Get rid of __ prefix for internal frand state
[fio.git] / backend.c
CommitLineData
2e1df07d
JA
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., 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>
e43606c2 33#include <inttypes.h>
2e1df07d
JA
34#include <sys/stat.h>
35#include <sys/wait.h>
36#include <sys/ipc.h>
2e1df07d
JA
37#include <sys/mman.h>
38
39#include "fio.h"
a5e0ee11
O
40#ifndef FIO_NO_HAVE_SHM_H
41#include <sys/shm.h>
42#endif
2e1df07d
JA
43#include "hash.h"
44#include "smalloc.h"
45#include "verify.h"
46#include "trim.h"
47#include "diskutil.h"
48#include "cgroup.h"
49#include "profile.h"
50#include "lib/rand.h"
51#include "memalign.h"
52#include "server.h"
44404c5a 53#include "lib/getrusage.h"
f2a2ce0e 54#include "idletime.h"
002fe734 55#include "err.h"
78d55e72 56#include "lib/tp.h"
2e1df07d 57
5ddc6707
JA
58static pthread_t helper_thread;
59static pthread_mutex_t helper_lock;
60pthread_cond_t helper_cond;
61int helper_do_stat = 0;
8aab824f 62
2e1df07d 63static struct fio_mutex *startup_mutex;
2e1df07d
JA
64static struct flist_head *cgroup_list;
65static char *cgroup_mnt;
66static int exit_value;
67static volatile int fio_abort;
3a5f6bde
JA
68static unsigned int nr_process = 0;
69static unsigned int nr_thread = 0;
2e1df07d 70
6eaf09d6 71struct io_log *agg_io_log[DDIR_RWDIR_CNT];
2e1df07d 72
a3efc919
JA
73int groupid = 0;
74unsigned int thread_number = 0;
108fea77 75unsigned int stat_number = 0;
a3efc919
JA
76int shm_id = 0;
77int temp_stall_ts;
78unsigned long done_secs = 0;
5ddc6707 79volatile int helper_exit = 0;
a3efc919 80
2e1df07d 81#define PAGE_ALIGN(buf) \
e43606c2 82 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
2e1df07d
JA
83
84#define JOB_START_TIMEOUT (5 * 1000)
85
86static void sig_int(int sig)
87{
88 if (threads) {
89 if (is_backend)
90 fio_server_got_signal(sig);
91 else {
92 log_info("\nfio: terminating on signal %d\n", sig);
e411c301 93 log_info_flush();
2e1df07d
JA
94 exit_value = 128;
95 }
96
97 fio_terminate_threads(TERMINATE_ALL);
98 }
99}
100
4c6d91e8
JA
101static void sig_show_status(int sig)
102{
103 show_running_run_stats();
104}
105
2e1df07d
JA
106static void set_sig_handlers(void)
107{
108 struct sigaction act;
109
110 memset(&act, 0, sizeof(act));
111 act.sa_handler = sig_int;
112 act.sa_flags = SA_RESTART;
113 sigaction(SIGINT, &act, NULL);
114
115 memset(&act, 0, sizeof(act));
116 act.sa_handler = sig_int;
117 act.sa_flags = SA_RESTART;
118 sigaction(SIGTERM, &act, NULL);
119
2f694507
BC
120/* Windows uses SIGBREAK as a quit signal from other applications */
121#ifdef WIN32
122 memset(&act, 0, sizeof(act));
123 act.sa_handler = sig_int;
124 act.sa_flags = SA_RESTART;
125 sigaction(SIGBREAK, &act, NULL);
126#endif
127
4c6d91e8
JA
128 memset(&act, 0, sizeof(act));
129 act.sa_handler = sig_show_status;
130 act.sa_flags = SA_RESTART;
131 sigaction(SIGUSR1, &act, NULL);
132
2e1df07d
JA
133 if (is_backend) {
134 memset(&act, 0, sizeof(act));
135 act.sa_handler = sig_int;
136 act.sa_flags = SA_RESTART;
137 sigaction(SIGPIPE, &act, NULL);
138 }
139}
140
141/*
142 * Check if we are above the minimum rate given.
143 */
144static int __check_min_rate(struct thread_data *td, struct timeval *now,
145 enum fio_ddir ddir)
146{
147 unsigned long long bytes = 0;
148 unsigned long iops = 0;
149 unsigned long spent;
150 unsigned long rate;
151 unsigned int ratemin = 0;
152 unsigned int rate_iops = 0;
153 unsigned int rate_iops_min = 0;
154
155 assert(ddir_rw(ddir));
156
157 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
158 return 0;
159
160 /*
161 * allow a 2 second settle period in the beginning
162 */
163 if (mtime_since(&td->start, now) < 2000)
164 return 0;
165
166 iops += td->this_io_blocks[ddir];
167 bytes += td->this_io_bytes[ddir];
168 ratemin += td->o.ratemin[ddir];
169 rate_iops += td->o.rate_iops[ddir];
170 rate_iops_min += td->o.rate_iops_min[ddir];
171
172 /*
173 * if rate blocks is set, sample is running
174 */
175 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
176 spent = mtime_since(&td->lastrate[ddir], now);
177 if (spent < td->o.ratecycle)
178 return 0;
179
180 if (td->o.rate[ddir]) {
181 /*
182 * check bandwidth specified rate
183 */
184 if (bytes < td->rate_bytes[ddir]) {
185 log_err("%s: min rate %u not met\n", td->o.name,
186 ratemin);
187 return 1;
188 } else {
49cba9b3
JA
189 if (spent)
190 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
191 else
192 rate = 0;
193
2e1df07d
JA
194 if (rate < ratemin ||
195 bytes < td->rate_bytes[ddir]) {
196 log_err("%s: min rate %u not met, got"
197 " %luKB/sec\n", td->o.name,
198 ratemin, rate);
199 return 1;
200 }
201 }
202 } else {
203 /*
204 * checks iops specified rate
205 */
206 if (iops < rate_iops) {
207 log_err("%s: min iops rate %u not met\n",
208 td->o.name, rate_iops);
209 return 1;
210 } else {
4c707a3b
JA
211 if (spent)
212 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
213 else
214 rate = 0;
215
2e1df07d
JA
216 if (rate < rate_iops_min ||
217 iops < td->rate_blocks[ddir]) {
218 log_err("%s: min iops rate %u not met,"
219 " got %lu\n", td->o.name,
220 rate_iops_min, rate);
221 }
222 }
223 }
224 }
225
226 td->rate_bytes[ddir] = bytes;
227 td->rate_blocks[ddir] = iops;
228 memcpy(&td->lastrate[ddir], now, sizeof(*now));
229 return 0;
230}
231
232static int check_min_rate(struct thread_data *td, struct timeval *now,
100f49f1 233 uint64_t *bytes_done)
2e1df07d
JA
234{
235 int ret = 0;
236
6eaf09d6
SL
237 if (bytes_done[DDIR_READ])
238 ret |= __check_min_rate(td, now, DDIR_READ);
239 if (bytes_done[DDIR_WRITE])
240 ret |= __check_min_rate(td, now, DDIR_WRITE);
241 if (bytes_done[DDIR_TRIM])
242 ret |= __check_min_rate(td, now, DDIR_TRIM);
2e1df07d
JA
243
244 return ret;
245}
246
247/*
248 * When job exits, we can cancel the in-flight IO if we are using async
249 * io. Attempt to do so.
250 */
251static void cleanup_pending_aio(struct thread_data *td)
252{
2e1df07d
JA
253 int r;
254
255 /*
256 * get immediately available events, if any
257 */
258 r = io_u_queued_complete(td, 0, NULL);
259 if (r < 0)
260 return;
261
262 /*
263 * now cancel remaining active events
264 */
265 if (td->io_ops->cancel) {
2ae0b204
JA
266 struct io_u *io_u;
267 int i;
2e1df07d 268
2ae0b204
JA
269 io_u_qiter(&td->io_u_all, io_u, i) {
270 if (io_u->flags & IO_U_F_FLIGHT) {
2e1df07d
JA
271 r = td->io_ops->cancel(td, io_u);
272 if (!r)
273 put_io_u(td, io_u);
274 }
275 }
276 }
277
278 if (td->cur_depth)
279 r = io_u_queued_complete(td, td->cur_depth, NULL);
280}
281
282/*
283 * Helper to handle the final sync of a file. Works just like the normal
284 * io path, just does everything sync.
285 */
286static int fio_io_sync(struct thread_data *td, struct fio_file *f)
287{
288 struct io_u *io_u = __get_io_u(td);
289 int ret;
290
291 if (!io_u)
292 return 1;
293
294 io_u->ddir = DDIR_SYNC;
295 io_u->file = f;
296
297 if (td_io_prep(td, io_u)) {
298 put_io_u(td, io_u);
299 return 1;
300 }
301
302requeue:
303 ret = td_io_queue(td, io_u);
304 if (ret < 0) {
305 td_verror(td, io_u->error, "td_io_queue");
306 put_io_u(td, io_u);
307 return 1;
308 } else if (ret == FIO_Q_QUEUED) {
309 if (io_u_queued_complete(td, 1, NULL) < 0)
310 return 1;
311 } else if (ret == FIO_Q_COMPLETED) {
312 if (io_u->error) {
313 td_verror(td, io_u->error, "td_io_queue");
314 return 1;
315 }
316
317 if (io_u_sync_complete(td, io_u, NULL) < 0)
318 return 1;
319 } else if (ret == FIO_Q_BUSY) {
320 if (td_io_commit(td))
321 return 1;
322 goto requeue;
323 }
324
325 return 0;
326}
a3efc919 327
61ee0f86
JA
328static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
329{
330 int ret;
331
332 if (fio_file_open(f))
333 return fio_io_sync(td, f);
334
335 if (td_io_open_file(td, f))
336 return 1;
337
338 ret = fio_io_sync(td, f);
339 td_io_close_file(td, f);
340 return ret;
341}
342
2e1df07d
JA
343static inline void __update_tv_cache(struct thread_data *td)
344{
345 fio_gettime(&td->tv_cache, NULL);
346}
347
348static inline void update_tv_cache(struct thread_data *td)
349{
350 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
351 __update_tv_cache(td);
352}
353
354static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
355{
356 if (in_ramp_time(td))
357 return 0;
358 if (!td->o.timeout)
359 return 0;
0de5b26f 360 if (utime_since(&td->epoch, t) >= td->o.timeout)
2e1df07d
JA
361 return 1;
362
363 return 0;
364}
365
366static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
367 int *retptr)
368{
369 int ret = *retptr;
370
371 if (ret < 0 || td->error) {
8b28bd41
DM
372 int err = td->error;
373 enum error_type_bit eb;
2e1df07d
JA
374
375 if (ret < 0)
376 err = -ret;
2e1df07d 377
8b28bd41
DM
378 eb = td_error_type(ddir, err);
379 if (!(td->o.continue_on_error & (1 << eb)))
2e1df07d
JA
380 return 1;
381
8b28bd41 382 if (td_non_fatal_error(td, eb, err)) {
2e1df07d
JA
383 /*
384 * Continue with the I/Os in case of
385 * a non fatal error.
386 */
387 update_error_count(td, err);
388 td_clear_error(td);
389 *retptr = 0;
390 return 0;
391 } else if (td->o.fill_device && err == ENOSPC) {
392 /*
393 * We expect to hit this error if
394 * fill_device option is set.
395 */
396 td_clear_error(td);
ebea2133 397 fio_mark_td_terminate(td);
2e1df07d
JA
398 return 1;
399 } else {
400 /*
401 * Stop the I/O in case of a fatal
402 * error.
403 */
404 update_error_count(td, err);
405 return 1;
406 }
407 }
408
409 return 0;
410}
411
c97f1ad6
JA
412static void check_update_rusage(struct thread_data *td)
413{
414 if (td->update_rusage) {
415 td->update_rusage = 0;
416 update_rusage_stat(td);
417 fio_mutex_up(td->rusage_sem);
418 }
419}
420
2e1df07d
JA
421/*
422 * The main verify engine. Runs over the writes we previously submitted,
423 * reads the blocks back in, and checks the crc/md5 of the data.
424 */
100f49f1 425static void do_verify(struct thread_data *td, uint64_t verify_bytes)
2e1df07d 426{
100f49f1 427 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
2e1df07d
JA
428 struct fio_file *f;
429 struct io_u *io_u;
430 int ret, min_events;
431 unsigned int i;
432
433 dprint(FD_VERIFY, "starting loop\n");
434
435 /*
436 * sync io first and invalidate cache, to make sure we really
437 * read from disk.
438 */
439 for_each_file(td, f, i) {
440 if (!fio_file_open(f))
441 continue;
442 if (fio_io_sync(td, f))
443 break;
444 if (file_invalidate_cache(td, f))
445 break;
446 }
447
c97f1ad6
JA
448 check_update_rusage(td);
449
2e1df07d
JA
450 if (td->error)
451 return;
452
453 td_set_runstate(td, TD_VERIFYING);
454
455 io_u = NULL;
456 while (!td->terminate) {
fbccf46c 457 enum fio_ddir ddir;
2e1df07d
JA
458 int ret2, full;
459
460 update_tv_cache(td);
c97f1ad6 461 check_update_rusage(td);
2e1df07d
JA
462
463 if (runtime_exceeded(td, &td->tv_cache)) {
464 __update_tv_cache(td);
465 if (runtime_exceeded(td, &td->tv_cache)) {
ebea2133 466 fio_mark_td_terminate(td);
2e1df07d
JA
467 break;
468 }
469 }
470
9e684a49
DE
471 if (flow_threshold_exceeded(td))
472 continue;
473
44cbc6da
JA
474 if (!td->o.experimental_verify) {
475 io_u = __get_io_u(td);
476 if (!io_u)
477 break;
2e1df07d 478
44cbc6da
JA
479 if (get_next_verify(td, io_u)) {
480 put_io_u(td, io_u);
481 break;
482 }
2e1df07d 483
44cbc6da
JA
484 if (td_io_prep(td, io_u)) {
485 put_io_u(td, io_u);
486 break;
487 }
488 } else {
100f49f1
JA
489 if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
490 break;
491
bcd5abfa 492 while ((io_u = get_io_u(td)) != NULL) {
002fe734
JA
493 if (IS_ERR(io_u)) {
494 io_u = NULL;
495 ret = FIO_Q_BUSY;
496 goto reap;
497 }
498
bcd5abfa
JA
499 /*
500 * We are only interested in the places where
501 * we wrote or trimmed IOs. Turn those into
502 * reads for verification purposes.
503 */
504 if (io_u->ddir == DDIR_READ) {
505 /*
506 * Pretend we issued it for rwmix
507 * accounting
508 */
509 td->io_issues[DDIR_READ]++;
510 put_io_u(td, io_u);
511 continue;
512 } else if (io_u->ddir == DDIR_TRIM) {
513 io_u->ddir = DDIR_READ;
514 io_u->flags |= IO_U_F_TRIMMED;
515 break;
516 } else if (io_u->ddir == DDIR_WRITE) {
517 io_u->ddir = DDIR_READ;
518 break;
519 } else {
520 put_io_u(td, io_u);
521 continue;
522 }
523 }
44cbc6da 524
bcd5abfa 525 if (!io_u)
44cbc6da 526 break;
2e1df07d
JA
527 }
528
529 if (td->o.verify_async)
530 io_u->end_io = verify_io_u_async;
531 else
532 io_u->end_io = verify_io_u;
533
fbccf46c
JA
534 ddir = io_u->ddir;
535
2e1df07d
JA
536 ret = td_io_queue(td, io_u);
537 switch (ret) {
538 case FIO_Q_COMPLETED:
539 if (io_u->error) {
540 ret = -io_u->error;
541 clear_io_u(td, io_u);
542 } else if (io_u->resid) {
543 int bytes = io_u->xfer_buflen - io_u->resid;
544
545 /*
546 * zero read, fail
547 */
548 if (!bytes) {
549 td_verror(td, EIO, "full resid");
550 put_io_u(td, io_u);
551 break;
552 }
553
554 io_u->xfer_buflen = io_u->resid;
555 io_u->xfer_buf += bytes;
556 io_u->offset += bytes;
557
558 if (ddir_rw(io_u->ddir))
559 td->ts.short_io_u[io_u->ddir]++;
560
561 f = io_u->file;
562 if (io_u->offset == f->real_file_size)
563 goto sync_done;
564
565 requeue_io_u(td, &io_u);
566 } else {
567sync_done:
100f49f1 568 ret = io_u_sync_complete(td, io_u, bytes_done);
2e1df07d
JA
569 if (ret < 0)
570 break;
571 }
572 continue;
573 case FIO_Q_QUEUED:
574 break;
575 case FIO_Q_BUSY:
576 requeue_io_u(td, &io_u);
577 ret2 = td_io_commit(td);
578 if (ret2 < 0)
579 ret = ret2;
580 break;
581 default:
582 assert(ret < 0);
583 td_verror(td, -ret, "td_io_queue");
584 break;
585 }
586
fbccf46c 587 if (break_on_this_error(td, ddir, &ret))
2e1df07d
JA
588 break;
589
590 /*
591 * if we can queue more, do so. but check if there are
592 * completed io_u's first. Note that we can get BUSY even
593 * without IO queued, if the system is resource starved.
594 */
002fe734 595reap:
2e1df07d
JA
596 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
597 if (full || !td->o.iodepth_batch_complete) {
598 min_events = min(td->o.iodepth_batch_complete,
599 td->cur_depth);
8a74b56d
JA
600 /*
601 * if the queue is full, we MUST reap at least 1 event
602 */
603 if (full && !min_events)
2e1df07d
JA
604 min_events = 1;
605
606 do {
607 /*
608 * Reap required number of io units, if any,
609 * and do the verification on them through
610 * the callback handler
611 */
100f49f1 612 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
2e1df07d
JA
613 ret = -1;
614 break;
615 }
616 } while (full && (td->cur_depth > td->o.iodepth_low));
617 }
618 if (ret < 0)
619 break;
620 }
621
c97f1ad6
JA
622 check_update_rusage(td);
623
2e1df07d
JA
624 if (!td->error) {
625 min_events = td->cur_depth;
626
627 if (min_events)
628 ret = io_u_queued_complete(td, min_events, NULL);
629 } else
630 cleanup_pending_aio(td);
631
632 td_set_runstate(td, TD_RUNNING);
633
634 dprint(FD_VERIFY, "exiting loop\n");
635}
636
3939fe85
JA
637static unsigned int exceeds_number_ios(struct thread_data *td)
638{
639 unsigned long long number_ios;
640
641 if (!td->o.number_ios)
642 return 0;
643
644 number_ios = ddir_rw_sum(td->this_io_blocks);
645 number_ios += td->io_u_queued + td->io_u_in_flight;
646
647 return number_ios >= td->o.number_ios;
648}
649
f7078f7b
JA
650static int io_bytes_exceeded(struct thread_data *td)
651{
77731b29 652 unsigned long long bytes, limit;
f7078f7b
JA
653
654 if (td_rw(td))
6eaf09d6 655 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
f7078f7b 656 else if (td_write(td))
6eaf09d6
SL
657 bytes = td->this_io_bytes[DDIR_WRITE];
658 else if (td_read(td))
659 bytes = td->this_io_bytes[DDIR_READ];
f7078f7b 660 else
6eaf09d6 661 bytes = td->this_io_bytes[DDIR_TRIM];
f7078f7b 662
77731b29
JA
663 if (td->o.io_limit)
664 limit = td->o.io_limit;
665 else
666 limit = td->o.size;
667
668 return bytes >= limit || exceeds_number_ios(td);
f7078f7b
JA
669}
670
2e1df07d
JA
671/*
672 * Main IO worker function. It retrieves io_u's to process and queues
673 * and reaps them, checking for rate and errors along the way.
100f49f1
JA
674 *
675 * Returns number of bytes written and trimmed.
2e1df07d 676 */
100f49f1 677static uint64_t do_io(struct thread_data *td)
2e1df07d 678{
100f49f1 679 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
2e1df07d
JA
680 unsigned int i;
681 int ret = 0;
c2703bf3 682 uint64_t total_bytes, bytes_issued = 0;
2e1df07d
JA
683
684 if (in_ramp_time(td))
685 td_set_runstate(td, TD_RAMP);
686 else
687 td_set_runstate(td, TD_RUNNING);
688
3e260a46
JA
689 lat_target_init(td);
690
78a6469c
JA
691 /*
692 * If verify_backlog is enabled, we'll run the verify in this
693 * handler as well. For that case, we may need up to twice the
694 * amount of bytes.
695 */
c2703bf3 696 total_bytes = td->o.size;
78a6469c
JA
697 if (td->o.verify != VERIFY_NONE &&
698 (td_write(td) && td->o.verify_backlog))
c2703bf3
JA
699 total_bytes += td->o.size;
700
f7078f7b 701 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
c04e4661
DE
702 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
703 td->o.time_based) {
2e1df07d 704 struct timeval comp_time;
2e1df07d
JA
705 int min_evts = 0;
706 struct io_u *io_u;
707 int ret2, full;
708 enum fio_ddir ddir;
709
c97f1ad6
JA
710 check_update_rusage(td);
711
7d7803fa 712 if (td->terminate || td->done)
2e1df07d
JA
713 break;
714
715 update_tv_cache(td);
716
717 if (runtime_exceeded(td, &td->tv_cache)) {
718 __update_tv_cache(td);
719 if (runtime_exceeded(td, &td->tv_cache)) {
ebea2133 720 fio_mark_td_terminate(td);
2e1df07d
JA
721 break;
722 }
723 }
724
9e684a49
DE
725 if (flow_threshold_exceeded(td))
726 continue;
727
c2703bf3 728 if (bytes_issued >= total_bytes)
20876c53
JC
729 break;
730
2e1df07d 731 io_u = get_io_u(td);
002fe734
JA
732 if (IS_ERR_OR_NULL(io_u)) {
733 int err = PTR_ERR(io_u);
734
735 io_u = NULL;
736 if (err == -EBUSY) {
737 ret = FIO_Q_BUSY;
738 goto reap;
739 }
3e260a46
JA
740 if (td->o.latency_target)
741 goto reap;
2e1df07d 742 break;
3e260a46 743 }
2e1df07d
JA
744
745 ddir = io_u->ddir;
746
747 /*
82af2a7c
JA
748 * Add verification end_io handler if:
749 * - Asked to verify (!td_rw(td))
750 * - Or the io_u is from our verify list (mixed write/ver)
2e1df07d
JA
751 */
752 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
82af2a7c 753 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
c4b6117b
PV
754
755 if (!td->o.verify_pattern_bytes) {
d6b72507 756 io_u->rand_seed = __rand(&td->verify_state);
c4b6117b 757 if (sizeof(int) != sizeof(long *))
d6b72507 758 io_u->rand_seed *= __rand(&td->verify_state);
c4b6117b
PV
759 }
760
2e1df07d
JA
761 if (td->o.verify_async)
762 io_u->end_io = verify_io_u_async;
763 else
764 io_u->end_io = verify_io_u;
765 td_set_runstate(td, TD_VERIFYING);
766 } else if (in_ramp_time(td))
767 td_set_runstate(td, TD_RAMP);
768 else
769 td_set_runstate(td, TD_RUNNING);
770
9a50c5c5 771 /*
f9401285
JA
772 * Always log IO before it's issued, so we know the specific
773 * order of it. The logged unit will track when the IO has
774 * completed.
9a50c5c5 775 */
c4b6117b
PV
776 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
777 td->o.do_verify &&
778 td->o.verify != VERIFY_NONE &&
f9401285 779 !td->o.experimental_verify)
c4b6117b
PV
780 log_io_piece(td, io_u);
781
2e1df07d
JA
782 ret = td_io_queue(td, io_u);
783 switch (ret) {
784 case FIO_Q_COMPLETED:
785 if (io_u->error) {
786 ret = -io_u->error;
890b6656 787 unlog_io_piece(td, io_u);
2e1df07d
JA
788 clear_io_u(td, io_u);
789 } else if (io_u->resid) {
790 int bytes = io_u->xfer_buflen - io_u->resid;
791 struct fio_file *f = io_u->file;
792
20876c53 793 bytes_issued += bytes;
890b6656
JA
794
795 trim_io_piece(td, io_u);
796
2e1df07d
JA
797 /*
798 * zero read, fail
799 */
800 if (!bytes) {
890b6656 801 unlog_io_piece(td, io_u);
2e1df07d
JA
802 td_verror(td, EIO, "full resid");
803 put_io_u(td, io_u);
804 break;
805 }
806
807 io_u->xfer_buflen = io_u->resid;
808 io_u->xfer_buf += bytes;
809 io_u->offset += bytes;
810
811 if (ddir_rw(io_u->ddir))
812 td->ts.short_io_u[io_u->ddir]++;
813
814 if (io_u->offset == f->real_file_size)
815 goto sync_done;
816
817 requeue_io_u(td, &io_u);
818 } else {
819sync_done:
6eaf09d6
SL
820 if (__should_check_rate(td, DDIR_READ) ||
821 __should_check_rate(td, DDIR_WRITE) ||
822 __should_check_rate(td, DDIR_TRIM))
2e1df07d
JA
823 fio_gettime(&comp_time, NULL);
824
825 ret = io_u_sync_complete(td, io_u, bytes_done);
826 if (ret < 0)
827 break;
20876c53 828 bytes_issued += io_u->xfer_buflen;
2e1df07d
JA
829 }
830 break;
831 case FIO_Q_QUEUED:
832 /*
833 * if the engine doesn't have a commit hook,
834 * the io_u is really queued. if it does have such
835 * a hook, it has to call io_u_queued() itself.
836 */
837 if (td->io_ops->commit == NULL)
838 io_u_queued(td, io_u);
20876c53 839 bytes_issued += io_u->xfer_buflen;
2e1df07d
JA
840 break;
841 case FIO_Q_BUSY:
890b6656 842 unlog_io_piece(td, io_u);
2e1df07d
JA
843 requeue_io_u(td, &io_u);
844 ret2 = td_io_commit(td);
845 if (ret2 < 0)
846 ret = ret2;
847 break;
848 default:
849 assert(ret < 0);
850 put_io_u(td, io_u);
851 break;
852 }
853
854 if (break_on_this_error(td, ddir, &ret))
855 break;
856
857 /*
858 * See if we need to complete some commands. Note that we
859 * can get BUSY even without IO queued, if the system is
860 * resource starved.
861 */
3e260a46 862reap:
2e1df07d
JA
863 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
864 if (full || !td->o.iodepth_batch_complete) {
865 min_evts = min(td->o.iodepth_batch_complete,
866 td->cur_depth);
8a74b56d
JA
867 /*
868 * if the queue is full, we MUST reap at least 1 event
869 */
870 if (full && !min_evts)
2e1df07d
JA
871 min_evts = 1;
872
6eaf09d6
SL
873 if (__should_check_rate(td, DDIR_READ) ||
874 __should_check_rate(td, DDIR_WRITE) ||
875 __should_check_rate(td, DDIR_TRIM))
2e1df07d
JA
876 fio_gettime(&comp_time, NULL);
877
878 do {
879 ret = io_u_queued_complete(td, min_evts, bytes_done);
880 if (ret < 0)
881 break;
882
883 } while (full && (td->cur_depth > td->o.iodepth_low));
884 }
885
886 if (ret < 0)
887 break;
d5abee06 888 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
2e1df07d
JA
889 continue;
890
891 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
892 if (check_min_rate(td, &comp_time, bytes_done)) {
893 if (exitall_on_terminate)
894 fio_terminate_threads(td->groupid);
895 td_verror(td, EIO, "check_min_rate");
896 break;
897 }
898 }
3e260a46
JA
899 if (!in_ramp_time(td) && td->o.latency_target)
900 lat_target_check(td);
2e1df07d
JA
901
902 if (td->o.thinktime) {
903 unsigned long long b;
904
342f4be4 905 b = ddir_rw_sum(td->io_blocks);
2e1df07d
JA
906 if (!(b % td->o.thinktime_blocks)) {
907 int left;
908
002e7183
JA
909 io_u_quiesce(td);
910
2e1df07d
JA
911 if (td->o.thinktime_spin)
912 usec_spin(td->o.thinktime_spin);
913
914 left = td->o.thinktime - td->o.thinktime_spin;
915 if (left)
916 usec_sleep(td, left);
917 }
918 }
919 }
920
c97f1ad6
JA
921 check_update_rusage(td);
922
2e1df07d 923 if (td->trim_entries)
4e0a8fa2 924 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
2e1df07d
JA
925
926 if (td->o.fill_device && td->error == ENOSPC) {
927 td->error = 0;
ebea2133 928 fio_mark_td_terminate(td);
2e1df07d
JA
929 }
930 if (!td->error) {
931 struct fio_file *f;
932
933 i = td->cur_depth;
934 if (i) {
5bd5f71a 935 ret = io_u_queued_complete(td, i, bytes_done);
2e1df07d
JA
936 if (td->o.fill_device && td->error == ENOSPC)
937 td->error = 0;
938 }
939
940 if (should_fsync(td) && td->o.end_fsync) {
941 td_set_runstate(td, TD_FSYNCING);
942
943 for_each_file(td, f, i) {
61ee0f86 944 if (!fio_file_fsync(td, f))
2e1df07d 945 continue;
61ee0f86
JA
946
947 log_err("fio: end_fsync failed for file %s\n",
948 f->file_name);
2e1df07d
JA
949 }
950 }
951 } else
952 cleanup_pending_aio(td);
953
954 /*
955 * stop job if we failed doing any IO
956 */
342f4be4 957 if (!ddir_rw_sum(td->this_io_bytes))
2e1df07d 958 td->done = 1;
100f49f1
JA
959
960 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
2e1df07d
JA
961}
962
963static void cleanup_io_u(struct thread_data *td)
964{
2e1df07d
JA
965 struct io_u *io_u;
966
2ae0b204 967 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
c73ed246
JA
968
969 if (td->io_ops->io_u_free)
970 td->io_ops->io_u_free(td, io_u);
971
2e1df07d
JA
972 fio_memfree(io_u, sizeof(*io_u));
973 }
974
975 free_io_mem(td);
2ae0b204
JA
976
977 io_u_rexit(&td->io_u_requeues);
978 io_u_qexit(&td->io_u_freelist);
979 io_u_qexit(&td->io_u_all);
2e1df07d
JA
980}
981
982static int init_io_u(struct thread_data *td)
983{
984 struct io_u *io_u;
9c42684e 985 unsigned int max_bs, min_write;
2e1df07d 986 int cl_align, i, max_units;
2ae0b204 987 int data_xfer = 1, err;
2e1df07d
JA
988 char *p;
989
990 max_units = td->o.iodepth;
74f4b020 991 max_bs = td_max_bs(td);
9c42684e 992 min_write = td->o.min_bs[DDIR_WRITE];
2e1df07d
JA
993 td->orig_buffer_size = (unsigned long long) max_bs
994 * (unsigned long long) max_units;
995
88045e04 996 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
59d8d0f5
JA
997 data_xfer = 0;
998
2ae0b204
JA
999 err = 0;
1000 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1001 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1002 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1003
1004 if (err) {
1005 log_err("fio: failed setting up IO queues\n");
1006 return 1;
1007 }
1008
fd8a09b8 1009 /*
1010 * if we may later need to do address alignment, then add any
1011 * possible adjustment here so that we don't cause a buffer
1012 * overflow later. this adjustment may be too much if we get
1013 * lucky and the allocator gives us an aligned address.
1014 */
d01612f3
CM
1015 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1016 (td->io_ops->flags & FIO_RAWIO))
fd8a09b8 1017 td->orig_buffer_size += page_mask + td->o.mem_align;
1018
2e1df07d
JA
1019 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1020 unsigned long bs;
1021
1022 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1023 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1024 }
1025
1026 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1027 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1028 return 1;
1029 }
1030
59d8d0f5 1031 if (data_xfer && allocate_io_mem(td))
2e1df07d
JA
1032 return 1;
1033
d01612f3 1034 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
2e1df07d
JA
1035 (td->io_ops->flags & FIO_RAWIO))
1036 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1037 else
1038 p = td->orig_buffer;
1039
1040 cl_align = os_cache_line_size();
1041
1042 for (i = 0; i < max_units; i++) {
1043 void *ptr;
1044
1045 if (td->terminate)
1046 return 1;
1047
1048 ptr = fio_memalign(cl_align, sizeof(*io_u));
1049 if (!ptr) {
1050 log_err("fio: unable to allocate aligned memory\n");
1051 break;
1052 }
1053
1054 io_u = ptr;
1055 memset(io_u, 0, sizeof(*io_u));
2ae0b204 1056 INIT_FLIST_HEAD(&io_u->verify_list);
2e1df07d
JA
1057 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1058
59d8d0f5 1059 if (data_xfer) {
2e1df07d
JA
1060 io_u->buf = p;
1061 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1062
1063 if (td_write(td))
9c42684e 1064 io_u_fill_buffer(td, io_u, min_write, max_bs);
2e1df07d
JA
1065 if (td_write(td) && td->o.verify_pattern_bytes) {
1066 /*
1067 * Fill the buffer with the pattern if we are
1068 * going to be doing writes.
1069 */
ce35b1ec 1070 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
2e1df07d
JA
1071 }
1072 }
1073
1074 io_u->index = i;
1075 io_u->flags = IO_U_F_FREE;
2ae0b204
JA
1076 io_u_qpush(&td->io_u_freelist, io_u);
1077
1078 /*
1079 * io_u never leaves this stack, used for iteration of all
1080 * io_u buffers.
1081 */
1082 io_u_qpush(&td->io_u_all, io_u);
c73ed246
JA
1083
1084 if (td->io_ops->io_u_init) {
1085 int ret = td->io_ops->io_u_init(td, io_u);
1086
1087 if (ret) {
1088 log_err("fio: failed to init engine data: %d\n", ret);
1089 return 1;
1090 }
1091 }
1092
2e1df07d
JA
1093 p += max_bs;
1094 }
1095
1096 return 0;
1097}
1098
1099static int switch_ioscheduler(struct thread_data *td)
1100{
1101 char tmp[256], tmp2[128];
1102 FILE *f;
1103 int ret;
1104
1105 if (td->io_ops->flags & FIO_DISKLESSIO)
1106 return 0;
1107
1108 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1109
1110 f = fopen(tmp, "r+");
1111 if (!f) {
1112 if (errno == ENOENT) {
1113 log_err("fio: os or kernel doesn't support IO scheduler"
1114 " switching\n");
1115 return 0;
1116 }
1117 td_verror(td, errno, "fopen iosched");
1118 return 1;
1119 }
1120
1121 /*
1122 * Set io scheduler.
1123 */
1124 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1125 if (ferror(f) || ret != 1) {
1126 td_verror(td, errno, "fwrite");
1127 fclose(f);
1128 return 1;
1129 }
1130
1131 rewind(f);
1132
1133 /*
1134 * Read back and check that the selected scheduler is now the default.
1135 */
49c6f33d 1136 ret = fread(tmp, sizeof(tmp), 1, f);
2e1df07d
JA
1137 if (ferror(f) || ret < 0) {
1138 td_verror(td, errno, "fread");
1139 fclose(f);
1140 return 1;
1141 }
49c6f33d
JA
1142 tmp[sizeof(tmp) - 1] = '\0';
1143
2e1df07d
JA
1144
1145 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1146 if (!strstr(tmp, tmp2)) {
1147 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1148 td_verror(td, EINVAL, "iosched_switch");
1149 fclose(f);
1150 return 1;
1151 }
1152
1153 fclose(f);
1154 return 0;
1155}
1156
1157static int keep_running(struct thread_data *td)
1158{
77731b29
JA
1159 unsigned long long limit;
1160
2e1df07d
JA
1161 if (td->done)
1162 return 0;
1163 if (td->o.time_based)
1164 return 1;
1165 if (td->o.loops) {
1166 td->o.loops--;
1167 return 1;
1168 }
3939fe85
JA
1169 if (exceeds_number_ios(td))
1170 return 0;
26251d8d 1171
77731b29
JA
1172 if (td->o.io_limit)
1173 limit = td->o.io_limit;
1174 else
1175 limit = td->o.size;
1176
1177 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
5bd5f71a
JA
1178 uint64_t diff;
1179
1180 /*
1181 * If the difference is less than the minimum IO size, we
1182 * are done.
1183 */
77731b29 1184 diff = limit - ddir_rw_sum(td->io_bytes);
74f4b020 1185 if (diff < td_max_bs(td))
5bd5f71a
JA
1186 return 0;
1187
002fe734
JA
1188 if (fio_files_done(td))
1189 return 0;
1190
2e1df07d 1191 return 1;
5bd5f71a 1192 }
2e1df07d
JA
1193
1194 return 0;
1195}
1196
ce486495 1197static int exec_string(struct thread_options *o, const char *string, const char *mode)
2e1df07d 1198{
ce486495 1199 int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
2e1df07d
JA
1200 char *str;
1201
1202 str = malloc(newlen);
ce486495 1203 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
2e1df07d 1204
ce486495 1205 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
2e1df07d
JA
1206 ret = system(str);
1207 if (ret == -1)
1208 log_err("fio: exec of cmd <%s> failed\n", str);
1209
1210 free(str);
1211 return ret;
1212}
1213
62167762
JC
1214/*
1215 * Dry run to compute correct state of numberio for verification.
1216 */
1217static uint64_t do_dry_run(struct thread_data *td)
1218{
1219 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1220
1221 td_set_runstate(td, TD_RUNNING);
1222
1223 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1224 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1225 struct io_u *io_u;
1226 int ret;
1227
1228 if (td->terminate || td->done)
1229 break;
1230
1231 io_u = get_io_u(td);
1232 if (!io_u)
1233 break;
1234
1235 io_u->flags |= IO_U_F_FLIGHT;
1236 io_u->error = 0;
1237 io_u->resid = 0;
1238 if (ddir_rw(acct_ddir(io_u)))
1239 td->io_issues[acct_ddir(io_u)]++;
1240 if (ddir_rw(io_u->ddir)) {
1241 io_u_mark_depth(td, 1);
1242 td->ts.total_io_u[io_u->ddir]++;
1243 }
1244
2e63e96b
PV
1245 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1246 td->o.do_verify &&
1247 td->o.verify != VERIFY_NONE &&
1248 !td->o.experimental_verify)
1249 log_io_piece(td, io_u);
1250
62167762
JC
1251 ret = io_u_sync_complete(td, io_u, bytes_done);
1252 (void) ret;
1253 }
1254
1255 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1256}
1257
2e1df07d
JA
1258/*
1259 * Entry point for the thread based jobs. The process based jobs end up
1260 * here as well, after a little setup.
1261 */
1262static void *thread_main(void *data)
1263{
1264 unsigned long long elapsed;
1265 struct thread_data *td = data;
4896473e 1266 struct thread_options *o = &td->o;
2e1df07d
JA
1267 pthread_condattr_t attr;
1268 int clear_state;
28727df7 1269 int ret;
2e1df07d 1270
4896473e 1271 if (!o->use_thread) {
2e1df07d
JA
1272 setsid();
1273 td->pid = getpid();
1274 } else
1275 td->pid = gettid();
1276
4896473e 1277 fio_local_clock_init(o->use_thread);
5d879392 1278
2e1df07d
JA
1279 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1280
122c7725
JA
1281 if (is_backend)
1282 fio_server_send_start(td);
1283
2e1df07d
JA
1284 INIT_FLIST_HEAD(&td->io_log_list);
1285 INIT_FLIST_HEAD(&td->io_hist_list);
1286 INIT_FLIST_HEAD(&td->verify_list);
1287 INIT_FLIST_HEAD(&td->trim_list);
1ae83d45 1288 INIT_FLIST_HEAD(&td->next_rand_list);
2e1df07d
JA
1289 pthread_mutex_init(&td->io_u_lock, NULL);
1290 td->io_hist_tree = RB_ROOT;
1291
1292 pthread_condattr_init(&attr);
1293 pthread_cond_init(&td->verify_cond, &attr);
1294 pthread_cond_init(&td->free_cond, &attr);
1295
1296 td_set_runstate(td, TD_INITIALIZED);
1297 dprint(FD_MUTEX, "up startup_mutex\n");
1298 fio_mutex_up(startup_mutex);
1299 dprint(FD_MUTEX, "wait on td->mutex\n");
1300 fio_mutex_down(td->mutex);
1301 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1302
2e1df07d
JA
1303 /*
1304 * A new gid requires privilege, so we need to do this before setting
1305 * the uid.
1306 */
4896473e 1307 if (o->gid != -1U && setgid(o->gid)) {
2e1df07d
JA
1308 td_verror(td, errno, "setgid");
1309 goto err;
1310 }
4896473e 1311 if (o->uid != -1U && setuid(o->uid)) {
2e1df07d
JA
1312 td_verror(td, errno, "setuid");
1313 goto err;
1314 }
1315
1316 /*
1317 * If we have a gettimeofday() thread, make sure we exclude that
1318 * thread from this job
1319 */
4896473e
JA
1320 if (o->gtod_cpu)
1321 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
2e1df07d
JA
1322
1323 /*
1324 * Set affinity first, in case it has an impact on the memory
1325 * allocations.
1326 */
4896473e 1327 if (o->cpumask_set) {
c2acfbac 1328 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
30cb4c65 1329 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
c2acfbac
JA
1330 if (!ret) {
1331 log_err("fio: no CPUs set\n");
1332 log_err("fio: Try increasing number of available CPUs\n");
1333 td_verror(td, EINVAL, "cpus_split");
1334 goto err;
1335 }
1336 }
28727df7
JA
1337 ret = fio_setaffinity(td->pid, o->cpumask);
1338 if (ret == -1) {
4896473e
JA
1339 td_verror(td, errno, "cpu_set_affinity");
1340 goto err;
1341 }
2e1df07d
JA
1342 }
1343
67bf9823 1344#ifdef CONFIG_LIBNUMA
d0b937ed 1345 /* numa node setup */
4896473e 1346 if (o->numa_cpumask_set || o->numa_memmask_set) {
43522848 1347 struct bitmask *mask;
d0b937ed
YR
1348 int ret;
1349
1350 if (numa_available() < 0) {
1351 td_verror(td, errno, "Does not support NUMA API\n");
1352 goto err;
1353 }
1354
4896473e 1355 if (o->numa_cpumask_set) {
43522848
DG
1356 mask = numa_parse_nodestring(o->numa_cpunodes);
1357 ret = numa_run_on_node_mask(mask);
1358 numa_free_nodemask(mask);
d0b937ed
YR
1359 if (ret == -1) {
1360 td_verror(td, errno, \
1361 "numa_run_on_node_mask failed\n");
1362 goto err;
1363 }
1364 }
1365
4896473e 1366 if (o->numa_memmask_set) {
d0b937ed 1367
43522848
DG
1368 mask = NULL;
1369 if (o->numa_memnodes)
1370 mask = numa_parse_nodestring(o->numa_memnodes);
1371
4896473e 1372 switch (o->numa_mem_mode) {
d0b937ed 1373 case MPOL_INTERLEAVE:
43522848 1374 numa_set_interleave_mask(mask);
d0b937ed
YR
1375 break;
1376 case MPOL_BIND:
43522848 1377 numa_set_membind(mask);
d0b937ed
YR
1378 break;
1379 case MPOL_LOCAL:
1380 numa_set_localalloc();
1381 break;
1382 case MPOL_PREFERRED:
4896473e 1383 numa_set_preferred(o->numa_mem_prefer_node);
d0b937ed
YR
1384 break;
1385 case MPOL_DEFAULT:
1386 default:
1387 break;
1388 }
1389
43522848
DG
1390 if (mask)
1391 numa_free_nodemask(mask);
1392
d0b937ed
YR
1393 }
1394 }
1395#endif
1396
9a3f1100
JA
1397 if (fio_pin_memory(td))
1398 goto err;
1399
2e1df07d
JA
1400 /*
1401 * May alter parameters that init_io_u() will use, so we need to
1402 * do this first.
1403 */
1404 if (init_iolog(td))
1405 goto err;
1406
1407 if (init_io_u(td))
1408 goto err;
1409
4896473e 1410 if (o->verify_async && verify_async_init(td))
2e1df07d
JA
1411 goto err;
1412
28727df7
JA
1413 if (o->ioprio) {
1414 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1415 if (ret == -1) {
2e1df07d
JA
1416 td_verror(td, errno, "ioprio_set");
1417 goto err;
1418 }
1419 }
1420
4896473e 1421 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
2e1df07d
JA
1422 goto err;
1423
649c10c5 1424 errno = 0;
4896473e 1425 if (nice(o->nice) == -1 && errno != 0) {
2e1df07d
JA
1426 td_verror(td, errno, "nice");
1427 goto err;
1428 }
1429
4896473e 1430 if (o->ioscheduler && switch_ioscheduler(td))
2e1df07d
JA
1431 goto err;
1432
4896473e 1433 if (!o->create_serialize && setup_files(td))
2e1df07d
JA
1434 goto err;
1435
1436 if (td_io_init(td))
1437 goto err;
1438
1439 if (init_random_map(td))
1440 goto err;
1441
ce486495 1442 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
4896473e 1443 goto err;
2e1df07d 1444
4896473e 1445 if (o->pre_read) {
2e1df07d
JA
1446 if (pre_read_files(td) < 0)
1447 goto err;
1448 }
1449
aee2ab67
JA
1450 if (td->flags & TD_F_COMPRESS_LOG)
1451 tp_init(&td->tp_data);
1452
dc5bfbb2
JA
1453 fio_verify_init(td);
1454
2e1df07d 1455 fio_gettime(&td->epoch, NULL);
44404c5a 1456 fio_getrusage(&td->ru_start);
2e1df07d
JA
1457 clear_state = 0;
1458 while (keep_running(td)) {
100f49f1
JA
1459 uint64_t verify_bytes;
1460
2e1df07d
JA
1461 fio_gettime(&td->start, NULL);
1462 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1463 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1464 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1465
4896473e
JA
1466 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1467 o->ratemin[DDIR_TRIM]) {
6eaf09d6 1468 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
2e1df07d 1469 sizeof(td->bw_sample_time));
6eaf09d6
SL
1470 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1471 sizeof(td->bw_sample_time));
1472 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
2e1df07d
JA
1473 sizeof(td->bw_sample_time));
1474 }
1475
1476 if (clear_state)
1477 clear_io_state(td);
1478
1479 prune_io_piece_log(td);
1480
62167762
JC
1481 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1482 verify_bytes = do_dry_run(td);
1483 else
1484 verify_bytes = do_io(td);
2e1df07d
JA
1485
1486 clear_state = 1;
1487
1488 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1489 elapsed = utime_since_now(&td->start);
1490 td->ts.runtime[DDIR_READ] += elapsed;
1491 }
1492 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1493 elapsed = utime_since_now(&td->start);
1494 td->ts.runtime[DDIR_WRITE] += elapsed;
1495 }
6eaf09d6
SL
1496 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1497 elapsed = utime_since_now(&td->start);
1498 td->ts.runtime[DDIR_TRIM] += elapsed;
1499 }
2e1df07d
JA
1500
1501 if (td->error || td->terminate)
1502 break;
1503
4896473e
JA
1504 if (!o->do_verify ||
1505 o->verify == VERIFY_NONE ||
2e1df07d
JA
1506 (td->io_ops->flags & FIO_UNIDIR))
1507 continue;
1508
1509 clear_io_state(td);
1510
1511 fio_gettime(&td->start, NULL);
1512
100f49f1 1513 do_verify(td, verify_bytes);
2e1df07d
JA
1514
1515 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1516
1517 if (td->error || td->terminate)
1518 break;
1519 }
1520
1521 update_rusage_stat(td);
6eaf09d6
SL
1522 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1523 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1524 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
2e1df07d 1525 td->ts.total_run_time = mtime_since_now(&td->epoch);
6eaf09d6
SL
1526 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1527 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1528 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2e1df07d 1529
9a3f1100
JA
1530 fio_unpin_memory(td);
1531
905e3d4f 1532 fio_writeout_logs(td);
2e1df07d 1533
aee2ab67
JA
1534 if (td->flags & TD_F_COMPRESS_LOG)
1535 tp_exit(&td->tp_data);
1536
4896473e 1537 if (o->exec_postrun)
ce486495 1538 exec_string(o, o->exec_postrun, (const char *)"postrun");
2e1df07d
JA
1539
1540 if (exitall_on_terminate)
1541 fio_terminate_threads(td->groupid);
1542
1543err:
1544 if (td->error)
1545 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1546 td->verror);
1547
4896473e 1548 if (o->verify_async)
2e1df07d
JA
1549 verify_async_exit(td);
1550
1551 close_and_free_files(td);
2e1df07d 1552 cleanup_io_u(td);
32dbca2c 1553 close_ioengine(td);
2e1df07d
JA
1554 cgroup_shutdown(td, &cgroup_mnt);
1555
4896473e
JA
1556 if (o->cpumask_set) {
1557 int ret = fio_cpuset_exit(&o->cpumask);
2e1df07d
JA
1558
1559 td_verror(td, ret, "fio_cpuset_exit");
1560 }
1561
1562 /*
1563 * do this very late, it will log file closing as well
1564 */
4896473e 1565 if (o->write_iolog_file)
2e1df07d
JA
1566 write_iolog_close(td);
1567
ea66e04f
JA
1568 fio_mutex_remove(td->mutex);
1569 td->mutex = NULL;
1570
2e1df07d 1571 td_set_runstate(td, TD_EXITED);
fda2cfac
JA
1572
1573 /*
1574 * Do this last after setting our runstate to exited, so we
1575 * know that the stat thread is signaled.
1576 */
1577 check_update_rusage(td);
1578
e43606c2 1579 return (void *) (uintptr_t) td->error;
2e1df07d
JA
1580}
1581
1582
1583/*
1584 * We cannot pass the td data into a forked process, so attach the td and
1585 * pass it to the thread worker.
1586 */
1587static int fork_main(int shmid, int offset)
1588{
1589 struct thread_data *td;
1590 void *data, *ret;
1591
c8931876 1592#if !defined(__hpux) && !defined(CONFIG_NO_SHM)
2e1df07d
JA
1593 data = shmat(shmid, NULL, 0);
1594 if (data == (void *) -1) {
1595 int __err = errno;
1596
1597 perror("shmat");
1598 return __err;
1599 }
1600#else
1601 /*
1602 * HP-UX inherits shm mappings?
1603 */
1604 data = threads;
1605#endif
1606
1607 td = data + offset * sizeof(struct thread_data);
1608 ret = thread_main(td);
1609 shmdt(data);
e43606c2 1610 return (int) (uintptr_t) ret;
2e1df07d
JA
1611}
1612
cba5460c
JA
1613static void dump_td_info(struct thread_data *td)
1614{
1615 log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1616 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1617 (unsigned long) time_since_now(&td->terminate_time));
1618}
1619
2e1df07d
JA
1620/*
1621 * Run over the job map and reap the threads that have exited, if any.
1622 */
1623static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1624 unsigned int *m_rate)
1625{
1626 struct thread_data *td;
1627 unsigned int cputhreads, realthreads, pending;
1628 int i, status, ret;
1629
1630 /*
1631 * reap exited threads (TD_EXITED -> TD_REAPED)
1632 */
1633 realthreads = pending = cputhreads = 0;
1634 for_each_td(td, i) {
1635 int flags = 0;
1636
1637 /*
1638 * ->io_ops is NULL for a thread that has closed its
1639 * io engine
1640 */
1641 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1642 cputhreads++;
1643 else
1644 realthreads++;
1645
1646 if (!td->pid) {
1647 pending++;
1648 continue;
1649 }
1650 if (td->runstate == TD_REAPED)
1651 continue;
1652 if (td->o.use_thread) {
1653 if (td->runstate == TD_EXITED) {
1654 td_set_runstate(td, TD_REAPED);
1655 goto reaped;
1656 }
1657 continue;
1658 }
1659
1660 flags = WNOHANG;
1661 if (td->runstate == TD_EXITED)
1662 flags = 0;
1663
1664 /*
1665 * check if someone quit or got killed in an unusual way
1666 */
1667 ret = waitpid(td->pid, &status, flags);
1668 if (ret < 0) {
1669 if (errno == ECHILD) {
1670 log_err("fio: pid=%d disappeared %d\n",
1671 (int) td->pid, td->runstate);
a5e371a6 1672 td->sig = ECHILD;
2e1df07d
JA
1673 td_set_runstate(td, TD_REAPED);
1674 goto reaped;
1675 }
1676 perror("waitpid");
1677 } else if (ret == td->pid) {
1678 if (WIFSIGNALED(status)) {
1679 int sig = WTERMSIG(status);
1680
36d80bc7 1681 if (sig != SIGTERM && sig != SIGUSR2)
2e1df07d
JA
1682 log_err("fio: pid=%d, got signal=%d\n",
1683 (int) td->pid, sig);
a5e371a6 1684 td->sig = sig;
2e1df07d
JA
1685 td_set_runstate(td, TD_REAPED);
1686 goto reaped;
1687 }
1688 if (WIFEXITED(status)) {
1689 if (WEXITSTATUS(status) && !td->error)
1690 td->error = WEXITSTATUS(status);
1691
1692 td_set_runstate(td, TD_REAPED);
1693 goto reaped;
1694 }
1695 }
1696
cba5460c
JA
1697 /*
1698 * If the job is stuck, do a forceful timeout of it and
1699 * move on.
1700 */
1701 if (td->terminate &&
1702 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1703 dump_td_info(td);
1704 td_set_runstate(td, TD_REAPED);
1705 goto reaped;
1706 }
1707
2e1df07d
JA
1708 /*
1709 * thread is not dead, continue
1710 */
1711 pending++;
1712 continue;
1713reaped:
1714 (*nr_running)--;
342f4be4
JA
1715 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1716 (*t_rate) -= ddir_rw_sum(td->o.rate);
2e1df07d
JA
1717 if (!td->pid)
1718 pending--;
1719
1720 if (td->error)
1721 exit_value++;
1722
1723 done_secs += mtime_since_now(&td->epoch) / 1000;
4a88752a 1724 profile_td_exit(td);
2e1df07d
JA
1725 }
1726
1727 if (*nr_running == cputhreads && !pending && realthreads)
1728 fio_terminate_threads(TERMINATE_ALL);
1729}
1730
06464907
JA
1731static void do_usleep(unsigned int usecs)
1732{
1733 check_for_running_stats();
1734 usleep(usecs);
1735}
1736
2e1df07d
JA
1737/*
1738 * Main function for kicking off and reaping jobs, as needed.
1739 */
1740static void run_threads(void)
1741{
1742 struct thread_data *td;
2e1df07d 1743 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
0de5b26f 1744 uint64_t spent;
2e1df07d 1745
2e1df07d
JA
1746 if (fio_gtod_offload && fio_start_gtod_thread())
1747 return;
334185e9 1748
f2a2ce0e 1749 fio_idle_prof_init();
2e1df07d
JA
1750
1751 set_sig_handlers();
1752
3a5f6bde
JA
1753 nr_thread = nr_process = 0;
1754 for_each_td(td, i) {
1755 if (td->o.use_thread)
1756 nr_thread++;
1757 else
1758 nr_process++;
1759 }
1760
f3afa57e 1761 if (output_format == FIO_OUTPUT_NORMAL) {
2e1df07d
JA
1762 log_info("Starting ");
1763 if (nr_thread)
1764 log_info("%d thread%s", nr_thread,
1765 nr_thread > 1 ? "s" : "");
1766 if (nr_process) {
1767 if (nr_thread)
1768 log_info(" and ");
1769 log_info("%d process%s", nr_process,
1770 nr_process > 1 ? "es" : "");
1771 }
1772 log_info("\n");
e411c301 1773 log_info_flush();
2e1df07d
JA
1774 }
1775
1776 todo = thread_number;
1777 nr_running = 0;
1778 nr_started = 0;
1779 m_rate = t_rate = 0;
1780
1781 for_each_td(td, i) {
1782 print_status_init(td->thread_number - 1);
1783
1784 if (!td->o.create_serialize)
1785 continue;
1786
1787 /*
1788 * do file setup here so it happens sequentially,
1789 * we don't want X number of threads getting their
1790 * client data interspersed on disk
1791 */
1792 if (setup_files(td)) {
1793 exit_value++;
1794 if (td->error)
1795 log_err("fio: pid=%d, err=%d/%s\n",
1796 (int) td->pid, td->error, td->verror);
1797 td_set_runstate(td, TD_REAPED);
1798 todo--;
1799 } else {
1800 struct fio_file *f;
1801 unsigned int j;
1802
1803 /*
1804 * for sharing to work, each job must always open
1805 * its own files. so close them, if we opened them
1806 * for creation
1807 */
1808 for_each_file(td, f, j) {
1809 if (fio_file_open(f))
1810 td_io_close_file(td, f);
1811 }
1812 }
1813 }
1814
f2a2ce0e
HL
1815 /* start idle threads before io threads start to run */
1816 fio_idle_prof_start();
1817
2e1df07d
JA
1818 set_genesis_time();
1819
1820 while (todo) {
1821 struct thread_data *map[REAL_MAX_JOBS];
1822 struct timeval this_start;
1823 int this_jobs = 0, left;
1824
1825 /*
1826 * create threads (TD_NOT_CREATED -> TD_CREATED)
1827 */
1828 for_each_td(td, i) {
1829 if (td->runstate != TD_NOT_CREATED)
1830 continue;
1831
1832 /*
1833 * never got a chance to start, killed by other
1834 * thread for some reason
1835 */
1836 if (td->terminate) {
1837 todo--;
1838 continue;
1839 }
1840
1841 if (td->o.start_delay) {
0de5b26f 1842 spent = utime_since_genesis();
2e1df07d 1843
74454ce4 1844 if (td->o.start_delay > spent)
2e1df07d
JA
1845 continue;
1846 }
1847
1848 if (td->o.stonewall && (nr_started || nr_running)) {
1849 dprint(FD_PROCESS, "%s: stonewall wait\n",
1850 td->o.name);
1851 break;
1852 }
1853
1854 init_disk_util(td);
1855
c97f1ad6
JA
1856 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1857 td->update_rusage = 0;
1858
2e1df07d
JA
1859 /*
1860 * Set state to created. Thread will transition
1861 * to TD_INITIALIZED when it's done setting up.
1862 */
1863 td_set_runstate(td, TD_CREATED);
1864 map[this_jobs++] = td;
1865 nr_started++;
1866
1867 if (td->o.use_thread) {
1868 int ret;
1869
1870 dprint(FD_PROCESS, "will pthread_create\n");
1871 ret = pthread_create(&td->thread, NULL,
1872 thread_main, td);
1873 if (ret) {
1874 log_err("pthread_create: %s\n",
1875 strerror(ret));
1876 nr_started--;
1877 break;
1878 }
1879 ret = pthread_detach(td->thread);
1880 if (ret)
1881 log_err("pthread_detach: %s",
1882 strerror(ret));
1883 } else {
1884 pid_t pid;
1885 dprint(FD_PROCESS, "will fork\n");
1886 pid = fork();
1887 if (!pid) {
1888 int ret = fork_main(shm_id, i);
1889
1890 _exit(ret);
1891 } else if (i == fio_debug_jobno)
1892 *fio_debug_jobp = pid;
1893 }
1894 dprint(FD_MUTEX, "wait on startup_mutex\n");
1895 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1896 log_err("fio: job startup hung? exiting.\n");
1897 fio_terminate_threads(TERMINATE_ALL);
1898 fio_abort = 1;
1899 nr_started--;
1900 break;
1901 }
1902 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1903 }
1904
1905 /*
1906 * Wait for the started threads to transition to
1907 * TD_INITIALIZED.
1908 */
1909 fio_gettime(&this_start, NULL);
1910 left = this_jobs;
1911 while (left && !fio_abort) {
1912 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1913 break;
1914
06464907 1915 do_usleep(100000);
2e1df07d
JA
1916
1917 for (i = 0; i < this_jobs; i++) {
1918 td = map[i];
1919 if (!td)
1920 continue;
1921 if (td->runstate == TD_INITIALIZED) {
1922 map[i] = NULL;
1923 left--;
1924 } else if (td->runstate >= TD_EXITED) {
1925 map[i] = NULL;
1926 left--;
1927 todo--;
1928 nr_running++; /* work-around... */
1929 }
1930 }
1931 }
1932
1933 if (left) {
4e87c37a
JA
1934 log_err("fio: %d job%s failed to start\n", left,
1935 left > 1 ? "s" : "");
2e1df07d
JA
1936 for (i = 0; i < this_jobs; i++) {
1937 td = map[i];
1938 if (!td)
1939 continue;
1940 kill(td->pid, SIGTERM);
1941 }
1942 break;
1943 }
1944
1945 /*
1946 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1947 */
1948 for_each_td(td, i) {
1949 if (td->runstate != TD_INITIALIZED)
1950 continue;
1951
1952 if (in_ramp_time(td))
1953 td_set_runstate(td, TD_RAMP);
1954 else
1955 td_set_runstate(td, TD_RUNNING);
1956 nr_running++;
1957 nr_started--;
342f4be4
JA
1958 m_rate += ddir_rw_sum(td->o.ratemin);
1959 t_rate += ddir_rw_sum(td->o.rate);
2e1df07d
JA
1960 todo--;
1961 fio_mutex_up(td->mutex);
1962 }
1963
1964 reap_threads(&nr_running, &t_rate, &m_rate);
1965
122c7725 1966 if (todo)
06464907 1967 do_usleep(100000);
2e1df07d
JA
1968 }
1969
1970 while (nr_running) {
1971 reap_threads(&nr_running, &t_rate, &m_rate);
06464907 1972 do_usleep(10000);
2e1df07d
JA
1973 }
1974
f2a2ce0e
HL
1975 fio_idle_prof_stop();
1976
2e1df07d 1977 update_io_ticks();
2e1df07d
JA
1978}
1979
5ddc6707 1980static void wait_for_helper_thread_exit(void)
9ec7779f 1981{
8aab824f
JA
1982 void *ret;
1983
98b4b0a2 1984 helper_exit = 1;
5ddc6707
JA
1985 pthread_cond_signal(&helper_cond);
1986 pthread_join(helper_thread, &ret);
9ec7779f
JA
1987}
1988
27357187
JA
1989static void free_disk_util(void)
1990{
27357187 1991 disk_util_prune_entries();
8aab824f 1992
5ddc6707 1993 pthread_cond_destroy(&helper_cond);
27357187
JA
1994}
1995
5ddc6707 1996static void *helper_thread_main(void *data)
2e1df07d 1997{
9ec7779f
JA
1998 int ret = 0;
1999
2e1df07d
JA
2000 fio_mutex_up(startup_mutex);
2001
8aab824f
JA
2002 while (!ret) {
2003 uint64_t sec = DISK_UTIL_MSEC / 1000;
2004 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2005 struct timespec ts;
2006 struct timeval tv;
2007
2008 gettimeofday(&tv, NULL);
2009 ts.tv_sec = tv.tv_sec + sec;
2010 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
2011 if (ts.tv_nsec > 1000000000ULL) {
2012 ts.tv_nsec -= 1000000000ULL;
2013 ts.tv_sec++;
2014 }
2015
5ddc6707 2016 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
8aab824f 2017
9ec7779f 2018 ret = update_io_ticks();
2e1df07d 2019
5ddc6707
JA
2020 if (helper_do_stat) {
2021 helper_do_stat = 0;
2022 __show_running_run_stats();
2023 }
2024
2e1df07d
JA
2025 if (!is_backend)
2026 print_thread_status();
2027 }
2028
2029 return NULL;
2030}
2031
5ddc6707 2032static int create_helper_thread(void)
2e1df07d
JA
2033{
2034 int ret;
2035
9ec7779f
JA
2036 setup_disk_util();
2037
5ddc6707
JA
2038 pthread_cond_init(&helper_cond, NULL);
2039 pthread_mutex_init(&helper_lock, NULL);
8aab824f 2040
5ddc6707 2041 ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2e1df07d 2042 if (ret) {
5ddc6707 2043 log_err("Can't create helper thread: %s\n", strerror(ret));
2e1df07d
JA
2044 return 1;
2045 }
2046
2e1df07d
JA
2047 dprint(FD_MUTEX, "wait on startup_mutex\n");
2048 fio_mutex_down(startup_mutex);
2049 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2050 return 0;
2051}
2052
2e1df07d
JA
2053int fio_backend(void)
2054{
2055 struct thread_data *td;
2056 int i;
2057
2058 if (exec_profile) {
2059 if (load_profile(exec_profile))
2060 return 1;
2061 free(exec_profile);
2062 exec_profile = NULL;
2063 }
2064 if (!thread_number)
2065 return 0;
2066
2067 if (write_bw_log) {
aee2ab67
JA
2068 struct log_params p = {
2069 .log_type = IO_LOG_TYPE_BW,
2070 };
2071
2072 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2073 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2074 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2e1df07d
JA
2075 }
2076
521da527 2077 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2e1df07d
JA
2078 if (startup_mutex == NULL)
2079 return 1;
2e1df07d
JA
2080
2081 set_genesis_time();
cef9175e 2082 stat_init();
5ddc6707 2083 create_helper_thread();
2e1df07d
JA
2084
2085 cgroup_list = smalloc(sizeof(*cgroup_list));
2086 INIT_FLIST_HEAD(cgroup_list);
2087
2088 run_threads();
2089
5ddc6707 2090 wait_for_helper_thread_exit();
8aab824f 2091
2e1df07d 2092 if (!fio_abort) {
83f7b64e 2093 __show_run_stats();
2e1df07d 2094 if (write_bw_log) {
cb7e0ace
JA
2095 int i;
2096
2097 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2098 struct io_log *log = agg_io_log[i];
2099
aee2ab67 2100 flush_log(log);
518dac09 2101 free_log(log);
cb7e0ace 2102 }
2e1df07d
JA
2103 }
2104 }
2105
fda2cfac 2106 for_each_td(td, i) {
2e1df07d 2107 fio_options_free(td);
fda2cfac
JA
2108 fio_mutex_remove(td->rusage_sem);
2109 td->rusage_sem = NULL;
2110 }
2e1df07d 2111
a462baef 2112 free_disk_util();
2e1df07d
JA
2113 cgroup_kill(cgroup_list);
2114 sfree(cgroup_list);
2115 sfree(cgroup_mnt);
2116
2117 fio_mutex_remove(startup_mutex);
cef9175e 2118 stat_exit();
2e1df07d
JA
2119 return exit_value;
2120}