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"
59 static pthread_t helper_thread;
60 static pthread_mutex_t helper_lock;
61 pthread_cond_t helper_cond;
62 int helper_do_stat = 0;
64 static struct fio_mutex *startup_mutex;
65 static struct flist_head *cgroup_list;
66 static char *cgroup_mnt;
67 static int exit_value;
68 static volatile int fio_abort;
69 static unsigned int nr_process = 0;
70 static unsigned int nr_thread = 0;
72 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
75 unsigned int thread_number = 0;
76 unsigned int stat_number = 0;
79 unsigned long done_secs = 0;
80 volatile int helper_exit = 0;
82 #define PAGE_ALIGN(buf) \
83 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
85 #define JOB_START_TIMEOUT (5 * 1000)
87 static void sig_int(int sig)
91 fio_server_got_signal(sig);
93 log_info("\nfio: terminating on signal %d\n", sig);
98 fio_terminate_threads(TERMINATE_ALL);
102 static void sig_show_status(int sig)
104 show_running_run_stats();
107 static void set_sig_handlers(void)
109 struct sigaction act;
111 memset(&act, 0, sizeof(act));
112 act.sa_handler = sig_int;
113 act.sa_flags = SA_RESTART;
114 sigaction(SIGINT, &act, NULL);
116 memset(&act, 0, sizeof(act));
117 act.sa_handler = sig_int;
118 act.sa_flags = SA_RESTART;
119 sigaction(SIGTERM, &act, NULL);
121 /* Windows uses SIGBREAK as a quit signal from other applications */
123 memset(&act, 0, sizeof(act));
124 act.sa_handler = sig_int;
125 act.sa_flags = SA_RESTART;
126 sigaction(SIGBREAK, &act, NULL);
129 memset(&act, 0, sizeof(act));
130 act.sa_handler = sig_show_status;
131 act.sa_flags = SA_RESTART;
132 sigaction(SIGUSR1, &act, NULL);
135 memset(&act, 0, sizeof(act));
136 act.sa_handler = sig_int;
137 act.sa_flags = SA_RESTART;
138 sigaction(SIGPIPE, &act, NULL);
143 * Check if we are above the minimum rate given.
145 static int __check_min_rate(struct thread_data *td, struct timeval *now,
148 unsigned long long bytes = 0;
149 unsigned long iops = 0;
152 unsigned int ratemin = 0;
153 unsigned int rate_iops = 0;
154 unsigned int rate_iops_min = 0;
156 assert(ddir_rw(ddir));
158 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
162 * allow a 2 second settle period in the beginning
164 if (mtime_since(&td->start, now) < 2000)
167 iops += td->this_io_blocks[ddir];
168 bytes += td->this_io_bytes[ddir];
169 ratemin += td->o.ratemin[ddir];
170 rate_iops += td->o.rate_iops[ddir];
171 rate_iops_min += td->o.rate_iops_min[ddir];
174 * if rate blocks is set, sample is running
176 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
177 spent = mtime_since(&td->lastrate[ddir], now);
178 if (spent < td->o.ratecycle)
181 if (td->o.rate[ddir]) {
183 * check bandwidth specified rate
185 if (bytes < td->rate_bytes[ddir]) {
186 log_err("%s: min rate %u not met\n", td->o.name,
191 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
195 if (rate < ratemin ||
196 bytes < td->rate_bytes[ddir]) {
197 log_err("%s: min rate %u not met, got"
198 " %luKB/sec\n", td->o.name,
205 * checks iops specified rate
207 if (iops < rate_iops) {
208 log_err("%s: min iops rate %u not met\n",
209 td->o.name, rate_iops);
213 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
217 if (rate < rate_iops_min ||
218 iops < td->rate_blocks[ddir]) {
219 log_err("%s: min iops rate %u not met,"
220 " got %lu\n", td->o.name,
221 rate_iops_min, rate);
227 td->rate_bytes[ddir] = bytes;
228 td->rate_blocks[ddir] = iops;
229 memcpy(&td->lastrate[ddir], now, sizeof(*now));
233 static int check_min_rate(struct thread_data *td, struct timeval *now)
237 if (td->bytes_done[DDIR_READ])
238 ret |= __check_min_rate(td, now, DDIR_READ);
239 if (td->bytes_done[DDIR_WRITE])
240 ret |= __check_min_rate(td, now, DDIR_WRITE);
241 if (td->bytes_done[DDIR_TRIM])
242 ret |= __check_min_rate(td, now, DDIR_TRIM);
248 * When job exits, we can cancel the in-flight IO if we are using async
249 * io. Attempt to do so.
251 static void cleanup_pending_aio(struct thread_data *td)
256 * get immediately available events, if any
258 r = io_u_queued_complete(td, 0);
263 * now cancel remaining active events
265 if (td->io_ops->cancel) {
269 io_u_qiter(&td->io_u_all, io_u, i) {
270 if (io_u->flags & IO_U_F_FLIGHT) {
271 r = td->io_ops->cancel(td, io_u);
279 r = io_u_queued_complete(td, td->cur_depth);
283 * Helper to handle the final sync of a file. Works just like the normal
284 * io path, just does everything sync.
286 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
288 struct io_u *io_u = __get_io_u(td);
294 io_u->ddir = DDIR_SYNC;
297 if (td_io_prep(td, io_u)) {
303 ret = td_io_queue(td, io_u);
305 td_verror(td, io_u->error, "td_io_queue");
308 } else if (ret == FIO_Q_QUEUED) {
309 if (io_u_queued_complete(td, 1) < 0)
311 } else if (ret == FIO_Q_COMPLETED) {
313 td_verror(td, io_u->error, "td_io_queue");
317 if (io_u_sync_complete(td, io_u) < 0)
319 } else if (ret == FIO_Q_BUSY) {
320 if (td_io_commit(td))
328 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
332 if (fio_file_open(f))
333 return fio_io_sync(td, f);
335 if (td_io_open_file(td, f))
338 ret = fio_io_sync(td, f);
339 td_io_close_file(td, f);
343 static inline void __update_tv_cache(struct thread_data *td)
345 fio_gettime(&td->tv_cache, NULL);
348 static inline void update_tv_cache(struct thread_data *td)
350 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
351 __update_tv_cache(td);
354 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
356 if (in_ramp_time(td))
360 if (utime_since(&td->epoch, t) >= td->o.timeout)
366 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
371 if (ret < 0 || td->error) {
373 enum error_type_bit eb;
378 eb = td_error_type(ddir, err);
379 if (!(td->o.continue_on_error & (1 << eb)))
382 if (td_non_fatal_error(td, eb, err)) {
384 * Continue with the I/Os in case of
387 update_error_count(td, err);
391 } else if (td->o.fill_device && err == ENOSPC) {
393 * We expect to hit this error if
394 * fill_device option is set.
397 fio_mark_td_terminate(td);
401 * Stop the I/O in case of a fatal
404 update_error_count(td, err);
412 static void check_update_rusage(struct thread_data *td)
414 if (td->update_rusage) {
415 td->update_rusage = 0;
416 update_rusage_stat(td);
417 fio_mutex_up(td->rusage_sem);
421 static int wait_for_completions(struct thread_data *td, struct timeval *time)
423 const int full = queue_full(td);
428 * if the queue is full, we MUST reap at least 1 event
430 min_evts = min(td->o.iodepth_batch_complete, td->cur_depth);
431 if (full && !min_evts)
434 if (time && (__should_check_rate(td, DDIR_READ) ||
435 __should_check_rate(td, DDIR_WRITE) ||
436 __should_check_rate(td, DDIR_TRIM)))
437 fio_gettime(time, NULL);
440 ret = io_u_queued_complete(td, min_evts);
443 } while (full && (td->cur_depth > td->o.iodepth_low));
448 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
449 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
450 struct timeval *comp_time)
455 case FIO_Q_COMPLETED:
458 clear_io_u(td, io_u);
459 } else if (io_u->resid) {
460 int bytes = io_u->xfer_buflen - io_u->resid;
461 struct fio_file *f = io_u->file;
464 *bytes_issued += bytes;
467 trim_io_piece(td, io_u);
474 unlog_io_piece(td, io_u);
475 td_verror(td, EIO, "full resid");
480 io_u->xfer_buflen = io_u->resid;
481 io_u->xfer_buf += bytes;
482 io_u->offset += bytes;
484 if (ddir_rw(io_u->ddir))
485 td->ts.short_io_u[io_u->ddir]++;
488 if (io_u->offset == f->real_file_size)
491 requeue_io_u(td, &io_u);
494 if (comp_time && (__should_check_rate(td, DDIR_READ) ||
495 __should_check_rate(td, DDIR_WRITE) ||
496 __should_check_rate(td, DDIR_TRIM)))
497 fio_gettime(comp_time, NULL);
499 *ret = io_u_sync_complete(td, io_u);
506 * if the engine doesn't have a commit hook,
507 * the io_u is really queued. if it does have such
508 * a hook, it has to call io_u_queued() itself.
510 if (td->io_ops->commit == NULL)
511 io_u_queued(td, io_u);
513 *bytes_issued += io_u->xfer_buflen;
517 unlog_io_piece(td, io_u);
518 requeue_io_u(td, &io_u);
519 ret2 = td_io_commit(td);
525 td_verror(td, -(*ret), "td_io_queue");
529 if (break_on_this_error(td, ddir, ret))
536 * The main verify engine. Runs over the writes we previously submitted,
537 * reads the blocks back in, and checks the crc/md5 of the data.
539 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
546 dprint(FD_VERIFY, "starting loop\n");
549 * sync io first and invalidate cache, to make sure we really
552 for_each_file(td, f, i) {
553 if (!fio_file_open(f))
555 if (fio_io_sync(td, f))
557 if (file_invalidate_cache(td, f))
561 check_update_rusage(td);
566 td_set_runstate(td, TD_VERIFYING);
569 while (!td->terminate) {
574 check_update_rusage(td);
576 if (runtime_exceeded(td, &td->tv_cache)) {
577 __update_tv_cache(td);
578 if (runtime_exceeded(td, &td->tv_cache)) {
579 fio_mark_td_terminate(td);
584 if (flow_threshold_exceeded(td))
587 if (!td->o.experimental_verify) {
588 io_u = __get_io_u(td);
592 if (get_next_verify(td, io_u)) {
597 if (td_io_prep(td, io_u)) {
602 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
605 while ((io_u = get_io_u(td)) != NULL) {
613 * We are only interested in the places where
614 * we wrote or trimmed IOs. Turn those into
615 * reads for verification purposes.
617 if (io_u->ddir == DDIR_READ) {
619 * Pretend we issued it for rwmix
622 td->io_issues[DDIR_READ]++;
625 } else if (io_u->ddir == DDIR_TRIM) {
626 io_u->ddir = DDIR_READ;
627 io_u_set(io_u, IO_U_F_TRIMMED);
629 } else if (io_u->ddir == DDIR_WRITE) {
630 io_u->ddir = DDIR_READ;
642 if (verify_state_should_stop(td, io_u)) {
647 if (td->o.verify_async)
648 io_u->end_io = verify_io_u_async;
650 io_u->end_io = verify_io_u;
653 if (!td->o.disable_slat)
654 fio_gettime(&io_u->start_time, NULL);
656 ret = td_io_queue(td, io_u);
658 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
662 * if we can queue more, do so. but check if there are
663 * completed io_u's first. Note that we can get BUSY even
664 * without IO queued, if the system is resource starved.
667 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
668 if (full || !td->o.iodepth_batch_complete)
669 ret = wait_for_completions(td, NULL);
675 check_update_rusage(td);
678 min_events = td->cur_depth;
681 ret = io_u_queued_complete(td, min_events);
683 cleanup_pending_aio(td);
685 td_set_runstate(td, TD_RUNNING);
687 dprint(FD_VERIFY, "exiting loop\n");
690 static unsigned int exceeds_number_ios(struct thread_data *td)
692 unsigned long long number_ios;
694 if (!td->o.number_ios)
697 number_ios = ddir_rw_sum(td->io_blocks);
698 number_ios += td->io_u_queued + td->io_u_in_flight;
700 return number_ios >= (td->o.number_ios * td->loops);
703 static int io_issue_bytes_exceeded(struct thread_data *td)
705 unsigned long long bytes, limit;
708 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
709 else if (td_write(td))
710 bytes = td->io_issue_bytes[DDIR_WRITE];
711 else if (td_read(td))
712 bytes = td->io_issue_bytes[DDIR_READ];
714 bytes = td->io_issue_bytes[DDIR_TRIM];
717 limit = td->o.io_limit;
722 return bytes >= limit || exceeds_number_ios(td);
725 static int io_complete_bytes_exceeded(struct thread_data *td)
727 unsigned long long bytes, limit;
730 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
731 else if (td_write(td))
732 bytes = td->this_io_bytes[DDIR_WRITE];
733 else if (td_read(td))
734 bytes = td->this_io_bytes[DDIR_READ];
736 bytes = td->this_io_bytes[DDIR_TRIM];
739 limit = td->o.io_limit;
744 return bytes >= limit || exceeds_number_ios(td);
748 * Main IO worker function. It retrieves io_u's to process and queues
749 * and reaps them, checking for rate and errors along the way.
751 * Returns number of bytes written and trimmed.
753 static uint64_t do_io(struct thread_data *td)
757 uint64_t total_bytes, bytes_issued = 0;
759 if (in_ramp_time(td))
760 td_set_runstate(td, TD_RAMP);
762 td_set_runstate(td, TD_RUNNING);
766 total_bytes = td->o.size;
768 * Allow random overwrite workloads to write up to io_limit
769 * before starting verification phase as 'size' doesn't apply.
771 if (td_write(td) && td_random(td) && td->o.norandommap)
772 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
774 * If verify_backlog is enabled, we'll run the verify in this
775 * handler as well. For that case, we may need up to twice the
778 if (td->o.verify != VERIFY_NONE &&
779 (td_write(td) && td->o.verify_backlog))
780 total_bytes += td->o.size;
782 /* In trimwrite mode, each byte is trimmed and then written, so
783 * allow total_bytes to be twice as big */
784 if (td_trimwrite(td))
785 total_bytes += td->total_io_size;
787 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
788 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
790 struct timeval comp_time;
795 check_update_rusage(td);
797 if (td->terminate || td->done)
802 if (runtime_exceeded(td, &td->tv_cache)) {
803 __update_tv_cache(td);
804 if (runtime_exceeded(td, &td->tv_cache)) {
805 fio_mark_td_terminate(td);
810 if (flow_threshold_exceeded(td))
813 if (bytes_issued >= total_bytes)
817 if (IS_ERR_OR_NULL(io_u)) {
818 int err = PTR_ERR(io_u);
825 if (td->o.latency_target)
833 * Add verification end_io handler if:
834 * - Asked to verify (!td_rw(td))
835 * - Or the io_u is from our verify list (mixed write/ver)
837 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
838 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
840 if (!td->o.verify_pattern_bytes) {
841 io_u->rand_seed = __rand(&td->verify_state);
842 if (sizeof(int) != sizeof(long *))
843 io_u->rand_seed *= __rand(&td->verify_state);
846 if (verify_state_should_stop(td, io_u)) {
851 if (td->o.verify_async)
852 io_u->end_io = verify_io_u_async;
854 io_u->end_io = verify_io_u;
855 td_set_runstate(td, TD_VERIFYING);
856 } else if (in_ramp_time(td))
857 td_set_runstate(td, TD_RAMP);
859 td_set_runstate(td, TD_RUNNING);
862 * Always log IO before it's issued, so we know the specific
863 * order of it. The logged unit will track when the IO has
866 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
868 td->o.verify != VERIFY_NONE &&
869 !td->o.experimental_verify)
870 log_io_piece(td, io_u);
872 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
875 ret = workqueue_enqueue(&td->io_wq, io_u);
877 ret = td_io_queue(td, io_u);
879 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 1, &comp_time))
883 * See if we need to complete some commands. Note that
884 * we can get BUSY even without IO queued, if the
885 * system is resource starved.
888 full = queue_full(td) ||
889 (ret == FIO_Q_BUSY && td->cur_depth);
890 if (full || !td->o.iodepth_batch_complete)
891 ret = wait_for_completions(td, &comp_time);
895 if (!ddir_rw_sum(td->bytes_done) &&
896 !(td->io_ops->flags & FIO_NOIO))
899 if (!in_ramp_time(td) && should_check_rate(td)) {
900 if (check_min_rate(td, &comp_time)) {
901 if (exitall_on_terminate)
902 fio_terminate_threads(td->groupid);
903 td_verror(td, EIO, "check_min_rate");
907 if (!in_ramp_time(td) && td->o.latency_target)
908 lat_target_check(td);
910 if (td->o.thinktime) {
911 unsigned long long b;
913 b = ddir_rw_sum(td->io_blocks);
914 if (!(b % td->o.thinktime_blocks)) {
919 if (td->o.thinktime_spin)
920 usec_spin(td->o.thinktime_spin);
922 left = td->o.thinktime - td->o.thinktime_spin;
924 usec_sleep(td, left);
929 check_update_rusage(td);
931 if (td->trim_entries)
932 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
934 if (td->o.fill_device && td->error == ENOSPC) {
936 fio_mark_td_terminate(td);
941 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
942 workqueue_flush(&td->io_wq);
948 ret = io_u_queued_complete(td, i);
949 if (td->o.fill_device && td->error == ENOSPC)
953 if (should_fsync(td) && td->o.end_fsync) {
954 td_set_runstate(td, TD_FSYNCING);
956 for_each_file(td, f, i) {
957 if (!fio_file_fsync(td, f))
960 log_err("fio: end_fsync failed for file %s\n",
965 cleanup_pending_aio(td);
968 * stop job if we failed doing any IO
970 if (!ddir_rw_sum(td->this_io_bytes))
973 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
976 static void cleanup_io_u(struct thread_data *td)
980 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
982 if (td->io_ops->io_u_free)
983 td->io_ops->io_u_free(td, io_u);
985 fio_memfree(io_u, sizeof(*io_u));
990 io_u_rexit(&td->io_u_requeues);
991 io_u_qexit(&td->io_u_freelist);
992 io_u_qexit(&td->io_u_all);
994 if (td->last_write_comp)
995 sfree(td->last_write_comp);
998 static int init_io_u(struct thread_data *td)
1001 unsigned int max_bs, min_write;
1002 int cl_align, i, max_units;
1003 int data_xfer = 1, err;
1006 max_units = td->o.iodepth;
1007 max_bs = td_max_bs(td);
1008 min_write = td->o.min_bs[DDIR_WRITE];
1009 td->orig_buffer_size = (unsigned long long) max_bs
1010 * (unsigned long long) max_units;
1012 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1016 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1017 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1018 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1021 log_err("fio: failed setting up IO queues\n");
1026 * if we may later need to do address alignment, then add any
1027 * possible adjustment here so that we don't cause a buffer
1028 * overflow later. this adjustment may be too much if we get
1029 * lucky and the allocator gives us an aligned address.
1031 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1032 (td->io_ops->flags & FIO_RAWIO))
1033 td->orig_buffer_size += page_mask + td->o.mem_align;
1035 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1038 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1039 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1042 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1043 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1047 if (data_xfer && allocate_io_mem(td))
1050 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1051 (td->io_ops->flags & FIO_RAWIO))
1052 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1054 p = td->orig_buffer;
1056 cl_align = os_cache_line_size();
1058 for (i = 0; i < max_units; i++) {
1064 ptr = fio_memalign(cl_align, sizeof(*io_u));
1066 log_err("fio: unable to allocate aligned memory\n");
1071 memset(io_u, 0, sizeof(*io_u));
1072 INIT_FLIST_HEAD(&io_u->verify_list);
1073 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1077 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1080 io_u_fill_buffer(td, io_u, min_write, max_bs);
1081 if (td_write(td) && td->o.verify_pattern_bytes) {
1083 * Fill the buffer with the pattern if we are
1084 * going to be doing writes.
1086 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1091 io_u->flags = IO_U_F_FREE;
1092 io_u_qpush(&td->io_u_freelist, io_u);
1095 * io_u never leaves this stack, used for iteration of all
1098 io_u_qpush(&td->io_u_all, io_u);
1100 if (td->io_ops->io_u_init) {
1101 int ret = td->io_ops->io_u_init(td, io_u);
1104 log_err("fio: failed to init engine data: %d\n", ret);
1112 if (td->o.verify != VERIFY_NONE) {
1113 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1114 if (!td->last_write_comp) {
1115 log_err("fio: failed to alloc write comp data\n");
1123 static int switch_ioscheduler(struct thread_data *td)
1125 char tmp[256], tmp2[128];
1129 if (td->io_ops->flags & FIO_DISKLESSIO)
1132 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1134 f = fopen(tmp, "r+");
1136 if (errno == ENOENT) {
1137 log_err("fio: os or kernel doesn't support IO scheduler"
1141 td_verror(td, errno, "fopen iosched");
1148 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1149 if (ferror(f) || ret != 1) {
1150 td_verror(td, errno, "fwrite");
1158 * Read back and check that the selected scheduler is now the default.
1160 ret = fread(tmp, sizeof(tmp), 1, f);
1161 if (ferror(f) || ret < 0) {
1162 td_verror(td, errno, "fread");
1166 tmp[sizeof(tmp) - 1] = '\0';
1169 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1170 if (!strstr(tmp, tmp2)) {
1171 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1172 td_verror(td, EINVAL, "iosched_switch");
1181 static int keep_running(struct thread_data *td)
1183 unsigned long long limit;
1187 if (td->o.time_based)
1193 if (exceeds_number_ios(td))
1197 limit = td->o.io_limit;
1201 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1205 * If the difference is less than the minimum IO size, we
1208 diff = limit - ddir_rw_sum(td->io_bytes);
1209 if (diff < td_max_bs(td))
1212 if (fio_files_done(td))
1221 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1223 int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1226 str = malloc(newlen);
1227 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1229 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1232 log_err("fio: exec of cmd <%s> failed\n", str);
1239 * Dry run to compute correct state of numberio for verification.
1241 static uint64_t do_dry_run(struct thread_data *td)
1243 td_set_runstate(td, TD_RUNNING);
1245 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1246 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1250 if (td->terminate || td->done)
1253 io_u = get_io_u(td);
1257 io_u_set(io_u, IO_U_F_FLIGHT);
1260 if (ddir_rw(acct_ddir(io_u)))
1261 td->io_issues[acct_ddir(io_u)]++;
1262 if (ddir_rw(io_u->ddir)) {
1263 io_u_mark_depth(td, 1);
1264 td->ts.total_io_u[io_u->ddir]++;
1267 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1269 td->o.verify != VERIFY_NONE &&
1270 !td->o.experimental_verify)
1271 log_io_piece(td, io_u);
1273 ret = io_u_sync_complete(td, io_u);
1277 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1280 static void io_workqueue_fn(struct thread_data *td, struct io_u *io_u)
1282 const enum fio_ddir ddir = io_u->ddir;
1285 dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
1287 io_u_set(io_u, IO_U_F_NO_FILE_PUT);
1291 ret = td_io_queue(td, io_u);
1293 dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
1295 io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
1297 if (ret == FIO_Q_QUEUED)
1298 ret = io_u_queued_complete(td, 1);
1304 * Entry point for the thread based jobs. The process based jobs end up
1305 * here as well, after a little setup.
1307 static void *thread_main(void *data)
1309 unsigned long long elapsed;
1310 struct thread_data *td = data;
1311 struct thread_options *o = &td->o;
1312 pthread_condattr_t attr;
1316 if (!o->use_thread) {
1322 fio_local_clock_init(o->use_thread);
1324 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1327 fio_server_send_start(td);
1329 INIT_FLIST_HEAD(&td->io_log_list);
1330 INIT_FLIST_HEAD(&td->io_hist_list);
1331 INIT_FLIST_HEAD(&td->verify_list);
1332 INIT_FLIST_HEAD(&td->trim_list);
1333 INIT_FLIST_HEAD(&td->next_rand_list);
1334 pthread_mutex_init(&td->io_u_lock, NULL);
1335 td->io_hist_tree = RB_ROOT;
1337 pthread_condattr_init(&attr);
1338 pthread_cond_init(&td->verify_cond, &attr);
1339 pthread_cond_init(&td->free_cond, &attr);
1341 td_set_runstate(td, TD_INITIALIZED);
1342 dprint(FD_MUTEX, "up startup_mutex\n");
1343 fio_mutex_up(startup_mutex);
1344 dprint(FD_MUTEX, "wait on td->mutex\n");
1345 fio_mutex_down(td->mutex);
1346 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1349 * A new gid requires privilege, so we need to do this before setting
1352 if (o->gid != -1U && setgid(o->gid)) {
1353 td_verror(td, errno, "setgid");
1356 if (o->uid != -1U && setuid(o->uid)) {
1357 td_verror(td, errno, "setuid");
1362 * If we have a gettimeofday() thread, make sure we exclude that
1363 * thread from this job
1366 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1369 * Set affinity first, in case it has an impact on the memory
1372 if (fio_option_is_set(o, cpumask)) {
1373 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1374 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1376 log_err("fio: no CPUs set\n");
1377 log_err("fio: Try increasing number of available CPUs\n");
1378 td_verror(td, EINVAL, "cpus_split");
1382 ret = fio_setaffinity(td->pid, o->cpumask);
1384 td_verror(td, errno, "cpu_set_affinity");
1389 #ifdef CONFIG_LIBNUMA
1390 /* numa node setup */
1391 if (fio_option_is_set(o, numa_cpunodes) ||
1392 fio_option_is_set(o, numa_memnodes)) {
1393 struct bitmask *mask;
1395 if (numa_available() < 0) {
1396 td_verror(td, errno, "Does not support NUMA API\n");
1400 if (fio_option_is_set(o, numa_cpunodes)) {
1401 mask = numa_parse_nodestring(o->numa_cpunodes);
1402 ret = numa_run_on_node_mask(mask);
1403 numa_free_nodemask(mask);
1405 td_verror(td, errno, \
1406 "numa_run_on_node_mask failed\n");
1411 if (fio_option_is_set(o, numa_memnodes)) {
1413 if (o->numa_memnodes)
1414 mask = numa_parse_nodestring(o->numa_memnodes);
1416 switch (o->numa_mem_mode) {
1417 case MPOL_INTERLEAVE:
1418 numa_set_interleave_mask(mask);
1421 numa_set_membind(mask);
1424 numa_set_localalloc();
1426 case MPOL_PREFERRED:
1427 numa_set_preferred(o->numa_mem_prefer_node);
1435 numa_free_nodemask(mask);
1441 if (fio_pin_memory(td))
1445 * May alter parameters that init_io_u() will use, so we need to
1454 if (o->verify_async && verify_async_init(td))
1457 if (fio_option_is_set(o, ioprio) ||
1458 fio_option_is_set(o, ioprio_class)) {
1459 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1461 td_verror(td, errno, "ioprio_set");
1466 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1470 if (nice(o->nice) == -1 && errno != 0) {
1471 td_verror(td, errno, "nice");
1475 if (o->ioscheduler && switch_ioscheduler(td))
1478 if (!o->create_serialize && setup_files(td))
1484 if (init_random_map(td))
1487 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1491 if (pre_read_files(td) < 0)
1495 if (td->flags & TD_F_COMPRESS_LOG)
1496 tp_init(&td->tp_data);
1498 fio_verify_init(td);
1500 if ((o->io_submit_mode == IO_MODE_OFFLOAD) &&
1501 workqueue_init(td, &td->io_wq, io_workqueue_fn, td->o.iodepth))
1504 fio_gettime(&td->epoch, NULL);
1505 fio_getrusage(&td->ru_start);
1507 while (keep_running(td)) {
1508 uint64_t verify_bytes;
1510 fio_gettime(&td->start, NULL);
1511 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1512 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1513 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1515 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1516 o->ratemin[DDIR_TRIM]) {
1517 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1518 sizeof(td->bw_sample_time));
1519 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1520 sizeof(td->bw_sample_time));
1521 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1522 sizeof(td->bw_sample_time));
1528 prune_io_piece_log(td);
1530 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1531 verify_bytes = do_dry_run(td);
1533 verify_bytes = do_io(td);
1538 * Make sure we've successfully updated the rusage stats
1539 * before waiting on the stat mutex. Otherwise we could have
1540 * the stat thread holding stat mutex and waiting for
1541 * the rusage_sem, which would never get upped because
1542 * this thread is waiting for the stat mutex.
1544 check_update_rusage(td);
1546 fio_mutex_down(stat_mutex);
1547 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1548 elapsed = mtime_since_now(&td->start);
1549 td->ts.runtime[DDIR_READ] += elapsed;
1551 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1552 elapsed = mtime_since_now(&td->start);
1553 td->ts.runtime[DDIR_WRITE] += elapsed;
1555 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1556 elapsed = mtime_since_now(&td->start);
1557 td->ts.runtime[DDIR_TRIM] += elapsed;
1559 fio_gettime(&td->start, NULL);
1560 fio_mutex_up(stat_mutex);
1562 if (td->error || td->terminate)
1565 if (!o->do_verify ||
1566 o->verify == VERIFY_NONE ||
1567 (td->io_ops->flags & FIO_UNIDIR))
1572 fio_gettime(&td->start, NULL);
1574 do_verify(td, verify_bytes);
1577 * See comment further up for why this is done here.
1579 check_update_rusage(td);
1581 fio_mutex_down(stat_mutex);
1582 td->ts.runtime[DDIR_READ] += mtime_since_now(&td->start);
1583 fio_gettime(&td->start, NULL);
1584 fio_mutex_up(stat_mutex);
1586 if (td->error || td->terminate)
1590 update_rusage_stat(td);
1591 td->ts.total_run_time = mtime_since_now(&td->epoch);
1592 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1593 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1594 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1596 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1597 (td->o.verify != VERIFY_NONE && td_write(td))) {
1598 struct all_io_list *state;
1601 state = get_all_io_list(td->thread_number, &sz);
1603 __verify_save_state(state, "local");
1608 fio_unpin_memory(td);
1610 fio_writeout_logs(td);
1612 if (o->io_submit_mode == IO_MODE_OFFLOAD)
1613 workqueue_exit(&td->io_wq);
1615 if (td->flags & TD_F_COMPRESS_LOG)
1616 tp_exit(&td->tp_data);
1618 if (o->exec_postrun)
1619 exec_string(o, o->exec_postrun, (const char *)"postrun");
1621 if (exitall_on_terminate)
1622 fio_terminate_threads(td->groupid);
1626 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1629 if (o->verify_async)
1630 verify_async_exit(td);
1632 close_and_free_files(td);
1635 cgroup_shutdown(td, &cgroup_mnt);
1636 verify_free_state(td);
1638 if (fio_option_is_set(o, cpumask)) {
1639 ret = fio_cpuset_exit(&o->cpumask);
1641 td_verror(td, ret, "fio_cpuset_exit");
1645 * do this very late, it will log file closing as well
1647 if (o->write_iolog_file)
1648 write_iolog_close(td);
1650 fio_mutex_remove(td->mutex);
1653 td_set_runstate(td, TD_EXITED);
1656 * Do this last after setting our runstate to exited, so we
1657 * know that the stat thread is signaled.
1659 check_update_rusage(td);
1661 return (void *) (uintptr_t) td->error;
1666 * We cannot pass the td data into a forked process, so attach the td and
1667 * pass it to the thread worker.
1669 static int fork_main(int shmid, int offset)
1671 struct thread_data *td;
1674 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1675 data = shmat(shmid, NULL, 0);
1676 if (data == (void *) -1) {
1684 * HP-UX inherits shm mappings?
1689 td = data + offset * sizeof(struct thread_data);
1690 ret = thread_main(td);
1692 return (int) (uintptr_t) ret;
1695 static void dump_td_info(struct thread_data *td)
1697 log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1698 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1699 (unsigned long) time_since_now(&td->terminate_time));
1703 * Run over the job map and reap the threads that have exited, if any.
1705 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1706 unsigned int *m_rate)
1708 struct thread_data *td;
1709 unsigned int cputhreads, realthreads, pending;
1713 * reap exited threads (TD_EXITED -> TD_REAPED)
1715 realthreads = pending = cputhreads = 0;
1716 for_each_td(td, i) {
1720 * ->io_ops is NULL for a thread that has closed its
1723 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1732 if (td->runstate == TD_REAPED)
1734 if (td->o.use_thread) {
1735 if (td->runstate == TD_EXITED) {
1736 td_set_runstate(td, TD_REAPED);
1743 if (td->runstate == TD_EXITED)
1747 * check if someone quit or got killed in an unusual way
1749 ret = waitpid(td->pid, &status, flags);
1751 if (errno == ECHILD) {
1752 log_err("fio: pid=%d disappeared %d\n",
1753 (int) td->pid, td->runstate);
1755 td_set_runstate(td, TD_REAPED);
1759 } else if (ret == td->pid) {
1760 if (WIFSIGNALED(status)) {
1761 int sig = WTERMSIG(status);
1763 if (sig != SIGTERM && sig != SIGUSR2)
1764 log_err("fio: pid=%d, got signal=%d\n",
1765 (int) td->pid, sig);
1767 td_set_runstate(td, TD_REAPED);
1770 if (WIFEXITED(status)) {
1771 if (WEXITSTATUS(status) && !td->error)
1772 td->error = WEXITSTATUS(status);
1774 td_set_runstate(td, TD_REAPED);
1780 * If the job is stuck, do a forceful timeout of it and
1783 if (td->terminate &&
1784 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1786 td_set_runstate(td, TD_REAPED);
1791 * thread is not dead, continue
1797 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1798 (*t_rate) -= ddir_rw_sum(td->o.rate);
1805 done_secs += mtime_since_now(&td->epoch) / 1000;
1806 profile_td_exit(td);
1809 if (*nr_running == cputhreads && !pending && realthreads)
1810 fio_terminate_threads(TERMINATE_ALL);
1813 static int __check_trigger_file(void)
1820 if (stat(trigger_file, &sb))
1823 if (unlink(trigger_file) < 0)
1824 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1830 static int trigger_timedout(void)
1832 if (trigger_timeout)
1833 return time_since_genesis() >= trigger_timeout;
1838 void exec_trigger(const char *cmd)
1847 log_err("fio: failed executing %s trigger\n", cmd);
1850 void check_trigger_file(void)
1852 if (__check_trigger_file() || trigger_timedout()) {
1854 fio_clients_send_trigger(trigger_remote_cmd);
1856 verify_save_state();
1857 fio_terminate_threads(TERMINATE_ALL);
1858 exec_trigger(trigger_cmd);
1863 static int fio_verify_load_state(struct thread_data *td)
1867 if (!td->o.verify_state)
1873 ret = fio_server_get_verify_state(td->o.name,
1874 td->thread_number - 1, &data);
1876 verify_convert_assign_state(td, data);
1878 ret = verify_load_state(td, "local");
1883 static void do_usleep(unsigned int usecs)
1885 check_for_running_stats();
1886 check_trigger_file();
1891 * Main function for kicking off and reaping jobs, as needed.
1893 static void run_threads(void)
1895 struct thread_data *td;
1896 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1899 if (fio_gtod_offload && fio_start_gtod_thread())
1902 fio_idle_prof_init();
1906 nr_thread = nr_process = 0;
1907 for_each_td(td, i) {
1908 if (td->o.use_thread)
1914 if (output_format == FIO_OUTPUT_NORMAL) {
1915 log_info("Starting ");
1917 log_info("%d thread%s", nr_thread,
1918 nr_thread > 1 ? "s" : "");
1922 log_info("%d process%s", nr_process,
1923 nr_process > 1 ? "es" : "");
1929 todo = thread_number;
1932 m_rate = t_rate = 0;
1934 for_each_td(td, i) {
1935 print_status_init(td->thread_number - 1);
1937 if (!td->o.create_serialize)
1940 if (fio_verify_load_state(td))
1944 * do file setup here so it happens sequentially,
1945 * we don't want X number of threads getting their
1946 * client data interspersed on disk
1948 if (setup_files(td)) {
1952 log_err("fio: pid=%d, err=%d/%s\n",
1953 (int) td->pid, td->error, td->verror);
1954 td_set_runstate(td, TD_REAPED);
1961 * for sharing to work, each job must always open
1962 * its own files. so close them, if we opened them
1965 for_each_file(td, f, j) {
1966 if (fio_file_open(f))
1967 td_io_close_file(td, f);
1972 /* start idle threads before io threads start to run */
1973 fio_idle_prof_start();
1978 struct thread_data *map[REAL_MAX_JOBS];
1979 struct timeval this_start;
1980 int this_jobs = 0, left;
1983 * create threads (TD_NOT_CREATED -> TD_CREATED)
1985 for_each_td(td, i) {
1986 if (td->runstate != TD_NOT_CREATED)
1990 * never got a chance to start, killed by other
1991 * thread for some reason
1993 if (td->terminate) {
1998 if (td->o.start_delay) {
1999 spent = utime_since_genesis();
2001 if (td->o.start_delay > spent)
2005 if (td->o.stonewall && (nr_started || nr_running)) {
2006 dprint(FD_PROCESS, "%s: stonewall wait\n",
2013 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2014 td->update_rusage = 0;
2017 * Set state to created. Thread will transition
2018 * to TD_INITIALIZED when it's done setting up.
2020 td_set_runstate(td, TD_CREATED);
2021 map[this_jobs++] = td;
2024 if (td->o.use_thread) {
2027 dprint(FD_PROCESS, "will pthread_create\n");
2028 ret = pthread_create(&td->thread, NULL,
2031 log_err("pthread_create: %s\n",
2036 ret = pthread_detach(td->thread);
2038 log_err("pthread_detach: %s",
2042 dprint(FD_PROCESS, "will fork\n");
2045 int ret = fork_main(shm_id, i);
2048 } else if (i == fio_debug_jobno)
2049 *fio_debug_jobp = pid;
2051 dprint(FD_MUTEX, "wait on startup_mutex\n");
2052 if (fio_mutex_down_timeout(startup_mutex, 10)) {
2053 log_err("fio: job startup hung? exiting.\n");
2054 fio_terminate_threads(TERMINATE_ALL);
2059 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2063 * Wait for the started threads to transition to
2066 fio_gettime(&this_start, NULL);
2068 while (left && !fio_abort) {
2069 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2074 for (i = 0; i < this_jobs; i++) {
2078 if (td->runstate == TD_INITIALIZED) {
2081 } else if (td->runstate >= TD_EXITED) {
2085 nr_running++; /* work-around... */
2091 log_err("fio: %d job%s failed to start\n", left,
2092 left > 1 ? "s" : "");
2093 for (i = 0; i < this_jobs; i++) {
2097 kill(td->pid, SIGTERM);
2103 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2105 for_each_td(td, i) {
2106 if (td->runstate != TD_INITIALIZED)
2109 if (in_ramp_time(td))
2110 td_set_runstate(td, TD_RAMP);
2112 td_set_runstate(td, TD_RUNNING);
2115 m_rate += ddir_rw_sum(td->o.ratemin);
2116 t_rate += ddir_rw_sum(td->o.rate);
2118 fio_mutex_up(td->mutex);
2121 reap_threads(&nr_running, &t_rate, &m_rate);
2127 while (nr_running) {
2128 reap_threads(&nr_running, &t_rate, &m_rate);
2132 fio_idle_prof_stop();
2137 static void wait_for_helper_thread_exit(void)
2142 pthread_cond_signal(&helper_cond);
2143 pthread_join(helper_thread, &ret);
2146 static void free_disk_util(void)
2148 disk_util_prune_entries();
2150 pthread_cond_destroy(&helper_cond);
2153 static void *helper_thread_main(void *data)
2157 fio_mutex_up(startup_mutex);
2160 uint64_t sec = DISK_UTIL_MSEC / 1000;
2161 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2165 gettimeofday(&tv, NULL);
2166 ts.tv_sec = tv.tv_sec + sec;
2167 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2169 if (ts.tv_nsec >= 1000000000ULL) {
2170 ts.tv_nsec -= 1000000000ULL;
2174 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2176 ret = update_io_ticks();
2178 if (helper_do_stat) {
2180 __show_running_run_stats();
2184 print_thread_status();
2190 static int create_helper_thread(void)
2196 pthread_cond_init(&helper_cond, NULL);
2197 pthread_mutex_init(&helper_lock, NULL);
2199 ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2201 log_err("Can't create helper thread: %s\n", strerror(ret));
2205 dprint(FD_MUTEX, "wait on startup_mutex\n");
2206 fio_mutex_down(startup_mutex);
2207 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2211 int fio_backend(void)
2213 struct thread_data *td;
2217 if (load_profile(exec_profile))
2220 exec_profile = NULL;
2226 struct log_params p = {
2227 .log_type = IO_LOG_TYPE_BW,
2230 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2231 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2232 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2235 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2236 if (startup_mutex == NULL)
2241 create_helper_thread();
2243 cgroup_list = smalloc(sizeof(*cgroup_list));
2244 INIT_FLIST_HEAD(cgroup_list);
2248 wait_for_helper_thread_exit();
2253 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2254 struct io_log *log = agg_io_log[i];
2262 for_each_td(td, i) {
2263 fio_options_free(td);
2264 if (td->rusage_sem) {
2265 fio_mutex_remove(td->rusage_sem);
2266 td->rusage_sem = NULL;
2271 cgroup_kill(cgroup_list);
2275 fio_mutex_remove(startup_mutex);