blob: 52791040ff21384d283b2a09de3ff9718e19d866 [file] [log] [blame]
Jens Axboe2e1df072012-02-09 11:15:02 +01001/*
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., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
22 *
23 */
24#include <unistd.h>
25#include <fcntl.h>
26#include <string.h>
27#include <limits.h>
28#include <signal.h>
29#include <time.h>
30#include <locale.h>
31#include <assert.h>
32#include <time.h>
Bruce Crane43606c2012-02-20 09:34:24 +010033#include <inttypes.h>
Jens Axboe2e1df072012-02-09 11:15:02 +010034#include <sys/stat.h>
35#include <sys/wait.h>
36#include <sys/ipc.h>
37#include <sys/shm.h>
38#include <sys/mman.h>
39
40#include "fio.h"
41#include "hash.h"
42#include "smalloc.h"
43#include "verify.h"
44#include "trim.h"
45#include "diskutil.h"
46#include "cgroup.h"
47#include "profile.h"
48#include "lib/rand.h"
49#include "memalign.h"
50#include "server.h"
51
52static pthread_t disk_util_thread;
53static struct fio_mutex *startup_mutex;
54static struct fio_mutex *writeout_mutex;
55static struct flist_head *cgroup_list;
56static char *cgroup_mnt;
57static int exit_value;
58static volatile int fio_abort;
59
60struct io_log *agg_io_log[2];
61
Jens Axboea3efc912012-02-09 11:25:24 +010062int groupid = 0;
63unsigned int thread_number = 0;
64unsigned int nr_process = 0;
65unsigned int nr_thread = 0;
66int shm_id = 0;
67int temp_stall_ts;
68unsigned long done_secs = 0;
69
Jens Axboe2e1df072012-02-09 11:15:02 +010070#define PAGE_ALIGN(buf) \
Bruce Crane43606c2012-02-20 09:34:24 +010071 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
Jens Axboe2e1df072012-02-09 11:15:02 +010072
73#define JOB_START_TIMEOUT (5 * 1000)
74
75static void sig_int(int sig)
76{
77 if (threads) {
78 if (is_backend)
79 fio_server_got_signal(sig);
80 else {
81 log_info("\nfio: terminating on signal %d\n", sig);
82 fflush(stdout);
83 exit_value = 128;
84 }
85
86 fio_terminate_threads(TERMINATE_ALL);
87 }
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 if (is_backend) {
105 memset(&act, 0, sizeof(act));
106 act.sa_handler = sig_int;
107 act.sa_flags = SA_RESTART;
108 sigaction(SIGPIPE, &act, NULL);
109 }
110}
111
112/*
113 * Check if we are above the minimum rate given.
114 */
115static int __check_min_rate(struct thread_data *td, struct timeval *now,
116 enum fio_ddir ddir)
117{
118 unsigned long long bytes = 0;
119 unsigned long iops = 0;
120 unsigned long spent;
121 unsigned long rate;
122 unsigned int ratemin = 0;
123 unsigned int rate_iops = 0;
124 unsigned int rate_iops_min = 0;
125
126 assert(ddir_rw(ddir));
127
128 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
129 return 0;
130
131 /*
132 * allow a 2 second settle period in the beginning
133 */
134 if (mtime_since(&td->start, now) < 2000)
135 return 0;
136
137 iops += td->this_io_blocks[ddir];
138 bytes += td->this_io_bytes[ddir];
139 ratemin += td->o.ratemin[ddir];
140 rate_iops += td->o.rate_iops[ddir];
141 rate_iops_min += td->o.rate_iops_min[ddir];
142
143 /*
144 * if rate blocks is set, sample is running
145 */
146 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
147 spent = mtime_since(&td->lastrate[ddir], now);
148 if (spent < td->o.ratecycle)
149 return 0;
150
151 if (td->o.rate[ddir]) {
152 /*
153 * check bandwidth specified rate
154 */
155 if (bytes < td->rate_bytes[ddir]) {
156 log_err("%s: min rate %u not met\n", td->o.name,
157 ratemin);
158 return 1;
159 } else {
160 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
161 if (rate < ratemin ||
162 bytes < td->rate_bytes[ddir]) {
163 log_err("%s: min rate %u not met, got"
164 " %luKB/sec\n", td->o.name,
165 ratemin, rate);
166 return 1;
167 }
168 }
169 } else {
170 /*
171 * checks iops specified rate
172 */
173 if (iops < rate_iops) {
174 log_err("%s: min iops rate %u not met\n",
175 td->o.name, rate_iops);
176 return 1;
177 } else {
178 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
179 if (rate < rate_iops_min ||
180 iops < td->rate_blocks[ddir]) {
181 log_err("%s: min iops rate %u not met,"
182 " got %lu\n", td->o.name,
183 rate_iops_min, rate);
184 }
185 }
186 }
187 }
188
189 td->rate_bytes[ddir] = bytes;
190 td->rate_blocks[ddir] = iops;
191 memcpy(&td->lastrate[ddir], now, sizeof(*now));
192 return 0;
193}
194
195static int check_min_rate(struct thread_data *td, struct timeval *now,
196 unsigned long *bytes_done)
197{
198 int ret = 0;
199
200 if (bytes_done[0])
201 ret |= __check_min_rate(td, now, 0);
202 if (bytes_done[1])
203 ret |= __check_min_rate(td, now, 1);
204
205 return ret;
206}
207
208/*
209 * When job exits, we can cancel the in-flight IO if we are using async
210 * io. Attempt to do so.
211 */
212static void cleanup_pending_aio(struct thread_data *td)
213{
214 struct flist_head *entry, *n;
215 struct io_u *io_u;
216 int r;
217
218 /*
219 * get immediately available events, if any
220 */
221 r = io_u_queued_complete(td, 0, NULL);
222 if (r < 0)
223 return;
224
225 /*
226 * now cancel remaining active events
227 */
228 if (td->io_ops->cancel) {
229 flist_for_each_safe(entry, n, &td->io_u_busylist) {
230 io_u = flist_entry(entry, struct io_u, list);
231
232 /*
233 * if the io_u isn't in flight, then that generally
234 * means someone leaked an io_u. complain but fix
235 * it up, so we don't stall here.
236 */
237 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
238 log_err("fio: non-busy IO on busy list\n");
239 put_io_u(td, io_u);
240 } else {
241 r = td->io_ops->cancel(td, io_u);
242 if (!r)
243 put_io_u(td, io_u);
244 }
245 }
246 }
247
248 if (td->cur_depth)
249 r = io_u_queued_complete(td, td->cur_depth, NULL);
250}
251
252/*
253 * Helper to handle the final sync of a file. Works just like the normal
254 * io path, just does everything sync.
255 */
256static int fio_io_sync(struct thread_data *td, struct fio_file *f)
257{
258 struct io_u *io_u = __get_io_u(td);
259 int ret;
260
261 if (!io_u)
262 return 1;
263
264 io_u->ddir = DDIR_SYNC;
265 io_u->file = f;
266
267 if (td_io_prep(td, io_u)) {
268 put_io_u(td, io_u);
269 return 1;
270 }
271
272requeue:
273 ret = td_io_queue(td, io_u);
274 if (ret < 0) {
275 td_verror(td, io_u->error, "td_io_queue");
276 put_io_u(td, io_u);
277 return 1;
278 } else if (ret == FIO_Q_QUEUED) {
279 if (io_u_queued_complete(td, 1, NULL) < 0)
280 return 1;
281 } else if (ret == FIO_Q_COMPLETED) {
282 if (io_u->error) {
283 td_verror(td, io_u->error, "td_io_queue");
284 return 1;
285 }
286
287 if (io_u_sync_complete(td, io_u, NULL) < 0)
288 return 1;
289 } else if (ret == FIO_Q_BUSY) {
290 if (td_io_commit(td))
291 return 1;
292 goto requeue;
293 }
294
295 return 0;
296}
Jens Axboea3efc912012-02-09 11:25:24 +0100297
Jens Axboe2e1df072012-02-09 11:15:02 +0100298static inline void __update_tv_cache(struct thread_data *td)
299{
300 fio_gettime(&td->tv_cache, NULL);
301}
302
303static inline void update_tv_cache(struct thread_data *td)
304{
305 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
306 __update_tv_cache(td);
307}
308
309static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
310{
311 if (in_ramp_time(td))
312 return 0;
313 if (!td->o.timeout)
314 return 0;
315 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
316 return 1;
317
318 return 0;
319}
320
321static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
322 int *retptr)
323{
324 int ret = *retptr;
325
326 if (ret < 0 || td->error) {
327 int err;
328
329 if (ret < 0)
330 err = -ret;
331 else
332 err = td->error;
333
334 if (!(td->o.continue_on_error & td_error_type(ddir, err)))
335 return 1;
336
337 if (td_non_fatal_error(err)) {
338 /*
339 * Continue with the I/Os in case of
340 * a non fatal error.
341 */
342 update_error_count(td, err);
343 td_clear_error(td);
344 *retptr = 0;
345 return 0;
346 } else if (td->o.fill_device && err == ENOSPC) {
347 /*
348 * We expect to hit this error if
349 * fill_device option is set.
350 */
351 td_clear_error(td);
352 td->terminate = 1;
353 return 1;
354 } else {
355 /*
356 * Stop the I/O in case of a fatal
357 * error.
358 */
359 update_error_count(td, err);
360 return 1;
361 }
362 }
363
364 return 0;
365}
366
Jens Axboe2e1df072012-02-09 11:15:02 +0100367/*
368 * The main verify engine. Runs over the writes we previously submitted,
369 * reads the blocks back in, and checks the crc/md5 of the data.
370 */
371static void do_verify(struct thread_data *td)
372{
373 struct fio_file *f;
374 struct io_u *io_u;
375 int ret, min_events;
376 unsigned int i;
377
378 dprint(FD_VERIFY, "starting loop\n");
379
380 /*
381 * sync io first and invalidate cache, to make sure we really
382 * read from disk.
383 */
384 for_each_file(td, f, i) {
385 if (!fio_file_open(f))
386 continue;
387 if (fio_io_sync(td, f))
388 break;
389 if (file_invalidate_cache(td, f))
390 break;
391 }
392
393 if (td->error)
394 return;
395
396 td_set_runstate(td, TD_VERIFYING);
397
398 io_u = NULL;
399 while (!td->terminate) {
400 int ret2, full;
401
402 update_tv_cache(td);
403
404 if (runtime_exceeded(td, &td->tv_cache)) {
405 __update_tv_cache(td);
406 if (runtime_exceeded(td, &td->tv_cache)) {
407 td->terminate = 1;
408 break;
409 }
410 }
411
Dan Ehrenberg9e684a42012-02-20 11:05:14 +0100412 if (flow_threshold_exceeded(td))
413 continue;
414
Jens Axboe2e1df072012-02-09 11:15:02 +0100415 io_u = __get_io_u(td);
416 if (!io_u)
417 break;
418
419 if (get_next_verify(td, io_u)) {
420 put_io_u(td, io_u);
421 break;
422 }
423
424 if (td_io_prep(td, io_u)) {
425 put_io_u(td, io_u);
426 break;
427 }
428
429 if (td->o.verify_async)
430 io_u->end_io = verify_io_u_async;
431 else
432 io_u->end_io = verify_io_u;
433
434 ret = td_io_queue(td, io_u);
435 switch (ret) {
436 case FIO_Q_COMPLETED:
437 if (io_u->error) {
438 ret = -io_u->error;
439 clear_io_u(td, io_u);
440 } else if (io_u->resid) {
441 int bytes = io_u->xfer_buflen - io_u->resid;
442
443 /*
444 * zero read, fail
445 */
446 if (!bytes) {
447 td_verror(td, EIO, "full resid");
448 put_io_u(td, io_u);
449 break;
450 }
451
452 io_u->xfer_buflen = io_u->resid;
453 io_u->xfer_buf += bytes;
454 io_u->offset += bytes;
455
456 if (ddir_rw(io_u->ddir))
457 td->ts.short_io_u[io_u->ddir]++;
458
459 f = io_u->file;
460 if (io_u->offset == f->real_file_size)
461 goto sync_done;
462
463 requeue_io_u(td, &io_u);
464 } else {
465sync_done:
466 ret = io_u_sync_complete(td, io_u, NULL);
467 if (ret < 0)
468 break;
469 }
470 continue;
471 case FIO_Q_QUEUED:
472 break;
473 case FIO_Q_BUSY:
474 requeue_io_u(td, &io_u);
475 ret2 = td_io_commit(td);
476 if (ret2 < 0)
477 ret = ret2;
478 break;
479 default:
480 assert(ret < 0);
481 td_verror(td, -ret, "td_io_queue");
482 break;
483 }
484
485 if (break_on_this_error(td, io_u->ddir, &ret))
486 break;
487
488 /*
489 * if we can queue more, do so. but check if there are
490 * completed io_u's first. Note that we can get BUSY even
491 * without IO queued, if the system is resource starved.
492 */
493 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
494 if (full || !td->o.iodepth_batch_complete) {
495 min_events = min(td->o.iodepth_batch_complete,
496 td->cur_depth);
497 if (full && !min_events && td->o.iodepth_batch_complete != 0)
498 min_events = 1;
499
500 do {
501 /*
502 * Reap required number of io units, if any,
503 * and do the verification on them through
504 * the callback handler
505 */
506 if (io_u_queued_complete(td, min_events, NULL) < 0) {
507 ret = -1;
508 break;
509 }
510 } while (full && (td->cur_depth > td->o.iodepth_low));
511 }
512 if (ret < 0)
513 break;
514 }
515
516 if (!td->error) {
517 min_events = td->cur_depth;
518
519 if (min_events)
520 ret = io_u_queued_complete(td, min_events, NULL);
521 } else
522 cleanup_pending_aio(td);
523
524 td_set_runstate(td, TD_RUNNING);
525
526 dprint(FD_VERIFY, "exiting loop\n");
527}
528
529/*
530 * Main IO worker function. It retrieves io_u's to process and queues
531 * and reaps them, checking for rate and errors along the way.
532 */
533static void do_io(struct thread_data *td)
534{
535 unsigned int i;
536 int ret = 0;
537
538 if (in_ramp_time(td))
539 td_set_runstate(td, TD_RAMP);
540 else
541 td_set_runstate(td, TD_RUNNING);
542
543 while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
544 (!flist_empty(&td->trim_list)) ||
545 ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) ) {
546 struct timeval comp_time;
547 unsigned long bytes_done[2] = { 0, 0 };
548 int min_evts = 0;
549 struct io_u *io_u;
550 int ret2, full;
551 enum fio_ddir ddir;
552
553 if (td->terminate)
554 break;
555
556 update_tv_cache(td);
557
558 if (runtime_exceeded(td, &td->tv_cache)) {
559 __update_tv_cache(td);
560 if (runtime_exceeded(td, &td->tv_cache)) {
561 td->terminate = 1;
562 break;
563 }
564 }
565
Dan Ehrenberg9e684a42012-02-20 11:05:14 +0100566 if (flow_threshold_exceeded(td))
567 continue;
568
Jens Axboe2e1df072012-02-09 11:15:02 +0100569 io_u = get_io_u(td);
570 if (!io_u)
571 break;
572
573 ddir = io_u->ddir;
574
575 /*
576 * Add verification end_io handler, if asked to verify
577 * a previously written file.
578 */
579 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
580 !td_rw(td)) {
581 if (td->o.verify_async)
582 io_u->end_io = verify_io_u_async;
583 else
584 io_u->end_io = verify_io_u;
585 td_set_runstate(td, TD_VERIFYING);
586 } else if (in_ramp_time(td))
587 td_set_runstate(td, TD_RAMP);
588 else
589 td_set_runstate(td, TD_RUNNING);
590
591 ret = td_io_queue(td, io_u);
592 switch (ret) {
593 case FIO_Q_COMPLETED:
594 if (io_u->error) {
595 ret = -io_u->error;
596 clear_io_u(td, io_u);
597 } else if (io_u->resid) {
598 int bytes = io_u->xfer_buflen - io_u->resid;
599 struct fio_file *f = io_u->file;
600
601 /*
602 * zero read, fail
603 */
604 if (!bytes) {
605 td_verror(td, EIO, "full resid");
606 put_io_u(td, io_u);
607 break;
608 }
609
610 io_u->xfer_buflen = io_u->resid;
611 io_u->xfer_buf += bytes;
612 io_u->offset += bytes;
613
614 if (ddir_rw(io_u->ddir))
615 td->ts.short_io_u[io_u->ddir]++;
616
617 if (io_u->offset == f->real_file_size)
618 goto sync_done;
619
620 requeue_io_u(td, &io_u);
621 } else {
622sync_done:
623 if (__should_check_rate(td, 0) ||
624 __should_check_rate(td, 1))
625 fio_gettime(&comp_time, NULL);
626
627 ret = io_u_sync_complete(td, io_u, bytes_done);
628 if (ret < 0)
629 break;
630 }
631 break;
632 case FIO_Q_QUEUED:
633 /*
634 * if the engine doesn't have a commit hook,
635 * the io_u is really queued. if it does have such
636 * a hook, it has to call io_u_queued() itself.
637 */
638 if (td->io_ops->commit == NULL)
639 io_u_queued(td, io_u);
640 break;
641 case FIO_Q_BUSY:
642 requeue_io_u(td, &io_u);
643 ret2 = td_io_commit(td);
644 if (ret2 < 0)
645 ret = ret2;
646 break;
647 default:
648 assert(ret < 0);
649 put_io_u(td, io_u);
650 break;
651 }
652
653 if (break_on_this_error(td, ddir, &ret))
654 break;
655
656 /*
657 * See if we need to complete some commands. Note that we
658 * can get BUSY even without IO queued, if the system is
659 * resource starved.
660 */
661 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
662 if (full || !td->o.iodepth_batch_complete) {
663 min_evts = min(td->o.iodepth_batch_complete,
664 td->cur_depth);
665 if (full && !min_evts && td->o.iodepth_batch_complete != 0)
666 min_evts = 1;
667
668 if (__should_check_rate(td, 0) ||
669 __should_check_rate(td, 1))
670 fio_gettime(&comp_time, NULL);
671
672 do {
673 ret = io_u_queued_complete(td, min_evts, bytes_done);
674 if (ret < 0)
675 break;
676
677 } while (full && (td->cur_depth > td->o.iodepth_low));
678 }
679
680 if (ret < 0)
681 break;
682 if (!(bytes_done[0] + bytes_done[1]))
683 continue;
684
685 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
686 if (check_min_rate(td, &comp_time, bytes_done)) {
687 if (exitall_on_terminate)
688 fio_terminate_threads(td->groupid);
689 td_verror(td, EIO, "check_min_rate");
690 break;
691 }
692 }
693
694 if (td->o.thinktime) {
695 unsigned long long b;
696
697 b = td->io_blocks[0] + td->io_blocks[1];
698 if (!(b % td->o.thinktime_blocks)) {
699 int left;
700
701 if (td->o.thinktime_spin)
702 usec_spin(td->o.thinktime_spin);
703
704 left = td->o.thinktime - td->o.thinktime_spin;
705 if (left)
706 usec_sleep(td, left);
707 }
708 }
709 }
710
711 if (td->trim_entries)
712 log_err("fio: %d trim entries leaked?\n", td->trim_entries);
713
714 if (td->o.fill_device && td->error == ENOSPC) {
715 td->error = 0;
716 td->terminate = 1;
717 }
718 if (!td->error) {
719 struct fio_file *f;
720
721 i = td->cur_depth;
722 if (i) {
723 ret = io_u_queued_complete(td, i, NULL);
724 if (td->o.fill_device && td->error == ENOSPC)
725 td->error = 0;
726 }
727
728 if (should_fsync(td) && td->o.end_fsync) {
729 td_set_runstate(td, TD_FSYNCING);
730
731 for_each_file(td, f, i) {
732 if (!fio_file_open(f))
733 continue;
734 fio_io_sync(td, f);
735 }
736 }
737 } else
738 cleanup_pending_aio(td);
739
740 /*
741 * stop job if we failed doing any IO
742 */
743 if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
744 td->done = 1;
745}
746
747static void cleanup_io_u(struct thread_data *td)
748{
749 struct flist_head *entry, *n;
750 struct io_u *io_u;
751
752 flist_for_each_safe(entry, n, &td->io_u_freelist) {
753 io_u = flist_entry(entry, struct io_u, list);
754
755 flist_del(&io_u->list);
756 fio_memfree(io_u, sizeof(*io_u));
757 }
758
759 free_io_mem(td);
760}
761
762static int init_io_u(struct thread_data *td)
763{
764 struct io_u *io_u;
765 unsigned int max_bs;
766 int cl_align, i, max_units;
767 char *p;
768
769 max_units = td->o.iodepth;
770 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
771 td->orig_buffer_size = (unsigned long long) max_bs
772 * (unsigned long long) max_units;
773
774 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
775 unsigned long bs;
776
777 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
778 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
779 }
780
781 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
782 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
783 return 1;
784 }
785
786 if (allocate_io_mem(td))
787 return 1;
788
789 if (td->o.odirect || td->o.mem_align ||
790 (td->io_ops->flags & FIO_RAWIO))
791 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
792 else
793 p = td->orig_buffer;
794
795 cl_align = os_cache_line_size();
796
797 for (i = 0; i < max_units; i++) {
798 void *ptr;
799
800 if (td->terminate)
801 return 1;
802
803 ptr = fio_memalign(cl_align, sizeof(*io_u));
804 if (!ptr) {
805 log_err("fio: unable to allocate aligned memory\n");
806 break;
807 }
808
809 io_u = ptr;
810 memset(io_u, 0, sizeof(*io_u));
811 INIT_FLIST_HEAD(&io_u->list);
812 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
813
814 if (!(td->io_ops->flags & FIO_NOIO)) {
815 io_u->buf = p;
816 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
817
818 if (td_write(td))
819 io_u_fill_buffer(td, io_u, max_bs);
820 if (td_write(td) && td->o.verify_pattern_bytes) {
821 /*
822 * Fill the buffer with the pattern if we are
823 * going to be doing writes.
824 */
825 fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
826 }
827 }
828
829 io_u->index = i;
830 io_u->flags = IO_U_F_FREE;
831 flist_add(&io_u->list, &td->io_u_freelist);
832 p += max_bs;
833 }
834
835 return 0;
836}
837
838static int switch_ioscheduler(struct thread_data *td)
839{
840 char tmp[256], tmp2[128];
841 FILE *f;
842 int ret;
843
844 if (td->io_ops->flags & FIO_DISKLESSIO)
845 return 0;
846
847 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
848
849 f = fopen(tmp, "r+");
850 if (!f) {
851 if (errno == ENOENT) {
852 log_err("fio: os or kernel doesn't support IO scheduler"
853 " switching\n");
854 return 0;
855 }
856 td_verror(td, errno, "fopen iosched");
857 return 1;
858 }
859
860 /*
861 * Set io scheduler.
862 */
863 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
864 if (ferror(f) || ret != 1) {
865 td_verror(td, errno, "fwrite");
866 fclose(f);
867 return 1;
868 }
869
870 rewind(f);
871
872 /*
873 * Read back and check that the selected scheduler is now the default.
874 */
875 ret = fread(tmp, 1, sizeof(tmp), f);
876 if (ferror(f) || ret < 0) {
877 td_verror(td, errno, "fread");
878 fclose(f);
879 return 1;
880 }
881
882 sprintf(tmp2, "[%s]", td->o.ioscheduler);
883 if (!strstr(tmp, tmp2)) {
884 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
885 td_verror(td, EINVAL, "iosched_switch");
886 fclose(f);
887 return 1;
888 }
889
890 fclose(f);
891 return 0;
892}
893
894static int keep_running(struct thread_data *td)
895{
896 unsigned long long io_done;
897
898 if (td->done)
899 return 0;
900 if (td->o.time_based)
901 return 1;
902 if (td->o.loops) {
903 td->o.loops--;
904 return 1;
905 }
906
907 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
908 + td->io_skip_bytes;
909 if (io_done < td->o.size)
910 return 1;
911
912 return 0;
913}
914
915static int exec_string(const char *string)
916{
917 int ret, newlen = strlen(string) + 1 + 8;
918 char *str;
919
920 str = malloc(newlen);
921 sprintf(str, "sh -c %s", string);
922
923 ret = system(str);
924 if (ret == -1)
925 log_err("fio: exec of cmd <%s> failed\n", str);
926
927 free(str);
928 return ret;
929}
930
931/*
932 * Entry point for the thread based jobs. The process based jobs end up
933 * here as well, after a little setup.
934 */
935static void *thread_main(void *data)
936{
937 unsigned long long elapsed;
938 struct thread_data *td = data;
939 pthread_condattr_t attr;
940 int clear_state;
941
942 if (!td->o.use_thread) {
943 setsid();
944 td->pid = getpid();
945 } else
946 td->pid = gettid();
947
948 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
949
950 INIT_FLIST_HEAD(&td->io_u_freelist);
951 INIT_FLIST_HEAD(&td->io_u_busylist);
952 INIT_FLIST_HEAD(&td->io_u_requeues);
953 INIT_FLIST_HEAD(&td->io_log_list);
954 INIT_FLIST_HEAD(&td->io_hist_list);
955 INIT_FLIST_HEAD(&td->verify_list);
956 INIT_FLIST_HEAD(&td->trim_list);
957 pthread_mutex_init(&td->io_u_lock, NULL);
958 td->io_hist_tree = RB_ROOT;
959
960 pthread_condattr_init(&attr);
961 pthread_cond_init(&td->verify_cond, &attr);
962 pthread_cond_init(&td->free_cond, &attr);
963
964 td_set_runstate(td, TD_INITIALIZED);
965 dprint(FD_MUTEX, "up startup_mutex\n");
966 fio_mutex_up(startup_mutex);
967 dprint(FD_MUTEX, "wait on td->mutex\n");
968 fio_mutex_down(td->mutex);
969 dprint(FD_MUTEX, "done waiting on td->mutex\n");
970
971 /*
972 * the ->mutex mutex is now no longer used, close it to avoid
973 * eating a file descriptor
974 */
975 fio_mutex_remove(td->mutex);
976
977 /*
978 * A new gid requires privilege, so we need to do this before setting
979 * the uid.
980 */
981 if (td->o.gid != -1U && setgid(td->o.gid)) {
982 td_verror(td, errno, "setgid");
983 goto err;
984 }
985 if (td->o.uid != -1U && setuid(td->o.uid)) {
986 td_verror(td, errno, "setuid");
987 goto err;
988 }
989
990 /*
991 * If we have a gettimeofday() thread, make sure we exclude that
992 * thread from this job
993 */
994 if (td->o.gtod_cpu)
995 fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
996
997 /*
998 * Set affinity first, in case it has an impact on the memory
999 * allocations.
1000 */
1001 if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
1002 td_verror(td, errno, "cpu_set_affinity");
1003 goto err;
1004 }
1005
1006 /*
1007 * May alter parameters that init_io_u() will use, so we need to
1008 * do this first.
1009 */
1010 if (init_iolog(td))
1011 goto err;
1012
1013 if (init_io_u(td))
1014 goto err;
1015
1016 if (td->o.verify_async && verify_async_init(td))
1017 goto err;
1018
1019 if (td->ioprio_set) {
1020 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1021 td_verror(td, errno, "ioprio_set");
1022 goto err;
1023 }
1024 }
1025
1026 if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
1027 goto err;
1028
Bruce Cran649c10c2012-02-20 07:59:06 +01001029 errno = 0;
1030 if (nice(td->o.nice) == -1 && errno != 0) {
Jens Axboe2e1df072012-02-09 11:15:02 +01001031 td_verror(td, errno, "nice");
1032 goto err;
1033 }
1034
1035 if (td->o.ioscheduler && switch_ioscheduler(td))
1036 goto err;
1037
1038 if (!td->o.create_serialize && setup_files(td))
1039 goto err;
1040
1041 if (td_io_init(td))
1042 goto err;
1043
1044 if (init_random_map(td))
1045 goto err;
1046
1047 if (td->o.exec_prerun) {
1048 if (exec_string(td->o.exec_prerun))
1049 goto err;
1050 }
1051
1052 if (td->o.pre_read) {
1053 if (pre_read_files(td) < 0)
1054 goto err;
1055 }
1056
1057 fio_gettime(&td->epoch, NULL);
1058 getrusage(RUSAGE_SELF, &td->ru_start);
1059
1060 clear_state = 0;
1061 while (keep_running(td)) {
1062 fio_gettime(&td->start, NULL);
1063 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1064 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1065 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1066
1067 if (td->o.ratemin[0] || td->o.ratemin[1]) {
1068 memcpy(&td->lastrate[0], &td->bw_sample_time,
1069 sizeof(td->bw_sample_time));
1070 memcpy(&td->lastrate[1], &td->bw_sample_time,
1071 sizeof(td->bw_sample_time));
1072 }
1073
1074 if (clear_state)
1075 clear_io_state(td);
1076
1077 prune_io_piece_log(td);
1078
1079 do_io(td);
1080
1081 clear_state = 1;
1082
1083 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1084 elapsed = utime_since_now(&td->start);
1085 td->ts.runtime[DDIR_READ] += elapsed;
1086 }
1087 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1088 elapsed = utime_since_now(&td->start);
1089 td->ts.runtime[DDIR_WRITE] += elapsed;
1090 }
1091
1092 if (td->error || td->terminate)
1093 break;
1094
1095 if (!td->o.do_verify ||
1096 td->o.verify == VERIFY_NONE ||
1097 (td->io_ops->flags & FIO_UNIDIR))
1098 continue;
1099
1100 clear_io_state(td);
1101
1102 fio_gettime(&td->start, NULL);
1103
1104 do_verify(td);
1105
1106 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1107
1108 if (td->error || td->terminate)
1109 break;
1110 }
1111
1112 update_rusage_stat(td);
1113 td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
1114 td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
1115 td->ts.total_run_time = mtime_since_now(&td->epoch);
1116 td->ts.io_bytes[0] = td->io_bytes[0];
1117 td->ts.io_bytes[1] = td->io_bytes[1];
1118
1119 fio_mutex_down(writeout_mutex);
1120 if (td->bw_log) {
1121 if (td->o.bw_log_file) {
1122 finish_log_named(td, td->bw_log,
1123 td->o.bw_log_file, "bw");
1124 } else
1125 finish_log(td, td->bw_log, "bw");
1126 }
1127 if (td->lat_log) {
1128 if (td->o.lat_log_file) {
1129 finish_log_named(td, td->lat_log,
1130 td->o.lat_log_file, "lat");
1131 } else
1132 finish_log(td, td->lat_log, "lat");
1133 }
1134 if (td->slat_log) {
1135 if (td->o.lat_log_file) {
1136 finish_log_named(td, td->slat_log,
1137 td->o.lat_log_file, "slat");
1138 } else
1139 finish_log(td, td->slat_log, "slat");
1140 }
1141 if (td->clat_log) {
1142 if (td->o.lat_log_file) {
1143 finish_log_named(td, td->clat_log,
1144 td->o.lat_log_file, "clat");
1145 } else
1146 finish_log(td, td->clat_log, "clat");
1147 }
1148 if (td->iops_log) {
1149 if (td->o.iops_log_file) {
1150 finish_log_named(td, td->iops_log,
1151 td->o.iops_log_file, "iops");
1152 } else
1153 finish_log(td, td->iops_log, "iops");
1154 }
1155
1156 fio_mutex_up(writeout_mutex);
1157 if (td->o.exec_postrun)
1158 exec_string(td->o.exec_postrun);
1159
1160 if (exitall_on_terminate)
1161 fio_terminate_threads(td->groupid);
1162
1163err:
1164 if (td->error)
1165 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1166 td->verror);
1167
1168 if (td->o.verify_async)
1169 verify_async_exit(td);
1170
1171 close_and_free_files(td);
1172 close_ioengine(td);
1173 cleanup_io_u(td);
1174 cgroup_shutdown(td, &cgroup_mnt);
1175
1176 if (td->o.cpumask_set) {
1177 int ret = fio_cpuset_exit(&td->o.cpumask);
1178
1179 td_verror(td, ret, "fio_cpuset_exit");
1180 }
1181
1182 /*
1183 * do this very late, it will log file closing as well
1184 */
1185 if (td->o.write_iolog_file)
1186 write_iolog_close(td);
1187
1188 td_set_runstate(td, TD_EXITED);
Bruce Crane43606c2012-02-20 09:34:24 +01001189 return (void *) (uintptr_t) td->error;
Jens Axboe2e1df072012-02-09 11:15:02 +01001190}
1191
1192
1193/*
1194 * We cannot pass the td data into a forked process, so attach the td and
1195 * pass it to the thread worker.
1196 */
1197static int fork_main(int shmid, int offset)
1198{
1199 struct thread_data *td;
1200 void *data, *ret;
1201
1202#ifndef __hpux
1203 data = shmat(shmid, NULL, 0);
1204 if (data == (void *) -1) {
1205 int __err = errno;
1206
1207 perror("shmat");
1208 return __err;
1209 }
1210#else
1211 /*
1212 * HP-UX inherits shm mappings?
1213 */
1214 data = threads;
1215#endif
1216
1217 td = data + offset * sizeof(struct thread_data);
1218 ret = thread_main(td);
1219 shmdt(data);
Bruce Crane43606c2012-02-20 09:34:24 +01001220 return (int) (uintptr_t) ret;
Jens Axboe2e1df072012-02-09 11:15:02 +01001221}
1222
1223/*
1224 * Run over the job map and reap the threads that have exited, if any.
1225 */
1226static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1227 unsigned int *m_rate)
1228{
1229 struct thread_data *td;
1230 unsigned int cputhreads, realthreads, pending;
1231 int i, status, ret;
1232
1233 /*
1234 * reap exited threads (TD_EXITED -> TD_REAPED)
1235 */
1236 realthreads = pending = cputhreads = 0;
1237 for_each_td(td, i) {
1238 int flags = 0;
1239
1240 /*
1241 * ->io_ops is NULL for a thread that has closed its
1242 * io engine
1243 */
1244 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1245 cputhreads++;
1246 else
1247 realthreads++;
1248
1249 if (!td->pid) {
1250 pending++;
1251 continue;
1252 }
1253 if (td->runstate == TD_REAPED)
1254 continue;
1255 if (td->o.use_thread) {
1256 if (td->runstate == TD_EXITED) {
1257 td_set_runstate(td, TD_REAPED);
1258 goto reaped;
1259 }
1260 continue;
1261 }
1262
1263 flags = WNOHANG;
1264 if (td->runstate == TD_EXITED)
1265 flags = 0;
1266
1267 /*
1268 * check if someone quit or got killed in an unusual way
1269 */
1270 ret = waitpid(td->pid, &status, flags);
1271 if (ret < 0) {
1272 if (errno == ECHILD) {
1273 log_err("fio: pid=%d disappeared %d\n",
1274 (int) td->pid, td->runstate);
1275 td_set_runstate(td, TD_REAPED);
1276 goto reaped;
1277 }
1278 perror("waitpid");
1279 } else if (ret == td->pid) {
1280 if (WIFSIGNALED(status)) {
1281 int sig = WTERMSIG(status);
1282
1283 if (sig != SIGTERM)
1284 log_err("fio: pid=%d, got signal=%d\n",
1285 (int) td->pid, sig);
1286 td_set_runstate(td, TD_REAPED);
1287 goto reaped;
1288 }
1289 if (WIFEXITED(status)) {
1290 if (WEXITSTATUS(status) && !td->error)
1291 td->error = WEXITSTATUS(status);
1292
1293 td_set_runstate(td, TD_REAPED);
1294 goto reaped;
1295 }
1296 }
1297
1298 /*
1299 * thread is not dead, continue
1300 */
1301 pending++;
1302 continue;
1303reaped:
1304 (*nr_running)--;
1305 (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
1306 (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
1307 if (!td->pid)
1308 pending--;
1309
1310 if (td->error)
1311 exit_value++;
1312
1313 done_secs += mtime_since_now(&td->epoch) / 1000;
1314 }
1315
1316 if (*nr_running == cputhreads && !pending && realthreads)
1317 fio_terminate_threads(TERMINATE_ALL);
1318}
1319
Jens Axboe2e1df072012-02-09 11:15:02 +01001320/*
1321 * Main function for kicking off and reaping jobs, as needed.
1322 */
1323static void run_threads(void)
1324{
1325 struct thread_data *td;
1326 unsigned long spent;
1327 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
1328
1329 if (fio_pin_memory())
1330 return;
1331
1332 if (fio_gtod_offload && fio_start_gtod_thread())
1333 return;
1334
1335 set_sig_handlers();
1336
1337 if (!terse_output) {
1338 log_info("Starting ");
1339 if (nr_thread)
1340 log_info("%d thread%s", nr_thread,
1341 nr_thread > 1 ? "s" : "");
1342 if (nr_process) {
1343 if (nr_thread)
1344 log_info(" and ");
1345 log_info("%d process%s", nr_process,
1346 nr_process > 1 ? "es" : "");
1347 }
1348 log_info("\n");
1349 fflush(stdout);
1350 }
1351
1352 todo = thread_number;
1353 nr_running = 0;
1354 nr_started = 0;
1355 m_rate = t_rate = 0;
1356
1357 for_each_td(td, i) {
1358 print_status_init(td->thread_number - 1);
1359
1360 if (!td->o.create_serialize)
1361 continue;
1362
1363 /*
1364 * do file setup here so it happens sequentially,
1365 * we don't want X number of threads getting their
1366 * client data interspersed on disk
1367 */
1368 if (setup_files(td)) {
1369 exit_value++;
1370 if (td->error)
1371 log_err("fio: pid=%d, err=%d/%s\n",
1372 (int) td->pid, td->error, td->verror);
1373 td_set_runstate(td, TD_REAPED);
1374 todo--;
1375 } else {
1376 struct fio_file *f;
1377 unsigned int j;
1378
1379 /*
1380 * for sharing to work, each job must always open
1381 * its own files. so close them, if we opened them
1382 * for creation
1383 */
1384 for_each_file(td, f, j) {
1385 if (fio_file_open(f))
1386 td_io_close_file(td, f);
1387 }
1388 }
1389 }
1390
1391 set_genesis_time();
1392
1393 while (todo) {
1394 struct thread_data *map[REAL_MAX_JOBS];
1395 struct timeval this_start;
1396 int this_jobs = 0, left;
1397
1398 /*
1399 * create threads (TD_NOT_CREATED -> TD_CREATED)
1400 */
1401 for_each_td(td, i) {
1402 if (td->runstate != TD_NOT_CREATED)
1403 continue;
1404
1405 /*
1406 * never got a chance to start, killed by other
1407 * thread for some reason
1408 */
1409 if (td->terminate) {
1410 todo--;
1411 continue;
1412 }
1413
1414 if (td->o.start_delay) {
1415 spent = mtime_since_genesis();
1416
1417 if (td->o.start_delay * 1000 > spent)
1418 continue;
1419 }
1420
1421 if (td->o.stonewall && (nr_started || nr_running)) {
1422 dprint(FD_PROCESS, "%s: stonewall wait\n",
1423 td->o.name);
1424 break;
1425 }
1426
1427 init_disk_util(td);
1428
1429 /*
1430 * Set state to created. Thread will transition
1431 * to TD_INITIALIZED when it's done setting up.
1432 */
1433 td_set_runstate(td, TD_CREATED);
1434 map[this_jobs++] = td;
1435 nr_started++;
1436
1437 if (td->o.use_thread) {
1438 int ret;
1439
1440 dprint(FD_PROCESS, "will pthread_create\n");
1441 ret = pthread_create(&td->thread, NULL,
1442 thread_main, td);
1443 if (ret) {
1444 log_err("pthread_create: %s\n",
1445 strerror(ret));
1446 nr_started--;
1447 break;
1448 }
1449 ret = pthread_detach(td->thread);
1450 if (ret)
1451 log_err("pthread_detach: %s",
1452 strerror(ret));
1453 } else {
1454 pid_t pid;
1455 dprint(FD_PROCESS, "will fork\n");
1456 pid = fork();
1457 if (!pid) {
1458 int ret = fork_main(shm_id, i);
1459
1460 _exit(ret);
1461 } else if (i == fio_debug_jobno)
1462 *fio_debug_jobp = pid;
1463 }
1464 dprint(FD_MUTEX, "wait on startup_mutex\n");
1465 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1466 log_err("fio: job startup hung? exiting.\n");
1467 fio_terminate_threads(TERMINATE_ALL);
1468 fio_abort = 1;
1469 nr_started--;
1470 break;
1471 }
1472 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1473 }
1474
1475 /*
1476 * Wait for the started threads to transition to
1477 * TD_INITIALIZED.
1478 */
1479 fio_gettime(&this_start, NULL);
1480 left = this_jobs;
1481 while (left && !fio_abort) {
1482 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1483 break;
1484
1485 usleep(100000);
1486
1487 for (i = 0; i < this_jobs; i++) {
1488 td = map[i];
1489 if (!td)
1490 continue;
1491 if (td->runstate == TD_INITIALIZED) {
1492 map[i] = NULL;
1493 left--;
1494 } else if (td->runstate >= TD_EXITED) {
1495 map[i] = NULL;
1496 left--;
1497 todo--;
1498 nr_running++; /* work-around... */
1499 }
1500 }
1501 }
1502
1503 if (left) {
Jens Axboe4e87c372012-02-15 14:27:08 +01001504 log_err("fio: %d job%s failed to start\n", left,
1505 left > 1 ? "s" : "");
Jens Axboe2e1df072012-02-09 11:15:02 +01001506 for (i = 0; i < this_jobs; i++) {
1507 td = map[i];
1508 if (!td)
1509 continue;
1510 kill(td->pid, SIGTERM);
1511 }
1512 break;
1513 }
1514
1515 /*
1516 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1517 */
1518 for_each_td(td, i) {
1519 if (td->runstate != TD_INITIALIZED)
1520 continue;
1521
1522 if (in_ramp_time(td))
1523 td_set_runstate(td, TD_RAMP);
1524 else
1525 td_set_runstate(td, TD_RUNNING);
1526 nr_running++;
1527 nr_started--;
1528 m_rate += td->o.ratemin[0] + td->o.ratemin[1];
1529 t_rate += td->o.rate[0] + td->o.rate[1];
1530 todo--;
1531 fio_mutex_up(td->mutex);
1532 }
1533
1534 reap_threads(&nr_running, &t_rate, &m_rate);
1535
1536 if (todo) {
1537 if (is_backend)
1538 fio_server_idle_loop();
1539 else
1540 usleep(100000);
1541 }
1542 }
1543
1544 while (nr_running) {
1545 reap_threads(&nr_running, &t_rate, &m_rate);
1546
1547 if (is_backend)
1548 fio_server_idle_loop();
1549 else
1550 usleep(10000);
1551 }
1552
1553 update_io_ticks();
1554 fio_unpin_memory();
1555}
1556
1557static void *disk_thread_main(void *data)
1558{
1559 fio_mutex_up(startup_mutex);
1560
1561 while (threads) {
1562 usleep(DISK_UTIL_MSEC * 1000);
1563 if (!threads)
1564 break;
1565 update_io_ticks();
1566
1567 if (!is_backend)
1568 print_thread_status();
1569 }
1570
1571 return NULL;
1572}
1573
1574static int create_disk_util_thread(void)
1575{
1576 int ret;
1577
1578 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1579 if (ret) {
1580 log_err("Can't create disk util thread: %s\n", strerror(ret));
1581 return 1;
1582 }
1583
1584 ret = pthread_detach(disk_util_thread);
1585 if (ret) {
1586 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1587 return 1;
1588 }
1589
1590 dprint(FD_MUTEX, "wait on startup_mutex\n");
1591 fio_mutex_down(startup_mutex);
1592 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1593 return 0;
1594}
1595
Jens Axboe2e1df072012-02-09 11:15:02 +01001596int fio_backend(void)
1597{
1598 struct thread_data *td;
1599 int i;
1600
1601 if (exec_profile) {
1602 if (load_profile(exec_profile))
1603 return 1;
1604 free(exec_profile);
1605 exec_profile = NULL;
1606 }
1607 if (!thread_number)
1608 return 0;
1609
1610 if (write_bw_log) {
1611 setup_log(&agg_io_log[DDIR_READ], 0);
1612 setup_log(&agg_io_log[DDIR_WRITE], 0);
1613 }
1614
1615 startup_mutex = fio_mutex_init(0);
1616 if (startup_mutex == NULL)
1617 return 1;
1618 writeout_mutex = fio_mutex_init(1);
1619 if (writeout_mutex == NULL)
1620 return 1;
1621
1622 set_genesis_time();
1623 create_disk_util_thread();
1624
1625 cgroup_list = smalloc(sizeof(*cgroup_list));
1626 INIT_FLIST_HEAD(cgroup_list);
1627
1628 run_threads();
1629
1630 if (!fio_abort) {
1631 show_run_stats();
1632 if (write_bw_log) {
1633 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
1634 __finish_log(agg_io_log[DDIR_WRITE],
1635 "agg-write_bw.log");
1636 }
1637 }
1638
1639 for_each_td(td, i)
1640 fio_options_free(td);
1641
1642 cgroup_kill(cgroup_list);
1643 sfree(cgroup_list);
1644 sfree(cgroup_mnt);
1645
1646 fio_mutex_remove(startup_mutex);
1647 fio_mutex_remove(writeout_mutex);
1648 return exit_value;
1649}