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.
40 #include "lib/memalign.h"
42 #include "lib/getrusage.h"
45 #include "workqueue.h"
46 #include "lib/mountcheck.h"
47 #include "rate-submit.h"
48 #include "helper_thread.h"
51 static struct fio_sem *startup_sem;
52 static struct flist_head *cgroup_list;
53 static char *cgroup_mnt;
54 static int exit_value;
55 static volatile int fio_abort;
56 static unsigned int nr_process = 0;
57 static unsigned int nr_thread = 0;
59 struct io_log *agg_io_log[DDIR_RWDIR_CNT];
62 unsigned int thread_number = 0;
63 unsigned int stat_number = 0;
66 unsigned long done_secs = 0;
68 #define JOB_START_TIMEOUT (5 * 1000)
70 static void sig_int(int sig)
74 fio_server_got_signal(sig);
76 log_info("\nfio: terminating on signal %d\n", sig);
81 fio_terminate_threads(TERMINATE_ALL);
85 void sig_show_status(int sig)
87 show_running_run_stats();
90 static void set_sig_handlers(void)
94 memset(&act, 0, sizeof(act));
95 act.sa_handler = sig_int;
96 act.sa_flags = SA_RESTART;
97 sigaction(SIGINT, &act, NULL);
99 memset(&act, 0, sizeof(act));
100 act.sa_handler = sig_int;
101 act.sa_flags = SA_RESTART;
102 sigaction(SIGTERM, &act, NULL);
104 /* Windows uses SIGBREAK as a quit signal from other applications */
106 memset(&act, 0, sizeof(act));
107 act.sa_handler = sig_int;
108 act.sa_flags = SA_RESTART;
109 sigaction(SIGBREAK, &act, NULL);
112 memset(&act, 0, sizeof(act));
113 act.sa_handler = sig_show_status;
114 act.sa_flags = SA_RESTART;
115 sigaction(SIGUSR1, &act, NULL);
118 memset(&act, 0, sizeof(act));
119 act.sa_handler = sig_int;
120 act.sa_flags = SA_RESTART;
121 sigaction(SIGPIPE, &act, NULL);
126 * Check if we are above the minimum rate given.
128 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
131 unsigned long long bytes = 0;
132 unsigned long iops = 0;
135 unsigned int ratemin = 0;
136 unsigned int rate_iops = 0;
137 unsigned int rate_iops_min = 0;
139 assert(ddir_rw(ddir));
141 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
145 * allow a 2 second settle period in the beginning
147 if (mtime_since(&td->start, now) < 2000)
150 iops += td->this_io_blocks[ddir];
151 bytes += td->this_io_bytes[ddir];
152 ratemin += td->o.ratemin[ddir];
153 rate_iops += td->o.rate_iops[ddir];
154 rate_iops_min += td->o.rate_iops_min[ddir];
157 * if rate blocks is set, sample is running
159 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
160 spent = mtime_since(&td->lastrate[ddir], now);
161 if (spent < td->o.ratecycle)
164 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
166 * check bandwidth specified rate
168 if (bytes < td->rate_bytes[ddir]) {
169 log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
170 td->o.name, ratemin, bytes);
174 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
178 if (rate < ratemin ||
179 bytes < td->rate_bytes[ddir]) {
180 log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
181 td->o.name, ratemin, rate);
187 * checks iops specified rate
189 if (iops < rate_iops) {
190 log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
191 td->o.name, rate_iops, iops);
195 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
199 if (rate < rate_iops_min ||
200 iops < td->rate_blocks[ddir]) {
201 log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
202 td->o.name, rate_iops_min, rate);
209 td->rate_bytes[ddir] = bytes;
210 td->rate_blocks[ddir] = iops;
211 memcpy(&td->lastrate[ddir], now, sizeof(*now));
215 static bool check_min_rate(struct thread_data *td, struct timespec *now)
219 if (td->bytes_done[DDIR_READ])
220 ret |= __check_min_rate(td, now, DDIR_READ);
221 if (td->bytes_done[DDIR_WRITE])
222 ret |= __check_min_rate(td, now, DDIR_WRITE);
223 if (td->bytes_done[DDIR_TRIM])
224 ret |= __check_min_rate(td, now, DDIR_TRIM);
230 * When job exits, we can cancel the in-flight IO if we are using async
231 * io. Attempt to do so.
233 static void cleanup_pending_aio(struct thread_data *td)
238 * get immediately available events, if any
240 r = io_u_queued_complete(td, 0);
245 * now cancel remaining active events
247 if (td->io_ops->cancel) {
251 io_u_qiter(&td->io_u_all, io_u, i) {
252 if (io_u->flags & IO_U_F_FLIGHT) {
253 r = td->io_ops->cancel(td, io_u);
261 r = io_u_queued_complete(td, td->cur_depth);
265 * Helper to handle the final sync of a file. Works just like the normal
266 * io path, just does everything sync.
268 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
270 struct io_u *io_u = __get_io_u(td);
276 io_u->ddir = DDIR_SYNC;
279 if (td_io_prep(td, io_u)) {
285 ret = td_io_queue(td, io_u);
287 td_verror(td, io_u->error, "td_io_queue");
290 } else if (ret == FIO_Q_QUEUED) {
291 if (td_io_commit(td))
293 if (io_u_queued_complete(td, 1) < 0)
295 } else if (ret == FIO_Q_COMPLETED) {
297 td_verror(td, io_u->error, "td_io_queue");
301 if (io_u_sync_complete(td, io_u) < 0)
303 } else if (ret == FIO_Q_BUSY) {
304 if (td_io_commit(td))
312 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
316 if (fio_file_open(f))
317 return fio_io_sync(td, f);
319 if (td_io_open_file(td, f))
322 ret = fio_io_sync(td, f);
323 td_io_close_file(td, f);
327 static inline void __update_ts_cache(struct thread_data *td)
329 fio_gettime(&td->ts_cache, NULL);
332 static inline void update_ts_cache(struct thread_data *td)
334 if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
335 __update_ts_cache(td);
338 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
340 if (in_ramp_time(td))
344 if (utime_since(&td->epoch, t) >= td->o.timeout)
351 * We need to update the runtime consistently in ms, but keep a running
352 * tally of the current elapsed time in microseconds for sub millisecond
355 static inline void update_runtime(struct thread_data *td,
356 unsigned long long *elapsed_us,
357 const enum fio_ddir ddir)
359 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
362 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
363 elapsed_us[ddir] += utime_since_now(&td->start);
364 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
367 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
372 if (ret < 0 || td->error) {
374 enum error_type_bit eb;
379 eb = td_error_type(ddir, err);
380 if (!(td->o.continue_on_error & (1 << eb)))
383 if (td_non_fatal_error(td, eb, err)) {
385 * Continue with the I/Os in case of
388 update_error_count(td, err);
392 } else if (td->o.fill_device && err == ENOSPC) {
394 * We expect to hit this error if
395 * fill_device option is set.
398 fio_mark_td_terminate(td);
402 * Stop the I/O in case of a fatal
405 update_error_count(td, err);
413 static void check_update_rusage(struct thread_data *td)
415 if (td->update_rusage) {
416 td->update_rusage = 0;
417 update_rusage_stat(td);
418 fio_sem_up(td->rusage_sem);
422 static int wait_for_completions(struct thread_data *td, struct timespec *time)
424 const int full = queue_full(td);
428 if (td->flags & TD_F_REGROW_LOGS)
429 return io_u_quiesce(td);
432 * if the queue is full, we MUST reap at least 1 event
434 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
435 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
438 if (time && (__should_check_rate(td, DDIR_READ) ||
439 __should_check_rate(td, DDIR_WRITE) ||
440 __should_check_rate(td, DDIR_TRIM)))
441 fio_gettime(time, NULL);
444 ret = io_u_queued_complete(td, min_evts);
447 } while (full && (td->cur_depth > td->o.iodepth_low));
452 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
453 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
454 struct timespec *comp_time)
459 case FIO_Q_COMPLETED:
462 clear_io_u(td, io_u);
463 } else if (io_u->resid) {
464 int bytes = io_u->xfer_buflen - io_u->resid;
465 struct fio_file *f = io_u->file;
468 *bytes_issued += bytes;
471 trim_io_piece(td, io_u);
478 unlog_io_piece(td, io_u);
479 td_verror(td, EIO, "full resid");
484 io_u->xfer_buflen = io_u->resid;
485 io_u->xfer_buf += bytes;
486 io_u->offset += bytes;
488 if (ddir_rw(io_u->ddir))
489 td->ts.short_io_u[io_u->ddir]++;
491 if (io_u->offset == f->real_file_size)
494 requeue_io_u(td, &io_u);
497 if (comp_time && (__should_check_rate(td, DDIR_READ) ||
498 __should_check_rate(td, DDIR_WRITE) ||
499 __should_check_rate(td, DDIR_TRIM)))
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);
533 ret2 = td_io_commit(td);
539 td_verror(td, -(*ret), "td_io_queue");
543 if (break_on_this_error(td, ddir, ret))
549 static inline bool io_in_polling(struct thread_data *td)
551 return !td->o.iodepth_batch_complete_min &&
552 !td->o.iodepth_batch_complete_max;
555 * Unlinks files from thread data fio_file structure
557 static int unlink_all_files(struct thread_data *td)
563 for_each_file(td, f, i) {
564 if (f->filetype != FIO_TYPE_FILE)
566 ret = td_io_unlink_file(td, f);
572 td_verror(td, ret, "unlink_all_files");
578 * Check if io_u will overlap an in-flight IO in the queue
580 static bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
583 struct io_u *check_io_u;
584 unsigned long long x1, x2, y1, y2;
588 x2 = io_u->offset + io_u->buflen;
590 io_u_qiter(q, check_io_u, i) {
591 if (check_io_u->flags & IO_U_F_FLIGHT) {
592 y1 = check_io_u->offset;
593 y2 = check_io_u->offset + check_io_u->buflen;
595 if (x1 < y2 && y1 < x2) {
597 dprint(FD_IO, "in-flight overlap: %llu/%lu, %llu/%lu\n",
599 y1, check_io_u->buflen);
608 static int io_u_submit(struct thread_data *td, struct io_u *io_u)
611 * Check for overlap if the user asked us to, and we have
612 * at least one IO in flight besides this one.
614 if (td->o.serialize_overlap && td->cur_depth > 1 &&
615 in_flight_overlap(&td->io_u_all, io_u))
618 return td_io_queue(td, io_u);
622 * The main verify engine. Runs over the writes we previously submitted,
623 * reads the blocks back in, and checks the crc/md5 of the data.
625 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
632 dprint(FD_VERIFY, "starting loop\n");
635 * sync io first and invalidate cache, to make sure we really
638 for_each_file(td, f, i) {
639 if (!fio_file_open(f))
641 if (fio_io_sync(td, f))
643 if (file_invalidate_cache(td, f))
647 check_update_rusage(td);
653 * verify_state needs to be reset before verification
654 * proceeds so that expected random seeds match actual
655 * random seeds in headers. The main loop will reset
656 * all random number generators if randrepeat is set.
658 if (!td->o.rand_repeatable)
659 td_fill_verify_state_seed(td);
661 td_set_runstate(td, TD_VERIFYING);
664 while (!td->terminate) {
669 check_update_rusage(td);
671 if (runtime_exceeded(td, &td->ts_cache)) {
672 __update_ts_cache(td);
673 if (runtime_exceeded(td, &td->ts_cache)) {
674 fio_mark_td_terminate(td);
679 if (flow_threshold_exceeded(td))
682 if (!td->o.experimental_verify) {
683 io_u = __get_io_u(td);
687 if (get_next_verify(td, io_u)) {
692 if (td_io_prep(td, io_u)) {
697 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
700 while ((io_u = get_io_u(td)) != NULL) {
701 if (IS_ERR_OR_NULL(io_u)) {
708 * We are only interested in the places where
709 * we wrote or trimmed IOs. Turn those into
710 * reads for verification purposes.
712 if (io_u->ddir == DDIR_READ) {
714 * Pretend we issued it for rwmix
717 td->io_issues[DDIR_READ]++;
720 } else if (io_u->ddir == DDIR_TRIM) {
721 io_u->ddir = DDIR_READ;
722 io_u_set(td, io_u, IO_U_F_TRIMMED);
724 } else if (io_u->ddir == DDIR_WRITE) {
725 io_u->ddir = DDIR_READ;
737 if (verify_state_should_stop(td, io_u)) {
742 if (td->o.verify_async)
743 io_u->end_io = verify_io_u_async;
745 io_u->end_io = verify_io_u;
748 if (!td->o.disable_slat)
749 fio_gettime(&io_u->start_time, NULL);
751 ret = io_u_submit(td, io_u);
753 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
757 * if we can queue more, do so. but check if there are
758 * completed io_u's first. Note that we can get BUSY even
759 * without IO queued, if the system is resource starved.
762 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
763 if (full || io_in_polling(td))
764 ret = wait_for_completions(td, NULL);
770 check_update_rusage(td);
773 min_events = td->cur_depth;
776 ret = io_u_queued_complete(td, min_events);
778 cleanup_pending_aio(td);
780 td_set_runstate(td, TD_RUNNING);
782 dprint(FD_VERIFY, "exiting loop\n");
785 static bool exceeds_number_ios(struct thread_data *td)
787 unsigned long long number_ios;
789 if (!td->o.number_ios)
792 number_ios = ddir_rw_sum(td->io_blocks);
793 number_ios += td->io_u_queued + td->io_u_in_flight;
795 return number_ios >= (td->o.number_ios * td->loops);
798 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
800 unsigned long long bytes, limit;
803 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
804 else if (td_write(td))
805 bytes = this_bytes[DDIR_WRITE];
806 else if (td_read(td))
807 bytes = this_bytes[DDIR_READ];
809 bytes = this_bytes[DDIR_TRIM];
812 limit = td->o.io_size;
817 return bytes >= limit || exceeds_number_ios(td);
820 static bool io_issue_bytes_exceeded(struct thread_data *td)
822 return io_bytes_exceeded(td, td->io_issue_bytes);
825 static bool io_complete_bytes_exceeded(struct thread_data *td)
827 return io_bytes_exceeded(td, td->this_io_bytes);
831 * used to calculate the next io time for rate control
834 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
836 uint64_t bps = td->rate_bps[ddir];
838 assert(!(td->flags & TD_F_CHILD));
840 if (td->o.rate_process == RATE_PROCESS_POISSON) {
843 iops = bps / td->o.bs[ddir];
844 val = (int64_t) (1000000 / iops) *
845 -logf(__rand_0_1(&td->poisson_state[ddir]));
847 dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
848 (unsigned long long) 1000000 / val,
851 td->last_usec[ddir] += val;
852 return td->last_usec[ddir];
854 uint64_t bytes = td->rate_io_issue_bytes[ddir];
855 uint64_t secs = bytes / bps;
856 uint64_t remainder = bytes % bps;
858 return remainder * 1000000 / bps + secs * 1000000;
864 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
866 unsigned long long b;
870 b = ddir_rw_sum(td->io_blocks);
871 if (b % td->o.thinktime_blocks)
877 if (td->o.thinktime_spin)
878 total = usec_spin(td->o.thinktime_spin);
880 left = td->o.thinktime - total;
882 total += usec_sleep(td, left);
885 * If we're ignoring thinktime for the rate, add the number of bytes
886 * we would have done while sleeping, minus one block to ensure we
887 * start issuing immediately after the sleep.
889 if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
890 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
891 uint64_t bs = td->o.min_bs[ddir];
892 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
895 if (usperop <= total)
898 over = (usperop - total) / usperop * -bs;
900 td->rate_io_issue_bytes[ddir] += (missed - over);
905 * Main IO worker function. It retrieves io_u's to process and queues
906 * and reaps them, checking for rate and errors along the way.
908 * Returns number of bytes written and trimmed.
910 static void do_io(struct thread_data *td, uint64_t *bytes_done)
914 uint64_t total_bytes, bytes_issued = 0;
916 for (i = 0; i < DDIR_RWDIR_CNT; i++)
917 bytes_done[i] = td->bytes_done[i];
919 if (in_ramp_time(td))
920 td_set_runstate(td, TD_RAMP);
922 td_set_runstate(td, TD_RUNNING);
926 total_bytes = td->o.size;
928 * Allow random overwrite workloads to write up to io_size
929 * before starting verification phase as 'size' doesn't apply.
931 if (td_write(td) && td_random(td) && td->o.norandommap)
932 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
934 * If verify_backlog is enabled, we'll run the verify in this
935 * handler as well. For that case, we may need up to twice the
938 if (td->o.verify != VERIFY_NONE &&
939 (td_write(td) && td->o.verify_backlog))
940 total_bytes += td->o.size;
942 /* In trimwrite mode, each byte is trimmed and then written, so
943 * allow total_bytes to be twice as big */
944 if (td_trimwrite(td))
945 total_bytes += td->total_io_size;
947 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
948 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
950 struct timespec comp_time;
955 check_update_rusage(td);
957 if (td->terminate || td->done)
962 if (runtime_exceeded(td, &td->ts_cache)) {
963 __update_ts_cache(td);
964 if (runtime_exceeded(td, &td->ts_cache)) {
965 fio_mark_td_terminate(td);
970 if (flow_threshold_exceeded(td))
974 * Break if we exceeded the bytes. The exception is time
975 * based runs, but we still need to break out of the loop
976 * for those to run verification, if enabled.
978 if (bytes_issued >= total_bytes &&
979 (!td->o.time_based ||
980 (td->o.time_based && td->o.verify != VERIFY_NONE)))
984 if (IS_ERR_OR_NULL(io_u)) {
985 int err = PTR_ERR(io_u);
993 if (td->o.latency_target)
1001 * Add verification end_io handler if:
1002 * - Asked to verify (!td_rw(td))
1003 * - Or the io_u is from our verify list (mixed write/ver)
1005 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1006 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1008 if (!td->o.verify_pattern_bytes) {
1009 io_u->rand_seed = __rand(&td->verify_state);
1010 if (sizeof(int) != sizeof(long *))
1011 io_u->rand_seed *= __rand(&td->verify_state);
1014 if (verify_state_should_stop(td, io_u)) {
1019 if (td->o.verify_async)
1020 io_u->end_io = verify_io_u_async;
1022 io_u->end_io = verify_io_u;
1023 td_set_runstate(td, TD_VERIFYING);
1024 } else if (in_ramp_time(td))
1025 td_set_runstate(td, TD_RAMP);
1027 td_set_runstate(td, TD_RUNNING);
1030 * Always log IO before it's issued, so we know the specific
1031 * order of it. The logged unit will track when the IO has
1034 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1036 td->o.verify != VERIFY_NONE &&
1037 !td->o.experimental_verify)
1038 log_io_piece(td, io_u);
1040 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1041 const unsigned long blen = io_u->xfer_buflen;
1042 const enum fio_ddir ddir = acct_ddir(io_u);
1047 workqueue_enqueue(&td->io_wq, &io_u->work);
1050 if (ddir_rw(ddir)) {
1051 td->io_issues[ddir]++;
1052 td->io_issue_bytes[ddir] += blen;
1053 td->rate_io_issue_bytes[ddir] += blen;
1056 if (should_check_rate(td))
1057 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1060 ret = io_u_submit(td, io_u);
1062 if (should_check_rate(td))
1063 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1065 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1069 * See if we need to complete some commands. Note that
1070 * we can get BUSY even without IO queued, if the
1071 * system is resource starved.
1074 full = queue_full(td) ||
1075 (ret == FIO_Q_BUSY && td->cur_depth);
1076 if (full || io_in_polling(td))
1077 ret = wait_for_completions(td, &comp_time);
1081 if (!ddir_rw_sum(td->bytes_done) &&
1082 !td_ioengine_flagged(td, FIO_NOIO))
1085 if (!in_ramp_time(td) && should_check_rate(td)) {
1086 if (check_min_rate(td, &comp_time)) {
1087 if (exitall_on_terminate || td->o.exitall_error)
1088 fio_terminate_threads(td->groupid);
1089 td_verror(td, EIO, "check_min_rate");
1093 if (!in_ramp_time(td) && td->o.latency_target)
1094 lat_target_check(td);
1096 if (ddir_rw(ddir) && td->o.thinktime)
1097 handle_thinktime(td, ddir);
1100 check_update_rusage(td);
1102 if (td->trim_entries)
1103 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1105 if (td->o.fill_device && td->error == ENOSPC) {
1107 fio_mark_td_terminate(td);
1112 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1113 workqueue_flush(&td->io_wq);
1119 ret = io_u_queued_complete(td, i);
1120 if (td->o.fill_device && td->error == ENOSPC)
1124 if (should_fsync(td) && td->o.end_fsync) {
1125 td_set_runstate(td, TD_FSYNCING);
1127 for_each_file(td, f, i) {
1128 if (!fio_file_fsync(td, f))
1131 log_err("fio: end_fsync failed for file %s\n",
1136 cleanup_pending_aio(td);
1139 * stop job if we failed doing any IO
1141 if (!ddir_rw_sum(td->this_io_bytes))
1144 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1145 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1148 static void free_file_completion_logging(struct thread_data *td)
1153 for_each_file(td, f, i) {
1154 if (!f->last_write_comp)
1156 sfree(f->last_write_comp);
1160 static int init_file_completion_logging(struct thread_data *td,
1166 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1169 for_each_file(td, f, i) {
1170 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1171 if (!f->last_write_comp)
1178 free_file_completion_logging(td);
1179 log_err("fio: failed to alloc write comp data\n");
1183 static void cleanup_io_u(struct thread_data *td)
1187 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1189 if (td->io_ops->io_u_free)
1190 td->io_ops->io_u_free(td, io_u);
1192 fio_memfree(io_u, sizeof(*io_u));
1197 io_u_rexit(&td->io_u_requeues);
1198 io_u_qexit(&td->io_u_freelist);
1199 io_u_qexit(&td->io_u_all);
1201 free_file_completion_logging(td);
1204 static int init_io_u(struct thread_data *td)
1207 unsigned int max_bs, min_write;
1208 int cl_align, i, max_units;
1209 int data_xfer = 1, err;
1212 max_units = td->o.iodepth;
1213 max_bs = td_max_bs(td);
1214 min_write = td->o.min_bs[DDIR_WRITE];
1215 td->orig_buffer_size = (unsigned long long) max_bs
1216 * (unsigned long long) max_units;
1218 if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1222 err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1223 err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1224 err += !io_u_qinit(&td->io_u_all, td->o.iodepth);
1227 log_err("fio: failed setting up IO queues\n");
1232 * if we may later need to do address alignment, then add any
1233 * possible adjustment here so that we don't cause a buffer
1234 * overflow later. this adjustment may be too much if we get
1235 * lucky and the allocator gives us an aligned address.
1237 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1238 td_ioengine_flagged(td, FIO_RAWIO))
1239 td->orig_buffer_size += page_mask + td->o.mem_align;
1241 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1244 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1245 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1248 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1249 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1253 if (data_xfer && allocate_io_mem(td))
1256 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1257 td_ioengine_flagged(td, FIO_RAWIO))
1258 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1260 p = td->orig_buffer;
1262 cl_align = os_cache_line_size();
1264 for (i = 0; i < max_units; i++) {
1270 ptr = fio_memalign(cl_align, sizeof(*io_u));
1272 log_err("fio: unable to allocate aligned memory\n");
1277 memset(io_u, 0, sizeof(*io_u));
1278 INIT_FLIST_HEAD(&io_u->verify_list);
1279 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1283 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1286 io_u_fill_buffer(td, io_u, min_write, max_bs);
1287 if (td_write(td) && td->o.verify_pattern_bytes) {
1289 * Fill the buffer with the pattern if we are
1290 * going to be doing writes.
1292 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1297 io_u->flags = IO_U_F_FREE;
1298 io_u_qpush(&td->io_u_freelist, io_u);
1301 * io_u never leaves this stack, used for iteration of all
1304 io_u_qpush(&td->io_u_all, io_u);
1306 if (td->io_ops->io_u_init) {
1307 int ret = td->io_ops->io_u_init(td, io_u);
1310 log_err("fio: failed to init engine data: %d\n", ret);
1318 if (init_file_completion_logging(td, max_units))
1325 * This function is Linux specific.
1326 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1328 static int switch_ioscheduler(struct thread_data *td)
1330 #ifdef FIO_HAVE_IOSCHED_SWITCH
1331 char tmp[256], tmp2[128];
1335 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1338 assert(td->files && td->files[0]);
1339 sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1341 f = fopen(tmp, "r+");
1343 if (errno == ENOENT) {
1344 log_err("fio: os or kernel doesn't support IO scheduler"
1348 td_verror(td, errno, "fopen iosched");
1355 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1356 if (ferror(f) || ret != 1) {
1357 td_verror(td, errno, "fwrite");
1365 * Read back and check that the selected scheduler is now the default.
1367 memset(tmp, 0, sizeof(tmp));
1368 ret = fread(tmp, sizeof(tmp), 1, f);
1369 if (ferror(f) || ret < 0) {
1370 td_verror(td, errno, "fread");
1375 * either a list of io schedulers or "none\n" is expected.
1377 tmp[strlen(tmp) - 1] = '\0';
1380 * Write to "none" entry doesn't fail, so check the result here.
1382 if (!strcmp(tmp, "none")) {
1383 log_err("fio: io scheduler is not tunable\n");
1388 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1389 if (!strstr(tmp, tmp2)) {
1390 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1391 td_verror(td, EINVAL, "iosched_switch");
1403 static bool keep_running(struct thread_data *td)
1405 unsigned long long limit;
1411 if (td->o.time_based)
1417 if (exceeds_number_ios(td))
1421 limit = td->o.io_size;
1425 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1429 * If the difference is less than the maximum IO size, we
1432 diff = limit - ddir_rw_sum(td->io_bytes);
1433 if (diff < td_max_bs(td))
1436 if (fio_files_done(td) && !td->o.io_size)
1445 static int exec_string(struct thread_options *o, const char *string, const char *mode)
1447 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1451 str = malloc(newlen);
1452 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1454 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1457 log_err("fio: exec of cmd <%s> failed\n", str);
1464 * Dry run to compute correct state of numberio for verification.
1466 static uint64_t do_dry_run(struct thread_data *td)
1468 td_set_runstate(td, TD_RUNNING);
1470 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1471 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1475 if (td->terminate || td->done)
1478 io_u = get_io_u(td);
1479 if (IS_ERR_OR_NULL(io_u))
1482 io_u_set(td, io_u, IO_U_F_FLIGHT);
1485 if (ddir_rw(acct_ddir(io_u)))
1486 td->io_issues[acct_ddir(io_u)]++;
1487 if (ddir_rw(io_u->ddir)) {
1488 io_u_mark_depth(td, 1);
1489 td->ts.total_io_u[io_u->ddir]++;
1492 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1494 td->o.verify != VERIFY_NONE &&
1495 !td->o.experimental_verify)
1496 log_io_piece(td, io_u);
1498 ret = io_u_sync_complete(td, io_u);
1502 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1506 struct thread_data *td;
1507 struct sk_out *sk_out;
1511 * Entry point for the thread based jobs. The process based jobs end up
1512 * here as well, after a little setup.
1514 static void *thread_main(void *data)
1516 struct fork_data *fd = data;
1517 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1518 struct thread_data *td = fd->td;
1519 struct thread_options *o = &td->o;
1520 struct sk_out *sk_out = fd->sk_out;
1521 uint64_t bytes_done[DDIR_RWDIR_CNT];
1522 int deadlock_loop_cnt;
1523 bool clear_state, did_some_io;
1526 sk_out_assign(sk_out);
1529 if (!o->use_thread) {
1535 fio_local_clock_init(o->use_thread);
1537 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1540 fio_server_send_start(td);
1542 INIT_FLIST_HEAD(&td->io_log_list);
1543 INIT_FLIST_HEAD(&td->io_hist_list);
1544 INIT_FLIST_HEAD(&td->verify_list);
1545 INIT_FLIST_HEAD(&td->trim_list);
1546 INIT_FLIST_HEAD(&td->next_rand_list);
1547 td->io_hist_tree = RB_ROOT;
1549 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1551 td_verror(td, ret, "mutex_cond_init_pshared");
1554 ret = cond_init_pshared(&td->verify_cond);
1556 td_verror(td, ret, "mutex_cond_pshared");
1560 td_set_runstate(td, TD_INITIALIZED);
1561 dprint(FD_MUTEX, "up startup_sem\n");
1562 fio_sem_up(startup_sem);
1563 dprint(FD_MUTEX, "wait on td->sem\n");
1564 fio_sem_down(td->sem);
1565 dprint(FD_MUTEX, "done waiting on td->sem\n");
1568 * A new gid requires privilege, so we need to do this before setting
1571 if (o->gid != -1U && setgid(o->gid)) {
1572 td_verror(td, errno, "setgid");
1575 if (o->uid != -1U && setuid(o->uid)) {
1576 td_verror(td, errno, "setuid");
1581 * Do this early, we don't want the compress threads to be limited
1582 * to the same CPUs as the IO workers. So do this before we set
1583 * any potential CPU affinity
1585 if (iolog_compress_init(td, sk_out))
1589 * If we have a gettimeofday() thread, make sure we exclude that
1590 * thread from this job
1593 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1596 * Set affinity first, in case it has an impact on the memory
1599 if (fio_option_is_set(o, cpumask)) {
1600 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1601 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1603 log_err("fio: no CPUs set\n");
1604 log_err("fio: Try increasing number of available CPUs\n");
1605 td_verror(td, EINVAL, "cpus_split");
1609 ret = fio_setaffinity(td->pid, o->cpumask);
1611 td_verror(td, errno, "cpu_set_affinity");
1616 #ifdef CONFIG_LIBNUMA
1617 /* numa node setup */
1618 if (fio_option_is_set(o, numa_cpunodes) ||
1619 fio_option_is_set(o, numa_memnodes)) {
1620 struct bitmask *mask;
1622 if (numa_available() < 0) {
1623 td_verror(td, errno, "Does not support NUMA API\n");
1627 if (fio_option_is_set(o, numa_cpunodes)) {
1628 mask = numa_parse_nodestring(o->numa_cpunodes);
1629 ret = numa_run_on_node_mask(mask);
1630 numa_free_nodemask(mask);
1632 td_verror(td, errno, \
1633 "numa_run_on_node_mask failed\n");
1638 if (fio_option_is_set(o, numa_memnodes)) {
1640 if (o->numa_memnodes)
1641 mask = numa_parse_nodestring(o->numa_memnodes);
1643 switch (o->numa_mem_mode) {
1644 case MPOL_INTERLEAVE:
1645 numa_set_interleave_mask(mask);
1648 numa_set_membind(mask);
1651 numa_set_localalloc();
1653 case MPOL_PREFERRED:
1654 numa_set_preferred(o->numa_mem_prefer_node);
1662 numa_free_nodemask(mask);
1668 if (fio_pin_memory(td))
1672 * May alter parameters that init_io_u() will use, so we need to
1681 if (o->verify_async && verify_async_init(td))
1684 if (fio_option_is_set(o, ioprio) ||
1685 fio_option_is_set(o, ioprio_class)) {
1686 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1688 td_verror(td, errno, "ioprio_set");
1693 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1697 if (nice(o->nice) == -1 && errno != 0) {
1698 td_verror(td, errno, "nice");
1702 if (o->ioscheduler && switch_ioscheduler(td))
1705 if (!o->create_serialize && setup_files(td))
1711 if (!init_random_map(td))
1714 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1717 if (o->pre_read && !pre_read_files(td))
1720 fio_verify_init(td);
1722 if (rate_submit_init(td, sk_out))
1725 set_epoch_time(td, o->log_unix_epoch);
1726 fio_getrusage(&td->ru_start);
1727 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1728 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1729 memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1731 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1732 o->ratemin[DDIR_TRIM]) {
1733 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1734 sizeof(td->bw_sample_time));
1735 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1736 sizeof(td->bw_sample_time));
1737 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1738 sizeof(td->bw_sample_time));
1741 memset(bytes_done, 0, sizeof(bytes_done));
1742 clear_state = false;
1743 did_some_io = false;
1745 while (keep_running(td)) {
1746 uint64_t verify_bytes;
1748 fio_gettime(&td->start, NULL);
1749 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1752 clear_io_state(td, 0);
1754 if (o->unlink_each_loop && unlink_all_files(td))
1758 prune_io_piece_log(td);
1760 if (td->o.verify_only && td_write(td))
1761 verify_bytes = do_dry_run(td);
1763 do_io(td, bytes_done);
1765 if (!ddir_rw_sum(bytes_done)) {
1766 fio_mark_td_terminate(td);
1769 verify_bytes = bytes_done[DDIR_WRITE] +
1770 bytes_done[DDIR_TRIM];
1775 * If we took too long to shut down, the main thread could
1776 * already consider us reaped/exited. If that happens, break
1779 if (td->runstate >= TD_EXITED)
1785 * Make sure we've successfully updated the rusage stats
1786 * before waiting on the stat mutex. Otherwise we could have
1787 * the stat thread holding stat mutex and waiting for
1788 * the rusage_sem, which would never get upped because
1789 * this thread is waiting for the stat mutex.
1791 deadlock_loop_cnt = 0;
1793 check_update_rusage(td);
1794 if (!fio_sem_down_trylock(stat_sem))
1797 if (deadlock_loop_cnt++ > 5000) {
1798 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1799 td->error = EDEADLK;
1804 if (td_read(td) && td->io_bytes[DDIR_READ])
1805 update_runtime(td, elapsed_us, DDIR_READ);
1806 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1807 update_runtime(td, elapsed_us, DDIR_WRITE);
1808 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1809 update_runtime(td, elapsed_us, DDIR_TRIM);
1810 fio_gettime(&td->start, NULL);
1811 fio_sem_up(stat_sem);
1813 if (td->error || td->terminate)
1816 if (!o->do_verify ||
1817 o->verify == VERIFY_NONE ||
1818 td_ioengine_flagged(td, FIO_UNIDIR))
1821 if (ddir_rw_sum(bytes_done))
1824 clear_io_state(td, 0);
1826 fio_gettime(&td->start, NULL);
1828 do_verify(td, verify_bytes);
1831 * See comment further up for why this is done here.
1833 check_update_rusage(td);
1835 fio_sem_down(stat_sem);
1836 update_runtime(td, elapsed_us, DDIR_READ);
1837 fio_gettime(&td->start, NULL);
1838 fio_sem_up(stat_sem);
1840 if (td->error || td->terminate)
1845 * If td ended up with no I/O when it should have had,
1846 * then something went wrong unless FIO_NOIO or FIO_DISKLESSIO.
1847 * (Are we not missing other flags that can be ignored ?)
1849 if ((td->o.size || td->o.io_size) && !ddir_rw_sum(bytes_done) &&
1850 !did_some_io && !td->o.create_only &&
1851 !(td_ioengine_flagged(td, FIO_NOIO) ||
1852 td_ioengine_flagged(td, FIO_DISKLESSIO)))
1853 log_err("%s: No I/O performed by %s, "
1854 "perhaps try --debug=io option for details?\n",
1855 td->o.name, td->io_ops->name);
1857 td_set_runstate(td, TD_FINISHING);
1859 update_rusage_stat(td);
1860 td->ts.total_run_time = mtime_since_now(&td->epoch);
1861 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1862 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1863 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
1865 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
1866 (td->o.verify != VERIFY_NONE && td_write(td)))
1867 verify_save_state(td->thread_number);
1869 fio_unpin_memory(td);
1871 td_writeout_logs(td, true);
1873 iolog_compress_exit(td);
1874 rate_submit_exit(td);
1876 if (o->exec_postrun)
1877 exec_string(o, o->exec_postrun, (const char *)"postrun");
1879 if (exitall_on_terminate || (o->exitall_error && td->error))
1880 fio_terminate_threads(td->groupid);
1884 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1887 if (o->verify_async)
1888 verify_async_exit(td);
1890 close_and_free_files(td);
1893 cgroup_shutdown(td, &cgroup_mnt);
1894 verify_free_state(td);
1896 if (td->zone_state_index) {
1899 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1900 free(td->zone_state_index[i]);
1901 free(td->zone_state_index);
1902 td->zone_state_index = NULL;
1905 if (fio_option_is_set(o, cpumask)) {
1906 ret = fio_cpuset_exit(&o->cpumask);
1908 td_verror(td, ret, "fio_cpuset_exit");
1912 * do this very late, it will log file closing as well
1914 if (o->write_iolog_file)
1915 write_iolog_close(td);
1917 td_set_runstate(td, TD_EXITED);
1920 * Do this last after setting our runstate to exited, so we
1921 * know that the stat thread is signaled.
1923 check_update_rusage(td);
1926 return (void *) (uintptr_t) td->error;
1930 * Run over the job map and reap the threads that have exited, if any.
1932 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1935 struct thread_data *td;
1936 unsigned int cputhreads, realthreads, pending;
1940 * reap exited threads (TD_EXITED -> TD_REAPED)
1942 realthreads = pending = cputhreads = 0;
1943 for_each_td(td, i) {
1946 if (!strcmp(td->o.ioengine, "cpuio"))
1955 if (td->runstate == TD_REAPED)
1957 if (td->o.use_thread) {
1958 if (td->runstate == TD_EXITED) {
1959 td_set_runstate(td, TD_REAPED);
1966 if (td->runstate == TD_EXITED)
1970 * check if someone quit or got killed in an unusual way
1972 ret = waitpid(td->pid, &status, flags);
1974 if (errno == ECHILD) {
1975 log_err("fio: pid=%d disappeared %d\n",
1976 (int) td->pid, td->runstate);
1978 td_set_runstate(td, TD_REAPED);
1982 } else if (ret == td->pid) {
1983 if (WIFSIGNALED(status)) {
1984 int sig = WTERMSIG(status);
1986 if (sig != SIGTERM && sig != SIGUSR2)
1987 log_err("fio: pid=%d, got signal=%d\n",
1988 (int) td->pid, sig);
1990 td_set_runstate(td, TD_REAPED);
1993 if (WIFEXITED(status)) {
1994 if (WEXITSTATUS(status) && !td->error)
1995 td->error = WEXITSTATUS(status);
1997 td_set_runstate(td, TD_REAPED);
2003 * If the job is stuck, do a forceful timeout of it and
2006 if (td->terminate &&
2007 td->runstate < TD_FSYNCING &&
2008 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2009 log_err("fio: job '%s' (state=%d) hasn't exited in "
2010 "%lu seconds, it appears to be stuck. Doing "
2011 "forceful exit of this job.\n",
2012 td->o.name, td->runstate,
2013 (unsigned long) time_since_now(&td->terminate_time));
2014 td_set_runstate(td, TD_REAPED);
2019 * thread is not dead, continue
2025 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2026 (*t_rate) -= ddir_rw_sum(td->o.rate);
2033 done_secs += mtime_since_now(&td->epoch) / 1000;
2034 profile_td_exit(td);
2037 if (*nr_running == cputhreads && !pending && realthreads)
2038 fio_terminate_threads(TERMINATE_ALL);
2041 static bool __check_trigger_file(void)
2048 if (stat(trigger_file, &sb))
2051 if (unlink(trigger_file) < 0)
2052 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2058 static bool trigger_timedout(void)
2060 if (trigger_timeout)
2061 if (time_since_genesis() >= trigger_timeout) {
2062 trigger_timeout = 0;
2069 void exec_trigger(const char *cmd)
2073 if (!cmd || cmd[0] == '\0')
2078 log_err("fio: failed executing %s trigger\n", cmd);
2081 void check_trigger_file(void)
2083 if (__check_trigger_file() || trigger_timedout()) {
2085 fio_clients_send_trigger(trigger_remote_cmd);
2087 verify_save_state(IO_LIST_ALL);
2088 fio_terminate_threads(TERMINATE_ALL);
2089 exec_trigger(trigger_cmd);
2094 static int fio_verify_load_state(struct thread_data *td)
2098 if (!td->o.verify_state)
2104 ret = fio_server_get_verify_state(td->o.name,
2105 td->thread_number - 1, &data);
2107 verify_assign_state(td, data);
2109 ret = verify_load_state(td, "local");
2114 static void do_usleep(unsigned int usecs)
2116 check_for_running_stats();
2117 check_trigger_file();
2121 static bool check_mount_writes(struct thread_data *td)
2126 if (!td_write(td) || td->o.allow_mounted_write)
2130 * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2131 * are mkfs'd and mounted.
2133 for_each_file(td, f, i) {
2134 #ifdef FIO_HAVE_CHARDEV_SIZE
2135 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2137 if (f->filetype != FIO_TYPE_BLOCK)
2140 if (device_is_mounted(f->file_name))
2146 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2150 static bool waitee_running(struct thread_data *me)
2152 const char *waitee = me->o.wait_for;
2153 const char *self = me->o.name;
2154 struct thread_data *td;
2160 for_each_td(td, i) {
2161 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2164 if (td->runstate < TD_EXITED) {
2165 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2167 runstate_to_name(td->runstate));
2172 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2177 * Main function for kicking off and reaping jobs, as needed.
2179 static void run_threads(struct sk_out *sk_out)
2181 struct thread_data *td;
2182 unsigned int i, todo, nr_running, nr_started;
2183 uint64_t m_rate, t_rate;
2186 if (fio_gtod_offload && fio_start_gtod_thread())
2189 fio_idle_prof_init();
2193 nr_thread = nr_process = 0;
2194 for_each_td(td, i) {
2195 if (check_mount_writes(td))
2197 if (td->o.use_thread)
2203 if (output_format & FIO_OUTPUT_NORMAL) {
2204 log_info("Starting ");
2206 log_info("%d thread%s", nr_thread,
2207 nr_thread > 1 ? "s" : "");
2211 log_info("%d process%s", nr_process,
2212 nr_process > 1 ? "es" : "");
2218 todo = thread_number;
2221 m_rate = t_rate = 0;
2223 for_each_td(td, i) {
2224 print_status_init(td->thread_number - 1);
2226 if (!td->o.create_serialize)
2229 if (fio_verify_load_state(td))
2233 * do file setup here so it happens sequentially,
2234 * we don't want X number of threads getting their
2235 * client data interspersed on disk
2237 if (setup_files(td)) {
2241 log_err("fio: pid=%d, err=%d/%s\n",
2242 (int) td->pid, td->error, td->verror);
2243 td_set_runstate(td, TD_REAPED);
2250 * for sharing to work, each job must always open
2251 * its own files. so close them, if we opened them
2254 for_each_file(td, f, j) {
2255 if (fio_file_open(f))
2256 td_io_close_file(td, f);
2261 /* start idle threads before io threads start to run */
2262 fio_idle_prof_start();
2267 struct thread_data *map[REAL_MAX_JOBS];
2268 struct timespec this_start;
2269 int this_jobs = 0, left;
2270 struct fork_data *fd;
2273 * create threads (TD_NOT_CREATED -> TD_CREATED)
2275 for_each_td(td, i) {
2276 if (td->runstate != TD_NOT_CREATED)
2280 * never got a chance to start, killed by other
2281 * thread for some reason
2283 if (td->terminate) {
2288 if (td->o.start_delay) {
2289 spent = utime_since_genesis();
2291 if (td->o.start_delay > spent)
2295 if (td->o.stonewall && (nr_started || nr_running)) {
2296 dprint(FD_PROCESS, "%s: stonewall wait\n",
2301 if (waitee_running(td)) {
2302 dprint(FD_PROCESS, "%s: waiting for %s\n",
2303 td->o.name, td->o.wait_for);
2309 td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2310 td->update_rusage = 0;
2313 * Set state to created. Thread will transition
2314 * to TD_INITIALIZED when it's done setting up.
2316 td_set_runstate(td, TD_CREATED);
2317 map[this_jobs++] = td;
2320 fd = calloc(1, sizeof(*fd));
2322 fd->sk_out = sk_out;
2324 if (td->o.use_thread) {
2327 dprint(FD_PROCESS, "will pthread_create\n");
2328 ret = pthread_create(&td->thread, NULL,
2331 log_err("pthread_create: %s\n",
2338 ret = pthread_detach(td->thread);
2340 log_err("pthread_detach: %s",
2344 dprint(FD_PROCESS, "will fork\n");
2349 ret = (int)(uintptr_t)thread_main(fd);
2351 } else if (i == fio_debug_jobno)
2352 *fio_debug_jobp = pid;
2354 dprint(FD_MUTEX, "wait on startup_sem\n");
2355 if (fio_sem_down_timeout(startup_sem, 10000)) {
2356 log_err("fio: job startup hung? exiting.\n");
2357 fio_terminate_threads(TERMINATE_ALL);
2363 dprint(FD_MUTEX, "done waiting on startup_sem\n");
2367 * Wait for the started threads to transition to
2370 fio_gettime(&this_start, NULL);
2372 while (left && !fio_abort) {
2373 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2378 for (i = 0; i < this_jobs; i++) {
2382 if (td->runstate == TD_INITIALIZED) {
2385 } else if (td->runstate >= TD_EXITED) {
2389 nr_running++; /* work-around... */
2395 log_err("fio: %d job%s failed to start\n", left,
2396 left > 1 ? "s" : "");
2397 for (i = 0; i < this_jobs; i++) {
2401 kill(td->pid, SIGTERM);
2407 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2409 for_each_td(td, i) {
2410 if (td->runstate != TD_INITIALIZED)
2413 if (in_ramp_time(td))
2414 td_set_runstate(td, TD_RAMP);
2416 td_set_runstate(td, TD_RUNNING);
2419 m_rate += ddir_rw_sum(td->o.ratemin);
2420 t_rate += ddir_rw_sum(td->o.rate);
2422 fio_sem_up(td->sem);
2425 reap_threads(&nr_running, &t_rate, &m_rate);
2431 while (nr_running) {
2432 reap_threads(&nr_running, &t_rate, &m_rate);
2436 fio_idle_prof_stop();
2441 static void free_disk_util(void)
2443 disk_util_prune_entries();
2444 helper_thread_destroy();
2447 int fio_backend(struct sk_out *sk_out)
2449 struct thread_data *td;
2453 if (load_profile(exec_profile))
2456 exec_profile = NULL;
2462 struct log_params p = {
2463 .log_type = IO_LOG_TYPE_BW,
2466 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2467 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2468 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2471 startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2472 if (startup_sem == NULL)
2477 helper_thread_create(startup_sem, sk_out);
2479 cgroup_list = smalloc(sizeof(*cgroup_list));
2481 INIT_FLIST_HEAD(cgroup_list);
2483 run_threads(sk_out);
2485 helper_thread_exit();
2490 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2491 struct io_log *log = agg_io_log[i];
2493 flush_log(log, false);
2499 for_each_td(td, i) {
2500 steadystate_free(td);
2501 fio_options_free(td);
2502 if (td->rusage_sem) {
2503 fio_sem_remove(td->rusage_sem);
2504 td->rusage_sem = NULL;
2506 fio_sem_remove(td->sem);
2512 cgroup_kill(cgroup_list);
2517 fio_sem_remove(startup_sem);