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, td->cur_depth);
450 if ((full && !min_evts) || !td->o.iodepth_batch_complete)
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))
555 * The main verify engine. Runs over the writes we previously submitted,
556 * reads the blocks back in, and checks the crc/md5 of the data.
558 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
565 dprint(FD_VERIFY, "starting loop\n");
568 * sync io first and invalidate cache, to make sure we really
571 for_each_file(td, f, i) {
572 if (!fio_file_open(f))
574 if (fio_io_sync(td, f))
576 if (file_invalidate_cache(td, f))
580 check_update_rusage(td);
585 td_set_runstate(td, TD_VERIFYING);
588 while (!td->terminate) {
593 check_update_rusage(td);
595 if (runtime_exceeded(td, &td->tv_cache)) {
596 __update_tv_cache(td);
597 if (runtime_exceeded(td, &td->tv_cache)) {
598 fio_mark_td_terminate(td);
603 if (flow_threshold_exceeded(td))
606 if (!td->o.experimental_verify) {
607 io_u = __get_io_u(td);
611 if (get_next_verify(td, io_u)) {
616 if (td_io_prep(td, io_u)) {
621 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
624 while ((io_u = get_io_u(td)) != NULL) {
632 * We are only interested in the places where
633 * we wrote or trimmed IOs. Turn those into
634 * reads for verification purposes.
636 if (io_u->ddir == DDIR_READ) {
638 * Pretend we issued it for rwmix
641 td->io_issues[DDIR_READ]++;
644 } else if (io_u->ddir == DDIR_TRIM) {
645 io_u->ddir = DDIR_READ;
646 io_u_set(io_u, IO_U_F_TRIMMED);
648 } else if (io_u->ddir == DDIR_WRITE) {
649 io_u->ddir = DDIR_READ;
661 if (verify_state_should_stop(td, io_u)) {
666 if (td->o.verify_async)
667 io_u->end_io = verify_io_u_async;
669 io_u->end_io = verify_io_u;
672 if (!td->o.disable_slat)
673 fio_gettime(&io_u->start_time, NULL);
675 ret = td_io_queue(td, io_u);
677 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
681 * if we can queue more, do so. but check if there are
682 * completed io_u's first. Note that we can get BUSY even
683 * without IO queued, if the system is resource starved.
686 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
687 if (full || !td->o.iodepth_batch_complete)
688 ret = wait_for_completions(td, NULL);
694 check_update_rusage(td);
697 min_events = td->cur_depth;
700 ret = io_u_queued_complete(td, min_events);
702 cleanup_pending_aio(td);
704 td_set_runstate(td, TD_RUNNING);
706 dprint(FD_VERIFY, "exiting loop\n");
709 static unsigned int exceeds_number_ios(struct thread_data *td)
711 unsigned long long number_ios;
713 if (!td->o.number_ios)
716 number_ios = ddir_rw_sum(td->io_blocks);
717 number_ios += td->io_u_queued + td->io_u_in_flight;
719 return number_ios >= (td->o.number_ios * td->loops);
722 static int io_issue_bytes_exceeded(struct thread_data *td)
724 unsigned long long bytes, limit;
727 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
728 else if (td_write(td))
729 bytes = td->io_issue_bytes[DDIR_WRITE];
730 else if (td_read(td))
731 bytes = td->io_issue_bytes[DDIR_READ];
733 bytes = td->io_issue_bytes[DDIR_TRIM];
736 limit = td->o.io_limit;
741 return bytes >= limit || exceeds_number_ios(td);
744 static int io_complete_bytes_exceeded(struct thread_data *td)
746 unsigned long long bytes, limit;
749 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
750 else if (td_write(td))
751 bytes = td->this_io_bytes[DDIR_WRITE];
752 else if (td_read(td))
753 bytes = td->this_io_bytes[DDIR_READ];
755 bytes = td->this_io_bytes[DDIR_TRIM];
758 limit = td->o.io_limit;
763 return bytes >= limit || exceeds_number_ios(td);
767 * used to calculate the next io time for rate control
770 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
772 uint64_t secs, remainder, bps, bytes;
774 assert(!(td->flags & TD_F_CHILD));
775 bytes = td->rate_io_issue_bytes[ddir];
776 bps = td->rate_bps[ddir];
779 remainder = bytes % bps;
780 return remainder * 1000000 / bps + secs * 1000000;
786 * Main IO worker function. It retrieves io_u's to process and queues
787 * and reaps them, checking for rate and errors along the way.
789 * Returns number of bytes written and trimmed.
791 static uint64_t do_io(struct thread_data *td)
795 uint64_t total_bytes, bytes_issued = 0;
797 if (in_ramp_time(td))
798 td_set_runstate(td, TD_RAMP);
800 td_set_runstate(td, TD_RUNNING);
804 total_bytes = td->o.size;
806 * Allow random overwrite workloads to write up to io_limit
807 * before starting verification phase as 'size' doesn't apply.
809 if (td_write(td) && td_random(td) && td->o.norandommap)
810 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
812 * If verify_backlog is enabled, we'll run the verify in this
813 * handler as well. For that case, we may need up to twice the
816 if (td->o.verify != VERIFY_NONE &&
817 (td_write(td) && td->o.verify_backlog))
818 total_bytes += td->o.size;
820 /* In trimwrite mode, each byte is trimmed and then written, so
821 * allow total_bytes to be twice as big */
822 if (td_trimwrite(td))
823 total_bytes += td->total_io_size;
825 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
826 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
828 struct timeval comp_time;
833 check_update_rusage(td);
835 if (td->terminate || td->done)
840 if (runtime_exceeded(td, &td->tv_cache)) {
841 __update_tv_cache(td);
842 if (runtime_exceeded(td, &td->tv_cache)) {
843 fio_mark_td_terminate(td);
848 if (flow_threshold_exceeded(td))
851 if (bytes_issued >= total_bytes)
855 if (IS_ERR_OR_NULL(io_u)) {
856 int err = PTR_ERR(io_u);
863 if (td->o.latency_target)
871 * Add verification end_io handler if:
872 * - Asked to verify (!td_rw(td))
873 * - Or the io_u is from our verify list (mixed write/ver)
875 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
876 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
878 if (!td->o.verify_pattern_bytes) {
879 io_u->rand_seed = __rand(&td->verify_state);
880 if (sizeof(int) != sizeof(long *))
881 io_u->rand_seed *= __rand(&td->verify_state);
884 if (verify_state_should_stop(td, io_u)) {
889 if (td->o.verify_async)
890 io_u->end_io = verify_io_u_async;
892 io_u->end_io = verify_io_u;
893 td_set_runstate(td, TD_VERIFYING);
894 } else if (in_ramp_time(td))
895 td_set_runstate(td, TD_RAMP);
897 td_set_runstate(td, TD_RUNNING);
900 * Always log IO before it's issued, so we know the specific
901 * order of it. The logged unit will track when the IO has
904 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
906 td->o.verify != VERIFY_NONE &&
907 !td->o.experimental_verify)
908 log_io_piece(td, io_u);
910 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
913 ret = workqueue_enqueue(&td->io_wq, io_u);
915 if (should_check_rate(td))
916 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
919 ret = td_io_queue(td, io_u);
921 if (should_check_rate(td))
922 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
924 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
928 * See if we need to complete some commands. Note that
929 * we can get BUSY even without IO queued, if the
930 * system is resource starved.
933 full = queue_full(td) ||
934 (ret == FIO_Q_BUSY && td->cur_depth);
935 if (full || !td->o.iodepth_batch_complete)
936 ret = wait_for_completions(td, &comp_time);
940 if (!ddir_rw_sum(td->bytes_done) &&
941 !(td->io_ops->flags & FIO_NOIO))
944 if (!in_ramp_time(td) && should_check_rate(td)) {
945 if (check_min_rate(td, &comp_time)) {
946 if (exitall_on_terminate)
947 fio_terminate_threads(td->groupid);
948 td_verror(td, EIO, "check_min_rate");
952 if (!in_ramp_time(td) && td->o.latency_target)
953 lat_target_check(td);
955 if (td->o.thinktime) {
956 unsigned long long b;
958 b = ddir_rw_sum(td->io_blocks);
959 if (!(b % td->o.thinktime_blocks)) {
964 if (td->o.thinktime_spin)
965 usec_spin(td->o.thinktime_spin);
967 left = td->o.thinktime - td->o.thinktime_spin;
969 usec_sleep(td, left);
974 check_update_rusage(td);
976 if (td->trim_entries)
977 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
979 if (td->o.fill_device && td->error == ENOSPC) {
981 fio_mark_td_terminate(td);
986 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
987 workqueue_flush(&td->io_wq);
993 ret = io_u_queued_complete(td, i);
994 if (td->o.fill_device && td->error == ENOSPC)
998 if (should_fsync(td) && td->o.end_fsync) {
999 td_set_runstate(td, TD_FSYNCING);
1001 for_each_file(td, f, i) {
1002 if (!fio_file_fsync(td, f))
1005 log_err("fio: end_fsync failed for file %s\n",
1010 cleanup_pending_aio(td);
1013 * stop job if we failed doing any IO
1015 if (!ddir_rw_sum(td->this_io_bytes))
1018 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1021 static void cleanup_io_u(struct thread_data *td)
1025 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1027 if (td->io_ops->io_u_free)
1028 td->io_ops->io_u_free(td, io_u);
1030 fio_memfree(io_u, sizeof(*io_u));
1035 io_u_rexit(&td->io_u_requeues);
1036 io_u_qexit(&td->io_u_freelist);
1037 io_u_qexit(&td->io_u_all);
1039 if (td->last_write_comp)
1040 sfree(td->last_write_comp);
1043 static int init_io_u(struct thread_data *td)
1046 unsigned int max_bs, min_write;
1047 int cl_align, i, max_units;
1048 int data_xfer = 1, err;
1051 max_units = td->o.iodepth;
1052 max_bs = td_max_bs(td);
1053 min_write = td->o.min_bs[DDIR_WRITE];
1054 td->orig_buffer_size = (unsigned long long) max_bs
1055 * (unsigned long long) max_units;
1057 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
1061 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1062 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1063 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1066 log_err("fio: failed setting up IO queues\n");
1071 * if we may later need to do address alignment, then add any
1072 * possible adjustment here so that we don't cause a buffer
1073 * overflow later. this adjustment may be too much if we get
1074 * lucky and the allocator gives us an aligned address.
1076 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1077 (td->io_ops->flags & FIO_RAWIO))
1078 td->orig_buffer_size += page_mask + td->o.mem_align;
1080 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1083 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1084 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1087 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1088 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1092 if (data_xfer && allocate_io_mem(td))
1095 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1096 (td->io_ops->flags & FIO_RAWIO))
1097 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1099 p = td->orig_buffer;
1101 cl_align = os_cache_line_size();
1103 for (i = 0; i < max_units; i++) {
1109 ptr = fio_memalign(cl_align, sizeof(*io_u));
1111 log_err("fio: unable to allocate aligned memory\n");
1116 memset(io_u, 0, sizeof(*io_u));
1117 INIT_FLIST_HEAD(&io_u->verify_list);
1118 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1122 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1125 io_u_fill_buffer(td, io_u, min_write, max_bs);
1126 if (td_write(td) && td->o.verify_pattern_bytes) {
1128 * Fill the buffer with the pattern if we are
1129 * going to be doing writes.
1131 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1136 io_u->flags = IO_U_F_FREE;
1137 io_u_qpush(&td->io_u_freelist, io_u);
1140 * io_u never leaves this stack, used for iteration of all
1143 io_u_qpush(&td->io_u_all, io_u);
1145 if (td->io_ops->io_u_init) {
1146 int ret = td->io_ops->io_u_init(td, io_u);
1149 log_err("fio: failed to init engine data: %d\n", ret);
1157 if (td->o.verify != VERIFY_NONE) {
1158 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1159 if (!td->last_write_comp) {
1160 log_err("fio: failed to alloc write comp data\n");
1168 static int switch_ioscheduler(struct thread_data *td)
1170 char tmp[256], tmp2[128];
1174 if (td->io_ops->flags & FIO_DISKLESSIO)
1177 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1179 f = fopen(tmp, "r+");
1181 if (errno == ENOENT) {
1182 log_err("fio: os or kernel doesn't support IO scheduler"
1186 td_verror(td, errno, "fopen iosched");
1193 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1194 if (ferror(f) || ret != 1) {
1195 td_verror(td, errno, "fwrite");
1203 * Read back and check that the selected scheduler is now the default.
1205 memset(tmp, 0, sizeof(tmp));
1206 ret = fread(tmp, sizeof(tmp), 1, f);
1207 if (ferror(f) || ret < 0) {
1208 td_verror(td, errno, "fread");
1213 * either a list of io schedulers or "none\n" is expected.
1215 tmp[strlen(tmp) - 1] = '\0';
1218 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1219 if (!strstr(tmp, tmp2)) {
1220 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1221 td_verror(td, EINVAL, "iosched_switch");
1230 static int keep_running(struct thread_data *td)
1232 unsigned long long limit;
1236 if (td->o.time_based)
1242 if (exceeds_number_ios(td))
1246 limit = td->o.io_limit;
1250 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1254 * If the difference is less than the minimum IO size, we
1257 diff = limit - ddir_rw_sum(td->io_bytes);
1258 if (diff < td_max_bs(td))
1261 if (fio_files_done(td))
1270 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1272 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1276 str = malloc(newlen);
1277 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1279 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1282 log_err("fio: exec of cmd <%s> failed\n", str);
1289 * Dry run to compute correct state of numberio for verification.
1291 static uint64_t do_dry_run(struct thread_data *td)
1293 td_set_runstate(td, TD_RUNNING);
1295 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1296 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1300 if (td->terminate || td->done)
1303 io_u = get_io_u(td);
1307 io_u_set(io_u, IO_U_F_FLIGHT);
1310 if (ddir_rw(acct_ddir(io_u)))
1311 td->io_issues[acct_ddir(io_u)]++;
1312 if (ddir_rw(io_u->ddir)) {
1313 io_u_mark_depth(td, 1);
1314 td->ts.total_io_u[io_u->ddir]++;
1317 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1319 td->o.verify != VERIFY_NONE &&
1320 !td->o.experimental_verify)
1321 log_io_piece(td, io_u);
1323 ret = io_u_sync_complete(td, io_u);
1327 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1330 static void io_workqueue_fn(struct thread_data *td, struct io_u *io_u)
1332 const enum fio_ddir ddir = io_u->ddir;
1335 dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
1337 io_u_set(io_u, IO_U_F_NO_FILE_PUT);
1341 ret = td_io_queue(td, io_u);
1343 dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
1345 io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
1347 if (ret == FIO_Q_QUEUED)
1348 ret = io_u_queued_complete(td, 1);
1354 * Entry point for the thread based jobs. The process based jobs end up
1355 * here as well, after a little setup.
1357 static void *thread_main(void *data)
1359 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1360 struct thread_data *td = data;
1361 struct thread_options *o = &td->o;
1362 pthread_condattr_t attr;
1366 if (!o->use_thread) {
1372 fio_local_clock_init(o->use_thread);
1374 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1377 fio_server_send_start(td);
1379 INIT_FLIST_HEAD(&td->io_log_list);
1380 INIT_FLIST_HEAD(&td->io_hist_list);
1381 INIT_FLIST_HEAD(&td->verify_list);
1382 INIT_FLIST_HEAD(&td->trim_list);
1383 INIT_FLIST_HEAD(&td->next_rand_list);
1384 pthread_mutex_init(&td->io_u_lock, NULL);
1385 td->io_hist_tree = RB_ROOT;
1387 pthread_condattr_init(&attr);
1388 pthread_cond_init(&td->verify_cond, &attr);
1389 pthread_cond_init(&td->free_cond, &attr);
1391 td_set_runstate(td, TD_INITIALIZED);
1392 dprint(FD_MUTEX, "up startup_mutex\n");
1393 fio_mutex_up(startup_mutex);
1394 dprint(FD_MUTEX, "wait on td->mutex\n");
1395 fio_mutex_down(td->mutex);
1396 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1399 * A new gid requires privilege, so we need to do this before setting
1402 if (o->gid != -1U && setgid(o->gid)) {
1403 td_verror(td, errno, "setgid");
1406 if (o->uid != -1U && setuid(o->uid)) {
1407 td_verror(td, errno, "setuid");
1412 * If we have a gettimeofday() thread, make sure we exclude that
1413 * thread from this job
1416 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1419 * Set affinity first, in case it has an impact on the memory
1422 if (fio_option_is_set(o, cpumask)) {
1423 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1424 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1426 log_err("fio: no CPUs set\n");
1427 log_err("fio: Try increasing number of available CPUs\n");
1428 td_verror(td, EINVAL, "cpus_split");
1432 ret = fio_setaffinity(td->pid, o->cpumask);
1434 td_verror(td, errno, "cpu_set_affinity");
1439 #ifdef CONFIG_LIBNUMA
1440 /* numa node setup */
1441 if (fio_option_is_set(o, numa_cpunodes) ||
1442 fio_option_is_set(o, numa_memnodes)) {
1443 struct bitmask *mask;
1445 if (numa_available() < 0) {
1446 td_verror(td, errno, "Does not support NUMA API\n");
1450 if (fio_option_is_set(o, numa_cpunodes)) {
1451 mask = numa_parse_nodestring(o->numa_cpunodes);
1452 ret = numa_run_on_node_mask(mask);
1453 numa_free_nodemask(mask);
1455 td_verror(td, errno, \
1456 "numa_run_on_node_mask failed\n");
1461 if (fio_option_is_set(o, numa_memnodes)) {
1463 if (o->numa_memnodes)
1464 mask = numa_parse_nodestring(o->numa_memnodes);
1466 switch (o->numa_mem_mode) {
1467 case MPOL_INTERLEAVE:
1468 numa_set_interleave_mask(mask);
1471 numa_set_membind(mask);
1474 numa_set_localalloc();
1476 case MPOL_PREFERRED:
1477 numa_set_preferred(o->numa_mem_prefer_node);
1485 numa_free_nodemask(mask);
1491 if (fio_pin_memory(td))
1495 * May alter parameters that init_io_u() will use, so we need to
1504 if (o->verify_async && verify_async_init(td))
1507 if (fio_option_is_set(o, ioprio) ||
1508 fio_option_is_set(o, ioprio_class)) {
1509 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1511 td_verror(td, errno, "ioprio_set");
1516 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1520 if (nice(o->nice) == -1 && errno != 0) {
1521 td_verror(td, errno, "nice");
1525 if (o->ioscheduler && switch_ioscheduler(td))
1528 if (!o->create_serialize && setup_files(td))
1534 if (init_random_map(td))
1537 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1541 if (pre_read_files(td) < 0)
1545 if (td->flags & TD_F_COMPRESS_LOG)
1546 tp_init(&td->tp_data);
1548 fio_verify_init(td);
1550 if ((o->io_submit_mode == IO_MODE_OFFLOAD) &&
1551 workqueue_init(td, &td->io_wq, io_workqueue_fn, td->o.iodepth))
1554 fio_gettime(&td->epoch, NULL);
1555 fio_getrusage(&td->ru_start);
1557 while (keep_running(td)) {
1558 uint64_t verify_bytes;
1560 fio_gettime(&td->start, NULL);
1561 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1562 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1563 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1565 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1566 o->ratemin[DDIR_TRIM]) {
1567 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1568 sizeof(td->bw_sample_time));
1569 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1570 sizeof(td->bw_sample_time));
1571 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1572 sizeof(td->bw_sample_time));
1578 prune_io_piece_log(td);
1580 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1581 verify_bytes = do_dry_run(td);
1583 verify_bytes = do_io(td);
1588 * Make sure we've successfully updated the rusage stats
1589 * before waiting on the stat mutex. Otherwise we could have
1590 * the stat thread holding stat mutex and waiting for
1591 * the rusage_sem, which would never get upped because
1592 * this thread is waiting for the stat mutex.
1594 check_update_rusage(td);
1596 fio_mutex_down(stat_mutex);
1597 if (td_read(td) && td->io_bytes[DDIR_READ])
1598 update_runtime(td, elapsed_us, DDIR_READ);
1599 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1600 update_runtime(td, elapsed_us, DDIR_WRITE);
1601 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1602 update_runtime(td, elapsed_us, DDIR_TRIM);
1603 fio_gettime(&td->start, NULL);
1604 fio_mutex_up(stat_mutex);
1606 if (td->error || td->terminate)
1609 if (!o->do_verify ||
1610 o->verify == VERIFY_NONE ||
1611 (td->io_ops->flags & FIO_UNIDIR))
1616 fio_gettime(&td->start, NULL);
1618 do_verify(td, verify_bytes);
1621 * See comment further up for why this is done here.
1623 check_update_rusage(td);
1625 fio_mutex_down(stat_mutex);
1626 update_runtime(td, elapsed_us, DDIR_READ);
1627 fio_gettime(&td->start, NULL);
1628 fio_mutex_up(stat_mutex);
1630 if (td->error || td->terminate)
1634 update_rusage_stat(td);
1635 td->ts.total_run_time = mtime_since_now(&td->epoch);
1636 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1637 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1638 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1640 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1641 (td->o.verify != VERIFY_NONE && td_write(td)))
1642 verify_save_state(td->thread_number);
1644 fio_unpin_memory(td);
1646 fio_writeout_logs(td);
1648 if (o->io_submit_mode == IO_MODE_OFFLOAD)
1649 workqueue_exit(&td->io_wq);
1651 if (td->flags & TD_F_COMPRESS_LOG)
1652 tp_exit(&td->tp_data);
1654 if (o->exec_postrun)
1655 exec_string(o, o->exec_postrun, (const char *)"postrun");
1657 if (exitall_on_terminate)
1658 fio_terminate_threads(td->groupid);
1662 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1665 if (o->verify_async)
1666 verify_async_exit(td);
1668 close_and_free_files(td);
1671 cgroup_shutdown(td, &cgroup_mnt);
1672 verify_free_state(td);
1674 if (fio_option_is_set(o, cpumask)) {
1675 ret = fio_cpuset_exit(&o->cpumask);
1677 td_verror(td, ret, "fio_cpuset_exit");
1681 * do this very late, it will log file closing as well
1683 if (o->write_iolog_file)
1684 write_iolog_close(td);
1686 fio_mutex_remove(td->mutex);
1689 td_set_runstate(td, TD_EXITED);
1692 * Do this last after setting our runstate to exited, so we
1693 * know that the stat thread is signaled.
1695 check_update_rusage(td);
1697 return (void *) (uintptr_t) td->error;
1702 * We cannot pass the td data into a forked process, so attach the td and
1703 * pass it to the thread worker.
1705 static int fork_main(int shmid, int offset)
1707 struct thread_data *td;
1710 #if !defined(__hpux) && !defined(CONFIG_NO_SHM)
1711 data = shmat(shmid, NULL, 0);
1712 if (data == (void *) -1) {
1720 * HP-UX inherits shm mappings?
1725 td = data + offset * sizeof(struct thread_data);
1726 ret = thread_main(td);
1728 return (int) (uintptr_t) ret;
1731 static void dump_td_info(struct thread_data *td)
1733 log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1734 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1735 (unsigned long) time_since_now(&td->terminate_time));
1739 * Run over the job map and reap the threads that have exited, if any.
1741 static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1742 unsigned int *m_rate)
1744 struct thread_data *td;
1745 unsigned int cputhreads, realthreads, pending;
1749 * reap exited threads (TD_EXITED -> TD_REAPED)
1751 realthreads = pending = cputhreads = 0;
1752 for_each_td(td, i) {
1756 * ->io_ops is NULL for a thread that has closed its
1759 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1768 if (td->runstate == TD_REAPED)
1770 if (td->o.use_thread) {
1771 if (td->runstate == TD_EXITED) {
1772 td_set_runstate(td, TD_REAPED);
1779 if (td->runstate == TD_EXITED)
1783 * check if someone quit or got killed in an unusual way
1785 ret = waitpid(td->pid, &status, flags);
1787 if (errno == ECHILD) {
1788 log_err("fio: pid=%d disappeared %d\n",
1789 (int) td->pid, td->runstate);
1791 td_set_runstate(td, TD_REAPED);
1795 } else if (ret == td->pid) {
1796 if (WIFSIGNALED(status)) {
1797 int sig = WTERMSIG(status);
1799 if (sig != SIGTERM && sig != SIGUSR2)
1800 log_err("fio: pid=%d, got signal=%d\n",
1801 (int) td->pid, sig);
1803 td_set_runstate(td, TD_REAPED);
1806 if (WIFEXITED(status)) {
1807 if (WEXITSTATUS(status) && !td->error)
1808 td->error = WEXITSTATUS(status);
1810 td_set_runstate(td, TD_REAPED);
1816 * If the job is stuck, do a forceful timeout of it and
1819 if (td->terminate &&
1820 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1822 td_set_runstate(td, TD_REAPED);
1827 * thread is not dead, continue
1833 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1834 (*t_rate) -= ddir_rw_sum(td->o.rate);
1841 done_secs += mtime_since_now(&td->epoch) / 1000;
1842 profile_td_exit(td);
1845 if (*nr_running == cputhreads && !pending && realthreads)
1846 fio_terminate_threads(TERMINATE_ALL);
1849 static int __check_trigger_file(void)
1856 if (stat(trigger_file, &sb))
1859 if (unlink(trigger_file) < 0)
1860 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1866 static int trigger_timedout(void)
1868 if (trigger_timeout)
1869 return time_since_genesis() >= trigger_timeout;
1874 void exec_trigger(const char *cmd)
1883 log_err("fio: failed executing %s trigger\n", cmd);
1886 void check_trigger_file(void)
1888 if (__check_trigger_file() || trigger_timedout()) {
1890 fio_clients_send_trigger(trigger_remote_cmd);
1892 verify_save_state(IO_LIST_ALL);
1893 fio_terminate_threads(TERMINATE_ALL);
1894 exec_trigger(trigger_cmd);
1899 static int fio_verify_load_state(struct thread_data *td)
1903 if (!td->o.verify_state)
1910 ret = fio_server_get_verify_state(td->o.name,
1911 td->thread_number - 1, &data, &ver);
1913 verify_convert_assign_state(td, data, ver);
1915 ret = verify_load_state(td, "local");
1920 static void do_usleep(unsigned int usecs)
1922 check_for_running_stats();
1923 check_trigger_file();
1927 static int check_mount_writes(struct thread_data *td)
1932 if (!td_write(td) || td->o.allow_mounted_write)
1935 for_each_file(td, f, i) {
1936 if (f->filetype != FIO_TYPE_BD)
1938 if (device_is_mounted(f->file_name))
1944 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
1949 * Main function for kicking off and reaping jobs, as needed.
1951 static void run_threads(void)
1953 struct thread_data *td;
1954 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1957 if (fio_gtod_offload && fio_start_gtod_thread())
1960 fio_idle_prof_init();
1964 nr_thread = nr_process = 0;
1965 for_each_td(td, i) {
1966 if (check_mount_writes(td))
1968 if (td->o.use_thread)
1974 if (output_format == FIO_OUTPUT_NORMAL) {
1975 log_info("Starting ");
1977 log_info("%d thread%s", nr_thread,
1978 nr_thread > 1 ? "s" : "");
1982 log_info("%d process%s", nr_process,
1983 nr_process > 1 ? "es" : "");
1989 todo = thread_number;
1992 m_rate = t_rate = 0;
1994 for_each_td(td, i) {
1995 print_status_init(td->thread_number - 1);
1997 if (!td->o.create_serialize)
2000 if (fio_verify_load_state(td))
2004 * do file setup here so it happens sequentially,
2005 * we don't want X number of threads getting their
2006 * client data interspersed on disk
2008 if (setup_files(td)) {
2012 log_err("fio: pid=%d, err=%d/%s\n",
2013 (int) td->pid, td->error, td->verror);
2014 td_set_runstate(td, TD_REAPED);
2021 * for sharing to work, each job must always open
2022 * its own files. so close them, if we opened them
2025 for_each_file(td, f, j) {
2026 if (fio_file_open(f))
2027 td_io_close_file(td, f);
2032 /* start idle threads before io threads start to run */
2033 fio_idle_prof_start();
2038 struct thread_data *map[REAL_MAX_JOBS];
2039 struct timeval this_start;
2040 int this_jobs = 0, left;
2043 * create threads (TD_NOT_CREATED -> TD_CREATED)
2045 for_each_td(td, i) {
2046 if (td->runstate != TD_NOT_CREATED)
2050 * never got a chance to start, killed by other
2051 * thread for some reason
2053 if (td->terminate) {
2058 if (td->o.start_delay) {
2059 spent = utime_since_genesis();
2061 if (td->o.start_delay > spent)
2065 if (td->o.stonewall && (nr_started || nr_running)) {
2066 dprint(FD_PROCESS, "%s: stonewall wait\n",
2073 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2074 td->update_rusage = 0;
2077 * Set state to created. Thread will transition
2078 * to TD_INITIALIZED when it's done setting up.
2080 td_set_runstate(td, TD_CREATED);
2081 map[this_jobs++] = td;
2084 if (td->o.use_thread) {
2087 dprint(FD_PROCESS, "will pthread_create\n");
2088 ret = pthread_create(&td->thread, NULL,
2091 log_err("pthread_create: %s\n",
2096 ret = pthread_detach(td->thread);
2098 log_err("pthread_detach: %s",
2102 dprint(FD_PROCESS, "will fork\n");
2105 int ret = fork_main(shm_id, i);
2108 } else if (i == fio_debug_jobno)
2109 *fio_debug_jobp = pid;
2111 dprint(FD_MUTEX, "wait on startup_mutex\n");
2112 if (fio_mutex_down_timeout(startup_mutex, 10)) {
2113 log_err("fio: job startup hung? exiting.\n");
2114 fio_terminate_threads(TERMINATE_ALL);
2119 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2123 * Wait for the started threads to transition to
2126 fio_gettime(&this_start, NULL);
2128 while (left && !fio_abort) {
2129 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2134 for (i = 0; i < this_jobs; i++) {
2138 if (td->runstate == TD_INITIALIZED) {
2141 } else if (td->runstate >= TD_EXITED) {
2145 nr_running++; /* work-around... */
2151 log_err("fio: %d job%s failed to start\n", left,
2152 left > 1 ? "s" : "");
2153 for (i = 0; i < this_jobs; i++) {
2157 kill(td->pid, SIGTERM);
2163 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2165 for_each_td(td, i) {
2166 if (td->runstate != TD_INITIALIZED)
2169 if (in_ramp_time(td))
2170 td_set_runstate(td, TD_RAMP);
2172 td_set_runstate(td, TD_RUNNING);
2175 m_rate += ddir_rw_sum(td->o.ratemin);
2176 t_rate += ddir_rw_sum(td->o.rate);
2178 fio_mutex_up(td->mutex);
2181 reap_threads(&nr_running, &t_rate, &m_rate);
2187 while (nr_running) {
2188 reap_threads(&nr_running, &t_rate, &m_rate);
2192 fio_idle_prof_stop();
2197 static void wait_for_helper_thread_exit(void)
2202 pthread_cond_signal(&helper_cond);
2203 pthread_join(helper_thread, &ret);
2206 static void free_disk_util(void)
2208 disk_util_prune_entries();
2210 pthread_cond_destroy(&helper_cond);
2213 static void *helper_thread_main(void *data)
2217 fio_mutex_up(startup_mutex);
2220 uint64_t sec = DISK_UTIL_MSEC / 1000;
2221 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2225 gettimeofday(&tv, NULL);
2226 ts.tv_sec = tv.tv_sec + sec;
2227 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2229 if (ts.tv_nsec >= 1000000000ULL) {
2230 ts.tv_nsec -= 1000000000ULL;
2234 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
2236 ret = update_io_ticks();
2238 if (helper_do_stat) {
2240 __show_running_run_stats();
2244 print_thread_status();
2250 static int create_helper_thread(void)
2256 pthread_cond_init(&helper_cond, NULL);
2257 pthread_mutex_init(&helper_lock, NULL);
2259 ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2261 log_err("Can't create helper thread: %s\n", strerror(ret));
2265 dprint(FD_MUTEX, "wait on startup_mutex\n");
2266 fio_mutex_down(startup_mutex);
2267 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2271 int fio_backend(void)
2273 struct thread_data *td;
2277 if (load_profile(exec_profile))
2280 exec_profile = NULL;
2286 struct log_params p = {
2287 .log_type = IO_LOG_TYPE_BW,
2290 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2291 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2292 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2295 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2296 if (startup_mutex == NULL)
2301 create_helper_thread();
2303 cgroup_list = smalloc(sizeof(*cgroup_list));
2304 INIT_FLIST_HEAD(cgroup_list);
2308 wait_for_helper_thread_exit();
2313 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2314 struct io_log *log = agg_io_log[i];
2322 for_each_td(td, i) {
2323 fio_options_free(td);
2324 if (td->rusage_sem) {
2325 fio_mutex_remove(td->rusage_sem);
2326 td->rusage_sem = NULL;
2331 cgroup_kill(cgroup_list);
2335 fio_mutex_remove(startup_mutex);