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., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
41 #include "lib/memalign.h"
43 #include "lib/getrusage.h"
46 #include "workqueue.h"
47 #include "lib/mountcheck.h"
48 #include "rate-submit.h"
49 #include "helper_thread.h"
51 #include "zone-dist.h"
53 static struct fio_sem *startup_sem;
54 static struct flist_head *cgroup_list;
55 static struct cgroup_mnt *cgroup_mnt;
56 static int exit_value;
57 static volatile bool fio_abort;
58 static unsigned int nr_process = 0;
59 static unsigned int nr_thread = 0;
61 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
64 unsigned int thread_number = 0;
65 unsigned int stat_number = 0;
68 unsigned long done_secs = 0;
69 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
71 #define JOB_START_TIMEOUT (5 * 1000)
73 static void sig_int(int sig)
77 fio_server_got_signal(sig);
79 log_info("\nfio: terminating on signal %d\n", sig);
84 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
88 void sig_show_status(int sig)
90 show_running_run_stats();
93 static void set_sig_handlers(void)
97 memset(&act, 0, sizeof(act));
98 act.sa_handler = sig_int;
99 act.sa_flags = SA_RESTART;
100 sigaction(SIGINT, &act, NULL);
102 memset(&act, 0, sizeof(act));
103 act.sa_handler = sig_int;
104 act.sa_flags = SA_RESTART;
105 sigaction(SIGTERM, &act, NULL);
107 /* Windows uses SIGBREAK as a quit signal from other applications */
109 memset(&act, 0, sizeof(act));
110 act.sa_handler = sig_int;
111 act.sa_flags = SA_RESTART;
112 sigaction(SIGBREAK, &act, NULL);
115 memset(&act, 0, sizeof(act));
116 act.sa_handler = sig_show_status;
117 act.sa_flags = SA_RESTART;
118 sigaction(SIGUSR1, &act, NULL);
121 memset(&act, 0, sizeof(act));
122 act.sa_handler = sig_int;
123 act.sa_flags = SA_RESTART;
124 sigaction(SIGPIPE, &act, NULL);
129 * Check if we are above the minimum rate given.
131 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
134 unsigned long long bytes = 0;
135 unsigned long iops = 0;
138 unsigned int ratemin = 0;
139 unsigned int rate_iops = 0;
140 unsigned int rate_iops_min = 0;
142 assert(ddir_rw(ddir));
144 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
148 * allow a 2 second settle period in the beginning
150 if (mtime_since(&td->start, now) < 2000)
153 iops += td->this_io_blocks[ddir];
154 bytes += td->this_io_bytes[ddir];
155 ratemin += td->o.ratemin[ddir];
156 rate_iops += td->o.rate_iops[ddir];
157 rate_iops_min += td->o.rate_iops_min[ddir];
160 * if rate blocks is set, sample is running
162 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
163 spent = mtime_since(&td->lastrate[ddir], now);
164 if (spent < td->o.ratecycle)
167 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
169 * check bandwidth specified rate
171 if (bytes < td->rate_bytes[ddir]) {
172 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
173 td->o.name, ratemin, bytes);
177 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
181 if (rate < ratemin ||
182 bytes < td->rate_bytes[ddir]) {
183 log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
184 td->o.name, ratemin, rate);
190 * checks iops specified rate
192 if (iops < rate_iops) {
193 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
194 td->o.name, rate_iops, iops);
198 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
202 if (rate < rate_iops_min ||
203 iops < td->rate_blocks[ddir]) {
204 log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
205 td->o.name, rate_iops_min, rate);
212 td->rate_bytes[ddir] = bytes;
213 td->rate_blocks[ddir] = iops;
214 memcpy(&td->lastrate[ddir], now, sizeof(*now));
218 static bool check_min_rate(struct thread_data *td, struct timespec *now)
222 if (td->bytes_done[DDIR_READ])
223 ret |= __check_min_rate(td, now, DDIR_READ);
224 if (td->bytes_done[DDIR_WRITE])
225 ret |= __check_min_rate(td, now, DDIR_WRITE);
226 if (td->bytes_done[DDIR_TRIM])
227 ret |= __check_min_rate(td, now, DDIR_TRIM);
233 * When job exits, we can cancel the in-flight IO if we are using async
234 * io. Attempt to do so.
236 static void cleanup_pending_aio(struct thread_data *td)
244 * get immediately available events, if any
246 r = io_u_queued_complete(td, 0);
251 * now cancel remaining active events
253 if (td->io_ops->cancel) {
257 io_u_qiter(&td->io_u_all, io_u, i) {
258 if (io_u->flags & IO_U_F_FLIGHT) {
259 r = td->io_ops->cancel(td, io_u);
267 r = io_u_queued_complete(td, td->cur_depth);
271 * Helper to handle the final sync of a file. Works just like the normal
272 * io path, just does everything sync.
274 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
276 struct io_u *io_u = __get_io_u(td);
277 enum fio_q_status ret;
282 io_u->ddir = DDIR_SYNC;
284 io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
286 if (td_io_prep(td, io_u)) {
292 ret = td_io_queue(td, io_u);
296 if (io_u_queued_complete(td, 1) < 0)
299 case FIO_Q_COMPLETED:
301 td_verror(td, io_u->error, "td_io_queue");
305 if (io_u_sync_complete(td, io_u) < 0)
316 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
320 if (fio_file_open(f))
321 return fio_io_sync(td, f);
323 if (td_io_open_file(td, f))
326 ret = fio_io_sync(td, f);
328 if (fio_file_open(f))
329 ret2 = td_io_close_file(td, f);
330 return (ret || ret2);
333 static inline void __update_ts_cache(struct thread_data *td)
335 fio_gettime(&td->ts_cache, NULL);
338 static inline void update_ts_cache(struct thread_data *td)
340 if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
341 __update_ts_cache(td);
344 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
346 if (in_ramp_time(td))
350 if (utime_since(&td->epoch, t) >= td->o.timeout)
357 * We need to update the runtime consistently in ms, but keep a running
358 * tally of the current elapsed time in microseconds for sub millisecond
361 static inline void update_runtime(struct thread_data *td,
362 unsigned long long *elapsed_us,
363 const enum fio_ddir ddir)
365 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
368 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
369 elapsed_us[ddir] += utime_since_now(&td->start);
370 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
373 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
378 if (ret < 0 || td->error) {
380 enum error_type_bit eb;
385 eb = td_error_type(ddir, err);
386 if (!(td->o.continue_on_error & (1 << eb)))
389 if (td_non_fatal_error(td, eb, err)) {
391 * Continue with the I/Os in case of
394 update_error_count(td, err);
398 } else if (td->o.fill_device && err == ENOSPC) {
400 * We expect to hit this error if
401 * fill_device option is set.
404 fio_mark_td_terminate(td);
408 * Stop the I/O in case of a fatal
411 update_error_count(td, err);
419 static void check_update_rusage(struct thread_data *td)
421 if (td->update_rusage) {
422 td->update_rusage = 0;
423 update_rusage_stat(td);
424 fio_sem_up(td->rusage_sem);
428 static int wait_for_completions(struct thread_data *td, struct timespec *time)
430 const int full = queue_full(td);
434 if (td->flags & TD_F_REGROW_LOGS)
435 return io_u_quiesce(td);
438 * if the queue is full, we MUST reap at least 1 event
440 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
441 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
444 if (time && __should_check_rate(td))
445 fio_gettime(time, NULL);
448 ret = io_u_queued_complete(td, min_evts);
451 } while (full && (td->cur_depth > td->o.iodepth_low));
456 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
457 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
458 struct timespec *comp_time)
461 case FIO_Q_COMPLETED:
464 clear_io_u(td, io_u);
465 } else if (io_u->resid) {
466 long long bytes = io_u->xfer_buflen - io_u->resid;
467 struct fio_file *f = io_u->file;
470 *bytes_issued += bytes;
480 unlog_io_piece(td, io_u);
481 td_verror(td, EIO, "full resid");
486 io_u->xfer_buflen = io_u->resid;
487 io_u->xfer_buf += bytes;
488 io_u->offset += bytes;
490 if (ddir_rw(io_u->ddir))
491 td->ts.short_io_u[io_u->ddir]++;
493 if (io_u->offset == f->real_file_size)
496 requeue_io_u(td, &io_u);
499 if (comp_time && __should_check_rate(td))
500 fio_gettime(comp_time, NULL);
502 *ret = io_u_sync_complete(td, io_u);
507 if (td->flags & TD_F_REGROW_LOGS)
511 * when doing I/O (not when verifying),
512 * check for any errors that are to be ignored
520 * if the engine doesn't have a commit hook,
521 * the io_u is really queued. if it does have such
522 * a hook, it has to call io_u_queued() itself.
524 if (td->io_ops->commit == NULL)
525 io_u_queued(td, io_u);
527 *bytes_issued += io_u->xfer_buflen;
531 unlog_io_piece(td, io_u);
532 requeue_io_u(td, &io_u);
537 td_verror(td, -(*ret), "td_io_queue");
541 if (break_on_this_error(td, ddir, ret))
547 static inline bool io_in_polling(struct thread_data *td)
549 return !td->o.iodepth_batch_complete_min &&
550 !td->o.iodepth_batch_complete_max;
553 * Unlinks files from thread data fio_file structure
555 static int unlink_all_files(struct thread_data *td)
561 for_each_file(td, f, i) {
562 if (f->filetype != FIO_TYPE_FILE)
564 ret = td_io_unlink_file(td, f);
570 td_verror(td, ret, "unlink_all_files");
576 * Check if io_u will overlap an in-flight IO in the queue
578 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
581 struct io_u *check_io_u;
582 unsigned long long x1, x2, y1, y2;
586 x2 = io_u->offset + io_u->buflen;
588 io_u_qiter(q, check_io_u, i) {
589 if (check_io_u->flags & IO_U_F_FLIGHT) {
590 y1 = check_io_u->offset;
591 y2 = check_io_u->offset + check_io_u->buflen;
593 if (x1 < y2 && y1 < x2) {
595 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
597 y1, check_io_u->buflen);
606 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
609 * Check for overlap if the user asked us to, and we have
610 * at least one IO in flight besides this one.
612 if (td->o.serialize_overlap && td->cur_depth > 1 &&
613 in_flight_overlap(&td->io_u_all, io_u))
616 return td_io_queue(td, io_u);
620 * The main verify engine. Runs over the writes we previously submitted,
621 * reads the blocks back in, and checks the crc/md5 of the data.
623 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
630 dprint(FD_VERIFY, "starting loop\n");
633 * sync io first and invalidate cache, to make sure we really
636 for_each_file(td, f, i) {
637 if (!fio_file_open(f))
639 if (fio_io_sync(td, f))
641 if (file_invalidate_cache(td, f))
645 check_update_rusage(td);
651 * verify_state needs to be reset before verification
652 * proceeds so that expected random seeds match actual
653 * random seeds in headers. The main loop will reset
654 * all random number generators if randrepeat is set.
656 if (!td->o.rand_repeatable)
657 td_fill_verify_state_seed(td);
659 td_set_runstate(td, TD_VERIFYING);
662 while (!td->terminate) {
667 check_update_rusage(td);
669 if (runtime_exceeded(td, &td->ts_cache)) {
670 __update_ts_cache(td);
671 if (runtime_exceeded(td, &td->ts_cache)) {
672 fio_mark_td_terminate(td);
677 if (flow_threshold_exceeded(td))
680 if (!td->o.experimental_verify) {
681 io_u = __get_io_u(td);
685 if (get_next_verify(td, io_u)) {
690 if (td_io_prep(td, io_u)) {
695 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
698 while ((io_u = get_io_u(td)) != NULL) {
699 if (IS_ERR_OR_NULL(io_u)) {
706 * We are only interested in the places where
707 * we wrote or trimmed IOs. Turn those into
708 * reads for verification purposes.
710 if (io_u->ddir == DDIR_READ) {
712 * Pretend we issued it for rwmix
715 td->io_issues[DDIR_READ]++;
718 } else if (io_u->ddir == DDIR_TRIM) {
719 io_u->ddir = DDIR_READ;
720 io_u_set(td, io_u, IO_U_F_TRIMMED);
722 } else if (io_u->ddir == DDIR_WRITE) {
723 io_u->ddir = DDIR_READ;
724 populate_verify_io_u(td, io_u);
736 if (verify_state_should_stop(td, io_u)) {
741 if (td->o.verify_async)
742 io_u->end_io = verify_io_u_async;
744 io_u->end_io = verify_io_u;
747 if (!td->o.disable_slat)
748 fio_gettime(&io_u->start_time, NULL);
750 ret = io_u_submit(td, io_u);
752 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
756 * if we can queue more, do so. but check if there are
757 * completed io_u's first. Note that we can get BUSY even
758 * without IO queued, if the system is resource starved.
761 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
762 if (full || io_in_polling(td))
763 ret = wait_for_completions(td, NULL);
769 check_update_rusage(td);
772 min_events = td->cur_depth;
775 ret = io_u_queued_complete(td, min_events);
777 cleanup_pending_aio(td);
779 td_set_runstate(td, TD_RUNNING);
781 dprint(FD_VERIFY, "exiting loop\n");
784 static bool exceeds_number_ios(struct thread_data *td)
786 unsigned long long number_ios;
788 if (!td->o.number_ios)
791 number_ios = ddir_rw_sum(td->io_blocks);
792 number_ios += td->io_u_queued + td->io_u_in_flight;
794 return number_ios >= (td->o.number_ios * td->loops);
797 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
799 unsigned long long bytes, limit;
802 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
803 else if (td_write(td))
804 bytes = this_bytes[DDIR_WRITE];
805 else if (td_read(td))
806 bytes = this_bytes[DDIR_READ];
808 bytes = this_bytes[DDIR_TRIM];
811 limit = td->o.io_size;
816 return bytes >= limit || exceeds_number_ios(td);
819 static bool io_issue_bytes_exceeded(struct thread_data *td)
821 return io_bytes_exceeded(td, td->io_issue_bytes);
824 static bool io_complete_bytes_exceeded(struct thread_data *td)
826 return io_bytes_exceeded(td, td->this_io_bytes);
830 * used to calculate the next io time for rate control
833 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
835 uint64_t bps = td->rate_bps[ddir];
837 assert(!(td->flags & TD_F_CHILD));
839 if (td->o.rate_process == RATE_PROCESS_POISSON) {
842 iops = bps / td->o.bs[ddir];
843 val = (int64_t) (1000000 / iops) *
844 -logf(__rand_0_1(&td->poisson_state[ddir]));
846 dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
847 (unsigned long long) 1000000 / val,
850 td->last_usec[ddir] += val;
851 return td->last_usec[ddir];
853 uint64_t bytes = td->rate_io_issue_bytes[ddir];
854 uint64_t secs = bytes / bps;
855 uint64_t remainder = bytes % bps;
857 return remainder * 1000000 / bps + secs * 1000000;
863 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
865 unsigned long long b;
869 b = ddir_rw_sum(td->io_blocks);
870 if (b % td->o.thinktime_blocks)
876 if (td->o.thinktime_spin)
877 total = usec_spin(td->o.thinktime_spin);
879 left = td->o.thinktime - total;
881 total += usec_sleep(td, left);
884 * If we're ignoring thinktime for the rate, add the number of bytes
885 * we would have done while sleeping, minus one block to ensure we
886 * start issuing immediately after the sleep.
888 if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
889 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
890 uint64_t bs = td->o.min_bs[ddir];
891 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
894 if (usperop <= total)
897 over = (usperop - total) / usperop * -bs;
899 td->rate_io_issue_bytes[ddir] += (missed - over);
900 /* adjust for rate_process=poisson */
901 td->last_usec[ddir] += total;
906 * Main IO worker function. It retrieves io_u's to process and queues
907 * and reaps them, checking for rate and errors along the way.
909 * Returns number of bytes written and trimmed.
911 static void do_io(struct thread_data *td, uint64_t *bytes_done)
915 uint64_t total_bytes, bytes_issued = 0;
917 for (i = 0; i < DDIR_RWDIR_CNT; i++)
918 bytes_done[i] = td->bytes_done[i];
920 if (in_ramp_time(td))
921 td_set_runstate(td, TD_RAMP);
923 td_set_runstate(td, TD_RUNNING);
927 total_bytes = td->o.size;
929 * Allow random overwrite workloads to write up to io_size
930 * before starting verification phase as 'size' doesn't apply.
932 if (td_write(td) && td_random(td) && td->o.norandommap)
933 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
935 * If verify_backlog is enabled, we'll run the verify in this
936 * handler as well. For that case, we may need up to twice the
939 if (td->o.verify != VERIFY_NONE &&
940 (td_write(td) && td->o.verify_backlog))
941 total_bytes += td->o.size;
943 /* In trimwrite mode, each byte is trimmed and then written, so
944 * allow total_bytes to be twice as big */
945 if (td_trimwrite(td))
946 total_bytes += td->total_io_size;
948 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
949 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
951 struct timespec comp_time;
956 check_update_rusage(td);
958 if (td->terminate || td->done)
963 if (runtime_exceeded(td, &td->ts_cache)) {
964 __update_ts_cache(td);
965 if (runtime_exceeded(td, &td->ts_cache)) {
966 fio_mark_td_terminate(td);
971 if (flow_threshold_exceeded(td))
975 * Break if we exceeded the bytes. The exception is time
976 * based runs, but we still need to break out of the loop
977 * for those to run verification, if enabled.
978 * Jobs read from iolog do not use this stop condition.
980 if (bytes_issued >= total_bytes &&
981 !td->o.read_iolog_file &&
982 (!td->o.time_based ||
983 (td->o.time_based && td->o.verify != VERIFY_NONE)))
987 if (IS_ERR_OR_NULL(io_u)) {
988 int err = PTR_ERR(io_u);
996 if (td->o.latency_target)
1001 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
1002 populate_verify_io_u(td, io_u);
1007 * Add verification end_io handler if:
1008 * - Asked to verify (!td_rw(td))
1009 * - Or the io_u is from our verify list (mixed write/ver)
1011 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1012 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1014 if (!td->o.verify_pattern_bytes) {
1015 io_u->rand_seed = __rand(&td->verify_state);
1016 if (sizeof(int) != sizeof(long *))
1017 io_u->rand_seed *= __rand(&td->verify_state);
1020 if (verify_state_should_stop(td, io_u)) {
1025 if (td->o.verify_async)
1026 io_u->end_io = verify_io_u_async;
1028 io_u->end_io = verify_io_u;
1029 td_set_runstate(td, TD_VERIFYING);
1030 } else if (in_ramp_time(td))
1031 td_set_runstate(td, TD_RAMP);
1033 td_set_runstate(td, TD_RUNNING);
1036 * Always log IO before it's issued, so we know the specific
1037 * order of it. The logged unit will track when the IO has
1040 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1042 td->o.verify != VERIFY_NONE &&
1043 !td->o.experimental_verify)
1044 log_io_piece(td, io_u);
1046 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1047 const unsigned long long blen = io_u->xfer_buflen;
1048 const enum fio_ddir __ddir = acct_ddir(io_u);
1053 workqueue_enqueue(&td->io_wq, &io_u->work);
1056 if (ddir_rw(__ddir)) {
1057 td->io_issues[__ddir]++;
1058 td->io_issue_bytes[__ddir] += blen;
1059 td->rate_io_issue_bytes[__ddir] += blen;
1062 if (should_check_rate(td))
1063 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1066 ret = io_u_submit(td, io_u);
1068 if (should_check_rate(td))
1069 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1071 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1075 * See if we need to complete some commands. Note that
1076 * we can get BUSY even without IO queued, if the
1077 * system is resource starved.
1080 full = queue_full(td) ||
1081 (ret == FIO_Q_BUSY && td->cur_depth);
1082 if (full || io_in_polling(td))
1083 ret = wait_for_completions(td, &comp_time);
1087 if (!ddir_rw_sum(td->bytes_done) &&
1088 !td_ioengine_flagged(td, FIO_NOIO))
1091 if (!in_ramp_time(td) && should_check_rate(td)) {
1092 if (check_min_rate(td, &comp_time)) {
1093 if (exitall_on_terminate || td->o.exitall_error)
1094 fio_terminate_threads(td->groupid, td->o.exit_what);
1095 td_verror(td, EIO, "check_min_rate");
1099 if (!in_ramp_time(td) && td->o.latency_target)
1100 lat_target_check(td);
1102 if (ddir_rw(ddir) && td->o.thinktime)
1103 handle_thinktime(td, ddir);
1106 check_update_rusage(td);
1108 if (td->trim_entries)
1109 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1111 if (td->o.fill_device && td->error == ENOSPC) {
1113 fio_mark_td_terminate(td);
1118 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1119 workqueue_flush(&td->io_wq);
1125 ret = io_u_queued_complete(td, i);
1126 if (td->o.fill_device && td->error == ENOSPC)
1130 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1131 td_set_runstate(td, TD_FSYNCING);
1133 for_each_file(td, f, i) {
1134 if (!fio_file_fsync(td, f))
1137 log_err("fio: end_fsync failed for file %s\n",
1142 cleanup_pending_aio(td);
1145 * stop job if we failed doing any IO
1147 if (!ddir_rw_sum(td->this_io_bytes))
1150 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1151 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1154 static void free_file_completion_logging(struct thread_data *td)
1159 for_each_file(td, f, i) {
1160 if (!f->last_write_comp)
1162 sfree(f->last_write_comp);
1166 static int init_file_completion_logging(struct thread_data *td,
1172 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1175 for_each_file(td, f, i) {
1176 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1177 if (!f->last_write_comp)
1184 free_file_completion_logging(td);
1185 log_err("fio: failed to alloc write comp data\n");
1189 static void cleanup_io_u(struct thread_data *td)
1193 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1195 if (td->io_ops->io_u_free)
1196 td->io_ops->io_u_free(td, io_u);
1198 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1203 io_u_rexit(&td->io_u_requeues);
1204 io_u_qexit(&td->io_u_freelist, false);
1205 io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1207 free_file_completion_logging(td);
1210 static int init_io_u(struct thread_data *td)
1213 int cl_align, i, max_units;
1216 max_units = td->o.iodepth;
1219 err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1220 err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1221 err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1224 log_err("fio: failed setting up IO queues\n");
1228 cl_align = os_cache_line_size();
1230 for (i = 0; i < max_units; i++) {
1236 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1238 log_err("fio: unable to allocate aligned memory\n");
1243 memset(io_u, 0, sizeof(*io_u));
1244 INIT_FLIST_HEAD(&io_u->verify_list);
1245 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1248 io_u->flags = IO_U_F_FREE;
1249 io_u_qpush(&td->io_u_freelist, io_u);
1252 * io_u never leaves this stack, used for iteration of all
1255 io_u_qpush(&td->io_u_all, io_u);
1257 if (td->io_ops->io_u_init) {
1258 int ret = td->io_ops->io_u_init(td, io_u);
1261 log_err("fio: failed to init engine data: %d\n", ret);
1267 init_io_u_buffers(td);
1269 if (init_file_completion_logging(td, max_units))
1275 int init_io_u_buffers(struct thread_data *td)
1278 unsigned long long max_bs, min_write;
1283 max_units = td->o.iodepth;
1284 max_bs = td_max_bs(td);
1285 min_write = td->o.min_bs[DDIR_WRITE];
1286 td->orig_buffer_size = (unsigned long long) max_bs
1287 * (unsigned long long) max_units;
1289 if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1293 * if we may later need to do address alignment, then add any
1294 * possible adjustment here so that we don't cause a buffer
1295 * overflow later. this adjustment may be too much if we get
1296 * lucky and the allocator gives us an aligned address.
1298 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1299 td_ioengine_flagged(td, FIO_RAWIO))
1300 td->orig_buffer_size += page_mask + td->o.mem_align;
1302 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1303 unsigned long long bs;
1305 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1306 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1309 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1310 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1314 if (data_xfer && allocate_io_mem(td))
1317 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1318 td_ioengine_flagged(td, FIO_RAWIO))
1319 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1321 p = td->orig_buffer;
1323 for (i = 0; i < max_units; i++) {
1324 io_u = td->io_u_all.io_us[i];
1325 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1329 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1332 io_u_fill_buffer(td, io_u, min_write, max_bs);
1333 if (td_write(td) && td->o.verify_pattern_bytes) {
1335 * Fill the buffer with the pattern if we are
1336 * going to be doing writes.
1338 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1348 * This function is Linux specific.
1349 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1351 static int switch_ioscheduler(struct thread_data *td)
1353 #ifdef FIO_HAVE_IOSCHED_SWITCH
1354 char tmp[256], tmp2[128], *p;
1358 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1361 assert(td->files && td->files[0]);
1362 sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1364 f = fopen(tmp, "r+");
1366 if (errno == ENOENT) {
1367 log_err("fio: os or kernel doesn't support IO scheduler"
1371 td_verror(td, errno, "fopen iosched");
1378 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1379 if (ferror(f) || ret != 1) {
1380 td_verror(td, errno, "fwrite");
1388 * Read back and check that the selected scheduler is now the default.
1390 ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1391 if (ferror(f) || ret < 0) {
1392 td_verror(td, errno, "fread");
1398 * either a list of io schedulers or "none\n" is expected. Strip the
1405 * Write to "none" entry doesn't fail, so check the result here.
1407 if (!strcmp(tmp, "none")) {
1408 log_err("fio: io scheduler is not tunable\n");
1413 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1414 if (!strstr(tmp, tmp2)) {
1415 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1416 td_verror(td, EINVAL, "iosched_switch");
1428 static bool keep_running(struct thread_data *td)
1430 unsigned long long limit;
1436 if (td->o.time_based)
1442 if (exceeds_number_ios(td))
1446 limit = td->o.io_size;
1450 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1454 * If the difference is less than the maximum IO size, we
1457 diff = limit - ddir_rw_sum(td->io_bytes);
1458 if (diff < td_max_bs(td))
1461 if (fio_files_done(td) && !td->o.io_size)
1470 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1472 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 13 + 1;
1476 str = malloc(newlen);
1477 sprintf(str, "%s > %s.%s.txt 2>&1", string, o->name, mode);
1479 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1482 log_err("fio: exec of cmd <%s> failed\n", str);
1489 * Dry run to compute correct state of numberio for verification.
1491 static uint64_t do_dry_run(struct thread_data *td)
1493 td_set_runstate(td, TD_RUNNING);
1495 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1496 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1500 if (td->terminate || td->done)
1503 io_u = get_io_u(td);
1504 if (IS_ERR_OR_NULL(io_u))
1507 io_u_set(td, io_u, IO_U_F_FLIGHT);
1510 if (ddir_rw(acct_ddir(io_u)))
1511 td->io_issues[acct_ddir(io_u)]++;
1512 if (ddir_rw(io_u->ddir)) {
1513 io_u_mark_depth(td, 1);
1514 td->ts.total_io_u[io_u->ddir]++;
1517 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1519 td->o.verify != VERIFY_NONE &&
1520 !td->o.experimental_verify)
1521 log_io_piece(td, io_u);
1523 ret = io_u_sync_complete(td, io_u);
1527 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1531 struct thread_data *td;
1532 struct sk_out *sk_out;
1536 * Entry point for the thread based jobs. The process based jobs end up
1537 * here as well, after a little setup.
1539 static void *thread_main(void *data)
1541 struct fork_data *fd = data;
1542 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1543 struct thread_data *td = fd->td;
1544 struct thread_options *o = &td->o;
1545 struct sk_out *sk_out = fd->sk_out;
1546 uint64_t bytes_done[DDIR_RWDIR_CNT];
1547 int deadlock_loop_cnt;
1551 sk_out_assign(sk_out);
1554 if (!o->use_thread) {
1560 fio_local_clock_init();
1562 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1565 fio_server_send_start(td);
1567 INIT_FLIST_HEAD(&td->io_log_list);
1568 INIT_FLIST_HEAD(&td->io_hist_list);
1569 INIT_FLIST_HEAD(&td->verify_list);
1570 INIT_FLIST_HEAD(&td->trim_list);
1571 td->io_hist_tree = RB_ROOT;
1573 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1575 td_verror(td, ret, "mutex_cond_init_pshared");
1578 ret = cond_init_pshared(&td->verify_cond);
1580 td_verror(td, ret, "mutex_cond_pshared");
1584 td_set_runstate(td, TD_INITIALIZED);
1585 dprint(FD_MUTEX, "up startup_sem\n");
1586 fio_sem_up(startup_sem);
1587 dprint(FD_MUTEX, "wait on td->sem\n");
1588 fio_sem_down(td->sem);
1589 dprint(FD_MUTEX, "done waiting on td->sem\n");
1592 * A new gid requires privilege, so we need to do this before setting
1595 if (o->gid != -1U && setgid(o->gid)) {
1596 td_verror(td, errno, "setgid");
1599 if (o->uid != -1U && setuid(o->uid)) {
1600 td_verror(td, errno, "setuid");
1604 td_zone_gen_index(td);
1607 * Do this early, we don't want the compress threads to be limited
1608 * to the same CPUs as the IO workers. So do this before we set
1609 * any potential CPU affinity
1611 if (iolog_compress_init(td, sk_out))
1615 * If we have a gettimeofday() thread, make sure we exclude that
1616 * thread from this job
1619 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1622 * Set affinity first, in case it has an impact on the memory
1625 if (fio_option_is_set(o, cpumask)) {
1626 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1627 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1629 log_err("fio: no CPUs set\n");
1630 log_err("fio: Try increasing number of available CPUs\n");
1631 td_verror(td, EINVAL, "cpus_split");
1635 ret = fio_setaffinity(td->pid, o->cpumask);
1637 td_verror(td, errno, "cpu_set_affinity");
1642 #ifdef CONFIG_LIBNUMA
1643 /* numa node setup */
1644 if (fio_option_is_set(o, numa_cpunodes) ||
1645 fio_option_is_set(o, numa_memnodes)) {
1646 struct bitmask *mask;
1648 if (numa_available() < 0) {
1649 td_verror(td, errno, "Does not support NUMA API\n");
1653 if (fio_option_is_set(o, numa_cpunodes)) {
1654 mask = numa_parse_nodestring(o->numa_cpunodes);
1655 ret = numa_run_on_node_mask(mask);
1656 numa_free_nodemask(mask);
1658 td_verror(td, errno, \
1659 "numa_run_on_node_mask failed\n");
1664 if (fio_option_is_set(o, numa_memnodes)) {
1666 if (o->numa_memnodes)
1667 mask = numa_parse_nodestring(o->numa_memnodes);
1669 switch (o->numa_mem_mode) {
1670 case MPOL_INTERLEAVE:
1671 numa_set_interleave_mask(mask);
1674 numa_set_membind(mask);
1677 numa_set_localalloc();
1679 case MPOL_PREFERRED:
1680 numa_set_preferred(o->numa_mem_prefer_node);
1688 numa_free_nodemask(mask);
1694 if (fio_pin_memory(td))
1698 * May alter parameters that init_io_u() will use, so we need to
1701 if (!init_iolog(td))
1710 if (td->io_ops->post_init && td->io_ops->post_init(td))
1713 if (o->verify_async && verify_async_init(td))
1716 if (fio_option_is_set(o, ioprio) ||
1717 fio_option_is_set(o, ioprio_class)) {
1718 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1720 td_verror(td, errno, "ioprio_set");
1725 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1729 if (nice(o->nice) == -1 && errno != 0) {
1730 td_verror(td, errno, "nice");
1734 if (o->ioscheduler && switch_ioscheduler(td))
1737 if (!o->create_serialize && setup_files(td))
1740 if (!init_random_map(td))
1743 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1746 if (o->pre_read && !pre_read_files(td))
1749 fio_verify_init(td);
1751 if (rate_submit_init(td, sk_out))
1754 set_epoch_time(td, o->log_unix_epoch);
1755 fio_getrusage(&td->ru_start);
1756 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1757 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1758 memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1760 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1761 o->ratemin[DDIR_TRIM]) {
1762 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1763 sizeof(td->bw_sample_time));
1764 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1765 sizeof(td->bw_sample_time));
1766 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1767 sizeof(td->bw_sample_time));
1770 memset(bytes_done, 0, sizeof(bytes_done));
1771 clear_state = false;
1773 while (keep_running(td)) {
1774 uint64_t verify_bytes;
1776 fio_gettime(&td->start, NULL);
1777 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1780 clear_io_state(td, 0);
1782 if (o->unlink_each_loop && unlink_all_files(td))
1786 prune_io_piece_log(td);
1788 if (td->o.verify_only && td_write(td))
1789 verify_bytes = do_dry_run(td);
1791 do_io(td, bytes_done);
1793 if (!ddir_rw_sum(bytes_done)) {
1794 fio_mark_td_terminate(td);
1797 verify_bytes = bytes_done[DDIR_WRITE] +
1798 bytes_done[DDIR_TRIM];
1803 * If we took too long to shut down, the main thread could
1804 * already consider us reaped/exited. If that happens, break
1807 if (td->runstate >= TD_EXITED)
1813 * Make sure we've successfully updated the rusage stats
1814 * before waiting on the stat mutex. Otherwise we could have
1815 * the stat thread holding stat mutex and waiting for
1816 * the rusage_sem, which would never get upped because
1817 * this thread is waiting for the stat mutex.
1819 deadlock_loop_cnt = 0;
1821 check_update_rusage(td);
1822 if (!fio_sem_down_trylock(stat_sem))
1825 if (deadlock_loop_cnt++ > 5000) {
1826 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1827 td->error = EDEADLK;
1832 if (td_read(td) && td->io_bytes[DDIR_READ])
1833 update_runtime(td, elapsed_us, DDIR_READ);
1834 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1835 update_runtime(td, elapsed_us, DDIR_WRITE);
1836 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1837 update_runtime(td, elapsed_us, DDIR_TRIM);
1838 fio_gettime(&td->start, NULL);
1839 fio_sem_up(stat_sem);
1841 if (td->error || td->terminate)
1844 if (!o->do_verify ||
1845 o->verify == VERIFY_NONE ||
1846 td_ioengine_flagged(td, FIO_UNIDIR))
1849 clear_io_state(td, 0);
1851 fio_gettime(&td->start, NULL);
1853 do_verify(td, verify_bytes);
1856 * See comment further up for why this is done here.
1858 check_update_rusage(td);
1860 fio_sem_down(stat_sem);
1861 update_runtime(td, elapsed_us, DDIR_READ);
1862 fio_gettime(&td->start, NULL);
1863 fio_sem_up(stat_sem);
1865 if (td->error || td->terminate)
1870 * Acquire this lock if we were doing overlap checking in
1871 * offload mode so that we don't clean up this job while
1872 * another thread is checking its io_u's for overlap
1874 if (td_offload_overlap(td))
1875 pthread_mutex_lock(&overlap_check);
1876 td_set_runstate(td, TD_FINISHING);
1877 if (td_offload_overlap(td))
1878 pthread_mutex_unlock(&overlap_check);
1880 update_rusage_stat(td);
1881 td->ts.total_run_time = mtime_since_now(&td->epoch);
1882 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1883 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1884 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1886 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1887 (td->o.verify != VERIFY_NONE && td_write(td)))
1888 verify_save_state(td->thread_number);
1890 fio_unpin_memory(td);
1892 td_writeout_logs(td, true);
1894 iolog_compress_exit(td);
1895 rate_submit_exit(td);
1897 if (o->exec_postrun)
1898 exec_string(o, o->exec_postrun, (const char *)"postrun");
1900 if (exitall_on_terminate || (o->exitall_error && td->error))
1901 fio_terminate_threads(td->groupid, td->o.exit_what);
1905 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1908 if (o->verify_async)
1909 verify_async_exit(td);
1911 close_and_free_files(td);
1914 cgroup_shutdown(td, cgroup_mnt);
1915 verify_free_state(td);
1916 td_zone_free_index(td);
1918 if (fio_option_is_set(o, cpumask)) {
1919 ret = fio_cpuset_exit(&o->cpumask);
1921 td_verror(td, ret, "fio_cpuset_exit");
1925 * do this very late, it will log file closing as well
1927 if (o->write_iolog_file)
1928 write_iolog_close(td);
1929 if (td->io_log_rfile)
1930 fclose(td->io_log_rfile);
1932 td_set_runstate(td, TD_EXITED);
1935 * Do this last after setting our runstate to exited, so we
1936 * know that the stat thread is signaled.
1938 check_update_rusage(td);
1941 return (void *) (uintptr_t) td->error;
1945 * Run over the job map and reap the threads that have exited, if any.
1947 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1950 struct thread_data *td;
1951 unsigned int cputhreads, realthreads, pending;
1955 * reap exited threads (TD_EXITED -> TD_REAPED)
1957 realthreads = pending = cputhreads = 0;
1958 for_each_td(td, i) {
1961 if (!strcmp(td->o.ioengine, "cpuio"))
1970 if (td->runstate == TD_REAPED)
1972 if (td->o.use_thread) {
1973 if (td->runstate == TD_EXITED) {
1974 td_set_runstate(td, TD_REAPED);
1981 if (td->runstate == TD_EXITED)
1985 * check if someone quit or got killed in an unusual way
1987 ret = waitpid(td->pid, &status, flags);
1989 if (errno == ECHILD) {
1990 log_err("fio: pid=%d disappeared %d\n",
1991 (int) td->pid, td->runstate);
1993 td_set_runstate(td, TD_REAPED);
1997 } else if (ret == td->pid) {
1998 if (WIFSIGNALED(status)) {
1999 int sig = WTERMSIG(status);
2001 if (sig != SIGTERM && sig != SIGUSR2)
2002 log_err("fio: pid=%d, got signal=%d\n",
2003 (int) td->pid, sig);
2005 td_set_runstate(td, TD_REAPED);
2008 if (WIFEXITED(status)) {
2009 if (WEXITSTATUS(status) && !td->error)
2010 td->error = WEXITSTATUS(status);
2012 td_set_runstate(td, TD_REAPED);
2018 * If the job is stuck, do a forceful timeout of it and
2021 if (td->terminate &&
2022 td->runstate < TD_FSYNCING &&
2023 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2024 log_err("fio: job '%s' (state=%d) hasn't exited in "
2025 "%lu seconds, it appears to be stuck. Doing "
2026 "forceful exit of this job.\n",
2027 td->o.name, td->runstate,
2028 (unsigned long) time_since_now(&td->terminate_time));
2029 td_set_runstate(td, TD_REAPED);
2034 * thread is not dead, continue
2040 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2041 (*t_rate) -= ddir_rw_sum(td->o.rate);
2048 done_secs += mtime_since_now(&td->epoch) / 1000;
2049 profile_td_exit(td);
2052 if (*nr_running == cputhreads && !pending && realthreads)
2053 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2056 static bool __check_trigger_file(void)
2063 if (stat(trigger_file, &sb))
2066 if (unlink(trigger_file) < 0)
2067 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2073 static bool trigger_timedout(void)
2075 if (trigger_timeout)
2076 if (time_since_genesis() >= trigger_timeout) {
2077 trigger_timeout = 0;
2084 void exec_trigger(const char *cmd)
2088 if (!cmd || cmd[0] == '\0')
2093 log_err("fio: failed executing %s trigger\n", cmd);
2096 void check_trigger_file(void)
2098 if (__check_trigger_file() || trigger_timedout()) {
2100 fio_clients_send_trigger(trigger_remote_cmd);
2102 verify_save_state(IO_LIST_ALL);
2103 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2104 exec_trigger(trigger_cmd);
2109 static int fio_verify_load_state(struct thread_data *td)
2113 if (!td->o.verify_state)
2119 ret = fio_server_get_verify_state(td->o.name,
2120 td->thread_number - 1, &data);
2122 verify_assign_state(td, data);
2124 ret = verify_load_state(td, "local");
2129 static void do_usleep(unsigned int usecs)
2131 check_for_running_stats();
2132 check_trigger_file();
2136 static bool check_mount_writes(struct thread_data *td)
2141 if (!td_write(td) || td->o.allow_mounted_write)
2145 * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2146 * are mkfs'd and mounted.
2148 for_each_file(td, f, i) {
2149 #ifdef FIO_HAVE_CHARDEV_SIZE
2150 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2152 if (f->filetype != FIO_TYPE_BLOCK)
2155 if (device_is_mounted(f->file_name))
2161 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2165 static bool waitee_running(struct thread_data *me)
2167 const char *waitee = me->o.wait_for;
2168 const char *self = me->o.name;
2169 struct thread_data *td;
2175 for_each_td(td, i) {
2176 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2179 if (td->runstate < TD_EXITED) {
2180 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2182 runstate_to_name(td->runstate));
2187 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2192 * Main function for kicking off and reaping jobs, as needed.
2194 static void run_threads(struct sk_out *sk_out)
2196 struct thread_data *td;
2197 unsigned int i, todo, nr_running, nr_started;
2198 uint64_t m_rate, t_rate;
2201 if (fio_gtod_offload && fio_start_gtod_thread())
2204 fio_idle_prof_init();
2208 nr_thread = nr_process = 0;
2209 for_each_td(td, i) {
2210 if (check_mount_writes(td))
2212 if (td->o.use_thread)
2218 if (output_format & FIO_OUTPUT_NORMAL) {
2219 struct buf_output out;
2221 buf_output_init(&out);
2222 __log_buf(&out, "Starting ");
2224 __log_buf(&out, "%d thread%s", nr_thread,
2225 nr_thread > 1 ? "s" : "");
2228 __log_buf(&out, " and ");
2229 __log_buf(&out, "%d process%s", nr_process,
2230 nr_process > 1 ? "es" : "");
2232 __log_buf(&out, "\n");
2233 log_info_buf(out.buf, out.buflen);
2234 buf_output_free(&out);
2237 todo = thread_number;
2240 m_rate = t_rate = 0;
2242 for_each_td(td, i) {
2243 print_status_init(td->thread_number - 1);
2245 if (!td->o.create_serialize)
2248 if (fio_verify_load_state(td))
2252 * do file setup here so it happens sequentially,
2253 * we don't want X number of threads getting their
2254 * client data interspersed on disk
2256 if (setup_files(td)) {
2260 log_err("fio: pid=%d, err=%d/%s\n",
2261 (int) td->pid, td->error, td->verror);
2262 td_set_runstate(td, TD_REAPED);
2269 * for sharing to work, each job must always open
2270 * its own files. so close them, if we opened them
2273 for_each_file(td, f, j) {
2274 if (fio_file_open(f))
2275 td_io_close_file(td, f);
2280 /* start idle threads before io threads start to run */
2281 fio_idle_prof_start();
2286 struct thread_data *map[REAL_MAX_JOBS];
2287 struct timespec this_start;
2288 int this_jobs = 0, left;
2289 struct fork_data *fd;
2292 * create threads (TD_NOT_CREATED -> TD_CREATED)
2294 for_each_td(td, i) {
2295 if (td->runstate != TD_NOT_CREATED)
2299 * never got a chance to start, killed by other
2300 * thread for some reason
2302 if (td->terminate) {
2307 if (td->o.start_delay) {
2308 spent = utime_since_genesis();
2310 if (td->o.start_delay > spent)
2314 if (td->o.stonewall && (nr_started || nr_running)) {
2315 dprint(FD_PROCESS, "%s: stonewall wait\n",
2320 if (waitee_running(td)) {
2321 dprint(FD_PROCESS, "%s: waiting for %s\n",
2322 td->o.name, td->o.wait_for);
2328 td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2329 td->update_rusage = 0;
2332 * Set state to created. Thread will transition
2333 * to TD_INITIALIZED when it's done setting up.
2335 td_set_runstate(td, TD_CREATED);
2336 map[this_jobs++] = td;
2339 fd = calloc(1, sizeof(*fd));
2341 fd->sk_out = sk_out;
2343 if (td->o.use_thread) {
2346 dprint(FD_PROCESS, "will pthread_create\n");
2347 ret = pthread_create(&td->thread, NULL,
2350 log_err("pthread_create: %s\n",
2357 ret = pthread_detach(td->thread);
2359 log_err("pthread_detach: %s",
2363 dprint(FD_PROCESS, "will fork\n");
2368 ret = (int)(uintptr_t)thread_main(fd);
2370 } else if (i == fio_debug_jobno)
2371 *fio_debug_jobp = pid;
2373 dprint(FD_MUTEX, "wait on startup_sem\n");
2374 if (fio_sem_down_timeout(startup_sem, 10000)) {
2375 log_err("fio: job startup hung? exiting.\n");
2376 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2382 dprint(FD_MUTEX, "done waiting on startup_sem\n");
2386 * Wait for the started threads to transition to
2389 fio_gettime(&this_start, NULL);
2391 while (left && !fio_abort) {
2392 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2397 for (i = 0; i < this_jobs; i++) {
2401 if (td->runstate == TD_INITIALIZED) {
2404 } else if (td->runstate >= TD_EXITED) {
2408 nr_running++; /* work-around... */
2414 log_err("fio: %d job%s failed to start\n", left,
2415 left > 1 ? "s" : "");
2416 for (i = 0; i < this_jobs; i++) {
2420 kill(td->pid, SIGTERM);
2426 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2428 for_each_td(td, i) {
2429 if (td->runstate != TD_INITIALIZED)
2432 if (in_ramp_time(td))
2433 td_set_runstate(td, TD_RAMP);
2435 td_set_runstate(td, TD_RUNNING);
2438 m_rate += ddir_rw_sum(td->o.ratemin);
2439 t_rate += ddir_rw_sum(td->o.rate);
2441 fio_sem_up(td->sem);
2444 reap_threads(&nr_running, &t_rate, &m_rate);
2450 while (nr_running) {
2451 reap_threads(&nr_running, &t_rate, &m_rate);
2455 fio_idle_prof_stop();
2460 static void free_disk_util(void)
2462 disk_util_prune_entries();
2463 helper_thread_destroy();
2466 int fio_backend(struct sk_out *sk_out)
2468 struct thread_data *td;
2472 if (load_profile(exec_profile))
2475 exec_profile = NULL;
2481 struct log_params p = {
2482 .log_type = IO_LOG_TYPE_BW,
2485 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2486 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2487 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2490 startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2492 is_local_backend = true;
2493 if (startup_sem == NULL)
2498 helper_thread_create(startup_sem, sk_out);
2500 cgroup_list = smalloc(sizeof(*cgroup_list));
2502 INIT_FLIST_HEAD(cgroup_list);
2504 run_threads(sk_out);
2506 helper_thread_exit();
2511 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2512 struct io_log *log = agg_io_log[i];
2514 flush_log(log, false);
2520 for_each_td(td, i) {
2521 steadystate_free(td);
2522 fio_options_free(td);
2523 if (td->rusage_sem) {
2524 fio_sem_remove(td->rusage_sem);
2525 td->rusage_sem = NULL;
2527 fio_sem_remove(td->sem);
2533 cgroup_kill(cgroup_list);
2537 fio_sem_remove(startup_sem);