projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
fallocate: use 'offset' parameter
[fio.git] / backend.c
... / ...
CommitLineData
1/*
2 * fio - the flexible io tester
3 *
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
6 *
7 * The license below covers all files distributed with fio unless otherwise
8 * noted in the file itself.
9 *
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.
13 *
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.
18 *
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.
22 *
23 */
24#include <unistd.h>
25#include <string.h>
26#include <signal.h>
27#include <assert.h>
28#include <inttypes.h>
29#include <sys/stat.h>
30#include <sys/wait.h>
31#include <math.h>
32
33#include "fio.h"
34#include "smalloc.h"
35#include "verify.h"
36#include "diskutil.h"
37#include "cgroup.h"
38#include "profile.h"
39#include "lib/rand.h"
40#include "lib/memalign.h"
41#include "server.h"
42#include "lib/getrusage.h"
43#include "idletime.h"
44#include "err.h"
45#include "workqueue.h"
46#include "lib/mountcheck.h"
47#include "rate-submit.h"
48#include "helper_thread.h"
49#include "pshared.h"
50
51static struct fio_sem *startup_sem;
52static struct flist_head *cgroup_list;
53static char *cgroup_mnt;
54static int exit_value;
55static volatile int fio_abort;
56static unsigned int nr_process = 0;
57static unsigned int nr_thread = 0;
58
59struct io_log *agg_io_log[DDIR_RWDIR_CNT];
60
61int groupid = 0;
62unsigned int thread_number = 0;
63unsigned int stat_number = 0;
64int shm_id = 0;
65int temp_stall_ts;
66unsigned long done_secs = 0;
67
68#define JOB_START_TIMEOUT (5 * 1000)
69
70static void sig_int(int sig)
71{
72 if (threads) {
73 if (is_backend)
74 fio_server_got_signal(sig);
75 else {
76 log_info("\nfio: terminating on signal %d\n", sig);
77 log_info_flush();
78 exit_value = 128;
79 }
80
81 fio_terminate_threads(TERMINATE_ALL);
82 }
83}
84
85void sig_show_status(int sig)
86{
87 show_running_run_stats();
88}
89
90static void set_sig_handlers(void)
91{
92 struct sigaction act;
93
94 memset(&act, 0, sizeof(act));
95 act.sa_handler = sig_int;
96 act.sa_flags = SA_RESTART;
97 sigaction(SIGINT, &act, NULL);
98
99 memset(&act, 0, sizeof(act));
100 act.sa_handler = sig_int;
101 act.sa_flags = SA_RESTART;
102 sigaction(SIGTERM, &act, NULL);
103
104/* Windows uses SIGBREAK as a quit signal from other applications */
105#ifdef WIN32
106 memset(&act, 0, sizeof(act));
107 act.sa_handler = sig_int;
108 act.sa_flags = SA_RESTART;
109 sigaction(SIGBREAK, &act, NULL);
110#endif
111
112 memset(&act, 0, sizeof(act));
113 act.sa_handler = sig_show_status;
114 act.sa_flags = SA_RESTART;
115 sigaction(SIGUSR1, &act, NULL);
116
117 if (is_backend) {
118 memset(&act, 0, sizeof(act));
119 act.sa_handler = sig_int;
120 act.sa_flags = SA_RESTART;
121 sigaction(SIGPIPE, &act, NULL);
122 }
123}
124
125/*
126 * Check if we are above the minimum rate given.
127 */
128static bool __check_min_rate(struct thread_data *td, struct timespec *now,
129 enum fio_ddir ddir)
130{
131 unsigned long long bytes = 0;
132 unsigned long iops = 0;
133 unsigned long spent;
134 unsigned long rate;
135 unsigned int ratemin = 0;
136 unsigned int rate_iops = 0;
137 unsigned int rate_iops_min = 0;
138
139 assert(ddir_rw(ddir));
140
141 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
142 return false;
143
144 /*
145 * allow a 2 second settle period in the beginning
146 */
147 if (mtime_since(&td->start, now) < 2000)
148 return false;
149
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];
155
156 /*
157 * if rate blocks is set, sample is running
158 */
159 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
160 spent = mtime_since(&td->lastrate[ddir], now);
161 if (spent < td->o.ratecycle)
162 return false;
163
164 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
165 /*
166 * check bandwidth specified rate
167 */
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);
171 return true;
172 } else {
173 if (spent)
174 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
175 else
176 rate = 0;
177
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);
182 return true;
183 }
184 }
185 } else {
186 /*
187 * checks iops specified rate
188 */
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);
192 return true;
193 } else {
194 if (spent)
195 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
196 else
197 rate = 0;
198
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);
203 return true;
204 }
205 }
206 }
207 }
208
209 td->rate_bytes[ddir] = bytes;
210 td->rate_blocks[ddir] = iops;
211 memcpy(&td->lastrate[ddir], now, sizeof(*now));
212 return false;
213}
214
215static bool check_min_rate(struct thread_data *td, struct timespec *now)
216{
217 bool ret = false;
218
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);
225
226 return ret;
227}
228
229/*
230 * When job exits, we can cancel the in-flight IO if we are using async
231 * io. Attempt to do so.
232 */
233static void cleanup_pending_aio(struct thread_data *td)
234{
235 int r;
236
237 /*
238 * get immediately available events, if any
239 */
240 r = io_u_queued_complete(td, 0);
241 if (r < 0)
242 return;
243
244 /*
245 * now cancel remaining active events
246 */
247 if (td->io_ops->cancel) {
248 struct io_u *io_u;
249 int i;
250
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);
254 if (!r)
255 put_io_u(td, io_u);
256 }
257 }
258 }
259
260 if (td->cur_depth)
261 r = io_u_queued_complete(td, td->cur_depth);
262}
263
264/*
265 * Helper to handle the final sync of a file. Works just like the normal
266 * io path, just does everything sync.
267 */
268static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
269{
270 struct io_u *io_u = __get_io_u(td);
271 enum fio_q_status ret;
272
273 if (!io_u)
274 return true;
275
276 io_u->ddir = DDIR_SYNC;
277 io_u->file = f;
278
279 if (td_io_prep(td, io_u)) {
280 put_io_u(td, io_u);
281 return true;
282 }
283
284requeue:
285 ret = td_io_queue(td, io_u);
286 switch (ret) {
287 case FIO_Q_QUEUED:
288 td_io_commit(td);
289 if (io_u_queued_complete(td, 1) < 0)
290 return true;
291 break;
292 case FIO_Q_COMPLETED:
293 if (io_u->error) {
294 td_verror(td, io_u->error, "td_io_queue");
295 return true;
296 }
297
298 if (io_u_sync_complete(td, io_u) < 0)
299 return true;
300 break;
301 case FIO_Q_BUSY:
302 td_io_commit(td);
303 goto requeue;
304 }
305
306 return false;
307}
308
309static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
310{
311 int ret;
312
313 if (fio_file_open(f))
314 return fio_io_sync(td, f);
315
316 if (td_io_open_file(td, f))
317 return 1;
318
319 ret = fio_io_sync(td, f);
320 td_io_close_file(td, f);
321 return ret;
322}
323
324static inline void __update_ts_cache(struct thread_data *td)
325{
326 fio_gettime(&td->ts_cache, NULL);
327}
328
329static inline void update_ts_cache(struct thread_data *td)
330{
331 if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
332 __update_ts_cache(td);
333}
334
335static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
336{
337 if (in_ramp_time(td))
338 return false;
339 if (!td->o.timeout)
340 return false;
341 if (utime_since(&td->epoch, t) >= td->o.timeout)
342 return true;
343
344 return false;
345}
346
347/*
348 * We need to update the runtime consistently in ms, but keep a running
349 * tally of the current elapsed time in microseconds for sub millisecond
350 * updates.
351 */
352static inline void update_runtime(struct thread_data *td,
353 unsigned long long *elapsed_us,
354 const enum fio_ddir ddir)
355{
356 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
357 return;
358
359 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
360 elapsed_us[ddir] += utime_since_now(&td->start);
361 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
362}
363
364static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
365 int *retptr)
366{
367 int ret = *retptr;
368
369 if (ret < 0 || td->error) {
370 int err = td->error;
371 enum error_type_bit eb;
372
373 if (ret < 0)
374 err = -ret;
375
376 eb = td_error_type(ddir, err);
377 if (!(td->o.continue_on_error & (1 << eb)))
378 return true;
379
380 if (td_non_fatal_error(td, eb, err)) {
381 /*
382 * Continue with the I/Os in case of
383 * a non fatal error.
384 */
385 update_error_count(td, err);
386 td_clear_error(td);
387 *retptr = 0;
388 return false;
389 } else if (td->o.fill_device && err == ENOSPC) {
390 /*
391 * We expect to hit this error if
392 * fill_device option is set.
393 */
394 td_clear_error(td);
395 fio_mark_td_terminate(td);
396 return true;
397 } else {
398 /*
399 * Stop the I/O in case of a fatal
400 * error.
401 */
402 update_error_count(td, err);
403 return true;
404 }
405 }
406
407 return false;
408}
409
410static void check_update_rusage(struct thread_data *td)
411{
412 if (td->update_rusage) {
413 td->update_rusage = 0;
414 update_rusage_stat(td);
415 fio_sem_up(td->rusage_sem);
416 }
417}
418
419static int wait_for_completions(struct thread_data *td, struct timespec *time)
420{
421 const int full = queue_full(td);
422 int min_evts = 0;
423 int ret;
424
425 if (td->flags & TD_F_REGROW_LOGS)
426 return io_u_quiesce(td);
427
428 /*
429 * if the queue is full, we MUST reap at least 1 event
430 */
431 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
432 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
433 min_evts = 1;
434
435 if (time && (__should_check_rate(td, DDIR_READ) ||
436 __should_check_rate(td, DDIR_WRITE) ||
437 __should_check_rate(td, DDIR_TRIM)))
438 fio_gettime(time, NULL);
439
440 do {
441 ret = io_u_queued_complete(td, min_evts);
442 if (ret < 0)
443 break;
444 } while (full && (td->cur_depth > td->o.iodepth_low));
445
446 return ret;
447}
448
449int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
450 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
451 struct timespec *comp_time)
452{
453 switch (*ret) {
454 case FIO_Q_COMPLETED:
455 if (io_u->error) {
456 *ret = -io_u->error;
457 clear_io_u(td, io_u);
458 } else if (io_u->resid) {
459 int bytes = io_u->xfer_buflen - io_u->resid;
460 struct fio_file *f = io_u->file;
461
462 if (bytes_issued)
463 *bytes_issued += bytes;
464
465 if (!from_verify)
466 trim_io_piece(td, io_u);
467
468 /*
469 * zero read, fail
470 */
471 if (!bytes) {
472 if (!from_verify)
473 unlog_io_piece(td, io_u);
474 td_verror(td, EIO, "full resid");
475 put_io_u(td, io_u);
476 break;
477 }
478
479 io_u->xfer_buflen = io_u->resid;
480 io_u->xfer_buf += bytes;
481 io_u->offset += bytes;
482
483 if (ddir_rw(io_u->ddir))
484 td->ts.short_io_u[io_u->ddir]++;
485
486 if (io_u->offset == f->real_file_size)
487 goto sync_done;
488
489 requeue_io_u(td, &io_u);
490 } else {
491sync_done:
492 if (comp_time && (__should_check_rate(td, DDIR_READ) ||
493 __should_check_rate(td, DDIR_WRITE) ||
494 __should_check_rate(td, DDIR_TRIM)))
495 fio_gettime(comp_time, NULL);
496
497 *ret = io_u_sync_complete(td, io_u);
498 if (*ret < 0)
499 break;
500 }
501
502 if (td->flags & TD_F_REGROW_LOGS)
503 regrow_logs(td);
504
505 /*
506 * when doing I/O (not when verifying),
507 * check for any errors that are to be ignored
508 */
509 if (!from_verify)
510 break;
511
512 return 0;
513 case FIO_Q_QUEUED:
514 /*
515 * if the engine doesn't have a commit hook,
516 * the io_u is really queued. if it does have such
517 * a hook, it has to call io_u_queued() itself.
518 */
519 if (td->io_ops->commit == NULL)
520 io_u_queued(td, io_u);
521 if (bytes_issued)
522 *bytes_issued += io_u->xfer_buflen;
523 break;
524 case FIO_Q_BUSY:
525 if (!from_verify)
526 unlog_io_piece(td, io_u);
527 requeue_io_u(td, &io_u);
528 td_io_commit(td);
529 break;
530 default:
531 assert(*ret < 0);
532 td_verror(td, -(*ret), "td_io_queue");
533 break;
534 }
535
536 if (break_on_this_error(td, ddir, ret))
537 return 1;
538
539 return 0;
540}
541
542static inline bool io_in_polling(struct thread_data *td)
543{
544 return !td->o.iodepth_batch_complete_min &&
545 !td->o.iodepth_batch_complete_max;
546}
547/*
548 * Unlinks files from thread data fio_file structure
549 */
550static int unlink_all_files(struct thread_data *td)
551{
552 struct fio_file *f;
553 unsigned int i;
554 int ret = 0;
555
556 for_each_file(td, f, i) {
557 if (f->filetype != FIO_TYPE_FILE)
558 continue;
559 ret = td_io_unlink_file(td, f);
560 if (ret)
561 break;
562 }
563
564 if (ret)
565 td_verror(td, ret, "unlink_all_files");
566
567 return ret;
568}
569
570/*
571 * Check if io_u will overlap an in-flight IO in the queue
572 */
573static bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
574{
575 bool overlap;
576 struct io_u *check_io_u;
577 unsigned long long x1, x2, y1, y2;
578 int i;
579
580 x1 = io_u->offset;
581 x2 = io_u->offset + io_u->buflen;
582 overlap = false;
583 io_u_qiter(q, check_io_u, i) {
584 if (check_io_u->flags & IO_U_F_FLIGHT) {
585 y1 = check_io_u->offset;
586 y2 = check_io_u->offset + check_io_u->buflen;
587
588 if (x1 < y2 && y1 < x2) {
589 overlap = true;
590 dprint(FD_IO, "in-flight overlap: %llu/%lu, %llu/%lu\n",
591 x1, io_u->buflen,
592 y1, check_io_u->buflen);
593 break;
594 }
595 }
596 }
597
598 return overlap;
599}
600
601static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
602{
603 /*
604 * Check for overlap if the user asked us to, and we have
605 * at least one IO in flight besides this one.
606 */
607 if (td->o.serialize_overlap && td->cur_depth > 1 &&
608 in_flight_overlap(&td->io_u_all, io_u))
609 return FIO_Q_BUSY;
610
611 return td_io_queue(td, io_u);
612}
613
614/*
615 * The main verify engine. Runs over the writes we previously submitted,
616 * reads the blocks back in, and checks the crc/md5 of the data.
617 */
618static void do_verify(struct thread_data *td, uint64_t verify_bytes)
619{
620 struct fio_file *f;
621 struct io_u *io_u;
622 int ret, min_events;
623 unsigned int i;
624
625 dprint(FD_VERIFY, "starting loop\n");
626
627 /*
628 * sync io first and invalidate cache, to make sure we really
629 * read from disk.
630 */
631 for_each_file(td, f, i) {
632 if (!fio_file_open(f))
633 continue;
634 if (fio_io_sync(td, f))
635 break;
636 if (file_invalidate_cache(td, f))
637 break;
638 }
639
640 check_update_rusage(td);
641
642 if (td->error)
643 return;
644
645 /*
646 * verify_state needs to be reset before verification
647 * proceeds so that expected random seeds match actual
648 * random seeds in headers. The main loop will reset
649 * all random number generators if randrepeat is set.
650 */
651 if (!td->o.rand_repeatable)
652 td_fill_verify_state_seed(td);
653
654 td_set_runstate(td, TD_VERIFYING);
655
656 io_u = NULL;
657 while (!td->terminate) {
658 enum fio_ddir ddir;
659 int full;
660
661 update_ts_cache(td);
662 check_update_rusage(td);
663
664 if (runtime_exceeded(td, &td->ts_cache)) {
665 __update_ts_cache(td);
666 if (runtime_exceeded(td, &td->ts_cache)) {
667 fio_mark_td_terminate(td);
668 break;
669 }
670 }
671
672 if (flow_threshold_exceeded(td))
673 continue;
674
675 if (!td->o.experimental_verify) {
676 io_u = __get_io_u(td);
677 if (!io_u)
678 break;
679
680 if (get_next_verify(td, io_u)) {
681 put_io_u(td, io_u);
682 break;
683 }
684
685 if (td_io_prep(td, io_u)) {
686 put_io_u(td, io_u);
687 break;
688 }
689 } else {
690 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
691 break;
692
693 while ((io_u = get_io_u(td)) != NULL) {
694 if (IS_ERR_OR_NULL(io_u)) {
695 io_u = NULL;
696 ret = FIO_Q_BUSY;
697 goto reap;
698 }
699
700 /*
701 * We are only interested in the places where
702 * we wrote or trimmed IOs. Turn those into
703 * reads for verification purposes.
704 */
705 if (io_u->ddir == DDIR_READ) {
706 /*
707 * Pretend we issued it for rwmix
708 * accounting
709 */
710 td->io_issues[DDIR_READ]++;
711 put_io_u(td, io_u);
712 continue;
713 } else if (io_u->ddir == DDIR_TRIM) {
714 io_u->ddir = DDIR_READ;
715 io_u_set(td, io_u, IO_U_F_TRIMMED);
716 break;
717 } else if (io_u->ddir == DDIR_WRITE) {
718 io_u->ddir = DDIR_READ;
719 populate_verify_io_u(td, io_u);
720 break;
721 } else {
722 put_io_u(td, io_u);
723 continue;
724 }
725 }
726
727 if (!io_u)
728 break;
729 }
730
731 if (verify_state_should_stop(td, io_u)) {
732 put_io_u(td, io_u);
733 break;
734 }
735
736 if (td->o.verify_async)
737 io_u->end_io = verify_io_u_async;
738 else
739 io_u->end_io = verify_io_u;
740
741 ddir = io_u->ddir;
742 if (!td->o.disable_slat)
743 fio_gettime(&io_u->start_time, NULL);
744
745 ret = io_u_submit(td, io_u);
746
747 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
748 break;
749
750 /*
751 * if we can queue more, do so. but check if there are
752 * completed io_u's first. Note that we can get BUSY even
753 * without IO queued, if the system is resource starved.
754 */
755reap:
756 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
757 if (full || io_in_polling(td))
758 ret = wait_for_completions(td, NULL);
759
760 if (ret < 0)
761 break;
762 }
763
764 check_update_rusage(td);
765
766 if (!td->error) {
767 min_events = td->cur_depth;
768
769 if (min_events)
770 ret = io_u_queued_complete(td, min_events);
771 } else
772 cleanup_pending_aio(td);
773
774 td_set_runstate(td, TD_RUNNING);
775
776 dprint(FD_VERIFY, "exiting loop\n");
777}
778
779static bool exceeds_number_ios(struct thread_data *td)
780{
781 unsigned long long number_ios;
782
783 if (!td->o.number_ios)
784 return false;
785
786 number_ios = ddir_rw_sum(td->io_blocks);
787 number_ios += td->io_u_queued + td->io_u_in_flight;
788
789 return number_ios >= (td->o.number_ios * td->loops);
790}
791
792static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
793{
794 unsigned long long bytes, limit;
795
796 if (td_rw(td))
797 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
798 else if (td_write(td))
799 bytes = this_bytes[DDIR_WRITE];
800 else if (td_read(td))
801 bytes = this_bytes[DDIR_READ];
802 else
803 bytes = this_bytes[DDIR_TRIM];
804
805 if (td->o.io_size)
806 limit = td->o.io_size;
807 else
808 limit = td->o.size;
809
810 limit *= td->loops;
811 return bytes >= limit || exceeds_number_ios(td);
812}
813
814static bool io_issue_bytes_exceeded(struct thread_data *td)
815{
816 return io_bytes_exceeded(td, td->io_issue_bytes);
817}
818
819static bool io_complete_bytes_exceeded(struct thread_data *td)
820{
821 return io_bytes_exceeded(td, td->this_io_bytes);
822}
823
824/*
825 * used to calculate the next io time for rate control
826 *
827 */
828static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
829{
830 uint64_t bps = td->rate_bps[ddir];
831
832 assert(!(td->flags & TD_F_CHILD));
833
834 if (td->o.rate_process == RATE_PROCESS_POISSON) {
835 uint64_t val, iops;
836
837 iops = bps / td->o.bs[ddir];
838 val = (int64_t) (1000000 / iops) *
839 -logf(__rand_0_1(&td->poisson_state[ddir]));
840 if (val) {
841 dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
842 (unsigned long long) 1000000 / val,
843 ddir);
844 }
845 td->last_usec[ddir] += val;
846 return td->last_usec[ddir];
847 } else if (bps) {
848 uint64_t bytes = td->rate_io_issue_bytes[ddir];
849 uint64_t secs = bytes / bps;
850 uint64_t remainder = bytes % bps;
851
852 return remainder * 1000000 / bps + secs * 1000000;
853 }
854
855 return 0;
856}
857
858static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
859{
860 unsigned long long b;
861 uint64_t total;
862 int left;
863
864 b = ddir_rw_sum(td->io_blocks);
865 if (b % td->o.thinktime_blocks)
866 return;
867
868 io_u_quiesce(td);
869
870 total = 0;
871 if (td->o.thinktime_spin)
872 total = usec_spin(td->o.thinktime_spin);
873
874 left = td->o.thinktime - total;
875 if (left)
876 total += usec_sleep(td, left);
877
878 /*
879 * If we're ignoring thinktime for the rate, add the number of bytes
880 * we would have done while sleeping, minus one block to ensure we
881 * start issuing immediately after the sleep.
882 */
883 if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
884 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
885 uint64_t bs = td->o.min_bs[ddir];
886 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
887 uint64_t over;
888
889 if (usperop <= total)
890 over = bs;
891 else
892 over = (usperop - total) / usperop * -bs;
893
894 td->rate_io_issue_bytes[ddir] += (missed - over);
895 /* adjust for rate_process=poisson */
896 td->last_usec[ddir] += total;
897 }
898}
899
900/*
901 * Main IO worker function. It retrieves io_u's to process and queues
902 * and reaps them, checking for rate and errors along the way.
903 *
904 * Returns number of bytes written and trimmed.
905 */
906static void do_io(struct thread_data *td, uint64_t *bytes_done)
907{
908 unsigned int i;
909 int ret = 0;
910 uint64_t total_bytes, bytes_issued = 0;
911
912 for (i = 0; i < DDIR_RWDIR_CNT; i++)
913 bytes_done[i] = td->bytes_done[i];
914
915 if (in_ramp_time(td))
916 td_set_runstate(td, TD_RAMP);
917 else
918 td_set_runstate(td, TD_RUNNING);
919
920 lat_target_init(td);
921
922 total_bytes = td->o.size;
923 /*
924 * Allow random overwrite workloads to write up to io_size
925 * before starting verification phase as 'size' doesn't apply.
926 */
927 if (td_write(td) && td_random(td) && td->o.norandommap)
928 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
929 /*
930 * If verify_backlog is enabled, we'll run the verify in this
931 * handler as well. For that case, we may need up to twice the
932 * amount of bytes.
933 */
934 if (td->o.verify != VERIFY_NONE &&
935 (td_write(td) && td->o.verify_backlog))
936 total_bytes += td->o.size;
937
938 /* In trimwrite mode, each byte is trimmed and then written, so
939 * allow total_bytes to be twice as big */
940 if (td_trimwrite(td))
941 total_bytes += td->total_io_size;
942
943 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
944 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
945 td->o.time_based) {
946 struct timespec comp_time;
947 struct io_u *io_u;
948 int full;
949 enum fio_ddir ddir;
950
951 check_update_rusage(td);
952
953 if (td->terminate || td->done)
954 break;
955
956 update_ts_cache(td);
957
958 if (runtime_exceeded(td, &td->ts_cache)) {
959 __update_ts_cache(td);
960 if (runtime_exceeded(td, &td->ts_cache)) {
961 fio_mark_td_terminate(td);
962 break;
963 }
964 }
965
966 if (flow_threshold_exceeded(td))
967 continue;
968
969 /*
970 * Break if we exceeded the bytes. The exception is time
971 * based runs, but we still need to break out of the loop
972 * for those to run verification, if enabled.
973 */
974 if (bytes_issued >= total_bytes &&
975 (!td->o.time_based ||
976 (td->o.time_based && td->o.verify != VERIFY_NONE)))
977 break;
978
979 io_u = get_io_u(td);
980 if (IS_ERR_OR_NULL(io_u)) {
981 int err = PTR_ERR(io_u);
982
983 io_u = NULL;
984 ddir = DDIR_INVAL;
985 if (err == -EBUSY) {
986 ret = FIO_Q_BUSY;
987 goto reap;
988 }
989 if (td->o.latency_target)
990 goto reap;
991 break;
992 }
993
994 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
995 populate_verify_io_u(td, io_u);
996
997 ddir = io_u->ddir;
998
999 /*
1000 * Add verification end_io handler if:
1001 * - Asked to verify (!td_rw(td))
1002 * - Or the io_u is from our verify list (mixed write/ver)
1003 */
1004 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
1005 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
1006
1007 if (!td->o.verify_pattern_bytes) {
1008 io_u->rand_seed = __rand(&td->verify_state);
1009 if (sizeof(int) != sizeof(long *))
1010 io_u->rand_seed *= __rand(&td->verify_state);
1011 }
1012
1013 if (verify_state_should_stop(td, io_u)) {
1014 put_io_u(td, io_u);
1015 break;
1016 }
1017
1018 if (td->o.verify_async)
1019 io_u->end_io = verify_io_u_async;
1020 else
1021 io_u->end_io = verify_io_u;
1022 td_set_runstate(td, TD_VERIFYING);
1023 } else if (in_ramp_time(td))
1024 td_set_runstate(td, TD_RAMP);
1025 else
1026 td_set_runstate(td, TD_RUNNING);
1027
1028 /*
1029 * Always log IO before it's issued, so we know the specific
1030 * order of it. The logged unit will track when the IO has
1031 * completed.
1032 */
1033 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1034 td->o.do_verify &&
1035 td->o.verify != VERIFY_NONE &&
1036 !td->o.experimental_verify)
1037 log_io_piece(td, io_u);
1038
1039 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1040 const unsigned long blen = io_u->xfer_buflen;
1041 const enum fio_ddir ddir = acct_ddir(io_u);
1042
1043 if (td->error)
1044 break;
1045
1046 workqueue_enqueue(&td->io_wq, &io_u->work);
1047 ret = FIO_Q_QUEUED;
1048
1049 if (ddir_rw(ddir)) {
1050 td->io_issues[ddir]++;
1051 td->io_issue_bytes[ddir] += blen;
1052 td->rate_io_issue_bytes[ddir] += blen;
1053 }
1054
1055 if (should_check_rate(td))
1056 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1057
1058 } else {
1059 ret = io_u_submit(td, io_u);
1060
1061 if (should_check_rate(td))
1062 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1063
1064 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
1065 break;
1066
1067 /*
1068 * See if we need to complete some commands. Note that
1069 * we can get BUSY even without IO queued, if the
1070 * system is resource starved.
1071 */
1072reap:
1073 full = queue_full(td) ||
1074 (ret == FIO_Q_BUSY && td->cur_depth);
1075 if (full || io_in_polling(td))
1076 ret = wait_for_completions(td, &comp_time);
1077 }
1078 if (ret < 0)
1079 break;
1080 if (!ddir_rw_sum(td->bytes_done) &&
1081 !td_ioengine_flagged(td, FIO_NOIO))
1082 continue;
1083
1084 if (!in_ramp_time(td) && should_check_rate(td)) {
1085 if (check_min_rate(td, &comp_time)) {
1086 if (exitall_on_terminate || td->o.exitall_error)
1087 fio_terminate_threads(td->groupid);
1088 td_verror(td, EIO, "check_min_rate");
1089 break;
1090 }
1091 }
1092 if (!in_ramp_time(td) && td->o.latency_target)
1093 lat_target_check(td);
1094
1095 if (ddir_rw(ddir) && td->o.thinktime)
1096 handle_thinktime(td, ddir);
1097 }
1098
1099 check_update_rusage(td);
1100
1101 if (td->trim_entries)
1102 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
1103
1104 if (td->o.fill_device && td->error == ENOSPC) {
1105 td->error = 0;
1106 fio_mark_td_terminate(td);
1107 }
1108 if (!td->error) {
1109 struct fio_file *f;
1110
1111 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1112 workqueue_flush(&td->io_wq);
1113 i = 0;
1114 } else
1115 i = td->cur_depth;
1116
1117 if (i) {
1118 ret = io_u_queued_complete(td, i);
1119 if (td->o.fill_device && td->error == ENOSPC)
1120 td->error = 0;
1121 }
1122
1123 if (should_fsync(td) && td->o.end_fsync) {
1124 td_set_runstate(td, TD_FSYNCING);
1125
1126 for_each_file(td, f, i) {
1127 if (!fio_file_fsync(td, f))
1128 continue;
1129
1130 log_err("fio: end_fsync failed for file %s\n",
1131 f->file_name);
1132 }
1133 }
1134 } else
1135 cleanup_pending_aio(td);
1136
1137 /*
1138 * stop job if we failed doing any IO
1139 */
1140 if (!ddir_rw_sum(td->this_io_bytes))
1141 td->done = 1;
1142
1143 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1144 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
1145}
1146
1147static void free_file_completion_logging(struct thread_data *td)
1148{
1149 struct fio_file *f;
1150 unsigned int i;
1151
1152 for_each_file(td, f, i) {
1153 if (!f->last_write_comp)
1154 break;
1155 sfree(f->last_write_comp);
1156 }
1157}
1158
1159static int init_file_completion_logging(struct thread_data *td,
1160 unsigned int depth)
1161{
1162 struct fio_file *f;
1163 unsigned int i;
1164
1165 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1166 return 0;
1167
1168 for_each_file(td, f, i) {
1169 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1170 if (!f->last_write_comp)
1171 goto cleanup;
1172 }
1173
1174 return 0;
1175
1176cleanup:
1177 free_file_completion_logging(td);
1178 log_err("fio: failed to alloc write comp data\n");
1179 return 1;
1180}
1181
1182static void cleanup_io_u(struct thread_data *td)
1183{
1184 struct io_u *io_u;
1185
1186 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
1187
1188 if (td->io_ops->io_u_free)
1189 td->io_ops->io_u_free(td, io_u);
1190
1191 fio_memfree(io_u, sizeof(*io_u));
1192 }
1193
1194 free_io_mem(td);
1195
1196 io_u_rexit(&td->io_u_requeues);
1197 io_u_qexit(&td->io_u_freelist);
1198 io_u_qexit(&td->io_u_all);
1199
1200 free_file_completion_logging(td);
1201}
1202
1203static int init_io_u(struct thread_data *td)
1204{
1205 struct io_u *io_u;
1206 unsigned int max_bs, min_write;
1207 int cl_align, i, max_units;
1208 int data_xfer = 1, err;
1209 char *p;
1210
1211 max_units = td->o.iodepth;
1212 max_bs = td_max_bs(td);
1213 min_write = td->o.min_bs[DDIR_WRITE];
1214 td->orig_buffer_size = (unsigned long long) max_bs
1215 * (unsigned long long) max_units;
1216
1217 if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1218 data_xfer = 0;
1219
1220 err = 0;
1221 err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1222 err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1223 err += !io_u_qinit(&td->io_u_all, td->o.iodepth);
1224
1225 if (err) {
1226 log_err("fio: failed setting up IO queues\n");
1227 return 1;
1228 }
1229
1230 /*
1231 * if we may later need to do address alignment, then add any
1232 * possible adjustment here so that we don't cause a buffer
1233 * overflow later. this adjustment may be too much if we get
1234 * lucky and the allocator gives us an aligned address.
1235 */
1236 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1237 td_ioengine_flagged(td, FIO_RAWIO))
1238 td->orig_buffer_size += page_mask + td->o.mem_align;
1239
1240 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1241 unsigned long bs;
1242
1243 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1244 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1245 }
1246
1247 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1248 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1249 return 1;
1250 }
1251
1252 if (data_xfer && allocate_io_mem(td))
1253 return 1;
1254
1255 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1256 td_ioengine_flagged(td, FIO_RAWIO))
1257 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
1258 else
1259 p = td->orig_buffer;
1260
1261 cl_align = os_cache_line_size();
1262
1263 for (i = 0; i < max_units; i++) {
1264 void *ptr;
1265
1266 if (td->terminate)
1267 return 1;
1268
1269 ptr = fio_memalign(cl_align, sizeof(*io_u));
1270 if (!ptr) {
1271 log_err("fio: unable to allocate aligned memory\n");
1272 break;
1273 }
1274
1275 io_u = ptr;
1276 memset(io_u, 0, sizeof(*io_u));
1277 INIT_FLIST_HEAD(&io_u->verify_list);
1278 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1279
1280 if (data_xfer) {
1281 io_u->buf = p;
1282 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1283
1284 if (td_write(td))
1285 io_u_fill_buffer(td, io_u, min_write, max_bs);
1286 if (td_write(td) && td->o.verify_pattern_bytes) {
1287 /*
1288 * Fill the buffer with the pattern if we are
1289 * going to be doing writes.
1290 */
1291 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
1292 }
1293 }
1294
1295 io_u->index = i;
1296 io_u->flags = IO_U_F_FREE;
1297 io_u_qpush(&td->io_u_freelist, io_u);
1298
1299 /*
1300 * io_u never leaves this stack, used for iteration of all
1301 * io_u buffers.
1302 */
1303 io_u_qpush(&td->io_u_all, io_u);
1304
1305 if (td->io_ops->io_u_init) {
1306 int ret = td->io_ops->io_u_init(td, io_u);
1307
1308 if (ret) {
1309 log_err("fio: failed to init engine data: %d\n", ret);
1310 return 1;
1311 }
1312 }
1313
1314 p += max_bs;
1315 }
1316
1317 if (init_file_completion_logging(td, max_units))
1318 return 1;
1319
1320 return 0;
1321}
1322
1323/*
1324 * This function is Linux specific.
1325 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1326 */
1327static int switch_ioscheduler(struct thread_data *td)
1328{
1329#ifdef FIO_HAVE_IOSCHED_SWITCH
1330 char tmp[256], tmp2[128], *p;
1331 FILE *f;
1332 int ret;
1333
1334 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1335 return 0;
1336
1337 assert(td->files && td->files[0]);
1338 sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
1339
1340 f = fopen(tmp, "r+");
1341 if (!f) {
1342 if (errno == ENOENT) {
1343 log_err("fio: os or kernel doesn't support IO scheduler"
1344 " switching\n");
1345 return 0;
1346 }
1347 td_verror(td, errno, "fopen iosched");
1348 return 1;
1349 }
1350
1351 /*
1352 * Set io scheduler.
1353 */
1354 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1355 if (ferror(f) || ret != 1) {
1356 td_verror(td, errno, "fwrite");
1357 fclose(f);
1358 return 1;
1359 }
1360
1361 rewind(f);
1362
1363 /*
1364 * Read back and check that the selected scheduler is now the default.
1365 */
1366 ret = fread(tmp, 1, sizeof(tmp) - 1, f);
1367 if (ferror(f) || ret < 0) {
1368 td_verror(td, errno, "fread");
1369 fclose(f);
1370 return 1;
1371 }
1372 tmp[ret] = '\0';
1373 /*
1374 * either a list of io schedulers or "none\n" is expected. Strip the
1375 * trailing newline.
1376 */
1377 p = tmp;
1378 strsep(&p, "\n");
1379
1380 /*
1381 * Write to "none" entry doesn't fail, so check the result here.
1382 */
1383 if (!strcmp(tmp, "none")) {
1384 log_err("fio: io scheduler is not tunable\n");
1385 fclose(f);
1386 return 0;
1387 }
1388
1389 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1390 if (!strstr(tmp, tmp2)) {
1391 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1392 td_verror(td, EINVAL, "iosched_switch");
1393 fclose(f);
1394 return 1;
1395 }
1396
1397 fclose(f);
1398 return 0;
1399#else
1400 return 0;
1401#endif
1402}
1403
1404static bool keep_running(struct thread_data *td)
1405{
1406 unsigned long long limit;
1407
1408 if (td->done)
1409 return false;
1410 if (td->terminate)
1411 return false;
1412 if (td->o.time_based)
1413 return true;
1414 if (td->o.loops) {
1415 td->o.loops--;
1416 return true;
1417 }
1418 if (exceeds_number_ios(td))
1419 return false;
1420
1421 if (td->o.io_size)
1422 limit = td->o.io_size;
1423 else
1424 limit = td->o.size;
1425
1426 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
1427 uint64_t diff;
1428
1429 /*
1430 * If the difference is less than the maximum IO size, we
1431 * are done.
1432 */
1433 diff = limit - ddir_rw_sum(td->io_bytes);
1434 if (diff < td_max_bs(td))
1435 return false;
1436
1437 if (fio_files_done(td) && !td->o.io_size)
1438 return false;
1439
1440 return true;
1441 }
1442
1443 return false;
1444}
1445
1446static int exec_string(struct thread_options *o, const char *string, const char *mode)
1447{
1448 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
1449 int ret;
1450 char *str;
1451
1452 str = malloc(newlen);
1453 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
1454
1455 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
1456 ret = system(str);
1457 if (ret == -1)
1458 log_err("fio: exec of cmd <%s> failed\n", str);
1459
1460 free(str);
1461 return ret;
1462}
1463
1464/*
1465 * Dry run to compute correct state of numberio for verification.
1466 */
1467static uint64_t do_dry_run(struct thread_data *td)
1468{
1469 td_set_runstate(td, TD_RUNNING);
1470
1471 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1472 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
1473 struct io_u *io_u;
1474 int ret;
1475
1476 if (td->terminate || td->done)
1477 break;
1478
1479 io_u = get_io_u(td);
1480 if (IS_ERR_OR_NULL(io_u))
1481 break;
1482
1483 io_u_set(td, io_u, IO_U_F_FLIGHT);
1484 io_u->error = 0;
1485 io_u->resid = 0;
1486 if (ddir_rw(acct_ddir(io_u)))
1487 td->io_issues[acct_ddir(io_u)]++;
1488 if (ddir_rw(io_u->ddir)) {
1489 io_u_mark_depth(td, 1);
1490 td->ts.total_io_u[io_u->ddir]++;
1491 }
1492
1493 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1494 td->o.do_verify &&
1495 td->o.verify != VERIFY_NONE &&
1496 !td->o.experimental_verify)
1497 log_io_piece(td, io_u);
1498
1499 ret = io_u_sync_complete(td, io_u);
1500 (void) ret;
1501 }
1502
1503 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
1504}
1505
1506struct fork_data {
1507 struct thread_data *td;
1508 struct sk_out *sk_out;
1509};
1510
1511/*
1512 * Entry point for the thread based jobs. The process based jobs end up
1513 * here as well, after a little setup.
1514 */
1515static void *thread_main(void *data)
1516{
1517 struct fork_data *fd = data;
1518 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
1519 struct thread_data *td = fd->td;
1520 struct thread_options *o = &td->o;
1521 struct sk_out *sk_out = fd->sk_out;
1522 uint64_t bytes_done[DDIR_RWDIR_CNT];
1523 int deadlock_loop_cnt;
1524 bool clear_state, did_some_io;
1525 int ret;
1526
1527 sk_out_assign(sk_out);
1528 free(fd);
1529
1530 if (!o->use_thread) {
1531 setsid();
1532 td->pid = getpid();
1533 } else
1534 td->pid = gettid();
1535
1536 fio_local_clock_init(o->use_thread);
1537
1538 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1539
1540 if (is_backend)
1541 fio_server_send_start(td);
1542
1543 INIT_FLIST_HEAD(&td->io_log_list);
1544 INIT_FLIST_HEAD(&td->io_hist_list);
1545 INIT_FLIST_HEAD(&td->verify_list);
1546 INIT_FLIST_HEAD(&td->trim_list);
1547 td->io_hist_tree = RB_ROOT;
1548
1549 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
1550 if (ret) {
1551 td_verror(td, ret, "mutex_cond_init_pshared");
1552 goto err;
1553 }
1554 ret = cond_init_pshared(&td->verify_cond);
1555 if (ret) {
1556 td_verror(td, ret, "mutex_cond_pshared");
1557 goto err;
1558 }
1559
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");
1566
1567 /*
1568 * A new gid requires privilege, so we need to do this before setting
1569 * the uid.
1570 */
1571 if (o->gid != -1U && setgid(o->gid)) {
1572 td_verror(td, errno, "setgid");
1573 goto err;
1574 }
1575 if (o->uid != -1U && setuid(o->uid)) {
1576 td_verror(td, errno, "setuid");
1577 goto err;
1578 }
1579
1580 /*
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
1584 */
1585 if (iolog_compress_init(td, sk_out))
1586 goto err;
1587
1588 /*
1589 * If we have a gettimeofday() thread, make sure we exclude that
1590 * thread from this job
1591 */
1592 if (o->gtod_cpu)
1593 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
1594
1595 /*
1596 * Set affinity first, in case it has an impact on the memory
1597 * allocations.
1598 */
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);
1602 if (!ret) {
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");
1606 goto err;
1607 }
1608 }
1609 ret = fio_setaffinity(td->pid, o->cpumask);
1610 if (ret == -1) {
1611 td_verror(td, errno, "cpu_set_affinity");
1612 goto err;
1613 }
1614 }
1615
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;
1621
1622 if (numa_available() < 0) {
1623 td_verror(td, errno, "Does not support NUMA API\n");
1624 goto err;
1625 }
1626
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);
1631 if (ret == -1) {
1632 td_verror(td, errno, \
1633 "numa_run_on_node_mask failed\n");
1634 goto err;
1635 }
1636 }
1637
1638 if (fio_option_is_set(o, numa_memnodes)) {
1639 mask = NULL;
1640 if (o->numa_memnodes)
1641 mask = numa_parse_nodestring(o->numa_memnodes);
1642
1643 switch (o->numa_mem_mode) {
1644 case MPOL_INTERLEAVE:
1645 numa_set_interleave_mask(mask);
1646 break;
1647 case MPOL_BIND:
1648 numa_set_membind(mask);
1649 break;
1650 case MPOL_LOCAL:
1651 numa_set_localalloc();
1652 break;
1653 case MPOL_PREFERRED:
1654 numa_set_preferred(o->numa_mem_prefer_node);
1655 break;
1656 case MPOL_DEFAULT:
1657 default:
1658 break;
1659 }
1660
1661 if (mask)
1662 numa_free_nodemask(mask);
1663
1664 }
1665 }
1666#endif
1667
1668 if (fio_pin_memory(td))
1669 goto err;
1670
1671 /*
1672 * May alter parameters that init_io_u() will use, so we need to
1673 * do this first.
1674 */
1675 if (!init_iolog(td))
1676 goto err;
1677
1678 if (init_io_u(td))
1679 goto err;
1680
1681 if (o->verify_async && verify_async_init(td))
1682 goto err;
1683
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);
1687 if (ret == -1) {
1688 td_verror(td, errno, "ioprio_set");
1689 goto err;
1690 }
1691 }
1692
1693 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1694 goto err;
1695
1696 errno = 0;
1697 if (nice(o->nice) == -1 && errno != 0) {
1698 td_verror(td, errno, "nice");
1699 goto err;
1700 }
1701
1702 if (o->ioscheduler && switch_ioscheduler(td))
1703 goto err;
1704
1705 if (!o->create_serialize && setup_files(td))
1706 goto err;
1707
1708 if (td_io_init(td))
1709 goto err;
1710
1711 if (!init_random_map(td))
1712 goto err;
1713
1714 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
1715 goto err;
1716
1717 if (o->pre_read && !pre_read_files(td))
1718 goto err;
1719
1720 fio_verify_init(td);
1721
1722 if (rate_submit_init(td, sk_out))
1723 goto err;
1724
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));
1730
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));
1739 }
1740
1741 memset(bytes_done, 0, sizeof(bytes_done));
1742 clear_state = false;
1743 did_some_io = false;
1744
1745 while (keep_running(td)) {
1746 uint64_t verify_bytes;
1747
1748 fio_gettime(&td->start, NULL);
1749 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
1750
1751 if (clear_state) {
1752 clear_io_state(td, 0);
1753
1754 if (o->unlink_each_loop && unlink_all_files(td))
1755 break;
1756 }
1757
1758 prune_io_piece_log(td);
1759
1760 if (td->o.verify_only && td_write(td))
1761 verify_bytes = do_dry_run(td);
1762 else {
1763 do_io(td, bytes_done);
1764
1765 if (!ddir_rw_sum(bytes_done)) {
1766 fio_mark_td_terminate(td);
1767 verify_bytes = 0;
1768 } else {
1769 verify_bytes = bytes_done[DDIR_WRITE] +
1770 bytes_done[DDIR_TRIM];
1771 }
1772 }
1773
1774 /*
1775 * If we took too long to shut down, the main thread could
1776 * already consider us reaped/exited. If that happens, break
1777 * out and clean up.
1778 */
1779 if (td->runstate >= TD_EXITED)
1780 break;
1781
1782 clear_state = true;
1783
1784 /*
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.
1790 */
1791 deadlock_loop_cnt = 0;
1792 do {
1793 check_update_rusage(td);
1794 if (!fio_sem_down_trylock(stat_sem))
1795 break;
1796 usleep(1000);
1797 if (deadlock_loop_cnt++ > 5000) {
1798 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
1799 td->error = EDEADLK;
1800 goto err;
1801 }
1802 } while (1);
1803
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);
1812
1813 if (td->error || td->terminate)
1814 break;
1815
1816 if (!o->do_verify ||
1817 o->verify == VERIFY_NONE ||
1818 td_ioengine_flagged(td, FIO_UNIDIR))
1819 continue;
1820
1821 if (ddir_rw_sum(bytes_done))
1822 did_some_io = true;
1823
1824 clear_io_state(td, 0);
1825
1826 fio_gettime(&td->start, NULL);
1827
1828 do_verify(td, verify_bytes);
1829
1830 /*
1831 * See comment further up for why this is done here.
1832 */
1833 check_update_rusage(td);
1834
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);
1839
1840 if (td->error || td->terminate)
1841 break;
1842 }
1843
1844 /*
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 ?)
1848 */
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);
1856
1857 td_set_runstate(td, TD_FINISHING);
1858
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];
1864
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);
1868
1869 fio_unpin_memory(td);
1870
1871 td_writeout_logs(td, true);
1872
1873 iolog_compress_exit(td);
1874 rate_submit_exit(td);
1875
1876 if (o->exec_postrun)
1877 exec_string(o, o->exec_postrun, (const char *)"postrun");
1878
1879 if (exitall_on_terminate || (o->exitall_error && td->error))
1880 fio_terminate_threads(td->groupid);
1881
1882err:
1883 if (td->error)
1884 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1885 td->verror);
1886
1887 if (o->verify_async)
1888 verify_async_exit(td);
1889
1890 close_and_free_files(td);
1891 cleanup_io_u(td);
1892 close_ioengine(td);
1893 cgroup_shutdown(td, &cgroup_mnt);
1894 verify_free_state(td);
1895
1896 if (td->zone_state_index) {
1897 int i;
1898
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;
1903 }
1904
1905 if (fio_option_is_set(o, cpumask)) {
1906 ret = fio_cpuset_exit(&o->cpumask);
1907 if (ret)
1908 td_verror(td, ret, "fio_cpuset_exit");
1909 }
1910
1911 /*
1912 * do this very late, it will log file closing as well
1913 */
1914 if (o->write_iolog_file)
1915 write_iolog_close(td);
1916
1917 td_set_runstate(td, TD_EXITED);
1918
1919 /*
1920 * Do this last after setting our runstate to exited, so we
1921 * know that the stat thread is signaled.
1922 */
1923 check_update_rusage(td);
1924
1925 sk_out_drop();
1926 return (void *) (uintptr_t) td->error;
1927}
1928
1929/*
1930 * Run over the job map and reap the threads that have exited, if any.
1931 */
1932static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
1933 uint64_t *m_rate)
1934{
1935 struct thread_data *td;
1936 unsigned int cputhreads, realthreads, pending;
1937 int i, status, ret;
1938
1939 /*
1940 * reap exited threads (TD_EXITED -> TD_REAPED)
1941 */
1942 realthreads = pending = cputhreads = 0;
1943 for_each_td(td, i) {
1944 int flags = 0;
1945
1946 if (!strcmp(td->o.ioengine, "cpuio"))
1947 cputhreads++;
1948 else
1949 realthreads++;
1950
1951 if (!td->pid) {
1952 pending++;
1953 continue;
1954 }
1955 if (td->runstate == TD_REAPED)
1956 continue;
1957 if (td->o.use_thread) {
1958 if (td->runstate == TD_EXITED) {
1959 td_set_runstate(td, TD_REAPED);
1960 goto reaped;
1961 }
1962 continue;
1963 }
1964
1965 flags = WNOHANG;
1966 if (td->runstate == TD_EXITED)
1967 flags = 0;
1968
1969 /*
1970 * check if someone quit or got killed in an unusual way
1971 */
1972 ret = waitpid(td->pid, &status, flags);
1973 if (ret < 0) {
1974 if (errno == ECHILD) {
1975 log_err("fio: pid=%d disappeared %d\n",
1976 (int) td->pid, td->runstate);
1977 td->sig = ECHILD;
1978 td_set_runstate(td, TD_REAPED);
1979 goto reaped;
1980 }
1981 perror("waitpid");
1982 } else if (ret == td->pid) {
1983 if (WIFSIGNALED(status)) {
1984 int sig = WTERMSIG(status);
1985
1986 if (sig != SIGTERM && sig != SIGUSR2)
1987 log_err("fio: pid=%d, got signal=%d\n",
1988 (int) td->pid, sig);
1989 td->sig = sig;
1990 td_set_runstate(td, TD_REAPED);
1991 goto reaped;
1992 }
1993 if (WIFEXITED(status)) {
1994 if (WEXITSTATUS(status) && !td->error)
1995 td->error = WEXITSTATUS(status);
1996
1997 td_set_runstate(td, TD_REAPED);
1998 goto reaped;
1999 }
2000 }
2001
2002 /*
2003 * If the job is stuck, do a forceful timeout of it and
2004 * move on.
2005 */
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);
2015 goto reaped;
2016 }
2017
2018 /*
2019 * thread is not dead, continue
2020 */
2021 pending++;
2022 continue;
2023reaped:
2024 (*nr_running)--;
2025 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2026 (*t_rate) -= ddir_rw_sum(td->o.rate);
2027 if (!td->pid)
2028 pending--;
2029
2030 if (td->error)
2031 exit_value++;
2032
2033 done_secs += mtime_since_now(&td->epoch) / 1000;
2034 profile_td_exit(td);
2035 }
2036
2037 if (*nr_running == cputhreads && !pending && realthreads)
2038 fio_terminate_threads(TERMINATE_ALL);
2039}
2040
2041static bool __check_trigger_file(void)
2042{
2043 struct stat sb;
2044
2045 if (!trigger_file)
2046 return false;
2047
2048 if (stat(trigger_file, &sb))
2049 return false;
2050
2051 if (unlink(trigger_file) < 0)
2052 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2053 strerror(errno));
2054
2055 return true;
2056}
2057
2058static bool trigger_timedout(void)
2059{
2060 if (trigger_timeout)
2061 if (time_since_genesis() >= trigger_timeout) {
2062 trigger_timeout = 0;
2063 return true;
2064 }
2065
2066 return false;
2067}
2068
2069void exec_trigger(const char *cmd)
2070{
2071 int ret;
2072
2073 if (!cmd || cmd[0] == '\0')
2074 return;
2075
2076 ret = system(cmd);
2077 if (ret == -1)
2078 log_err("fio: failed executing %s trigger\n", cmd);
2079}
2080
2081void check_trigger_file(void)
2082{
2083 if (__check_trigger_file() || trigger_timedout()) {
2084 if (nr_clients)
2085 fio_clients_send_trigger(trigger_remote_cmd);
2086 else {
2087 verify_save_state(IO_LIST_ALL);
2088 fio_terminate_threads(TERMINATE_ALL);
2089 exec_trigger(trigger_cmd);
2090 }
2091 }
2092}
2093
2094static int fio_verify_load_state(struct thread_data *td)
2095{
2096 int ret;
2097
2098 if (!td->o.verify_state)
2099 return 0;
2100
2101 if (is_backend) {
2102 void *data;
2103
2104 ret = fio_server_get_verify_state(td->o.name,
2105 td->thread_number - 1, &data);
2106 if (!ret)
2107 verify_assign_state(td, data);
2108 } else
2109 ret = verify_load_state(td, "local");
2110
2111 return ret;
2112}
2113
2114static void do_usleep(unsigned int usecs)
2115{
2116 check_for_running_stats();
2117 check_trigger_file();
2118 usleep(usecs);
2119}
2120
2121static bool check_mount_writes(struct thread_data *td)
2122{
2123 struct fio_file *f;
2124 unsigned int i;
2125
2126 if (!td_write(td) || td->o.allow_mounted_write)
2127 return false;
2128
2129 /*
2130 * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2131 * are mkfs'd and mounted.
2132 */
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)
2136#else
2137 if (f->filetype != FIO_TYPE_BLOCK)
2138#endif
2139 continue;
2140 if (device_is_mounted(f->file_name))
2141 goto mounted;
2142 }
2143
2144 return false;
2145mounted:
2146 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
2147 return true;
2148}
2149
2150static bool waitee_running(struct thread_data *me)
2151{
2152 const char *waitee = me->o.wait_for;
2153 const char *self = me->o.name;
2154 struct thread_data *td;
2155 int i;
2156
2157 if (!waitee)
2158 return false;
2159
2160 for_each_td(td, i) {
2161 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2162 continue;
2163
2164 if (td->runstate < TD_EXITED) {
2165 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2166 self, td->o.name,
2167 runstate_to_name(td->runstate));
2168 return true;
2169 }
2170 }
2171
2172 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2173 return false;
2174}
2175
2176/*
2177 * Main function for kicking off and reaping jobs, as needed.
2178 */
2179static void run_threads(struct sk_out *sk_out)
2180{
2181 struct thread_data *td;
2182 unsigned int i, todo, nr_running, nr_started;
2183 uint64_t m_rate, t_rate;
2184 uint64_t spent;
2185
2186 if (fio_gtod_offload && fio_start_gtod_thread())
2187 return;
2188
2189 fio_idle_prof_init();
2190
2191 set_sig_handlers();
2192
2193 nr_thread = nr_process = 0;
2194 for_each_td(td, i) {
2195 if (check_mount_writes(td))
2196 return;
2197 if (td->o.use_thread)
2198 nr_thread++;
2199 else
2200 nr_process++;
2201 }
2202
2203 if (output_format & FIO_OUTPUT_NORMAL) {
2204 log_info("Starting ");
2205 if (nr_thread)
2206 log_info("%d thread%s", nr_thread,
2207 nr_thread > 1 ? "s" : "");
2208 if (nr_process) {
2209 if (nr_thread)
2210 log_info(" and ");
2211 log_info("%d process%s", nr_process,
2212 nr_process > 1 ? "es" : "");
2213 }
2214 log_info("\n");
2215 log_info_flush();
2216 }
2217
2218 todo = thread_number;
2219 nr_running = 0;
2220 nr_started = 0;
2221 m_rate = t_rate = 0;
2222
2223 for_each_td(td, i) {
2224 print_status_init(td->thread_number - 1);
2225
2226 if (!td->o.create_serialize)
2227 continue;
2228
2229 if (fio_verify_load_state(td))
2230 goto reap;
2231
2232 /*
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
2236 */
2237 if (setup_files(td)) {
2238reap:
2239 exit_value++;
2240 if (td->error)
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);
2244 todo--;
2245 } else {
2246 struct fio_file *f;
2247 unsigned int j;
2248
2249 /*
2250 * for sharing to work, each job must always open
2251 * its own files. so close them, if we opened them
2252 * for creation
2253 */
2254 for_each_file(td, f, j) {
2255 if (fio_file_open(f))
2256 td_io_close_file(td, f);
2257 }
2258 }
2259 }
2260
2261 /* start idle threads before io threads start to run */
2262 fio_idle_prof_start();
2263
2264 set_genesis_time();
2265
2266 while (todo) {
2267 struct thread_data *map[REAL_MAX_JOBS];
2268 struct timespec this_start;
2269 int this_jobs = 0, left;
2270 struct fork_data *fd;
2271
2272 /*
2273 * create threads (TD_NOT_CREATED -> TD_CREATED)
2274 */
2275 for_each_td(td, i) {
2276 if (td->runstate != TD_NOT_CREATED)
2277 continue;
2278
2279 /*
2280 * never got a chance to start, killed by other
2281 * thread for some reason
2282 */
2283 if (td->terminate) {
2284 todo--;
2285 continue;
2286 }
2287
2288 if (td->o.start_delay) {
2289 spent = utime_since_genesis();
2290
2291 if (td->o.start_delay > spent)
2292 continue;
2293 }
2294
2295 if (td->o.stonewall && (nr_started || nr_running)) {
2296 dprint(FD_PROCESS, "%s: stonewall wait\n",
2297 td->o.name);
2298 break;
2299 }
2300
2301 if (waitee_running(td)) {
2302 dprint(FD_PROCESS, "%s: waiting for %s\n",
2303 td->o.name, td->o.wait_for);
2304 continue;
2305 }
2306
2307 init_disk_util(td);
2308
2309 td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
2310 td->update_rusage = 0;
2311
2312 /*
2313 * Set state to created. Thread will transition
2314 * to TD_INITIALIZED when it's done setting up.
2315 */
2316 td_set_runstate(td, TD_CREATED);
2317 map[this_jobs++] = td;
2318 nr_started++;
2319
2320 fd = calloc(1, sizeof(*fd));
2321 fd->td = td;
2322 fd->sk_out = sk_out;
2323
2324 if (td->o.use_thread) {
2325 int ret;
2326
2327 dprint(FD_PROCESS, "will pthread_create\n");
2328 ret = pthread_create(&td->thread, NULL,
2329 thread_main, fd);
2330 if (ret) {
2331 log_err("pthread_create: %s\n",
2332 strerror(ret));
2333 free(fd);
2334 nr_started--;
2335 break;
2336 }
2337 fd = NULL;
2338 ret = pthread_detach(td->thread);
2339 if (ret)
2340 log_err("pthread_detach: %s",
2341 strerror(ret));
2342 } else {
2343 pid_t pid;
2344 dprint(FD_PROCESS, "will fork\n");
2345 pid = fork();
2346 if (!pid) {
2347 int ret;
2348
2349 ret = (int)(uintptr_t)thread_main(fd);
2350 _exit(ret);
2351 } else if (i == fio_debug_jobno)
2352 *fio_debug_jobp = pid;
2353 }
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);
2358 fio_abort = 1;
2359 nr_started--;
2360 free(fd);
2361 break;
2362 }
2363 dprint(FD_MUTEX, "done waiting on startup_sem\n");
2364 }
2365
2366 /*
2367 * Wait for the started threads to transition to
2368 * TD_INITIALIZED.
2369 */
2370 fio_gettime(&this_start, NULL);
2371 left = this_jobs;
2372 while (left && !fio_abort) {
2373 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2374 break;
2375
2376 do_usleep(100000);
2377
2378 for (i = 0; i < this_jobs; i++) {
2379 td = map[i];
2380 if (!td)
2381 continue;
2382 if (td->runstate == TD_INITIALIZED) {
2383 map[i] = NULL;
2384 left--;
2385 } else if (td->runstate >= TD_EXITED) {
2386 map[i] = NULL;
2387 left--;
2388 todo--;
2389 nr_running++; /* work-around... */
2390 }
2391 }
2392 }
2393
2394 if (left) {
2395 log_err("fio: %d job%s failed to start\n", left,
2396 left > 1 ? "s" : "");
2397 for (i = 0; i < this_jobs; i++) {
2398 td = map[i];
2399 if (!td)
2400 continue;
2401 kill(td->pid, SIGTERM);
2402 }
2403 break;
2404 }
2405
2406 /*
2407 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2408 */
2409 for_each_td(td, i) {
2410 if (td->runstate != TD_INITIALIZED)
2411 continue;
2412
2413 if (in_ramp_time(td))
2414 td_set_runstate(td, TD_RAMP);
2415 else
2416 td_set_runstate(td, TD_RUNNING);
2417 nr_running++;
2418 nr_started--;
2419 m_rate += ddir_rw_sum(td->o.ratemin);
2420 t_rate += ddir_rw_sum(td->o.rate);
2421 todo--;
2422 fio_sem_up(td->sem);
2423 }
2424
2425 reap_threads(&nr_running, &t_rate, &m_rate);
2426
2427 if (todo)
2428 do_usleep(100000);
2429 }
2430
2431 while (nr_running) {
2432 reap_threads(&nr_running, &t_rate, &m_rate);
2433 do_usleep(10000);
2434 }
2435
2436 fio_idle_prof_stop();
2437
2438 update_io_ticks();
2439}
2440
2441static void free_disk_util(void)
2442{
2443 disk_util_prune_entries();
2444 helper_thread_destroy();
2445}
2446
2447int fio_backend(struct sk_out *sk_out)
2448{
2449 struct thread_data *td;
2450 int i;
2451
2452 if (exec_profile) {
2453 if (load_profile(exec_profile))
2454 return 1;
2455 free(exec_profile);
2456 exec_profile = NULL;
2457 }
2458 if (!thread_number)
2459 return 0;
2460
2461 if (write_bw_log) {
2462 struct log_params p = {
2463 .log_type = IO_LOG_TYPE_BW,
2464 };
2465
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");
2469 }
2470
2471 startup_sem = fio_sem_init(FIO_SEM_LOCKED);
2472 if (startup_sem == NULL)
2473 return 1;
2474
2475 set_genesis_time();
2476 stat_init();
2477 helper_thread_create(startup_sem, sk_out);
2478
2479 cgroup_list = smalloc(sizeof(*cgroup_list));
2480 if (cgroup_list)
2481 INIT_FLIST_HEAD(cgroup_list);
2482
2483 run_threads(sk_out);
2484
2485 helper_thread_exit();
2486
2487 if (!fio_abort) {
2488 __show_run_stats();
2489 if (write_bw_log) {
2490 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2491 struct io_log *log = agg_io_log[i];
2492
2493 flush_log(log, false);
2494 free_log(log);
2495 }
2496 }
2497 }
2498
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;
2505 }
2506 fio_sem_remove(td->sem);
2507 td->sem = NULL;
2508 }
2509
2510 free_disk_util();
2511 if (cgroup_list) {
2512 cgroup_kill(cgroup_list);
2513 sfree(cgroup_list);
2514 }
2515 sfree(cgroup_mnt);
2516
2517 fio_sem_remove(startup_sem);
2518 stat_exit();
2519 return exit_value;
2520}
fio - Flexible IO Tester