutex);
1632 update_runtime(td, elapsed_us, DDIR_READ);
1633 fio_gettime(&td->start, NULL);
1634 fio_mutex_up(stat_mutex);
1636 if (td->error || td->terminate)
1640 update_rusage_stat(td);
1641 td->ts.total_run_time = mtime_since_now(&td->epoch);
1642 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1643 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1644 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1646 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1647 (td->o.verify != VERIFY_NONE && td_write(td)))
1648 verify_save_state(td->thread_number);
1650 fio_unpin_memory(td);
1652 fio_writeout_logs(td);
1654 if (o->io_submit_mode == IO_MODE_OFFLOAD)
1655 workqueue_exit(&td->io_wq);
1657 if (td->flags & TD_F_COMPRESS_LOG)
1658 tp_exit(&td->tp_data);
1660 if (o->exec_postrun)
1661 exec_string(o, o->exec_postrun, (const char *)"postrun");
1663 if (exitall_on_terminate)
1664 fio_terminate_threads(td->groupid);
1668 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1671 if (o->verify_async)
1672 verify_async_exit(td);
1674 close_and_free_files(td);
1677 cgroup_shutdown(td, &cgroup_mnt);
1678 verify_free_state(td);
1680 if (fio_option_is_set(o, cpumask)) {
1681 ret = fio_cpuset_exit(&o->cpumask);
1683 td_verror(td, ret, "fio_cpuset_exit");
1687 * do this very late, it will log file closing as well
1689 if (o->write_iolog_file)
1690 write_iolog_close(td);
1692 fio_mutex_remove(td->mutex);
1695 td_set_runstate(td, TD_EXITED);
1698 * Do this last after setting our runstate to exited, so we
1699 * know that the stat thread is signaled.
1701 check_update_rusage(td);
1703 return (void *) (uintptr_t) td->error;
1708 * We cannot pass the td data into a forked process, so attach the td and
1709 * pass it to the thread worker.
1711 static int fork_main(int shmid, int offset)
1713 struct thread_data *td;
1716 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1717 data = shmat(shmid, NULL, 0);
1718 if (data == (void *) -1) {
1726 * HP-UX inherits shm mappings?
1731 td = data + offset * sizeof(struct thread_data);
1732 ret = thread_main(td);
1734 return (int) (uintptr_t) ret;
1737 static void dump_td_info(struct thread_data *td)
1739 log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1740 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1741 (unsigned long) time_since_now(&td->terminate_time));
1745 * Run over the job map and reap the threads that have exited, if any.
1747 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1748 unsigned int *m_rate)
1750 struct thread_data *td;
1751 unsigned int cputhreads, realthreads, pending;
1755 * reap exited threads (TD_EXITED -> TD_REAPED)
1757 realthreads = pending = cputhreads = 0;
1758 for_each_td(td, i) {
1762 * ->io_ops is NULL for a thread that has closed its
1765 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1774 if (td->runstate == TD_REAPED)
1776 if (td->o.use_thread) {
1777 if (td->runstate == TD_EXITED) {
1778 td_set_runstate(td, TD_REAPED);
1785 if (td->runstate == TD_EXITED)
1789 * check if someone quit or got killed in an unusual way
1791 ret = waitpid(td->pid, &status, flags);
1793 if (errno == ECHILD) {
1794 log_err("fio: pid=%d disappeared %d\n",
1795 (int) td->pid, td->runstate);
1797 td_set_runstate(td, TD_REAPED);
1801 } else if (ret == td->pid) {
1802 if (WIFSIGNALED(status)) {
1803 int sig = WTERMSIG(status);
1805 if (sig != SIGTERM && sig != SIGUSR2)
1806 log_err("fio: pid=%d, got signal=%d\n",
1807 (int) td->pid, sig);
1809 td_set_runstate(td, TD_REAPED);
1812 if (WIFEXITED(status)) {
1813 if (WEXITSTATUS(status) && !td->error)
1814 td->error = WEXITSTATUS(status);
1816 td_set_runstate(td, TD_REAPED);
1822 * If the job is stuck, do a forceful timeout of it and
1825 if (td->terminate &&
1826 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1828 td_set_runstate(td, TD_REAPED);
1833 * thread is not dead, continue
1839 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1840 (*t_rate) -= ddir_rw_sum(td->o.rate);
1847 done_secs += mtime_since_now(&td->epoch) / 1000;
1848 profile_td_exit(td);
1851 if (*nr_running == cputhreads && !pending && realthreads)
1852 fio_terminate_threads(TERMINATE_ALL);
1855 static int __check_trigger_file(void)
1862 if (stat(trigger_file, &sb))
1865 if (unlink(trigger_file) < 0)
1866 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1872 static int trigger_timedout(void)
1874 if (trigger_timeout)
1875 return time_since_genesis() >= trigger_timeout;
1880 void exec_trigger(const char *cmd)
1889 log_err("fio: failed executing %s trigger\n", cmd);
1892 void check_trigger_file(void)
1894 if (__check_trigger_file() || trigger_timedout()) {
1896 fio_clients_send_trigger(trigger_remote_cmd);
1898 verify_save_state(IO_LIST_ALL);
1899 fio_terminate_threads(TERMINATE_ALL);
1900 exec_trigger(trigger_cmd);
1905 static int fio_verify_load_state(struct thread_data *td)
1909 if (!td->o.verify_state)
1916 ret = fio_server_get_verify_state(td->o.name,
1917 td->thread_number - 1, &data, &ver);
1919 verify_convert_assign_state(td, data, ver);
1921 ret = verify_load_state(td, "local");
1926 static void do_usleep(unsigned int usecs)
1928 check_for_running_stats();
1929 check_trigger_file();
1933 static int check_mount_writes(struct thread_data *td)
1938 if (!td_write(td) || td->o.allow_mounted_write)
1941 for_each_file(td, f, i) {
1942 if (f->filetype != FIO_TYPE_BD)
1944 if (device_is_mounted(f->file_name))
1950 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
1955 * Main function for kicking off and reaping jobs, as needed.
1957 static void run_threads(void)
1959 struct thread_data *td;
1960 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1963 if (fio_gtod_offload && fio_start_gtod_thread())
1966 fio_idle_prof_init();
1970 nr_thread = nr_process = 0;
1971 for_each_td(td, i) {
1972 if (check_mount_writes(td))
1974 if (td->o.use_thread)
1980 if (output_format == FIO_OUTPUT_NORMAL) {
1981 log_info("Starting ");
1983 log_info("%d thread%s", nr_thread,
1984 nr_thread > 1 ? "s" : "");
1988 log_info("%d process%s", nr_process,
1989 nr_process > 1 ? "es" : "");
1995 todo = thread_number;
1998 m_rate = t_rate = 0;
2000 for_each_td(td, i) {
2001 print_status_init(td->thread_number - 1);
2003 if (!td->o.create_serialize)
2006 if (fio_verify_load_state(td))
2010 * do file setup here so it happens sequentially,
2011 * we don't want X number of threads getting their
2012 * client data interspersed on disk
2014 if (setup_files(td)) {
2018 log_err("fio: pid=%d, err=%d/%s\n",
2019 (int) td->pid, td->error, td->verror);
2020 td_set_runstate(td, TD_REAPED);
2027 * for sharing to work, each job must always open
2028 * its own files. so close them, if we opened them
2031 for_each_file(td, f, j) {
2032 if (fio_file_open(f))
2033 td_io_close_file(td, f);
2038 /* start idle threads before io threads start to run */
2039 fio_idle_prof_start();
2044 struct thread_data *map[REAL_MAX_JOBS];
2045 struct timeval this_start;
2046 int this_jobs = 0, left;
2049 * create threads (TD_NOT_CREATED -> TD_CREATED)
2051 for_each_td(td, i) {
2052 if (td->runstate != TD_NOT_CREATED)
2056 * never got a chance to start, killed by other
2057 * thread for some reason
2059 if (td->terminate) {
2064 if (td->o.start_delay) {
2065 spent = utime_since_genesis();
2067 if (td->o.start_delay > spent)
2071 if (td->o.stonewall && (nr_started || nr_running)) {
2072 dprint(FD_PROCESS, "%s: stonewall wait\n",
2079 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2080 td->update_rusage = 0;
2083 * Set state to created. Thread will transition
2084 * to TD_INITIALIZED when it's done setting up.
2086 td_set_runstate(td, TD_CREATED);
2087 map[this_jobs++] = td;
2090 if (td->o.use_thread) {
2093 dprint(FD_PROCESS, "will pthread_create\n");
2094 ret = pthread_create(&td->thread, NULL,
2097 log_err("pthread_create: %s\n",
2102 ret = pthread_detach(td->thread);
2104 log_err("pthread_detach: %s",
2108 dprint(FD_PROCESS, "will fork\n");
2111 int ret = fork_main(shm_id, i);
2114 } else if (i == fio_debug_jobno)
2115 *fio_debug_jobp = pid;
2117 dprint(FD_MUTEX, "wait on startup_mutex\n");
2118 if (fio_mutex_down_timeout(startup_mutex, 10)) {
2119 log_err("fio: job startup hung? exiting.\n");
2120 fio_terminate_threads(TERMINATE_ALL);
2125 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2129 * Wait for the started threads to transition to
2132 fio_gettime(&this_start, NULL);
2134 while (left && !fio_abort) {
2135 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2140 for (i = 0; i < this_jobs; i++) {
2144 if (td->runstate == TD_INITIALIZED) {
2147 } else if (td->runstate >= TD_EXITED) {
2151 nr_running++; /* work-around... */
2157 log_err("fio: %d job%s failed to start\n", left,
2158 left > 1 ? "s" : "");
2159 for (i = 0; i < this_jobs; i++) {
2163 kill(td->pid, SIGTERM);
2169 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2171 for_each_td(td, i) {
2172 if (td->runstate != TD_INITIALIZED)
2175 if (in_ramp_time(td))
2176 td_set_runstate(td, TD_RAMP);
2178 td_set_runstate(td, TD_RUNNING);
2181 m_rate += ddir_rw_sum(td->o.ratemin);
2182 t_rate += ddir_rw_sum(td->o.rate);
2184 fio_mutex_up(td->mutex);
2187 reap_threads(&nr_running, &t_rate, &m_rate);
2193 while (nr_running) {
2194 reap_threads(&nr_running, &t_rate, &m_rate);
2198 fio_idle_prof_stop();
2203 static void wait_for_helper_thread_exit(void)
2208 pthread_cond_signal(&helper_cond);
2209 pthread_join(helper_thread, &ret);
2212 static void free_disk_util(void)
2214 disk_util_prune_entries();
2216 pthread_cond_destroy(&helper_cond);
2219 static void *helper_thread_main(void *data)
2223 fio_mutex_up(startup_mutex);
2226 uint64_t sec = DISK_UTIL_MSEC / 1000;
2227 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2231 gettimeofday(&tv, NULL);
2232 ts.tv_sec = tv.tv_sec + sec;
2233 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2235 if (ts.tv_nsec >= 1000000000ULL) {
2236 ts.tv_nsec -= 1000000000ULL;
2240 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2242 ret = update_io_ticks();
2244 if (helper_do_stat) {
2246 __show_running_run_stats();
2250 print_thread_status();
2256 static int create_helper_thread(void)
2262 pthread_cond_init(&helper_cond, NULL);
2263 pthread_mutex_init(&helper_lock, NULL);
2265 ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2267 log_err("Can't create helper thread: %s\n", strerror(ret));
2271 dprint(FD_MUTEX, "wait on startup_mutex\n");
2272 fio_mutex_down(startup_mutex);
2273 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2277 int fio_backend(void)
2279 struct thread_data *td;
2283 if (load_profile(exec_profile))
2286 exec_profile = NULL;
2292 struct log_params p = {
2293 .log_type = IO_LOG_TYPE_BW,
2296 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2297 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2298 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2301 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2302 if (startup_mutex == NULL)
2307 create_helper_thread();
2309 cgroup_list = smalloc(sizeof(*cgroup_list));
2310 INIT_FLIST_HEAD(cgroup_list);
2314 wait_for_helper_thread_exit();
2319 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2320 struct io_log *log = agg_io_log[i];
2328 for_each_td(td, i) {
2329 fio_options_free(td);
2330 if (td->rusage_sem) {
2331 fio_mutex_remove(td->rusage_sem);
2332 td->rusage_sem = NULL;
2337 cgroup_kill(cgroup_list);
2341 fio_mutex_remove(startup_mutex);