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 nr_segments = 0;
66 unsigned int cur_segment = 0;
67 unsigned int stat_number = 0;
69 unsigned long done_secs = 0;
70 #ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP
71 pthread_mutex_t overlap_check = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
73 pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
76 #define JOB_START_TIMEOUT (5 * 1000)
78 static void sig_int(int sig)
82 fio_server_got_signal(sig);
84 log_info("\nfio: terminating on signal %d\n", sig);
89 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
94 static void sig_break(int sig)
99 * Windows terminates all job processes on SIGBREAK after the handler
100 * returns, so give them time to wrap-up and give stats
103 while (td->runstate < TD_EXITED)
109 void sig_show_status(int sig)
111 show_running_run_stats();
114 static void set_sig_handlers(void)
116 struct sigaction act;
118 memset(&act, 0, sizeof(act));
119 act.sa_handler = sig_int;
120 act.sa_flags = SA_RESTART;
121 sigaction(SIGINT, &act, NULL);
123 memset(&act, 0, sizeof(act));
124 act.sa_handler = sig_int;
125 act.sa_flags = SA_RESTART;
126 sigaction(SIGTERM, &act, NULL);
128 /* Windows uses SIGBREAK as a quit signal from other applications */
130 memset(&act, 0, sizeof(act));
131 act.sa_handler = sig_break;
132 act.sa_flags = SA_RESTART;
133 sigaction(SIGBREAK, &act, NULL);
136 memset(&act, 0, sizeof(act));
137 act.sa_handler = sig_show_status;
138 act.sa_flags = SA_RESTART;
139 sigaction(SIGUSR1, &act, NULL);
142 memset(&act, 0, sizeof(act));
143 act.sa_handler = sig_int;
144 act.sa_flags = SA_RESTART;
145 sigaction(SIGPIPE, &act, NULL);
150 * Check if we are above the minimum rate given.
152 static bool __check_min_rate(struct thread_data *td, struct timespec *now,
155 unsigned long long current_rate_check_bytes = td->this_io_bytes[ddir];
156 unsigned long current_rate_check_blocks = td->this_io_blocks[ddir];
157 unsigned long long option_rate_bytes_min = td->o.ratemin[ddir];
158 unsigned int option_rate_iops_min = td->o.rate_iops_min[ddir];
160 assert(ddir_rw(ddir));
162 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
166 * allow a 2 second settle period in the beginning
168 if (mtime_since(&td->start, now) < 2000)
172 * if last_rate_check_blocks or last_rate_check_bytes is set,
173 * we can compute a rate per ratecycle
175 if (td->last_rate_check_bytes[ddir] || td->last_rate_check_blocks[ddir]) {
176 unsigned long spent = mtime_since(&td->last_rate_check_time[ddir], now);
177 if (spent < td->o.ratecycle || spent==0)
180 if (td->o.ratemin[ddir]) {
182 * check bandwidth specified rate
184 unsigned long long current_rate_bytes =
185 ((current_rate_check_bytes - td->last_rate_check_bytes[ddir]) * 1000) / spent;
186 if (current_rate_bytes < option_rate_bytes_min) {
187 log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
188 td->o.name, option_rate_bytes_min, current_rate_bytes);
193 * checks iops specified rate
195 unsigned long long current_rate_iops =
196 ((current_rate_check_blocks - td->last_rate_check_blocks[ddir]) * 1000) / spent;
198 if (current_rate_iops < option_rate_iops_min) {
199 log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
200 td->o.name, option_rate_iops_min, current_rate_iops);
206 td->last_rate_check_bytes[ddir] = current_rate_check_bytes;
207 td->last_rate_check_blocks[ddir] = current_rate_check_blocks;
208 memcpy(&td->last_rate_check_time[ddir], now, sizeof(*now));
212 static bool check_min_rate(struct thread_data *td, struct timespec *now)
216 for_each_rw_ddir(ddir) {
217 if (td->bytes_done[ddir])
218 ret |= __check_min_rate(td, now, ddir);
225 * When job exits, we can cancel the in-flight IO if we are using async
226 * io. Attempt to do so.
228 static void cleanup_pending_aio(struct thread_data *td)
233 * get immediately available events, if any
235 r = io_u_queued_complete(td, 0);
238 * now cancel remaining active events
240 if (td->io_ops->cancel) {
244 io_u_qiter(&td->io_u_all, io_u, i) {
245 if (io_u->flags & IO_U_F_FLIGHT) {
246 r = td->io_ops->cancel(td, io_u);
254 r = io_u_queued_complete(td, td->cur_depth);
258 * Helper to handle the final sync of a file. Works just like the normal
259 * io path, just does everything sync.
261 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
263 struct io_u *io_u = __get_io_u(td);
264 enum fio_q_status ret;
269 io_u->ddir = DDIR_SYNC;
271 io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
273 if (td_io_prep(td, io_u)) {
279 ret = td_io_queue(td, io_u);
283 if (io_u_queued_complete(td, 1) < 0)
286 case FIO_Q_COMPLETED:
288 td_verror(td, io_u->error, "td_io_queue");
292 if (io_u_sync_complete(td, io_u) < 0)
303 static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
307 if (fio_file_open(f))
308 return fio_io_sync(td, f);
310 if (td_io_open_file(td, f))
313 ret = fio_io_sync(td, f);
315 if (fio_file_open(f))
316 ret2 = td_io_close_file(td, f);
317 return (ret || ret2);
320 static inline void __update_ts_cache(struct thread_data *td)
322 fio_gettime(&td->ts_cache, NULL);
325 static inline void update_ts_cache(struct thread_data *td)
327 if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
328 __update_ts_cache(td);
331 static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
333 if (in_ramp_time(td))
337 if (utime_since(&td->epoch, t) >= td->o.timeout)
344 * We need to update the runtime consistently in ms, but keep a running
345 * tally of the current elapsed time in microseconds for sub millisecond
348 static inline void update_runtime(struct thread_data *td,
349 unsigned long long *elapsed_us,
350 const enum fio_ddir ddir)
352 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
355 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
356 elapsed_us[ddir] += utime_since_now(&td->start);
357 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
360 static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
365 if (ret < 0 || td->error) {
367 enum error_type_bit eb;
372 eb = td_error_type(ddir, err);
373 if (!(td->o.continue_on_error & (1 << eb)))
376 if (td_non_fatal_error(td, eb, err)) {
378 * Continue with the I/Os in case of
381 update_error_count(td, err);
385 } else if (td->o.fill_device && (err == ENOSPC || err == EDQUOT)) {
387 * We expect to hit this error if
388 * fill_device option is set.
391 fio_mark_td_terminate(td);
395 * Stop the I/O in case of a fatal
398 update_error_count(td, err);
406 static void check_update_rusage(struct thread_data *td)
408 if (td->update_rusage) {
409 td->update_rusage = 0;
410 update_rusage_stat(td);
411 fio_sem_up(td->rusage_sem);
415 static int wait_for_completions(struct thread_data *td, struct timespec *time)
417 const int full = queue_full(td);
421 if (td->flags & TD_F_REGROW_LOGS)
422 return io_u_quiesce(td);
425 * if the queue is full, we MUST reap at least 1 event
427 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
428 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
431 if (time && should_check_rate(td))
432 fio_gettime(time, NULL);
435 ret = io_u_queued_complete(td, min_evts);
438 } while (full && (td->cur_depth > td->o.iodepth_low));
443 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
444 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
445 struct timespec *comp_time)
448 case FIO_Q_COMPLETED:
451 clear_io_u(td, io_u);
452 } else if (io_u->resid) {
453 long long bytes = io_u->xfer_buflen - io_u->resid;
454 struct fio_file *f = io_u->file;
457 *bytes_issued += bytes;
467 unlog_io_piece(td, io_u);
468 td_verror(td, EIO, "full resid");
469 clear_io_u(td, io_u);
473 io_u->xfer_buflen = io_u->resid;
474 io_u->xfer_buf += bytes;
475 io_u->offset += bytes;
477 if (ddir_rw(io_u->ddir))
478 td->ts.short_io_u[io_u->ddir]++;
480 if (io_u->offset == f->real_file_size)
483 requeue_io_u(td, &io_u);
486 if (comp_time && should_check_rate(td))
487 fio_gettime(comp_time, NULL);
489 *ret = io_u_sync_complete(td, io_u);
494 if (td->flags & TD_F_REGROW_LOGS)
498 * when doing I/O (not when verifying),
499 * check for any errors that are to be ignored
507 * if the engine doesn't have a commit hook,
508 * the io_u is really queued. if it does have such
509 * a hook, it has to call io_u_queued() itself.
511 if (td->io_ops->commit == NULL)
512 io_u_queued(td, io_u);
514 *bytes_issued += io_u->xfer_buflen;
518 unlog_io_piece(td, io_u);
519 requeue_io_u(td, &io_u);
524 td_verror(td, -(*ret), "td_io_queue");
528 if (break_on_this_error(td, ddir, ret))
534 static inline bool io_in_polling(struct thread_data *td)
536 return !td->o.iodepth_batch_complete_min &&
537 !td->o.iodepth_batch_complete_max;
540 * Unlinks files from thread data fio_file structure
542 static int unlink_all_files(struct thread_data *td)
548 for_each_file(td, f, i) {
549 if (f->filetype != FIO_TYPE_FILE)
551 ret = td_io_unlink_file(td, f);
557 td_verror(td, ret, "unlink_all_files");
563 * Check if io_u will overlap an in-flight IO in the queue
565 bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
568 struct io_u *check_io_u;
569 unsigned long long x1, x2, y1, y2;
573 x2 = io_u->offset + io_u->buflen;
575 io_u_qiter(q, check_io_u, i) {
576 if (check_io_u->flags & IO_U_F_FLIGHT) {
577 y1 = check_io_u->offset;
578 y2 = check_io_u->offset + check_io_u->buflen;
580 if (x1 < y2 && y1 < x2) {
582 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
584 y1, check_io_u->buflen);
593 static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
596 * Check for overlap if the user asked us to, and we have
597 * at least one IO in flight besides this one.
599 if (td->o.serialize_overlap && td->cur_depth > 1 &&
600 in_flight_overlap(&td->io_u_all, io_u))
603 return td_io_queue(td, io_u);
607 * The main verify engine. Runs over the writes we previously submitted,
608 * reads the blocks back in, and checks the crc/md5 of the data.
610 static void do_verify(struct thread_data *td, uint64_t verify_bytes)
617 dprint(FD_VERIFY, "starting loop\n");
620 * sync io first and invalidate cache, to make sure we really
623 for_each_file(td, f, i) {
624 if (!fio_file_open(f))
626 if (fio_io_sync(td, f))
628 if (file_invalidate_cache(td, f))
632 check_update_rusage(td);
637 td_set_runstate(td, TD_VERIFYING);
640 while (!td->terminate) {
645 check_update_rusage(td);
647 if (runtime_exceeded(td, &td->ts_cache)) {
648 __update_ts_cache(td);
649 if (runtime_exceeded(td, &td->ts_cache)) {
650 fio_mark_td_terminate(td);
655 if (flow_threshold_exceeded(td))
658 if (!td->o.experimental_verify) {
659 io_u = __get_io_u(td);
663 if (get_next_verify(td, io_u)) {
668 if (td_io_prep(td, io_u)) {
673 if (td->bytes_verified + td->o.rw_min_bs > verify_bytes)
676 while ((io_u = get_io_u(td)) != NULL) {
677 if (IS_ERR_OR_NULL(io_u)) {
684 * We are only interested in the places where
685 * we wrote or trimmed IOs. Turn those into
686 * reads for verification purposes.
688 if (io_u->ddir == DDIR_READ) {
690 * Pretend we issued it for rwmix
693 td->io_issues[DDIR_READ]++;
696 } else if (io_u->ddir == DDIR_TRIM) {
697 io_u->ddir = DDIR_READ;
698 io_u_set(td, io_u, IO_U_F_TRIMMED);
700 } else if (io_u->ddir == DDIR_WRITE) {
701 io_u->ddir = DDIR_READ;
702 io_u->numberio = td->verify_read_issues;
703 td->verify_read_issues++;
704 populate_verify_io_u(td, io_u);
716 if (verify_state_should_stop(td, io_u)) {
721 if (td->o.verify_async)
722 io_u->end_io = verify_io_u_async;
724 io_u->end_io = verify_io_u;
727 if (!td->o.disable_slat)
728 fio_gettime(&io_u->start_time, NULL);
730 ret = io_u_submit(td, io_u);
732 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
736 * if we can queue more, do so. but check if there are
737 * completed io_u's first. Note that we can get BUSY even
738 * without IO queued, if the system is resource starved.
741 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
742 if (full || io_in_polling(td))
743 ret = wait_for_completions(td, NULL);
749 check_update_rusage(td);
752 min_events = td->cur_depth;
755 ret = io_u_queued_complete(td, min_events);
757 cleanup_pending_aio(td);
759 td_set_runstate(td, TD_RUNNING);
761 dprint(FD_VERIFY, "exiting loop\n");
764 static bool exceeds_number_ios(struct thread_data *td)
766 unsigned long long number_ios;
768 if (!td->o.number_ios)
771 number_ios = ddir_rw_sum(td->io_blocks);
772 number_ios += td->io_u_queued + td->io_u_in_flight;
774 return number_ios >= (td->o.number_ios * td->loops);
777 static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
779 unsigned long long bytes, limit;
782 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
783 else if (td_write(td))
784 bytes = this_bytes[DDIR_WRITE];
785 else if (td_read(td))
786 bytes = this_bytes[DDIR_READ];
788 bytes = this_bytes[DDIR_TRIM];
791 limit = td->o.io_size;
796 return bytes >= limit || exceeds_number_ios(td);
799 static bool io_issue_bytes_exceeded(struct thread_data *td)
801 return io_bytes_exceeded(td, td->io_issue_bytes);
804 static bool io_complete_bytes_exceeded(struct thread_data *td)
806 return io_bytes_exceeded(td, td->this_io_bytes);
810 * used to calculate the next io time for rate control
813 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
815 uint64_t bps = td->rate_bps[ddir];
817 assert(!(td->flags & TD_F_CHILD));
819 if (td->o.rate_process == RATE_PROCESS_POISSON) {
822 iops = bps / td->o.min_bs[ddir];
823 val = (int64_t) (1000000 / iops) *
824 -logf(__rand_0_1(&td->poisson_state[ddir]));
826 dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
827 (unsigned long long) 1000000 / val,
830 td->last_usec[ddir] += val;
831 return td->last_usec[ddir];
833 uint64_t bytes = td->rate_io_issue_bytes[ddir];
834 uint64_t secs = bytes / bps;
835 uint64_t remainder = bytes % bps;
837 return remainder * 1000000 / bps + secs * 1000000;
843 static void init_thinktime(struct thread_data *td)
845 if (td->o.thinktime_blocks_type == THINKTIME_BLOCKS_TYPE_COMPLETE)
846 td->thinktime_blocks_counter = td->io_blocks;
848 td->thinktime_blocks_counter = td->io_issues;
849 td->last_thinktime = td->epoch;
850 td->last_thinktime_blocks = 0;
853 static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir,
854 struct timespec *time)
856 unsigned long long b;
857 unsigned long long runtime_left;
863 if (td->o.thinktime_iotime) {
864 fio_gettime(&now, NULL);
865 if (utime_since(&td->last_thinktime, &now)
866 >= td->o.thinktime_iotime) {
868 } else if (!fio_option_is_set(&td->o, thinktime_blocks)) {
870 * When thinktime_iotime is set and thinktime_blocks is
871 * not set, skip the thinktime_blocks check, since
872 * thinktime_blocks default value 1 does not work
873 * together with thinktime_iotime.
880 b = ddir_rw_sum(td->thinktime_blocks_counter);
881 if (b >= td->last_thinktime_blocks + td->o.thinktime_blocks)
889 left = td->o.thinktime_spin;
891 runtime_left = td->o.timeout - utime_since_now(&td->epoch);
892 if (runtime_left < (unsigned long long)left)
898 total = usec_spin(left);
901 * usec_spin() might run for slightly longer than intended in a VM
902 * where the vCPU could get descheduled or the hypervisor could steal
903 * CPU time. Ensure "left" doesn't become negative.
905 if (total < td->o.thinktime)
906 left = td->o.thinktime - total;
911 runtime_left = td->o.timeout - utime_since_now(&td->epoch);
912 if (runtime_left < (unsigned long long)left)
917 total += usec_sleep(td, left);
920 * If we're ignoring thinktime for the rate, add the number of bytes
921 * we would have done while sleeping, minus one block to ensure we
922 * start issuing immediately after the sleep.
924 if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
925 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
926 uint64_t bs = td->o.min_bs[ddir];
927 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
930 if (usperop <= total)
933 over = (usperop - total) / usperop * -bs;
935 td->rate_io_issue_bytes[ddir] += (missed - over);
936 /* adjust for rate_process=poisson */
937 td->last_usec[ddir] += total;
940 if (time && should_check_rate(td))
941 fio_gettime(time, NULL);
943 td->last_thinktime_blocks = b;
944 if (td->o.thinktime_iotime) {
945 fio_gettime(&now, NULL);
946 td->last_thinktime = now;
951 * Main IO worker function. It retrieves io_u's to process and queues
952 * and reaps them, checking for rate and errors along the way.
954 * Returns number of bytes written and trimmed.
956 static void do_io(struct thread_data *td, uint64_t *bytes_done)
960 uint64_t total_bytes, bytes_issued = 0;
962 for (i = 0; i < DDIR_RWDIR_CNT; i++)
963 bytes_done[i] = td->bytes_done[i];
965 if (in_ramp_time(td))
966 td_set_runstate(td, TD_RAMP);
968 td_set_runstate(td, TD_RUNNING);
972 total_bytes = td->o.size;
974 * Allow random overwrite workloads to write up to io_size
975 * before starting verification phase as 'size' doesn't apply.
977 if (td_write(td) && td_random(td) && td->o.norandommap)
978 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
980 * If verify_backlog is enabled, we'll run the verify in this
981 * handler as well. For that case, we may need up to twice the
984 if (td->o.verify != VERIFY_NONE &&
985 (td_write(td) && td->o.verify_backlog))
986 total_bytes += td->o.size;
988 /* In trimwrite mode, each byte is trimmed and then written, so
989 * allow total_bytes or number of ios to be twice as big */
990 if (td_trimwrite(td)) {
991 total_bytes += td->total_io_size;
992 td->o.number_ios *= 2;
995 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
996 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
998 struct timespec comp_time;
1003 check_update_rusage(td);
1005 if (td->terminate || td->done)
1008 update_ts_cache(td);
1010 if (runtime_exceeded(td, &td->ts_cache)) {
1011 __update_ts_cache(td);
1012 if (runtime_exceeded(td, &td->ts_cache)) {
1013 fio_mark_td_terminate(td);
1018 if (flow_threshold_exceeded(td))
1022 * Break if we exceeded the bytes. The exception is time
1023 * based runs, but we still need to break out of the loop
1024 * for those to run verification, if enabled.
1025 * Jobs read from iolog do not use this stop condition.
1027 if (bytes_issued >= total_bytes &&
1028 !td->o.read_iolog_file &&
1029 (!td->o.time_based ||
1030 (td->o.time_based && td->o.verify != VERIFY_NONE)))
1033 io_u = get_io_u(td);
1034 if (IS_ERR_OR_NULL(io_u)) {
1035 int err = PTR_ERR(io_u);
1039 if (err == -EBUSY) {
1043 if (td->o.latency_target)
1048 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY) {
1049 if (!(io_u->flags & IO_U_F_PATTERN_DONE)) {
1050 io_u_set(td, io_u, IO_U_F_PATTERN_DONE);
1051 io_u->numberio = td->io_issues[io_u->ddir];
1052 populate_verify_io_u(td, io_u);
1059 * Add verification end_io handler if:
1060 * - Asked to verify (!td_rw(td))
1061 * - Or the io_u is from our verify list (mixed write/ver)
1063 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1064 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1066 if (verify_state_should_stop(td, io_u)) {
1071 if (td->o.verify_async)
1072 io_u->end_io = verify_io_u_async;
1074 io_u->end_io = verify_io_u;
1075 td_set_runstate(td, TD_VERIFYING);
1076 } else if (in_ramp_time(td))
1077 td_set_runstate(td, TD_RAMP);
1079 td_set_runstate(td, TD_RUNNING);
1082 * Always log IO before it's issued, so we know the specific
1083 * order of it. The logged unit will track when the IO has
1086 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1088 td->o.verify != VERIFY_NONE &&
1089 !td->o.experimental_verify)
1090 log_io_piece(td, io_u);
1092 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1093 const unsigned long long blen = io_u->xfer_buflen;
1094 const enum fio_ddir __ddir = acct_ddir(io_u);
1099 workqueue_enqueue(&td->io_wq, &io_u->work);
1102 if (ddir_rw(__ddir)) {
1103 td->io_issues[__ddir]++;
1104 td->io_issue_bytes[__ddir] += blen;
1105 td->rate_io_issue_bytes[__ddir] += blen;
1108 if (should_check_rate(td)) {
1109 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
1110 fio_gettime(&comp_time, NULL);
1114 ret = io_u_submit(td, io_u);
1116 if (should_check_rate(td))
1117 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1119 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1123 * See if we need to complete some commands. Note that
1124 * we can get BUSY even without IO queued, if the
1125 * system is resource starved.
1128 full = queue_full(td) ||
1129 (ret == FIO_Q_BUSY && td->cur_depth);
1130 if (full || io_in_polling(td))
1131 ret = wait_for_completions(td, &comp_time);
1136 if (ddir_rw(ddir) && td->o.thinktime)
1137 handle_thinktime(td, ddir, &comp_time);
1139 if (!ddir_rw_sum(td->bytes_done) &&
1140 !td_ioengine_flagged(td, FIO_NOIO))
1143 if (!in_ramp_time(td) && should_check_rate(td)) {
1144 if (check_min_rate(td, &comp_time)) {
1145 if (exitall_on_terminate || td->o.exitall_error)
1146 fio_terminate_threads(td->groupid, td->o.exit_what);
1147 td_verror(td, EIO, "check_min_rate");
1151 if (!in_ramp_time(td) && td->o.latency_target)
1152 lat_target_check(td);
1155 check_update_rusage(td);
1157 if (td->trim_entries)
1158 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1160 if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
1162 fio_mark_td_terminate(td);
1167 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1168 workqueue_flush(&td->io_wq);
1174 ret = io_u_queued_complete(td, i);
1175 if (td->o.fill_device &&
1176 (td->error == ENOSPC || td->error == EDQUOT))
1180 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
1181 td_set_runstate(td, TD_FSYNCING);
1183 for_each_file(td, f, i) {
1184 if (!fio_file_fsync(td, f))
1187 log_err("fio: end_fsync failed for file %s\n",
1192 if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
1193 workqueue_flush(&td->io_wq);
1194 cleanup_pending_aio(td);
1198 * stop job if we failed doing any IO
1200 if (!ddir_rw_sum(td->this_io_bytes))
1203 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1204 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1207 static void free_file_completion_logging(struct thread_data *td)
1212 for_each_file(td, f, i) {
1213 if (!f->last_write_comp)
1215 sfree(f->last_write_comp);
1219 static int init_file_completion_logging(struct thread_data *td,
1225 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1228 for_each_file(td, f, i) {
1229 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1230 if (!f->last_write_comp)
1237 free_file_completion_logging(td);
1238 log_err("fio: failed to alloc write comp data\n");
1242 static void cleanup_io_u(struct thread_data *td)
1246 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1248 if (td->io_ops->io_u_free)
1249 td->io_ops->io_u_free(td, io_u);
1251 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
1256 io_u_rexit(&td->io_u_requeues);
1257 io_u_qexit(&td->io_u_freelist, false);
1258 io_u_qexit(&td->io_u_all, td_offload_overlap(td));
1260 free_file_completion_logging(td);
1263 static int init_io_u(struct thread_data *td)
1266 int cl_align, i, max_units;
1269 max_units = td->o.iodepth;
1272 err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1273 err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1274 err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
1277 log_err("fio: failed setting up IO queues\n");
1281 cl_align = os_cache_line_size();
1283 for (i = 0; i < max_units; i++) {
1289 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
1291 log_err("fio: unable to allocate aligned memory\n");
1296 memset(io_u, 0, sizeof(*io_u));
1297 INIT_FLIST_HEAD(&io_u->verify_list);
1298 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1301 io_u->flags = IO_U_F_FREE;
1302 io_u_qpush(&td->io_u_freelist, io_u);
1305 * io_u never leaves this stack, used for iteration of all
1308 io_u_qpush(&td->io_u_all, io_u);
1310 if (td->io_ops->io_u_init) {
1311 int ret = td->io_ops->io_u_init(td, io_u);
1314 log_err("fio: failed to init engine data: %d\n", ret);
1320 if (init_io_u_buffers(td))
1323 if (init_file_completion_logging(td, max_units))
1329 int init_io_u_buffers(struct thread_data *td)
1332 unsigned long long max_bs, min_write;
1337 max_units = td->o.iodepth;
1338 max_bs = td_max_bs(td);
1339 min_write = td->o.min_bs[DDIR_WRITE];
1340 td->orig_buffer_size = (unsigned long long) max_bs
1341 * (unsigned long long) max_units;
1343 if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1347 * if we may later need to do address alignment, then add any
1348 * possible adjustment here so that we don't cause a buffer
1349 * overflow later. this adjustment may be too much if we get
1350 * lucky and the allocator gives us an aligned address.
1352 if (td->o.odirect || td->o.mem_align ||
1353 td_ioengine_flagged(td, FIO_RAWIO))
1354 td->orig_buffer_size += page_mask + td->o.mem_align;
1356 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1357 unsigned long long bs;
1359 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1360 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1363 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1364 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1368 if (data_xfer && allocate_io_mem(td))
1371 if (td->o.odirect || td->o.mem_align ||
1372 td_ioengine_flagged(td, FIO_RAWIO))
1373 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1375 p = td->orig_buffer;
1377 for (i = 0; i < max_units; i++) {
1378 io_u = td->io_u_all.io_us[i];
1379 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1383 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1386 io_u_fill_buffer(td, io_u, min_write, max_bs);
1387 if (td_write(td) && td->o.verify_pattern_bytes) {
1389 * Fill the buffer with the pattern if we are
1390 * going to be doing writes.
1392 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1401 #ifdef FIO_HAVE_IOSCHED_SWITCH
1403 * These functions are Linux specific.
1404 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1406 static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
1408 char tmp[256], tmp2[128], *p;
1412 assert(file->du && file->du->sysfs_root);
1413 sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);
1415 f = fopen(tmp, "r+");
1417 if (errno == ENOENT) {
1418 log_err("fio: os or kernel doesn't support IO scheduler"
1422 td_verror(td, errno, "fopen iosched");
1429 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1430 if (ferror(f) || ret != 1) {
1431 td_verror(td, errno, "fwrite");
1439 * Read back and check that the selected scheduler is now the default.
1441 ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1442 if (ferror(f) || ret < 0) {
1443 td_verror(td, errno, "fread");
1449 * either a list of io schedulers or "none\n" is expected. Strip the
1456 * Write to "none" entry doesn't fail, so check the result here.
1458 if (!strcmp(tmp, "none")) {
1459 log_err("fio: io scheduler is not tunable\n");
1464 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1465 if (!strstr(tmp, tmp2)) {
1466 log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
1467 td_verror(td, EINVAL, "iosched_switch");
1476 static int switch_ioscheduler(struct thread_data *td)
1482 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1485 assert(td->files && td->files[0]);
1487 for_each_file(td, f, i) {
1489 /* Only consider regular files and block device files */
1490 switch (f->filetype) {
1492 case FIO_TYPE_BLOCK:
1494 * Make sure that the device hosting the file could
1506 ret = set_ioscheduler(td, f);
1516 static int switch_ioscheduler(struct thread_data *td)
1521 #endif /* FIO_HAVE_IOSCHED_SWITCH */
1523 static bool keep_running(struct thread_data *td)
1525 unsigned long long limit;
1531 if (td->o.time_based)
1537 if (exceeds_number_ios(td))
1541 limit = td->o.io_size;
1545 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1549 * If the difference is less than the maximum IO size, we
1552 diff = limit - ddir_rw_sum(td->io_bytes);
1553 if (diff < td_max_bs(td))
1556 if (fio_files_done(td) && !td->o.io_size)
1565 static int exec_string(struct thread_options *o, const char *string,
1571 if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1574 log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1578 log_err("fio: exec of cmd <%s> failed\n", str);
1585 * Dry run to compute correct state of numberio for verification.
1587 static uint64_t do_dry_run(struct thread_data *td)
1589 td_set_runstate(td, TD_RUNNING);
1591 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1592 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1596 if (td->terminate || td->done)
1599 io_u = get_io_u(td);
1600 if (IS_ERR_OR_NULL(io_u))
1603 io_u_set(td, io_u, IO_U_F_FLIGHT);
1606 if (ddir_rw(acct_ddir(io_u)))
1607 td->io_issues[acct_ddir(io_u)]++;
1608 if (ddir_rw(io_u->ddir)) {
1609 io_u_mark_depth(td, 1);
1610 td->ts.total_io_u[io_u->ddir]++;
1613 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1615 td->o.verify != VERIFY_NONE &&
1616 !td->o.experimental_verify)
1617 log_io_piece(td, io_u);
1619 ret = io_u_sync_complete(td, io_u);
1623 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1627 struct thread_data *td;
1628 struct sk_out *sk_out;
1632 * Entry point for the thread based jobs. The process based jobs end up
1633 * here as well, after a little setup.
1635 static void *thread_main(void *data)
1637 struct fork_data *fd = data;
1638 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1639 struct thread_data *td = fd->td;
1640 struct thread_options *o = &td->o;
1641 struct sk_out *sk_out = fd->sk_out;
1642 uint64_t bytes_done[DDIR_RWDIR_CNT];
1643 int deadlock_loop_cnt;
1647 sk_out_assign(sk_out);
1650 if (!o->use_thread) {
1656 fio_local_clock_init();
1658 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1661 fio_server_send_start(td);
1663 INIT_FLIST_HEAD(&td->io_log_list);
1664 INIT_FLIST_HEAD(&td->io_hist_list);
1665 INIT_FLIST_HEAD(&td->verify_list);
1666 INIT_FLIST_HEAD(&td->trim_list);
1667 td->io_hist_tree = RB_ROOT;
1669 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1671 td_verror(td, ret, "mutex_cond_init_pshared");
1674 ret = cond_init_pshared(&td->verify_cond);
1676 td_verror(td, ret, "mutex_cond_pshared");
1680 td_set_runstate(td, TD_INITIALIZED);
1681 dprint(FD_MUTEX, "up startup_sem\n");
1682 fio_sem_up(startup_sem);
1683 dprint(FD_MUTEX, "wait on td->sem\n");
1684 fio_sem_down(td->sem);
1685 dprint(FD_MUTEX, "done waiting on td->sem\n");
1688 * A new gid requires privilege, so we need to do this before setting
1691 if (o->gid != -1U && setgid(o->gid)) {
1692 td_verror(td, errno, "setgid");
1695 if (o->uid != -1U && setuid(o->uid)) {
1696 td_verror(td, errno, "setuid");
1700 td_zone_gen_index(td);
1703 * Do this early, we don't want the compress threads to be limited
1704 * to the same CPUs as the IO workers. So do this before we set
1705 * any potential CPU affinity
1707 if (iolog_compress_init(td, sk_out))
1711 * If we have a gettimeofday() thread, make sure we exclude that
1712 * thread from this job
1715 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1718 * Set affinity first, in case it has an impact on the memory
1721 if (fio_option_is_set(o, cpumask)) {
1722 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
1723 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
1725 log_err("fio: no CPUs set\n");
1726 log_err("fio: Try increasing number of available CPUs\n");
1727 td_verror(td, EINVAL, "cpus_split");
1731 ret = fio_setaffinity(td->pid, o->cpumask);
1733 td_verror(td, errno, "cpu_set_affinity");
1738 #ifdef CONFIG_LIBNUMA
1739 /* numa node setup */
1740 if (fio_option_is_set(o, numa_cpunodes) ||
1741 fio_option_is_set(o, numa_memnodes)) {
1742 struct bitmask *mask;
1744 if (numa_available() < 0) {
1745 td_verror(td, errno, "Does not support NUMA API\n");
1749 if (fio_option_is_set(o, numa_cpunodes)) {
1750 mask = numa_parse_nodestring(o->numa_cpunodes);
1751 ret = numa_run_on_node_mask(mask);
1752 numa_free_nodemask(mask);
1754 td_verror(td, errno, \
1755 "numa_run_on_node_mask failed\n");
1760 if (fio_option_is_set(o, numa_memnodes)) {
1762 if (o->numa_memnodes)
1763 mask = numa_parse_nodestring(o->numa_memnodes);
1765 switch (o->numa_mem_mode) {
1766 case MPOL_INTERLEAVE:
1767 numa_set_interleave_mask(mask);
1770 numa_set_membind(mask);
1773 numa_set_localalloc();
1775 case MPOL_PREFERRED:
1776 numa_set_preferred(o->numa_mem_prefer_node);
1784 numa_free_nodemask(mask);
1790 if (fio_pin_memory(td))
1794 * May alter parameters that init_io_u() will use, so we need to
1797 if (!init_iolog(td))
1800 /* ioprio_set() has to be done before td_io_init() */
1801 if (fio_option_is_set(o, ioprio) ||
1802 fio_option_is_set(o, ioprio_class) ||
1803 fio_option_is_set(o, ioprio_hint)) {
1804 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class,
1805 o->ioprio, o->ioprio_hint);
1807 td_verror(td, errno, "ioprio_set");
1810 td->ioprio = ioprio_value(o->ioprio_class, o->ioprio,
1812 td->ts.ioprio = td->ioprio;
1818 if (td_ioengine_flagged(td, FIO_SYNCIO) && td->o.iodepth > 1 && td->o.io_submit_mode != IO_MODE_OFFLOAD) {
1819 log_info("note: both iodepth >= 1 and synchronous I/O engine "
1820 "are selected, queue depth will be capped at 1\n");
1826 if (td->io_ops->post_init && td->io_ops->post_init(td))
1829 if (o->verify_async && verify_async_init(td))
1832 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1836 if (nice(o->nice) == -1 && errno != 0) {
1837 td_verror(td, errno, "nice");
1841 if (o->ioscheduler && switch_ioscheduler(td))
1844 if (!o->create_serialize && setup_files(td))
1847 if (!init_random_map(td))
1850 if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
1853 if (o->pre_read && !pre_read_files(td))
1856 fio_verify_init(td);
1858 if (rate_submit_init(td, sk_out))
1861 set_epoch_time(td, o->log_unix_epoch | o->log_alternate_epoch, o->log_alternate_epoch_clock_id);
1862 fio_getrusage(&td->ru_start);
1863 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1864 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1865 memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
1869 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1870 o->ratemin[DDIR_TRIM]) {
1871 memcpy(&td->last_rate_check_time[DDIR_READ], &td->bw_sample_time,
1872 sizeof(td->bw_sample_time));
1873 memcpy(&td->last_rate_check_time[DDIR_WRITE], &td->bw_sample_time,
1874 sizeof(td->bw_sample_time));
1875 memcpy(&td->last_rate_check_time[DDIR_TRIM], &td->bw_sample_time,
1876 sizeof(td->bw_sample_time));
1879 memset(bytes_done, 0, sizeof(bytes_done));
1880 clear_state = false;
1882 while (keep_running(td)) {
1883 uint64_t verify_bytes;
1885 fio_gettime(&td->start, NULL);
1886 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1889 clear_io_state(td, 0);
1891 if (o->unlink_each_loop && unlink_all_files(td))
1895 prune_io_piece_log(td);
1897 if (td->o.verify_only && td_write(td))
1898 verify_bytes = do_dry_run(td);
1900 if (!td->o.rand_repeatable)
1901 /* save verify rand state to replay hdr seeds later at verify */
1902 frand_copy(&td->verify_state_last_do_io, &td->verify_state);
1903 do_io(td, bytes_done);
1904 if (!td->o.rand_repeatable)
1905 frand_copy(&td->verify_state, &td->verify_state_last_do_io);
1906 if (!ddir_rw_sum(bytes_done)) {
1907 fio_mark_td_terminate(td);
1910 verify_bytes = bytes_done[DDIR_WRITE] +
1911 bytes_done[DDIR_TRIM];
1916 * If we took too long to shut down, the main thread could
1917 * already consider us reaped/exited. If that happens, break
1920 if (td->runstate >= TD_EXITED)
1926 * Make sure we've successfully updated the rusage stats
1927 * before waiting on the stat mutex. Otherwise we could have
1928 * the stat thread holding stat mutex and waiting for
1929 * the rusage_sem, which would never get upped because
1930 * this thread is waiting for the stat mutex.
1932 deadlock_loop_cnt = 0;
1934 check_update_rusage(td);
1935 if (!fio_sem_down_trylock(stat_sem))
1938 if (deadlock_loop_cnt++ > 5000) {
1939 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1940 td->error = EDEADLK;
1945 if (td->io_bytes[DDIR_READ] && (td_read(td) ||
1946 ((td->flags & TD_F_VER_BACKLOG) && td_write(td))))
1947 update_runtime(td, elapsed_us, DDIR_READ);
1948 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1949 update_runtime(td, elapsed_us, DDIR_WRITE);
1950 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1951 update_runtime(td, elapsed_us, DDIR_TRIM);
1952 fio_gettime(&td->start, NULL);
1953 fio_sem_up(stat_sem);
1955 if (td->error || td->terminate)
1958 if (!o->do_verify ||
1959 o->verify == VERIFY_NONE ||
1960 td_ioengine_flagged(td, FIO_UNIDIR))
1963 clear_io_state(td, 0);
1965 fio_gettime(&td->start, NULL);
1967 do_verify(td, verify_bytes);
1970 * See comment further up for why this is done here.
1972 check_update_rusage(td);
1974 fio_sem_down(stat_sem);
1975 update_runtime(td, elapsed_us, DDIR_READ);
1976 fio_gettime(&td->start, NULL);
1977 fio_sem_up(stat_sem);
1979 if (td->error || td->terminate)
1984 * Acquire this lock if we were doing overlap checking in
1985 * offload mode so that we don't clean up this job while
1986 * another thread is checking its io_u's for overlap
1988 if (td_offload_overlap(td)) {
1991 res = pthread_mutex_lock(&overlap_check);
1997 td_set_runstate(td, TD_FINISHING);
1998 if (td_offload_overlap(td)) {
2001 res = pthread_mutex_unlock(&overlap_check);
2008 update_rusage_stat(td);
2009 td->ts.total_run_time = mtime_since_now(&td->epoch);
2010 for_each_rw_ddir(ddir) {
2011 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
2014 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
2015 (td->o.verify != VERIFY_NONE && td_write(td)))
2016 verify_save_state(td->thread_number);
2018 fio_unpin_memory(td);
2020 td_writeout_logs(td, true);
2022 iolog_compress_exit(td);
2023 rate_submit_exit(td);
2025 if (o->exec_postrun)
2026 exec_string(o, o->exec_postrun, "postrun");
2028 if (exitall_on_terminate || (o->exitall_error && td->error))
2029 fio_terminate_threads(td->groupid, td->o.exit_what);
2033 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
2036 if (o->verify_async)
2037 verify_async_exit(td);
2039 close_and_free_files(td);
2042 cgroup_shutdown(td, cgroup_mnt);
2043 verify_free_state(td);
2044 td_zone_free_index(td);
2046 if (fio_option_is_set(o, cpumask)) {
2047 ret = fio_cpuset_exit(&o->cpumask);
2049 td_verror(td, ret, "fio_cpuset_exit");
2053 * do this very late, it will log file closing as well
2055 if (o->write_iolog_file)
2056 write_iolog_close(td);
2057 if (td->io_log_rfile)
2058 fclose(td->io_log_rfile);
2060 td_set_runstate(td, TD_EXITED);
2063 * Do this last after setting our runstate to exited, so we
2064 * know that the stat thread is signaled.
2066 check_update_rusage(td);
2069 return (void *) (uintptr_t) td->error;
2073 * Run over the job map and reap the threads that have exited, if any.
2075 static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
2078 unsigned int cputhreads, realthreads, pending;
2082 * reap exited threads (TD_EXITED -> TD_REAPED)
2084 realthreads = pending = cputhreads = 0;
2088 if (!strcmp(td->o.ioengine, "cpuio"))
2097 if (td->runstate == TD_REAPED)
2099 if (td->o.use_thread) {
2100 if (td->runstate == TD_EXITED) {
2101 td_set_runstate(td, TD_REAPED);
2108 if (td->runstate == TD_EXITED)
2112 * check if someone quit or got killed in an unusual way
2114 ret = waitpid(td->pid, &status, flags);
2116 if (errno == ECHILD) {
2117 log_err("fio: pid=%d disappeared %d\n",
2118 (int) td->pid, td->runstate);
2120 td_set_runstate(td, TD_REAPED);
2124 } else if (ret == td->pid) {
2125 if (WIFSIGNALED(status)) {
2126 int sig = WTERMSIG(status);
2128 if (sig != SIGTERM && sig != SIGUSR2)
2129 log_err("fio: pid=%d, got signal=%d\n",
2130 (int) td->pid, sig);
2132 td_set_runstate(td, TD_REAPED);
2135 if (WIFEXITED(status)) {
2136 if (WEXITSTATUS(status) && !td->error)
2137 td->error = WEXITSTATUS(status);
2139 td_set_runstate(td, TD_REAPED);
2145 * If the job is stuck, do a forceful timeout of it and
2148 if (td->terminate &&
2149 td->runstate < TD_FSYNCING &&
2150 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
2151 log_err("fio: job '%s' (state=%d) hasn't exited in "
2152 "%lu seconds, it appears to be stuck. Doing "
2153 "forceful exit of this job.\n",
2154 td->o.name, td->runstate,
2155 (unsigned long) time_since_now(&td->terminate_time));
2156 td_set_runstate(td, TD_REAPED);
2161 * thread is not dead, continue
2167 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2168 (*t_rate) -= ddir_rw_sum(td->o.rate);
2175 done_secs += mtime_since_now(&td->epoch) / 1000;
2176 profile_td_exit(td);
2180 if (*nr_running == cputhreads && !pending && realthreads)
2181 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2184 static bool __check_trigger_file(void)
2191 if (stat(trigger_file, &sb))
2194 if (unlink(trigger_file) < 0)
2195 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2201 static bool trigger_timedout(void)
2203 if (trigger_timeout)
2204 if (time_since_genesis() >= trigger_timeout) {
2205 trigger_timeout = 0;
2212 void exec_trigger(const char *cmd)
2216 if (!cmd || cmd[0] == '\0')
2221 log_err("fio: failed executing %s trigger\n", cmd);
2224 void check_trigger_file(void)
2226 if (__check_trigger_file() || trigger_timedout()) {
2228 fio_clients_send_trigger(trigger_remote_cmd);
2230 verify_save_state(IO_LIST_ALL);
2231 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2232 exec_trigger(trigger_cmd);
2237 static int fio_verify_load_state(struct thread_data *td)
2241 if (!td->o.verify_state)
2247 ret = fio_server_get_verify_state(td->o.name,
2248 td->thread_number - 1, &data);
2250 verify_assign_state(td, data);
2252 char prefix[PATH_MAX];
2255 sprintf(prefix, "%s%clocal", aux_path,
2256 FIO_OS_PATH_SEPARATOR);
2258 strcpy(prefix, "local");
2259 ret = verify_load_state(td, prefix);
2265 static void do_usleep(unsigned int usecs)
2267 check_for_running_stats();
2268 check_trigger_file();
2272 static bool check_mount_writes(struct thread_data *td)
2277 if (!td_write(td) || td->o.allow_mounted_write)
2281 * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2282 * are mkfs'd and mounted.
2284 for_each_file(td, f, i) {
2285 #ifdef FIO_HAVE_CHARDEV_SIZE
2286 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2288 if (f->filetype != FIO_TYPE_BLOCK)
2291 if (device_is_mounted(f->file_name))
2297 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2301 static bool waitee_running(struct thread_data *me)
2303 const char *waitee = me->o.wait_for;
2304 const char *self = me->o.name;
2310 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2313 if (td->runstate < TD_EXITED) {
2314 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2316 runstate_to_name(td->runstate));
2321 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2326 * Main function for kicking off and reaping jobs, as needed.
2328 static void run_threads(struct sk_out *sk_out)
2330 struct thread_data *td;
2331 unsigned int i, todo, nr_running, nr_started;
2332 uint64_t m_rate, t_rate;
2335 if (fio_gtod_offload && fio_start_gtod_thread())
2338 fio_idle_prof_init();
2342 nr_thread = nr_process = 0;
2344 if (check_mount_writes(td))
2346 if (td->o.use_thread)
2352 if (output_format & FIO_OUTPUT_NORMAL) {
2353 struct buf_output out;
2355 buf_output_init(&out);
2356 __log_buf(&out, "Starting ");
2358 __log_buf(&out, "%d thread%s", nr_thread,
2359 nr_thread > 1 ? "s" : "");
2362 __log_buf(&out, " and ");
2363 __log_buf(&out, "%d process%s", nr_process,
2364 nr_process > 1 ? "es" : "");
2366 __log_buf(&out, "\n");
2367 log_info_buf(out.buf, out.buflen);
2368 buf_output_free(&out);
2371 todo = thread_number;
2374 m_rate = t_rate = 0;
2377 print_status_init(td->thread_number - 1);
2379 if (!td->o.create_serialize)
2382 if (fio_verify_load_state(td))
2386 * do file setup here so it happens sequentially,
2387 * we don't want X number of threads getting their
2388 * client data interspersed on disk
2390 if (setup_files(td)) {
2394 log_err("fio: pid=%d, err=%d/%s\n",
2395 (int) td->pid, td->error, td->verror);
2396 td_set_runstate(td, TD_REAPED);
2403 * for sharing to work, each job must always open
2404 * its own files. so close them, if we opened them
2407 for_each_file(td, f, j) {
2408 if (fio_file_open(f))
2409 td_io_close_file(td, f);
2414 /* start idle threads before io threads start to run */
2415 fio_idle_prof_start();
2420 struct thread_data *map[REAL_MAX_JOBS];
2421 struct timespec this_start;
2422 int this_jobs = 0, left;
2423 struct fork_data *fd;
2426 * create threads (TD_NOT_CREATED -> TD_CREATED)
2429 if (td->runstate != TD_NOT_CREATED)
2433 * never got a chance to start, killed by other
2434 * thread for some reason
2436 if (td->terminate) {
2441 if (td->o.start_delay) {
2442 spent = utime_since_genesis();
2444 if (td->o.start_delay > spent)
2448 if (td->o.stonewall && (nr_started || nr_running)) {
2449 dprint(FD_PROCESS, "%s: stonewall wait\n",
2454 if (waitee_running(td)) {
2455 dprint(FD_PROCESS, "%s: waiting for %s\n",
2456 td->o.name, td->o.wait_for);
2462 td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2463 td->update_rusage = 0;
2466 * Set state to created. Thread will transition
2467 * to TD_INITIALIZED when it's done setting up.
2469 td_set_runstate(td, TD_CREATED);
2470 map[this_jobs++] = td;
2473 fd = calloc(1, sizeof(*fd));
2475 fd->sk_out = sk_out;
2477 if (td->o.use_thread) {
2480 dprint(FD_PROCESS, "will pthread_create\n");
2481 ret = pthread_create(&td->thread, NULL,
2484 log_err("pthread_create: %s\n",
2491 ret = pthread_detach(td->thread);
2493 log_err("pthread_detach: %s",
2498 dprint(FD_PROCESS, "will fork\n");
2505 ret = (int)(uintptr_t)thread_main(fd);
2507 } else if (__td_index == fio_debug_jobno)
2508 *fio_debug_jobp = pid;
2513 dprint(FD_MUTEX, "wait on startup_sem\n");
2514 if (fio_sem_down_timeout(startup_sem, 10000)) {
2515 log_err("fio: job startup hung? exiting.\n");
2516 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2522 dprint(FD_MUTEX, "done waiting on startup_sem\n");
2526 * Wait for the started threads to transition to
2529 fio_gettime(&this_start, NULL);
2531 while (left && !fio_abort) {
2532 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2537 for (i = 0; i < this_jobs; i++) {
2541 if (td->runstate == TD_INITIALIZED) {
2544 } else if (td->runstate >= TD_EXITED) {
2548 nr_running++; /* work-around... */
2554 log_err("fio: %d job%s failed to start\n", left,
2555 left > 1 ? "s" : "");
2556 for (i = 0; i < this_jobs; i++) {
2560 kill(td->pid, SIGTERM);
2566 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2569 if (td->runstate != TD_INITIALIZED)
2572 if (in_ramp_time(td))
2573 td_set_runstate(td, TD_RAMP);
2575 td_set_runstate(td, TD_RUNNING);
2578 m_rate += ddir_rw_sum(td->o.ratemin);
2579 t_rate += ddir_rw_sum(td->o.rate);
2581 fio_sem_up(td->sem);
2584 reap_threads(&nr_running, &t_rate, &m_rate);
2590 while (nr_running) {
2591 reap_threads(&nr_running, &t_rate, &m_rate);
2595 fio_idle_prof_stop();
2600 static void free_disk_util(void)
2602 disk_util_prune_entries();
2603 helper_thread_destroy();
2606 int fio_backend(struct sk_out *sk_out)
2610 if (load_profile(exec_profile))
2613 exec_profile = NULL;
2619 struct log_params p = {
2620 .log_type = IO_LOG_TYPE_BW,
2623 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2624 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2625 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2628 if (init_global_dedupe_working_set_seeds()) {
2629 log_err("fio: failed to initialize global dedupe working set\n");
2633 startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2635 is_local_backend = true;
2636 if (startup_sem == NULL)
2641 if (helper_thread_create(startup_sem, sk_out))
2642 log_err("fio: failed to create helper thread\n");
2644 cgroup_list = smalloc(sizeof(*cgroup_list));
2646 INIT_FLIST_HEAD(cgroup_list);
2648 run_threads(sk_out);
2650 helper_thread_exit();
2655 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2656 struct io_log *log = agg_io_log[i];
2658 flush_log(log, false);
2665 struct thread_stat *ts = &td->ts;
2667 free_clat_prio_stats(ts);
2668 steadystate_free(td);
2669 fio_options_free(td);
2670 fio_dump_options_free(td);
2671 if (td->rusage_sem) {
2672 fio_sem_remove(td->rusage_sem);
2673 td->rusage_sem = NULL;
2675 fio_sem_remove(td->sem);
2681 cgroup_kill(cgroup_list);
2685 fio_sem_remove(startup_sem);