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"
58 static pthread_t helper_thread;
59 static pthread_mutex_t helper_lock;
60 pthread_cond_t helper_cond;
61 int helper_do_stat = 0;
63 static struct fio_mutex *startup_mutex;
64 static struct flist_head *cgroup_list;
65 static char *cgroup_mnt;
66 static int exit_value;
67 static volatile int fio_abort;
68 static unsigned int nr_process = 0;
69 static unsigned int nr_thread = 0;
71 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
74 unsigned int thread_number = 0;
75 unsigned int stat_number = 0;
78 unsigned long done_secs = 0;
79 volatile int helper_exit = 0;
81 #define PAGE_ALIGN(buf) \
82 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
84 #define JOB_START_TIMEOUT (5 * 1000)
86 static void sig_int(int sig)
90 fio_server_got_signal(sig);
92 log_info("\nfio: terminating on signal %d\n", sig);
97 fio_terminate_threads(TERMINATE_ALL);
101 static void sig_show_status(int sig)
103 show_running_run_stats();
106 static void set_sig_handlers(void)
108 struct sigaction act;
110 memset(&act, 0, sizeof(act));
111 act.sa_handler = sig_int;
112 act.sa_flags = SA_RESTART;
113 sigaction(SIGINT, &act, NULL);
115 memset(&act, 0, sizeof(act));
116 act.sa_handler = sig_int;
117 act.sa_flags = SA_RESTART;
118 sigaction(SIGTERM, &act, NULL);
120 /* Windows uses SIGBREAK as a quit signal from other applications */
122 memset(&act, 0, sizeof(act));
123 act.sa_handler = sig_int;
124 act.sa_flags = SA_RESTART;
125 sigaction(SIGBREAK, &act, NULL);
128 memset(&act, 0, sizeof(act));
129 act.sa_handler = sig_show_status;
130 act.sa_flags = SA_RESTART;
131 sigaction(SIGUSR1, &act, NULL);
134 memset(&act, 0, sizeof(act));
135 act.sa_handler = sig_int;
136 act.sa_flags = SA_RESTART;
137 sigaction(SIGPIPE, &act, NULL);
142 * Check if we are above the minimum rate given.
144 static int __check_min_rate(struct thread_data *td, struct timeval *now,
147 unsigned long long bytes = 0;
148 unsigned long iops = 0;
151 unsigned int ratemin = 0;
152 unsigned int rate_iops = 0;
153 unsigned int rate_iops_min = 0;
155 assert(ddir_rw(ddir));
157 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
161 * allow a 2 second settle period in the beginning
163 if (mtime_since(&td->start, now) < 2000)
166 iops += td->this_io_blocks[ddir];
167 bytes += td->this_io_bytes[ddir];
168 ratemin += td->o.ratemin[ddir];
169 rate_iops += td->o.rate_iops[ddir];
170 rate_iops_min += td->o.rate_iops_min[ddir];
173 * if rate blocks is set, sample is running
175 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
176 spent = mtime_since(&td->lastrate[ddir], now);
177 if (spent < td->o.ratecycle)
180 if (td->o.rate[ddir]) {
182 * check bandwidth specified rate
184 if (bytes < td->rate_bytes[ddir]) {
185 log_err("%s: min rate %u not met\n", td->o.name,
190 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
194 if (rate < ratemin ||
195 bytes < td->rate_bytes[ddir]) {
196 log_err("%s: min rate %u not met, got"
197 " %luKB/sec\n", td->o.name,
204 * checks iops specified rate
206 if (iops < rate_iops) {
207 log_err("%s: min iops rate %u not met\n",
208 td->o.name, rate_iops);
212 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
216 if (rate < rate_iops_min ||
217 iops < td->rate_blocks[ddir]) {
218 log_err("%s: min iops rate %u not met,"
219 " got %lu\n", td->o.name,
220 rate_iops_min, rate);
226 td->rate_bytes[ddir] = bytes;
227 td->rate_blocks[ddir] = iops;
228 memcpy(&td->lastrate[ddir], now, sizeof(*now));
232 static int check_min_rate(struct thread_data *td, struct timeval *now)
236 if (td->bytes_done[DDIR_READ])
237 ret |= __check_min_rate(td, now, DDIR_READ);
238 if (td->bytes_done[DDIR_WRITE])
239 ret |= __check_min_rate(td, now, DDIR_WRITE);
240 if (td->bytes_done[DDIR_TRIM])
241 ret |= __check_min_rate(td, now, DDIR_TRIM);
247 * When job exits, we can cancel the in-flight IO if we are using async
248 * io. Attempt to do so.
250 static void cleanup_pending_aio(struct thread_data *td)
255 * get immediately available events, if any
257 r = io_u_queued_complete(td, 0);
262 * now cancel remaining active events
264 if (td->io_ops->cancel) {
268 io_u_qiter(&td->io_u_all, io_u, i) {
269 if (io_u->flags & IO_U_F_FLIGHT) {
270 r = td->io_ops->cancel(td, io_u);
278 r = io_u_queued_complete(td, td->cur_depth);
282 * Helper to handle the final sync of a file. Works just like the normal
283 * io path, just does everything sync.
285 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
287 struct io_u *io_u = __get_io_u(td);
293 io_u->ddir = DDIR_SYNC;
296 if (td_io_prep(td, io_u)) {
302 ret = td_io_queue(td, io_u);
304 td_verror(td, io_u->error, "td_io_queue");
307 } else if (ret == FIO_Q_QUEUED) {
308 if (io_u_queued_complete(td, 1) < 0)
310 } else if (ret == FIO_Q_COMPLETED) {
312 td_verror(td, io_u->error, "td_io_queue");
316 if (io_u_sync_complete(td, io_u) < 0)
318 } else if (ret == FIO_Q_BUSY) {
319 if (td_io_commit(td))
327 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
331 if (fio_file_open(f))
332 return fio_io_sync(td, f);
334 if (td_io_open_file(td, f))
337 ret = fio_io_sync(td, f);
338 td_io_close_file(td, f);
342 static inline void __update_tv_cache(struct thread_data *td)
344 fio_gettime(&td->tv_cache, NULL);
347 static inline void update_tv_cache(struct thread_data *td)
349 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
350 __update_tv_cache(td);
353 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
355 if (in_ramp_time(td))
359 if (utime_since(&td->epoch, t) >= td->o.timeout)
365 static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
370 if (ret < 0 || td->error) {
372 enum error_type_bit eb;
377 eb = td_error_type(ddir, err);
378 if (!(td->o.continue_on_error & (1 << eb)))
381 if (td_non_fatal_error(td, eb, err)) {
383 * Continue with the I/Os in case of
386 update_error_count(td, err);
390 } else if (td->o.fill_device && err == ENOSPC) {
392 * We expect to hit this error if
393 * fill_device option is set.
396 fio_mark_td_terminate(td);
400 * Stop the I/O in case of a fatal
403 update_error_count(td, err);
411 static void check_update_rusage(struct thread_data *td)
413 if (td->update_rusage) {
414 td->update_rusage = 0;
415 update_rusage_stat(td);
416 fio_mutex_up(td->rusage_sem);
420 static int wait_for_completions(struct thread_data *td, struct timeval *time)
422 const int full = queue_full(td);
427 * if the queue is full, we MUST reap at least 1 event
429 min_evts = min(td->o.iodepth_batch_complete, td->cur_depth);
430 if (full && !min_evts)
433 if (time && (__should_check_rate(td, DDIR_READ) ||
434 __should_check_rate(td, DDIR_WRITE) ||
435 __should_check_rate(td, DDIR_TRIM)))
436 fio_gettime(time, NULL);
439 ret = io_u_queued_complete(td, min_evts);
442 } while (full && (td->cur_depth > td->o.iodepth_low));
448 * The main verify engine. Runs over the writes we previously submitted,
449 * reads the blocks back in, and checks the crc/md5 of the data.
451 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
458 dprint(FD_VERIFY, "starting loop\n");
461 * sync io first and invalidate cache, to make sure we really
464 for_each_file(td, f, i) {
465 if (!fio_file_open(f))
467 if (fio_io_sync(td, f))
469 if (file_invalidate_cache(td, f))
473 check_update_rusage(td);
478 td_set_runstate(td, TD_VERIFYING);
481 while (!td->terminate) {
486 check_update_rusage(td);
488 if (runtime_exceeded(td, &td->tv_cache)) {
489 __update_tv_cache(td);
490 if (runtime_exceeded(td, &td->tv_cache)) {
491 fio_mark_td_terminate(td);
496 if (flow_threshold_exceeded(td))
499 if (!td->o.experimental_verify) {
500 io_u = __get_io_u(td);
504 if (get_next_verify(td, io_u)) {
509 if (td_io_prep(td, io_u)) {
514 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
517 while ((io_u = get_io_u(td)) != NULL) {
525 * We are only interested in the places where
526 * we wrote or trimmed IOs. Turn those into
527 * reads for verification purposes.
529 if (io_u->ddir == DDIR_READ) {
531 * Pretend we issued it for rwmix
534 td->io_issues[DDIR_READ]++;
537 } else if (io_u->ddir == DDIR_TRIM) {
538 io_u->ddir = DDIR_READ;
539 io_u->flags |= IO_U_F_TRIMMED;
541 } else if (io_u->ddir == DDIR_WRITE) {
542 io_u->ddir = DDIR_READ;
554 if (verify_state_should_stop(td, io_u)) {
559 if (td->o.verify_async)
560 io_u->end_io = verify_io_u_async;
562 io_u->end_io = verify_io_u;
565 if (!td->o.disable_slat)
566 fio_gettime(&io_u->start_time, NULL);
568 ret = td_io_queue(td, io_u);
570 case FIO_Q_COMPLETED:
573 clear_io_u(td, io_u);
574 } else if (io_u->resid) {
575 int bytes = io_u->xfer_buflen - io_u->resid;
581 td_verror(td, EIO, "full resid");
586 io_u->xfer_buflen = io_u->resid;
587 io_u->xfer_buf += bytes;
588 io_u->offset += bytes;
590 if (ddir_rw(io_u->ddir))
591 td->ts.short_io_u[io_u->ddir]++;
594 if (io_u->offset == f->real_file_size)
597 requeue_io_u(td, &io_u);
600 ret = io_u_sync_complete(td, io_u);
608 requeue_io_u(td, &io_u);
609 ret2 = td_io_commit(td);
615 td_verror(td, -ret, "td_io_queue");
619 if (break_on_this_error(td, ddir, &ret))
623 * if we can queue more, do so. but check if there are
624 * completed io_u's first. Note that we can get BUSY even
625 * without IO queued, if the system is resource starved.
628 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
629 if (full || !td->o.iodepth_batch_complete)
630 ret = wait_for_completions(td, NULL);
636 check_update_rusage(td);
639 min_events = td->cur_depth;
642 ret = io_u_queued_complete(td, min_events);
644 cleanup_pending_aio(td);
646 td_set_runstate(td, TD_RUNNING);
648 dprint(FD_VERIFY, "exiting loop\n");
651 static unsigned int exceeds_number_ios(struct thread_data *td)
653 unsigned long long number_ios;
655 if (!td->o.number_ios)
658 number_ios = ddir_rw_sum(td->io_blocks);
659 number_ios += td->io_u_queued + td->io_u_in_flight;
661 return number_ios >= (td->o.number_ios * td->loops);
664 static int io_issue_bytes_exceeded(struct thread_data *td)
666 unsigned long long bytes, limit;
669 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
670 else if (td_write(td))
671 bytes = td->io_issue_bytes[DDIR_WRITE];
672 else if (td_read(td))
673 bytes = td->io_issue_bytes[DDIR_READ];
675 bytes = td->io_issue_bytes[DDIR_TRIM];
678 limit = td->o.io_limit;
683 return bytes >= limit || exceeds_number_ios(td);
686 static int io_complete_bytes_exceeded(struct thread_data *td)
688 unsigned long long bytes, limit;
691 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
692 else if (td_write(td))
693 bytes = td->this_io_bytes[DDIR_WRITE];
694 else if (td_read(td))
695 bytes = td->this_io_bytes[DDIR_READ];
697 bytes = td->this_io_bytes[DDIR_TRIM];
700 limit = td->o.io_limit;
705 return bytes >= limit || exceeds_number_ios(td);
709 * Main IO worker function. It retrieves io_u's to process and queues
710 * and reaps them, checking for rate and errors along the way.
712 * Returns number of bytes written and trimmed.
714 static uint64_t do_io(struct thread_data *td)
718 uint64_t total_bytes, bytes_issued = 0;
720 if (in_ramp_time(td))
721 td_set_runstate(td, TD_RAMP);
723 td_set_runstate(td, TD_RUNNING);
727 total_bytes = td->o.size;
729 * Allow random overwrite workloads to write up to io_limit
730 * before starting verification phase as 'size' doesn't apply.
732 if (td_write(td) && td_random(td) && td->o.norandommap)
733 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
735 * If verify_backlog is enabled, we'll run the verify in this
736 * handler as well. For that case, we may need up to twice the
739 if (td->o.verify != VERIFY_NONE &&
740 (td_write(td) && td->o.verify_backlog))
741 total_bytes += td->o.size;
743 /* In trimwrite mode, each byte is trimmed and then written, so
744 * allow total_bytes to be twice as big */
745 if (td_trimwrite(td))
746 total_bytes += td->total_io_size;
748 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
749 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
751 struct timeval comp_time;
756 check_update_rusage(td);
758 if (td->terminate || td->done)
763 if (runtime_exceeded(td, &td->tv_cache)) {
764 __update_tv_cache(td);
765 if (runtime_exceeded(td, &td->tv_cache)) {
766 fio_mark_td_terminate(td);
771 if (flow_threshold_exceeded(td))
774 if (bytes_issued >= total_bytes)
778 if (IS_ERR_OR_NULL(io_u)) {
779 int err = PTR_ERR(io_u);
786 if (td->o.latency_target)
794 * Add verification end_io handler if:
795 * - Asked to verify (!td_rw(td))
796 * - Or the io_u is from our verify list (mixed write/ver)
798 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
799 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
801 if (!td->o.verify_pattern_bytes) {
802 io_u->rand_seed = __rand(&td->verify_state);
803 if (sizeof(int) != sizeof(long *))
804 io_u->rand_seed *= __rand(&td->verify_state);
807 if (verify_state_should_stop(td, io_u)) {
812 if (td->o.verify_async)
813 io_u->end_io = verify_io_u_async;
815 io_u->end_io = verify_io_u;
816 td_set_runstate(td, TD_VERIFYING);
817 } else if (in_ramp_time(td))
818 td_set_runstate(td, TD_RAMP);
820 td_set_runstate(td, TD_RUNNING);
823 * Always log IO before it's issued, so we know the specific
824 * order of it. The logged unit will track when the IO has
827 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
829 td->o.verify != VERIFY_NONE &&
830 !td->o.experimental_verify)
831 log_io_piece(td, io_u);
833 ret = td_io_queue(td, io_u);
835 case FIO_Q_COMPLETED:
838 unlog_io_piece(td, io_u);
839 clear_io_u(td, io_u);
840 } else if (io_u->resid) {
841 int bytes = io_u->xfer_buflen - io_u->resid;
842 struct fio_file *f = io_u->file;
844 bytes_issued += bytes;
846 trim_io_piece(td, io_u);
852 unlog_io_piece(td, io_u);
853 td_verror(td, EIO, "full resid");
858 io_u->xfer_buflen = io_u->resid;
859 io_u->xfer_buf += bytes;
860 io_u->offset += bytes;
862 if (ddir_rw(io_u->ddir))
863 td->ts.short_io_u[io_u->ddir]++;
865 if (io_u->offset == f->real_file_size)
868 requeue_io_u(td, &io_u);
871 if (__should_check_rate(td, DDIR_READ) ||
872 __should_check_rate(td, DDIR_WRITE) ||
873 __should_check_rate(td, DDIR_TRIM))
874 fio_gettime(&comp_time, NULL);
876 ret = io_u_sync_complete(td, io_u);
879 bytes_issued += io_u->xfer_buflen;
884 * if the engine doesn't have a commit hook,
885 * the io_u is really queued. if it does have such
886 * a hook, it has to call io_u_queued() itself.
888 if (td->io_ops->commit == NULL)
889 io_u_queued(td, io_u);
890 bytes_issued += io_u->xfer_buflen;
893 unlog_io_piece(td, io_u);
894 requeue_io_u(td, &io_u);
895 ret2 = td_io_commit(td);
905 if (break_on_this_error(td, ddir, &ret))
909 * See if we need to complete some commands. Note that we
910 * can get BUSY even without IO queued, if the system is
914 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
915 if (full || !td->o.iodepth_batch_complete)
916 ret = wait_for_completions(td, &comp_time);
919 if (!ddir_rw_sum(td->bytes_done) &&
920 !(td->io_ops->flags & FIO_NOIO))
923 if (!in_ramp_time(td) && should_check_rate(td)) {
924 if (check_min_rate(td, &comp_time)) {
925 if (exitall_on_terminate)
926 fio_terminate_threads(td->groupid);
927 td_verror(td, EIO, "check_min_rate");
931 if (!in_ramp_time(td) && td->o.latency_target)
932 lat_target_check(td);
934 if (td->o.thinktime) {
935 unsigned long long b;
937 b = ddir_rw_sum(td->io_blocks);
938 if (!(b % td->o.thinktime_blocks)) {
943 if (td->o.thinktime_spin)
944 usec_spin(td->o.thinktime_spin);
946 left = td->o.thinktime - td->o.thinktime_spin;
948 usec_sleep(td, left);
953 check_update_rusage(td);
955 if (td->trim_entries)
956 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
958 if (td->o.fill_device && td->error == ENOSPC) {
960 fio_mark_td_terminate(td);
967 ret = io_u_queued_complete(td, i);
968 if (td->o.fill_device && td->error == ENOSPC)
972 if (should_fsync(td) && td->o.end_fsync) {
973 td_set_runstate(td, TD_FSYNCING);
975 for_each_file(td, f, i) {
976 if (!fio_file_fsync(td, f))
979 log_err("fio: end_fsync failed for file %s\n",
984 cleanup_pending_aio(td);
987 * stop job if we failed doing any IO
989 if (!ddir_rw_sum(td->this_io_bytes))
992 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
995 static void cleanup_io_u(struct thread_data *td)
999 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1001 if (td->io_ops->io_u_free)
1002 td->io_ops->io_u_free(td, io_u);
1004 fio_memfree(io_u, sizeof(*io_u));
1009 io_u_rexit(&td->io_u_requeues);
1010 io_u_qexit(&td->io_u_freelist);
1011 io_u_qexit(&td->io_u_all);
1013 if (td->last_write_comp)
1014 sfree(td->last_write_comp);
1017 static int init_io_u(struct thread_data *td)
1020 unsigned int max_bs, min_write;
1021 int cl_align, i, max_units;
1022 int data_xfer = 1, err;
1025 max_units = td->o.iodepth;
1026 max_bs = td_max_bs(td);
1027 min_write = td->o.min_bs[DDIR_WRITE];
1028 td->orig_buffer_size = (unsigned long long) max_bs
1029 * (unsigned long long) max_units;
1031 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1035 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1036 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1037 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1040 log_err("fio: failed setting up IO queues\n");
1045 * if we may later need to do address alignment, then add any
1046 * possible adjustment here so that we don't cause a buffer
1047 * overflow later. this adjustment may be too much if we get
1048 * lucky and the allocator gives us an aligned address.
1050 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1051 (td->io_ops->flags & FIO_RAWIO))
1052 td->orig_buffer_size += page_mask + td->o.mem_align;
1054 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1057 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1058 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1061 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1062 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1066 if (data_xfer && allocate_io_mem(td))
1069 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1070 (td->io_ops->flags & FIO_RAWIO))
1071 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1073 p = td->orig_buffer;
1075 cl_align = os_cache_line_size();
1077 for (i = 0; i < max_units; i++) {
1083 ptr = fio_memalign(cl_align, sizeof(*io_u));
1085 log_err("fio: unable to allocate aligned memory\n");
1090 memset(io_u, 0, sizeof(*io_u));
1091 INIT_FLIST_HEAD(&io_u->verify_list);
1092 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1096 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1099 io_u_fill_buffer(td, io_u, min_write, max_bs);
1100 if (td_write(td) && td->o.verify_pattern_bytes) {
1102 * Fill the buffer with the pattern if we are
1103 * going to be doing writes.
1105 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1110 io_u->flags = IO_U_F_FREE;
1111 io_u_qpush(&td->io_u_freelist, io_u);
1114 * io_u never leaves this stack, used for iteration of all
1117 io_u_qpush(&td->io_u_all, io_u);
1119 if (td->io_ops->io_u_init) {
1120 int ret = td->io_ops->io_u_init(td, io_u);
1123 log_err("fio: failed to init engine data: %d\n", ret);
1131 if (td->o.verify != VERIFY_NONE) {
1132 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1133 if (!td->last_write_comp) {
1134 log_err("fio: failed to alloc write comp data\n");
1142 static int switch_ioscheduler(struct thread_data *td)
1144 char tmp[256], tmp2[128];
1148 if (td->io_ops->flags & FIO_DISKLESSIO)
1151 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1153 f = fopen(tmp, "r+");
1155 if (errno == ENOENT) {
1156 log_err("fio: os or kernel doesn't support IO scheduler"
1160 td_verror(td, errno, "fopen iosched");
1167 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1168 if (ferror(f) || ret != 1) {
1169 td_verror(td, errno, "fwrite");
1177 * Read back and check that the selected scheduler is now the default.
1179 ret = fread(tmp, sizeof(tmp), 1, f);
1180 if (ferror(f) || ret < 0) {
1181 td_verror(td, errno, "fread");
1185 tmp[sizeof(tmp) - 1] = '\0';
1188 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1189 if (!strstr(tmp, tmp2)) {
1190 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1191 td_verror(td, EINVAL, "iosched_switch");
1200 static int keep_running(struct thread_data *td)
1202 unsigned long long limit;
1206 if (td->o.time_based)
1212 if (exceeds_number_ios(td))
1216 limit = td->o.io_limit;
1220 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1224 * If the difference is less than the minimum IO size, we
1227 diff = limit - ddir_rw_sum(td->io_bytes);
1228 if (diff < td_max_bs(td))
1231 if (fio_files_done(td))
1240 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1242 int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1245 str = malloc(newlen);
1246 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1248 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1251 log_err("fio: exec of cmd <%s> failed\n", str);
1258 * Dry run to compute correct state of numberio for verification.
1260 static uint64_t do_dry_run(struct thread_data *td)
1262 td_set_runstate(td, TD_RUNNING);
1264 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1265 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1269 if (td->terminate || td->done)
1272 io_u = get_io_u(td);
1276 io_u->flags |= IO_U_F_FLIGHT;
1279 if (ddir_rw(acct_ddir(io_u)))
1280 td->io_issues[acct_ddir(io_u)]++;
1281 if (ddir_rw(io_u->ddir)) {
1282 io_u_mark_depth(td, 1);
1283 td->ts.total_io_u[io_u->ddir]++;
1286 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1288 td->o.verify != VERIFY_NONE &&
1289 !td->o.experimental_verify)
1290 log_io_piece(td, io_u);
1292 ret = io_u_sync_complete(td, io_u);
1296 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1300 * Entry point for the thread based jobs. The process based jobs end up
1301 * here as well, after a little setup.
1303 static void *thread_main(void *data)
1305 unsigned long long elapsed;
1306 struct thread_data *td = data;
1307 struct thread_options *o = &td->o;
1308 pthread_condattr_t attr;
1312 if (!o->use_thread) {
1318 fio_local_clock_init(o->use_thread);
1320 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1323 fio_server_send_start(td);
1325 INIT_FLIST_HEAD(&td->io_log_list);
1326 INIT_FLIST_HEAD(&td->io_hist_list);
1327 INIT_FLIST_HEAD(&td->verify_list);
1328 INIT_FLIST_HEAD(&td->trim_list);
1329 INIT_FLIST_HEAD(&td->next_rand_list);
1330 pthread_mutex_init(&td->io_u_lock, NULL);
1331 td->io_hist_tree = RB_ROOT;
1333 pthread_condattr_init(&attr);
1334 pthread_cond_init(&td->verify_cond, &attr);
1335 pthread_cond_init(&td->free_cond, &attr);
1337 td_set_runstate(td, TD_INITIALIZED);
1338 dprint(FD_MUTEX, "up startup_mutex\n");
1339 fio_mutex_up(startup_mutex);
1340 dprint(FD_MUTEX, "wait on td->mutex\n");
1341 fio_mutex_down(td->mutex);
1342 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1345 * A new gid requires privilege, so we need to do this before setting
1348 if (o->gid != -1U && setgid(o->gid)) {
1349 td_verror(td, errno, "setgid");
1352 if (o->uid != -1U && setuid(o->uid)) {
1353 td_verror(td, errno, "setuid");
1358 * If we have a gettimeofday() thread, make sure we exclude that
1359 * thread from this job
1362 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1365 * Set affinity first, in case it has an impact on the memory
1368 if (fio_option_is_set(o, cpumask)) {
1369 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1370 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1372 log_err("fio: no CPUs set\n");
1373 log_err("fio: Try increasing number of available CPUs\n");
1374 td_verror(td, EINVAL, "cpus_split");
1378 ret = fio_setaffinity(td->pid, o->cpumask);
1380 td_verror(td, errno, "cpu_set_affinity");
1385 #ifdef CONFIG_LIBNUMA
1386 /* numa node setup */
1387 if (fio_option_is_set(o, numa_cpunodes) ||
1388 fio_option_is_set(o, numa_memnodes)) {
1389 struct bitmask *mask;
1391 if (numa_available() < 0) {
1392 td_verror(td, errno, "Does not support NUMA API\n");
1396 if (fio_option_is_set(o, numa_cpunodes)) {
1397 mask = numa_parse_nodestring(o->numa_cpunodes);
1398 ret = numa_run_on_node_mask(mask);
1399 numa_free_nodemask(mask);
1401 td_verror(td, errno, \
1402 "numa_run_on_node_mask failed\n");
1407 if (fio_option_is_set(o, numa_memnodes)) {
1409 if (o->numa_memnodes)
1410 mask = numa_parse_nodestring(o->numa_memnodes);
1412 switch (o->numa_mem_mode) {
1413 case MPOL_INTERLEAVE:
1414 numa_set_interleave_mask(mask);
1417 numa_set_membind(mask);
1420 numa_set_localalloc();
1422 case MPOL_PREFERRED:
1423 numa_set_preferred(o->numa_mem_prefer_node);
1431 numa_free_nodemask(mask);
1437 if (fio_pin_memory(td))
1441 * May alter parameters that init_io_u() will use, so we need to
1450 if (o->verify_async && verify_async_init(td))
1453 if (fio_option_is_set(o, ioprio) ||
1454 fio_option_is_set(o, ioprio_class)) {
1455 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1457 td_verror(td, errno, "ioprio_set");
1462 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1466 if (nice(o->nice) == -1 && errno != 0) {
1467 td_verror(td, errno, "nice");
1471 if (o->ioscheduler && switch_ioscheduler(td))
1474 if (!o->create_serialize && setup_files(td))
1480 if (init_random_map(td))
1483 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1487 if (pre_read_files(td) < 0)
1491 if (td->flags & TD_F_COMPRESS_LOG)
1492 tp_init(&td->tp_data);
1494 fio_verify_init(td);
1496 fio_gettime(&td->epoch, NULL);
1497 fio_getrusage(&td->ru_start);
1499 while (keep_running(td)) {
1500 uint64_t verify_bytes;
1502 fio_gettime(&td->start, NULL);
1503 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1504 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1505 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1507 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1508 o->ratemin[DDIR_TRIM]) {
1509 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1510 sizeof(td->bw_sample_time));
1511 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1512 sizeof(td->bw_sample_time));
1513 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1514 sizeof(td->bw_sample_time));
1520 prune_io_piece_log(td);
1522 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1523 verify_bytes = do_dry_run(td);
1525 verify_bytes = do_io(td);
1530 * Make sure we've successfully updated the rusage stats
1531 * before waiting on the stat mutex. Otherwise we could have
1532 * the stat thread holding stat mutex and waiting for
1533 * the rusage_sem, which would never get upped because
1534 * this thread is waiting for the stat mutex.
1536 check_update_rusage(td);
1538 fio_mutex_down(stat_mutex);
1539 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1540 elapsed = mtime_since_now(&td->start);
1541 td->ts.runtime[DDIR_READ] += elapsed;
1543 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1544 elapsed = mtime_since_now(&td->start);
1545 td->ts.runtime[DDIR_WRITE] += elapsed;
1547 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1548 elapsed = mtime_since_now(&td->start);
1549 td->ts.runtime[DDIR_TRIM] += elapsed;
1551 fio_gettime(&td->start, NULL);
1552 fio_mutex_up(stat_mutex);
1554 if (td->error || td->terminate)
1557 if (!o->do_verify ||
1558 o->verify == VERIFY_NONE ||
1559 (td->io_ops->flags & FIO_UNIDIR))
1564 fio_gettime(&td->start, NULL);
1566 do_verify(td, verify_bytes);
1569 * See comment further up for why this is done here.
1571 check_update_rusage(td);
1573 fio_mutex_down(stat_mutex);
1574 td->ts.runtime[DDIR_READ] += mtime_since_now(&td->start);
1575 fio_gettime(&td->start, NULL);
1576 fio_mutex_up(stat_mutex);
1578 if (td->error || td->terminate)
1582 update_rusage_stat(td);
1583 td->ts.total_run_time = mtime_since_now(&td->epoch);
1584 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1585 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1586 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1588 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1589 (td->o.verify != VERIFY_NONE && td_write(td))) {
1590 struct all_io_list *state;
1593 state = get_all_io_list(td->thread_number, &sz);
1595 __verify_save_state(state, "local");
1600 fio_unpin_memory(td);
1602 fio_writeout_logs(td);
1604 if (td->flags & TD_F_COMPRESS_LOG)
1605 tp_exit(&td->tp_data);
1607 if (o->exec_postrun)
1608 exec_string(o, o->exec_postrun, (const char *)"postrun");
1610 if (exitall_on_terminate)
1611 fio_terminate_threads(td->groupid);
1615 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1618 if (o->verify_async)
1619 verify_async_exit(td);
1621 close_and_free_files(td);
1624 cgroup_shutdown(td, &cgroup_mnt);
1625 verify_free_state(td);
1627 if (fio_option_is_set(o, cpumask)) {
1628 ret = fio_cpuset_exit(&o->cpumask);
1630 td_verror(td, ret, "fio_cpuset_exit");
1634 * do this very late, it will log file closing as well
1636 if (o->write_iolog_file)
1637 write_iolog_close(td);
1639 fio_mutex_remove(td->mutex);
1642 td_set_runstate(td, TD_EXITED);
1645 * Do this last after setting our runstate to exited, so we
1646 * know that the stat thread is signaled.
1648 check_update_rusage(td);
1650 return (void *) (uintptr_t) td->error;
1655 * We cannot pass the td data into a forked process, so attach the td and
1656 * pass it to the thread worker.
1658 static int fork_main(int shmid, int offset)
1660 struct thread_data *td;
1663 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1664 data = shmat(shmid, NULL, 0);
1665 if (data == (void *) -1) {
1673 * HP-UX inherits shm mappings?
1678 td = data + offset * sizeof(struct thread_data);
1679 ret = thread_main(td);
1681 return (int) (uintptr_t) ret;
1684 static void dump_td_info(struct thread_data *td)
1686 log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1687 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1688 (unsigned long) time_since_now(&td->terminate_time));
1692 * Run over the job map and reap the threads that have exited, if any.
1694 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1695 unsigned int *m_rate)
1697 struct thread_data *td;
1698 unsigned int cputhreads, realthreads, pending;
1702 * reap exited threads (TD_EXITED -> TD_REAPED)
1704 realthreads = pending = cputhreads = 0;
1705 for_each_td(td, i) {
1709 * ->io_ops is NULL for a thread that has closed its
1712 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1721 if (td->runstate == TD_REAPED)
1723 if (td->o.use_thread) {
1724 if (td->runstate == TD_EXITED) {
1725 td_set_runstate(td, TD_REAPED);
1732 if (td->runstate == TD_EXITED)
1736 * check if someone quit or got killed in an unusual way
1738 ret = waitpid(td->pid, &status, flags);
1740 if (errno == ECHILD) {
1741 log_err("fio: pid=%d disappeared %d\n",
1742 (int) td->pid, td->runstate);
1744 td_set_runstate(td, TD_REAPED);
1748 } else if (ret == td->pid) {
1749 if (WIFSIGNALED(status)) {
1750 int sig = WTERMSIG(status);
1752 if (sig != SIGTERM && sig != SIGUSR2)
1753 log_err("fio: pid=%d, got signal=%d\n",
1754 (int) td->pid, sig);
1756 td_set_runstate(td, TD_REAPED);
1759 if (WIFEXITED(status)) {
1760 if (WEXITSTATUS(status) && !td->error)
1761 td->error = WEXITSTATUS(status);
1763 td_set_runstate(td, TD_REAPED);
1769 * If the job is stuck, do a forceful timeout of it and
1772 if (td->terminate &&
1773 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1775 td_set_runstate(td, TD_REAPED);
1780 * thread is not dead, continue
1786 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1787 (*t_rate) -= ddir_rw_sum(td->o.rate);
1794 done_secs += mtime_since_now(&td->epoch) / 1000;
1795 profile_td_exit(td);
1798 if (*nr_running == cputhreads && !pending && realthreads)
1799 fio_terminate_threads(TERMINATE_ALL);
1802 static int __check_trigger_file(void)
1809 if (stat(trigger_file, &sb))
1812 if (unlink(trigger_file) < 0)
1813 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1819 static int trigger_timedout(void)
1821 if (trigger_timeout)
1822 return time_since_genesis() >= trigger_timeout;
1827 void exec_trigger(const char *cmd)
1836 log_err("fio: failed executing %s trigger\n", cmd);
1839 void check_trigger_file(void)
1841 if (__check_trigger_file() || trigger_timedout()) {
1843 fio_clients_send_trigger(trigger_remote_cmd);
1845 verify_save_state();
1846 fio_terminate_threads(TERMINATE_ALL);
1847 exec_trigger(trigger_cmd);
1852 static int fio_verify_load_state(struct thread_data *td)
1856 if (!td->o.verify_state)
1862 ret = fio_server_get_verify_state(td->o.name,
1863 td->thread_number - 1, &data);
1865 verify_convert_assign_state(td, data);
1867 ret = verify_load_state(td, "local");
1872 static void do_usleep(unsigned int usecs)
1874 check_for_running_stats();
1875 check_trigger_file();
1880 * Main function for kicking off and reaping jobs, as needed.
1882 static void run_threads(void)
1884 struct thread_data *td;
1885 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1888 if (fio_gtod_offload && fio_start_gtod_thread())
1891 fio_idle_prof_init();
1895 nr_thread = nr_process = 0;
1896 for_each_td(td, i) {
1897 if (td->o.use_thread)
1903 if (output_format == FIO_OUTPUT_NORMAL) {
1904 log_info("Starting ");
1906 log_info("%d thread%s", nr_thread,
1907 nr_thread > 1 ? "s" : "");
1911 log_info("%d process%s", nr_process,
1912 nr_process > 1 ? "es" : "");
1918 todo = thread_number;
1921 m_rate = t_rate = 0;
1923 for_each_td(td, i) {
1924 print_status_init(td->thread_number - 1);
1926 if (!td->o.create_serialize)
1929 if (fio_verify_load_state(td))
1933 * do file setup here so it happens sequentially,
1934 * we don't want X number of threads getting their
1935 * client data interspersed on disk
1937 if (setup_files(td)) {
1941 log_err("fio: pid=%d, err=%d/%s\n",
1942 (int) td->pid, td->error, td->verror);
1943 td_set_runstate(td, TD_REAPED);
1950 * for sharing to work, each job must always open
1951 * its own files. so close them, if we opened them
1954 for_each_file(td, f, j) {
1955 if (fio_file_open(f))
1956 td_io_close_file(td, f);
1961 /* start idle threads before io threads start to run */
1962 fio_idle_prof_start();
1967 struct thread_data *map[REAL_MAX_JOBS];
1968 struct timeval this_start;
1969 int this_jobs = 0, left;
1972 * create threads (TD_NOT_CREATED -> TD_CREATED)
1974 for_each_td(td, i) {
1975 if (td->runstate != TD_NOT_CREATED)
1979 * never got a chance to start, killed by other
1980 * thread for some reason
1982 if (td->terminate) {
1987 if (td->o.start_delay) {
1988 spent = utime_since_genesis();
1990 if (td->o.start_delay > spent)
1994 if (td->o.stonewall && (nr_started || nr_running)) {
1995 dprint(FD_PROCESS, "%s: stonewall wait\n",
2002 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2003 td->update_rusage = 0;
2006 * Set state to created. Thread will transition
2007 * to TD_INITIALIZED when it's done setting up.
2009 td_set_runstate(td, TD_CREATED);
2010 map[this_jobs++] = td;
2013 if (td->o.use_thread) {
2016 dprint(FD_PROCESS, "will pthread_create\n");
2017 ret = pthread_create(&td->thread, NULL,
2020 log_err("pthread_create: %s\n",
2025 ret = pthread_detach(td->thread);
2027 log_err("pthread_detach: %s",
2031 dprint(FD_PROCESS, "will fork\n");
2034 int ret = fork_main(shm_id, i);
2037 } else if (i == fio_debug_jobno)
2038 *fio_debug_jobp = pid;
2040 dprint(FD_MUTEX, "wait on startup_mutex\n");
2041 if (fio_mutex_down_timeout(startup_mutex, 10)) {
2042 log_err("fio: job startup hung? exiting.\n");
2043 fio_terminate_threads(TERMINATE_ALL);
2048 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2052 * Wait for the started threads to transition to
2055 fio_gettime(&this_start, NULL);
2057 while (left && !fio_abort) {
2058 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2063 for (i = 0; i < this_jobs; i++) {
2067 if (td->runstate == TD_INITIALIZED) {
2070 } else if (td->runstate >= TD_EXITED) {
2074 nr_running++; /* work-around... */
2080 log_err("fio: %d job%s failed to start\n", left,
2081 left > 1 ? "s" : "");
2082 for (i = 0; i < this_jobs; i++) {
2086 kill(td->pid, SIGTERM);
2092 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2094 for_each_td(td, i) {
2095 if (td->runstate != TD_INITIALIZED)
2098 if (in_ramp_time(td))
2099 td_set_runstate(td, TD_RAMP);
2101 td_set_runstate(td, TD_RUNNING);
2104 m_rate += ddir_rw_sum(td->o.ratemin);
2105 t_rate += ddir_rw_sum(td->o.rate);
2107 fio_mutex_up(td->mutex);
2110 reap_threads(&nr_running, &t_rate, &m_rate);
2116 while (nr_running) {
2117 reap_threads(&nr_running, &t_rate, &m_rate);
2121 fio_idle_prof_stop();
2126 static void wait_for_helper_thread_exit(void)
2131 pthread_cond_signal(&helper_cond);
2132 pthread_join(helper_thread, &ret);
2135 static void free_disk_util(void)
2137 disk_util_prune_entries();
2139 pthread_cond_destroy(&helper_cond);
2142 static void *helper_thread_main(void *data)
2146 fio_mutex_up(startup_mutex);
2149 uint64_t sec = DISK_UTIL_MSEC / 1000;
2150 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2154 gettimeofday(&tv, NULL);
2155 ts.tv_sec = tv.tv_sec + sec;
2156 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2158 if (ts.tv_nsec >= 1000000000ULL) {
2159 ts.tv_nsec -= 1000000000ULL;
2163 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2165 ret = update_io_ticks();
2167 if (helper_do_stat) {
2169 __show_running_run_stats();
2173 print_thread_status();
2179 static int create_helper_thread(void)
2185 pthread_cond_init(&helper_cond, NULL);
2186 pthread_mutex_init(&helper_lock, NULL);
2188 ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2190 log_err("Can't create helper thread: %s\n", strerror(ret));
2194 dprint(FD_MUTEX, "wait on startup_mutex\n");
2195 fio_mutex_down(startup_mutex);
2196 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2200 int fio_backend(void)
2202 struct thread_data *td;
2206 if (load_profile(exec_profile))
2209 exec_profile = NULL;
2215 struct log_params p = {
2216 .log_type = IO_LOG_TYPE_BW,
2219 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2220 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2221 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2224 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2225 if (startup_mutex == NULL)
2230 create_helper_thread();
2232 cgroup_list = smalloc(sizeof(*cgroup_list));
2233 INIT_FLIST_HEAD(cgroup_list);
2237 wait_for_helper_thread_exit();
2242 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2243 struct io_log *log = agg_io_log[i];
2251 for_each_td(td, i) {
2252 fio_options_free(td);
2253 if (td->rusage_sem) {
2254 fio_mutex_remove(td->rusage_sem);
2255 td->rusage_sem = NULL;
2260 cgroup_kill(cgroup_list);
2264 fio_mutex_remove(startup_mutex);