Makefile: missing conversion, and oslib/ dep cleanup
[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 37#include <sys/mman.h>
ff6bb260 38#include <math.h>
2e1df07d
JA
39
40#include "fio.h"
a5e0ee11
O
41#ifndef FIO_NO_HAVE_SHM_H
42#include <sys/shm.h>
43#endif
2e1df07d
JA
44#include "hash.h"
45#include "smalloc.h"
46#include "verify.h"
47#include "trim.h"
48#include "diskutil.h"
49#include "cgroup.h"
50#include "profile.h"
51#include "lib/rand.h"
f7690c4a 52#include "lib/memalign.h"
2e1df07d 53#include "server.h"
44404c5a 54#include "lib/getrusage.h"
f2a2ce0e 55#include "idletime.h"
002fe734 56#include "err.h"
78d55e72 57#include "lib/tp.h"
a9da8ab2 58#include "workqueue.h"
e81ecca3 59#include "lib/mountcheck.h"
2e1df07d 60
5ddc6707
JA
61static pthread_t helper_thread;
62static pthread_mutex_t helper_lock;
63pthread_cond_t helper_cond;
64int helper_do_stat = 0;
8aab824f 65
2e1df07d 66static struct fio_mutex *startup_mutex;
2e1df07d
JA
67static struct flist_head *cgroup_list;
68static char *cgroup_mnt;
69static int exit_value;
70static volatile int fio_abort;
3a5f6bde
JA
71static unsigned int nr_process = 0;
72static unsigned int nr_thread = 0;
2e1df07d 73
6eaf09d6 74struct io_log *agg_io_log[DDIR_RWDIR_CNT];
2e1df07d 75
a3efc919
JA
76int groupid = 0;
77unsigned int thread_number = 0;
108fea77 78unsigned int stat_number = 0;
a3efc919
JA
79int shm_id = 0;
80int temp_stall_ts;
81unsigned long done_secs = 0;
5ddc6707 82volatile int helper_exit = 0;
a3efc919 83
2e1df07d 84#define PAGE_ALIGN(buf) \
e43606c2 85 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
2e1df07d
JA
86
87#define JOB_START_TIMEOUT (5 * 1000)
88
89static void sig_int(int sig)
90{
91 if (threads) {
92 if (is_backend)
93 fio_server_got_signal(sig);
94 else {
95 log_info("\nfio: terminating on signal %d\n", sig);
e411c301 96 log_info_flush();
2e1df07d
JA
97 exit_value = 128;
98 }
99
100 fio_terminate_threads(TERMINATE_ALL);
101 }
102}
103
d2235e56 104void sig_show_status(int sig)
4c6d91e8
JA
105{
106 show_running_run_stats();
107}
108
2e1df07d
JA
109static void set_sig_handlers(void)
110{
111 struct sigaction act;
112
113 memset(&act, 0, sizeof(act));
114 act.sa_handler = sig_int;
115 act.sa_flags = SA_RESTART;
116 sigaction(SIGINT, &act, NULL);
117
118 memset(&act, 0, sizeof(act));
119 act.sa_handler = sig_int;
120 act.sa_flags = SA_RESTART;
121 sigaction(SIGTERM, &act, NULL);
122
2f694507
BC
123/* Windows uses SIGBREAK as a quit signal from other applications */
124#ifdef WIN32
125 memset(&act, 0, sizeof(act));
126 act.sa_handler = sig_int;
127 act.sa_flags = SA_RESTART;
128 sigaction(SIGBREAK, &act, NULL);
129#endif
130
4c6d91e8
JA
131 memset(&act, 0, sizeof(act));
132 act.sa_handler = sig_show_status;
133 act.sa_flags = SA_RESTART;
134 sigaction(SIGUSR1, &act, NULL);
135
2e1df07d
JA
136 if (is_backend) {
137 memset(&act, 0, sizeof(act));
138 act.sa_handler = sig_int;
139 act.sa_flags = SA_RESTART;
140 sigaction(SIGPIPE, &act, NULL);
141 }
142}
143
144/*
145 * Check if we are above the minimum rate given.
146 */
147static int __check_min_rate(struct thread_data *td, struct timeval *now,
148 enum fio_ddir ddir)
149{
150 unsigned long long bytes = 0;
151 unsigned long iops = 0;
152 unsigned long spent;
153 unsigned long rate;
154 unsigned int ratemin = 0;
155 unsigned int rate_iops = 0;
156 unsigned int rate_iops_min = 0;
157
158 assert(ddir_rw(ddir));
159
160 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
161 return 0;
162
163 /*
164 * allow a 2 second settle period in the beginning
165 */
166 if (mtime_since(&td->start, now) < 2000)
167 return 0;
168
169 iops += td->this_io_blocks[ddir];
170 bytes += td->this_io_bytes[ddir];
171 ratemin += td->o.ratemin[ddir];
172 rate_iops += td->o.rate_iops[ddir];
173 rate_iops_min += td->o.rate_iops_min[ddir];
174
175 /*
176 * if rate blocks is set, sample is running
177 */
178 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
179 spent = mtime_since(&td->lastrate[ddir], now);
180 if (spent < td->o.ratecycle)
181 return 0;
182
20e37e4a 183 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
2e1df07d
JA
184 /*
185 * check bandwidth specified rate
186 */
187 if (bytes < td->rate_bytes[ddir]) {
188 log_err("%s: min rate %u not met\n", td->o.name,
189 ratemin);
190 return 1;
191 } else {
49cba9b3
JA
192 if (spent)
193 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
194 else
195 rate = 0;
196
2e1df07d
JA
197 if (rate < ratemin ||
198 bytes < td->rate_bytes[ddir]) {
199 log_err("%s: min rate %u not met, got"
200 " %luKB/sec\n", td->o.name,
201 ratemin, rate);
202 return 1;
203 }
204 }
205 } else {
206 /*
207 * checks iops specified rate
208 */
209 if (iops < rate_iops) {
210 log_err("%s: min iops rate %u not met\n",
211 td->o.name, rate_iops);
212 return 1;
213 } else {
4c707a3b
JA
214 if (spent)
215 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
216 else
217 rate = 0;
218
2e1df07d
JA
219 if (rate < rate_iops_min ||
220 iops < td->rate_blocks[ddir]) {
221 log_err("%s: min iops rate %u not met,"
222 " got %lu\n", td->o.name,
223 rate_iops_min, rate);
20e37e4a 224 return 1;
2e1df07d
JA
225 }
226 }
227 }
228 }
229
230 td->rate_bytes[ddir] = bytes;
231 td->rate_blocks[ddir] = iops;
232 memcpy(&td->lastrate[ddir], now, sizeof(*now));
233 return 0;
234}
235
55312f9f 236static int check_min_rate(struct thread_data *td, struct timeval *now)
2e1df07d
JA
237{
238 int ret = 0;
239
55312f9f 240 if (td->bytes_done[DDIR_READ])
6eaf09d6 241 ret |= __check_min_rate(td, now, DDIR_READ);
55312f9f 242 if (td->bytes_done[DDIR_WRITE])
6eaf09d6 243 ret |= __check_min_rate(td, now, DDIR_WRITE);
55312f9f 244 if (td->bytes_done[DDIR_TRIM])
6eaf09d6 245 ret |= __check_min_rate(td, now, DDIR_TRIM);
2e1df07d
JA
246
247 return ret;
248}
249
250/*
251 * When job exits, we can cancel the in-flight IO if we are using async
252 * io. Attempt to do so.
253 */
254static void cleanup_pending_aio(struct thread_data *td)
255{
2e1df07d
JA
256 int r;
257
258 /*
259 * get immediately available events, if any
260 */
55312f9f 261 r = io_u_queued_complete(td, 0);
2e1df07d
JA
262 if (r < 0)
263 return;
264
265 /*
266 * now cancel remaining active events
267 */
268 if (td->io_ops->cancel) {
2ae0b204
JA
269 struct io_u *io_u;
270 int i;
2e1df07d 271
2ae0b204
JA
272 io_u_qiter(&td->io_u_all, io_u, i) {
273 if (io_u->flags & IO_U_F_FLIGHT) {
2e1df07d
JA
274 r = td->io_ops->cancel(td, io_u);
275 if (!r)
276 put_io_u(td, io_u);
277 }
278 }
279 }
280
281 if (td->cur_depth)
55312f9f 282 r = io_u_queued_complete(td, td->cur_depth);
2e1df07d
JA
283}
284
285/*
286 * Helper to handle the final sync of a file. Works just like the normal
287 * io path, just does everything sync.
288 */
289static int fio_io_sync(struct thread_data *td, struct fio_file *f)
290{
291 struct io_u *io_u = __get_io_u(td);
292 int ret;
293
294 if (!io_u)
295 return 1;
296
297 io_u->ddir = DDIR_SYNC;
298 io_u->file = f;
299
300 if (td_io_prep(td, io_u)) {
301 put_io_u(td, io_u);
302 return 1;
303 }
304
305requeue:
306 ret = td_io_queue(td, io_u);
307 if (ret < 0) {
308 td_verror(td, io_u->error, "td_io_queue");
309 put_io_u(td, io_u);
310 return 1;
311 } else if (ret == FIO_Q_QUEUED) {
55312f9f 312 if (io_u_queued_complete(td, 1) < 0)
2e1df07d
JA
313 return 1;
314 } else if (ret == FIO_Q_COMPLETED) {
315 if (io_u->error) {
316 td_verror(td, io_u->error, "td_io_queue");
317 return 1;
318 }
319
55312f9f 320 if (io_u_sync_complete(td, io_u) < 0)
2e1df07d
JA
321 return 1;
322 } else if (ret == FIO_Q_BUSY) {
323 if (td_io_commit(td))
324 return 1;
325 goto requeue;
326 }
327
328 return 0;
329}
a3efc919 330
61ee0f86
JA
331static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
332{
333 int ret;
334
335 if (fio_file_open(f))
336 return fio_io_sync(td, f);
337
338 if (td_io_open_file(td, f))
339 return 1;
340
341 ret = fio_io_sync(td, f);
342 td_io_close_file(td, f);
343 return ret;
344}
345
2e1df07d
JA
346static inline void __update_tv_cache(struct thread_data *td)
347{
348 fio_gettime(&td->tv_cache, NULL);
349}
350
351static inline void update_tv_cache(struct thread_data *td)
352{
353 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
354 __update_tv_cache(td);
355}
356
357static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
358{
359 if (in_ramp_time(td))
360 return 0;
361 if (!td->o.timeout)
362 return 0;
0de5b26f 363 if (utime_since(&td->epoch, t) >= td->o.timeout)
2e1df07d
JA
364 return 1;
365
366 return 0;
367}
368
95603b74
BF
369/*
370 * We need to update the runtime consistently in ms, but keep a running
371 * tally of the current elapsed time in microseconds for sub millisecond
372 * updates.
373 */
374static inline void update_runtime(struct thread_data *td,
375 unsigned long long *elapsed_us,
376 const enum fio_ddir ddir)
377{
d5c8ea29
JA
378 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
379 return;
380
95603b74
BF
381 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
382 elapsed_us[ddir] += utime_since_now(&td->start);
383 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
384}
385
2e1df07d
JA
386static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
387 int *retptr)
388{
389 int ret = *retptr;
390
391 if (ret < 0 || td->error) {
8b28bd41
DM
392 int err = td->error;
393 enum error_type_bit eb;
2e1df07d
JA
394
395 if (ret < 0)
396 err = -ret;
2e1df07d 397
8b28bd41
DM
398 eb = td_error_type(ddir, err);
399 if (!(td->o.continue_on_error & (1 << eb)))
2e1df07d
JA
400 return 1;
401
8b28bd41 402 if (td_non_fatal_error(td, eb, err)) {
2e1df07d
JA
403 /*
404 * Continue with the I/Os in case of
405 * a non fatal error.
406 */
407 update_error_count(td, err);
408 td_clear_error(td);
409 *retptr = 0;
410 return 0;
411 } else if (td->o.fill_device && err == ENOSPC) {
412 /*
413 * We expect to hit this error if
414 * fill_device option is set.
415 */
416 td_clear_error(td);
ebea2133 417 fio_mark_td_terminate(td);
2e1df07d
JA
418 return 1;
419 } else {
420 /*
421 * Stop the I/O in case of a fatal
422 * error.
423 */
424 update_error_count(td, err);
425 return 1;
426 }
427 }
428
429 return 0;
430}
431
c97f1ad6
JA
432static void check_update_rusage(struct thread_data *td)
433{
434 if (td->update_rusage) {
435 td->update_rusage = 0;
436 update_rusage_stat(td);
437 fio_mutex_up(td->rusage_sem);
438 }
439}
440
55312f9f 441static int wait_for_completions(struct thread_data *td, struct timeval *time)
69fea81e
JA
442{
443 const int full = queue_full(td);
444 int min_evts = 0;
445 int ret;
446
447 /*
448 * if the queue is full, we MUST reap at least 1 event
449 */
82407585 450 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
4b157ac6 451 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
69fea81e
JA
452 min_evts = 1;
453
66b6c5ef 454 if (time && (__should_check_rate(td, DDIR_READ) ||
69fea81e
JA
455 __should_check_rate(td, DDIR_WRITE) ||
456 __should_check_rate(td, DDIR_TRIM)))
457 fio_gettime(time, NULL);
458
459 do {
55312f9f 460 ret = io_u_queued_complete(td, min_evts);
69fea81e
JA
461 if (ret < 0)
462 break;
463 } while (full && (td->cur_depth > td->o.iodepth_low));
464
465 return ret;
466}
467
e9d512d8
JA
468int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
469 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
470 struct timeval *comp_time)
471{
472 int ret2;
473
474 switch (*ret) {
475 case FIO_Q_COMPLETED:
476 if (io_u->error) {
477 *ret = -io_u->error;
478 clear_io_u(td, io_u);
479 } else if (io_u->resid) {
480 int bytes = io_u->xfer_buflen - io_u->resid;
481 struct fio_file *f = io_u->file;
482
483 if (bytes_issued)
484 *bytes_issued += bytes;
485
486 if (!from_verify)
487 trim_io_piece(td, io_u);
488
489 /*
490 * zero read, fail
491 */
492 if (!bytes) {
493 if (!from_verify)
494 unlog_io_piece(td, io_u);
495 td_verror(td, EIO, "full resid");
496 put_io_u(td, io_u);
497 break;
498 }
499
500 io_u->xfer_buflen = io_u->resid;
501 io_u->xfer_buf += bytes;
502 io_u->offset += bytes;
503
504 if (ddir_rw(io_u->ddir))
505 td->ts.short_io_u[io_u->ddir]++;
506
507 f = io_u->file;
508 if (io_u->offset == f->real_file_size)
509 goto sync_done;
510
511 requeue_io_u(td, &io_u);
512 } else {
513sync_done:
514 if (comp_time && (__should_check_rate(td, DDIR_READ) ||
515 __should_check_rate(td, DDIR_WRITE) ||
516 __should_check_rate(td, DDIR_TRIM)))
517 fio_gettime(comp_time, NULL);
518
519 *ret = io_u_sync_complete(td, io_u);
520 if (*ret < 0)
521 break;
522 }
523 return 0;
524 case FIO_Q_QUEUED:
525 /*
526 * if the engine doesn't have a commit hook,
527 * the io_u is really queued. if it does have such
528 * a hook, it has to call io_u_queued() itself.
529 */
530 if (td->io_ops->commit == NULL)
531 io_u_queued(td, io_u);
532 if (bytes_issued)
533 *bytes_issued += io_u->xfer_buflen;
534 break;
535 case FIO_Q_BUSY:
536 if (!from_verify)
537 unlog_io_piece(td, io_u);
538 requeue_io_u(td, &io_u);
539 ret2 = td_io_commit(td);
540 if (ret2 < 0)
541 *ret = ret2;
542 break;
543 default:
b0775325 544 assert(*ret < 0);
e9d512d8
JA
545 td_verror(td, -(*ret), "td_io_queue");
546 break;
547 }
548
549 if (break_on_this_error(td, ddir, ret))
550 return 1;
551
552 return 0;
553}
554
82407585
RP
555static inline int io_in_polling(struct thread_data *td)
556{
557 return !td->o.iodepth_batch_complete_min &&
558 !td->o.iodepth_batch_complete_max;
559}
560
2e1df07d
JA
561/*
562 * The main verify engine. Runs over the writes we previously submitted,
563 * reads the blocks back in, and checks the crc/md5 of the data.
564 */
100f49f1 565static void do_verify(struct thread_data *td, uint64_t verify_bytes)
2e1df07d
JA
566{
567 struct fio_file *f;
568 struct io_u *io_u;
569 int ret, min_events;
570 unsigned int i;
571
572 dprint(FD_VERIFY, "starting loop\n");
573
574 /*
575 * sync io first and invalidate cache, to make sure we really
576 * read from disk.
577 */
578 for_each_file(td, f, i) {
579 if (!fio_file_open(f))
580 continue;
581 if (fio_io_sync(td, f))
582 break;
583 if (file_invalidate_cache(td, f))
584 break;
585 }
586
c97f1ad6
JA
587 check_update_rusage(td);
588
2e1df07d
JA
589 if (td->error)
590 return;
591
592 td_set_runstate(td, TD_VERIFYING);
593
594 io_u = NULL;
595 while (!td->terminate) {
fbccf46c 596 enum fio_ddir ddir;
e9d512d8 597 int full;
2e1df07d
JA
598
599 update_tv_cache(td);
c97f1ad6 600 check_update_rusage(td);
2e1df07d
JA
601
602 if (runtime_exceeded(td, &td->tv_cache)) {
603 __update_tv_cache(td);
604 if (runtime_exceeded(td, &td->tv_cache)) {
ebea2133 605 fio_mark_td_terminate(td);
2e1df07d
JA
606 break;
607 }
608 }
609
9e684a49
DE
610 if (flow_threshold_exceeded(td))
611 continue;
612
44cbc6da
JA
613 if (!td->o.experimental_verify) {
614 io_u = __get_io_u(td);
615 if (!io_u)
616 break;
2e1df07d 617
44cbc6da
JA
618 if (get_next_verify(td, io_u)) {
619 put_io_u(td, io_u);
620 break;
621 }
2e1df07d 622
44cbc6da
JA
623 if (td_io_prep(td, io_u)) {
624 put_io_u(td, io_u);
625 break;
626 }
627 } else {
55312f9f 628 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
100f49f1
JA
629 break;
630
bcd5abfa 631 while ((io_u = get_io_u(td)) != NULL) {
002fe734
JA
632 if (IS_ERR(io_u)) {
633 io_u = NULL;
634 ret = FIO_Q_BUSY;
635 goto reap;
636 }
637
bcd5abfa
JA
638 /*
639 * We are only interested in the places where
640 * we wrote or trimmed IOs. Turn those into
641 * reads for verification purposes.
642 */
643 if (io_u->ddir == DDIR_READ) {
644 /*
645 * Pretend we issued it for rwmix
646 * accounting
647 */
648 td->io_issues[DDIR_READ]++;
649 put_io_u(td, io_u);
650 continue;
651 } else if (io_u->ddir == DDIR_TRIM) {
652 io_u->ddir = DDIR_READ;
a9da8ab2 653 io_u_set(io_u, IO_U_F_TRIMMED);
bcd5abfa
JA
654 break;
655 } else if (io_u->ddir == DDIR_WRITE) {
656 io_u->ddir = DDIR_READ;
657 break;
658 } else {
659 put_io_u(td, io_u);
660 continue;
661 }
662 }
44cbc6da 663
bcd5abfa 664 if (!io_u)
44cbc6da 665 break;
2e1df07d
JA
666 }
667
ca09be4b
JA
668 if (verify_state_should_stop(td, io_u)) {
669 put_io_u(td, io_u);
670 break;
671 }
672
2e1df07d
JA
673 if (td->o.verify_async)
674 io_u->end_io = verify_io_u_async;
675 else
676 io_u->end_io = verify_io_u;
677
fbccf46c 678 ddir = io_u->ddir;
10adc4a7
GG
679 if (!td->o.disable_slat)
680 fio_gettime(&io_u->start_time, NULL);
fbccf46c 681
2e1df07d 682 ret = td_io_queue(td, io_u);
2e1df07d 683
e9d512d8 684 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
2e1df07d
JA
685 break;
686
687 /*
688 * if we can queue more, do so. but check if there are
689 * completed io_u's first. Note that we can get BUSY even
690 * without IO queued, if the system is resource starved.
691 */
002fe734 692reap:
2e1df07d 693 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
82407585 694 if (full || io_in_polling(td))
55312f9f 695 ret = wait_for_completions(td, NULL);
2e1df07d 696
2e1df07d
JA
697 if (ret < 0)
698 break;
699 }
700
c97f1ad6
JA
701 check_update_rusage(td);
702
2e1df07d
JA
703 if (!td->error) {
704 min_events = td->cur_depth;
705
706 if (min_events)
55312f9f 707 ret = io_u_queued_complete(td, min_events);
2e1df07d
JA
708 } else
709 cleanup_pending_aio(td);
710
711 td_set_runstate(td, TD_RUNNING);
712
713 dprint(FD_VERIFY, "exiting loop\n");
714}
715
3939fe85
JA
716static unsigned int exceeds_number_ios(struct thread_data *td)
717{
718 unsigned long long number_ios;
719
720 if (!td->o.number_ios)
721 return 0;
722
cf8a46a7 723 number_ios = ddir_rw_sum(td->io_blocks);
3939fe85
JA
724 number_ios += td->io_u_queued + td->io_u_in_flight;
725
cf8a46a7 726 return number_ios >= (td->o.number_ios * td->loops);
3939fe85
JA
727}
728
e28dd2cf 729static int io_issue_bytes_exceeded(struct thread_data *td)
f7078f7b 730{
77731b29 731 unsigned long long bytes, limit;
f7078f7b
JA
732
733 if (td_rw(td))
74d6277f 734 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
f7078f7b 735 else if (td_write(td))
74d6277f 736 bytes = td->io_issue_bytes[DDIR_WRITE];
6eaf09d6 737 else if (td_read(td))
74d6277f 738 bytes = td->io_issue_bytes[DDIR_READ];
f7078f7b 739 else
74d6277f 740 bytes = td->io_issue_bytes[DDIR_TRIM];
f7078f7b 741
77731b29
JA
742 if (td->o.io_limit)
743 limit = td->o.io_limit;
744 else
745 limit = td->o.size;
746
cf8a46a7 747 limit *= td->loops;
77731b29 748 return bytes >= limit || exceeds_number_ios(td);
f7078f7b
JA
749}
750
e28dd2cf
JE
751static int io_complete_bytes_exceeded(struct thread_data *td)
752{
753 unsigned long long bytes, limit;
754
755 if (td_rw(td))
756 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
757 else if (td_write(td))
758 bytes = td->this_io_bytes[DDIR_WRITE];
759 else if (td_read(td))
760 bytes = td->this_io_bytes[DDIR_READ];
761 else
762 bytes = td->this_io_bytes[DDIR_TRIM];
763
764 if (td->o.io_limit)
765 limit = td->o.io_limit;
766 else
767 limit = td->o.size;
768
cf8a46a7 769 limit *= td->loops;
e28dd2cf
JE
770 return bytes >= limit || exceeds_number_ios(td);
771}
772
50a8ce86
D
773/*
774 * used to calculate the next io time for rate control
775 *
776 */
777static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
778{
ff6bb260 779 uint64_t secs, remainder, bps, bytes, iops;
50a8ce86
D
780
781 assert(!(td->flags & TD_F_CHILD));
782 bytes = td->rate_io_issue_bytes[ddir];
783 bps = td->rate_bps[ddir];
ff6bb260 784
6de65959 785 if (td->o.rate_process == RATE_PROCESS_POISSON) {
a77d139b 786 uint64_t val;
ff6bb260 787 iops = bps / td->o.bs[ddir];
a77d139b
JA
788 val = (int64_t) (1000000 / iops) *
789 -logf(__rand_0_1(&td->poisson_state));
790 if (val) {
791 dprint(FD_RATE, "poisson rate iops=%llu\n",
792 (unsigned long long) 1000000 / val);
793 }
794 td->last_usec += val;
ff6bb260
SL
795 return td->last_usec;
796 } else if (bps) {
50a8ce86
D
797 secs = bytes / bps;
798 remainder = bytes % bps;
799 return remainder * 1000000 / bps + secs * 1000000;
e7b24047
JA
800 }
801
802 return 0;
50a8ce86
D
803}
804
2e1df07d
JA
805/*
806 * Main IO worker function. It retrieves io_u's to process and queues
807 * and reaps them, checking for rate and errors along the way.
100f49f1
JA
808 *
809 * Returns number of bytes written and trimmed.
2e1df07d 810 */
100f49f1 811static uint64_t do_io(struct thread_data *td)
2e1df07d
JA
812{
813 unsigned int i;
814 int ret = 0;
c2703bf3 815 uint64_t total_bytes, bytes_issued = 0;
2e1df07d
JA
816
817 if (in_ramp_time(td))
818 td_set_runstate(td, TD_RAMP);
819 else
820 td_set_runstate(td, TD_RUNNING);
821
3e260a46
JA
822 lat_target_init(td);
823
1e564979
JE
824 total_bytes = td->o.size;
825 /*
826 * Allow random overwrite workloads to write up to io_limit
827 * before starting verification phase as 'size' doesn't apply.
828 */
829 if (td_write(td) && td_random(td) && td->o.norandommap)
830 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
78a6469c
JA
831 /*
832 * If verify_backlog is enabled, we'll run the verify in this
833 * handler as well. For that case, we may need up to twice the
834 * amount of bytes.
835 */
78a6469c
JA
836 if (td->o.verify != VERIFY_NONE &&
837 (td_write(td) && td->o.verify_backlog))
c2703bf3
JA
838 total_bytes += td->o.size;
839
82a90686 840 /* In trimwrite mode, each byte is trimmed and then written, so
0e4dd95c 841 * allow total_bytes to be twice as big */
82a90686 842 if (td_trimwrite(td))
0e4dd95c
DE
843 total_bytes += td->total_io_size;
844
f7078f7b 845 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
e28dd2cf 846 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
c04e4661 847 td->o.time_based) {
2e1df07d 848 struct timeval comp_time;
2e1df07d 849 struct io_u *io_u;
e9d512d8 850 int full;
2e1df07d
JA
851 enum fio_ddir ddir;
852
c97f1ad6
JA
853 check_update_rusage(td);
854
7d7803fa 855 if (td->terminate || td->done)
2e1df07d
JA
856 break;
857
858 update_tv_cache(td);
859
860 if (runtime_exceeded(td, &td->tv_cache)) {
861 __update_tv_cache(td);
862 if (runtime_exceeded(td, &td->tv_cache)) {
ebea2133 863 fio_mark_td_terminate(td);
2e1df07d
JA
864 break;
865 }
866 }
867
9e684a49
DE
868 if (flow_threshold_exceeded(td))
869 continue;
870
c82ea3d4 871 if (!td->o.time_based && bytes_issued >= total_bytes)
20876c53
JC
872 break;
873
2e1df07d 874 io_u = get_io_u(td);
002fe734
JA
875 if (IS_ERR_OR_NULL(io_u)) {
876 int err = PTR_ERR(io_u);
877
878 io_u = NULL;
879 if (err == -EBUSY) {
880 ret = FIO_Q_BUSY;
881 goto reap;
882 }
3e260a46
JA
883 if (td->o.latency_target)
884 goto reap;
2e1df07d 885 break;
3e260a46 886 }
2e1df07d
JA
887
888 ddir = io_u->ddir;
889
890 /*
82af2a7c
JA
891 * Add verification end_io handler if:
892 * - Asked to verify (!td_rw(td))
893 * - Or the io_u is from our verify list (mixed write/ver)
2e1df07d
JA
894 */
895 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
82af2a7c 896 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
c4b6117b
PV
897
898 if (!td->o.verify_pattern_bytes) {
d6b72507 899 io_u->rand_seed = __rand(&td->verify_state);
c4b6117b 900 if (sizeof(int) != sizeof(long *))
d6b72507 901 io_u->rand_seed *= __rand(&td->verify_state);
c4b6117b
PV
902 }
903
ca09be4b
JA
904 if (verify_state_should_stop(td, io_u)) {
905 put_io_u(td, io_u);
906 break;
907 }
908
2e1df07d
JA
909 if (td->o.verify_async)
910 io_u->end_io = verify_io_u_async;
911 else
912 io_u->end_io = verify_io_u;
913 td_set_runstate(td, TD_VERIFYING);
914 } else if (in_ramp_time(td))
915 td_set_runstate(td, TD_RAMP);
916 else
917 td_set_runstate(td, TD_RUNNING);
918
9a50c5c5 919 /*
f9401285
JA
920 * Always log IO before it's issued, so we know the specific
921 * order of it. The logged unit will track when the IO has
922 * completed.
9a50c5c5 923 */
c4b6117b
PV
924 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
925 td->o.do_verify &&
926 td->o.verify != VERIFY_NONE &&
f9401285 927 !td->o.experimental_verify)
c4b6117b
PV
928 log_io_piece(td, io_u);
929
a9da8ab2
JA
930 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
931 if (td->error)
932 break;
933 ret = workqueue_enqueue(&td->io_wq, io_u);
50a8ce86
D
934
935 if (should_check_rate(td))
936 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
937
a9da8ab2
JA
938 } else {
939 ret = td_io_queue(td, io_u);
2e1df07d 940
50a8ce86
D
941 if (should_check_rate(td))
942 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
943
fd727d9d 944 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
a9da8ab2 945 break;
2e1df07d 946
a9da8ab2
JA
947 /*
948 * See if we need to complete some commands. Note that
949 * we can get BUSY even without IO queued, if the
950 * system is resource starved.
951 */
3e260a46 952reap:
a9da8ab2
JA
953 full = queue_full(td) ||
954 (ret == FIO_Q_BUSY && td->cur_depth);
82407585 955 if (full || io_in_polling(td))
a9da8ab2
JA
956 ret = wait_for_completions(td, &comp_time);
957 }
2e1df07d
JA
958 if (ret < 0)
959 break;
55312f9f
JA
960 if (!ddir_rw_sum(td->bytes_done) &&
961 !(td->io_ops->flags & FIO_NOIO))
2e1df07d
JA
962 continue;
963
55312f9f
JA
964 if (!in_ramp_time(td) && should_check_rate(td)) {
965 if (check_min_rate(td, &comp_time)) {
2e1df07d
JA
966 if (exitall_on_terminate)
967 fio_terminate_threads(td->groupid);
968 td_verror(td, EIO, "check_min_rate");
969 break;
970 }
971 }
3e260a46
JA
972 if (!in_ramp_time(td) && td->o.latency_target)
973 lat_target_check(td);
e155cb64 974
2e1df07d
JA
975 if (td->o.thinktime) {
976 unsigned long long b;
977
342f4be4 978 b = ddir_rw_sum(td->io_blocks);
2e1df07d
JA
979 if (!(b % td->o.thinktime_blocks)) {
980 int left;
981
002e7183
JA
982 io_u_quiesce(td);
983
2e1df07d
JA
984 if (td->o.thinktime_spin)
985 usec_spin(td->o.thinktime_spin);
986
987 left = td->o.thinktime - td->o.thinktime_spin;
988 if (left)
989 usec_sleep(td, left);
990 }
991 }
992 }
993
c97f1ad6
JA
994 check_update_rusage(td);
995
2e1df07d 996 if (td->trim_entries)
4e0a8fa2 997 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
2e1df07d
JA
998
999 if (td->o.fill_device && td->error == ENOSPC) {
1000 td->error = 0;
ebea2133 1001 fio_mark_td_terminate(td);
2e1df07d
JA
1002 }
1003 if (!td->error) {
1004 struct fio_file *f;
1005
a9da8ab2
JA
1006 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1007 workqueue_flush(&td->io_wq);
1008 i = 0;
1009 } else
1010 i = td->cur_depth;
1011
2e1df07d 1012 if (i) {
55312f9f 1013 ret = io_u_queued_complete(td, i);
2e1df07d
JA
1014 if (td->o.fill_device && td->error == ENOSPC)
1015 td->error = 0;
1016 }
1017
1018 if (should_fsync(td) && td->o.end_fsync) {
1019 td_set_runstate(td, TD_FSYNCING);
1020
1021 for_each_file(td, f, i) {
61ee0f86 1022 if (!fio_file_fsync(td, f))
2e1df07d 1023 continue;
61ee0f86
JA
1024
1025 log_err("fio: end_fsync failed for file %s\n",
1026 f->file_name);
2e1df07d
JA
1027 }
1028 }
1029 } else
1030 cleanup_pending_aio(td);
1031
1032 /*
1033 * stop job if we failed doing any IO
1034 */
342f4be4 1035 if (!ddir_rw_sum(td->this_io_bytes))
2e1df07d 1036 td->done = 1;
100f49f1 1037
55312f9f 1038 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
2e1df07d
JA
1039}
1040
1041static void cleanup_io_u(struct thread_data *td)
1042{
2e1df07d
JA
1043 struct io_u *io_u;
1044
2ae0b204 1045 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
c73ed246
JA
1046
1047 if (td->io_ops->io_u_free)
1048 td->io_ops->io_u_free(td, io_u);
1049
2e1df07d
JA
1050 fio_memfree(io_u, sizeof(*io_u));
1051 }
1052
1053 free_io_mem(td);
2ae0b204
JA
1054
1055 io_u_rexit(&td->io_u_requeues);
1056 io_u_qexit(&td->io_u_freelist);
1057 io_u_qexit(&td->io_u_all);
ca09be4b
JA
1058
1059 if (td->last_write_comp)
1060 sfree(td->last_write_comp);
2e1df07d
JA
1061}
1062
1063static int init_io_u(struct thread_data *td)
1064{
1065 struct io_u *io_u;
9c42684e 1066 unsigned int max_bs, min_write;
2e1df07d 1067 int cl_align, i, max_units;
2ae0b204 1068 int data_xfer = 1, err;
2e1df07d
JA
1069 char *p;
1070
1071 max_units = td->o.iodepth;
74f4b020 1072 max_bs = td_max_bs(td);
9c42684e 1073 min_write = td->o.min_bs[DDIR_WRITE];
2e1df07d
JA
1074 td->orig_buffer_size = (unsigned long long) max_bs
1075 * (unsigned long long) max_units;
1076
88045e04 1077 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
59d8d0f5
JA
1078 data_xfer = 0;
1079
2ae0b204
JA
1080 err = 0;
1081 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1082 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1083 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1084
1085 if (err) {
1086 log_err("fio: failed setting up IO queues\n");
1087 return 1;
1088 }
1089
fd8a09b8 1090 /*
1091 * if we may later need to do address alignment, then add any
1092 * possible adjustment here so that we don't cause a buffer
1093 * overflow later. this adjustment may be too much if we get
1094 * lucky and the allocator gives us an aligned address.
1095 */
d01612f3
CM
1096 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1097 (td->io_ops->flags & FIO_RAWIO))
fd8a09b8 1098 td->orig_buffer_size += page_mask + td->o.mem_align;
1099
2e1df07d
JA
1100 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1101 unsigned long bs;
1102
1103 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1104 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1105 }
1106
1107 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1108 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1109 return 1;
1110 }
1111
59d8d0f5 1112 if (data_xfer && allocate_io_mem(td))
2e1df07d
JA
1113 return 1;
1114
d01612f3 1115 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
2e1df07d
JA
1116 (td->io_ops->flags & FIO_RAWIO))
1117 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1118 else
1119 p = td->orig_buffer;
1120
1121 cl_align = os_cache_line_size();
1122
1123 for (i = 0; i < max_units; i++) {
1124 void *ptr;
1125
1126 if (td->terminate)
1127 return 1;
1128
1129 ptr = fio_memalign(cl_align, sizeof(*io_u));
1130 if (!ptr) {
1131 log_err("fio: unable to allocate aligned memory\n");
1132 break;
1133 }
1134
1135 io_u = ptr;
1136 memset(io_u, 0, sizeof(*io_u));
2ae0b204 1137 INIT_FLIST_HEAD(&io_u->verify_list);
2e1df07d
JA
1138 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1139
59d8d0f5 1140 if (data_xfer) {
2e1df07d
JA
1141 io_u->buf = p;
1142 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1143
1144 if (td_write(td))
9c42684e 1145 io_u_fill_buffer(td, io_u, min_write, max_bs);
2e1df07d
JA
1146 if (td_write(td) && td->o.verify_pattern_bytes) {
1147 /*
1148 * Fill the buffer with the pattern if we are
1149 * going to be doing writes.
1150 */
ce35b1ec 1151 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
2e1df07d
JA
1152 }
1153 }
1154
1155 io_u->index = i;
1156 io_u->flags = IO_U_F_FREE;
2ae0b204
JA
1157 io_u_qpush(&td->io_u_freelist, io_u);
1158
1159 /*
1160 * io_u never leaves this stack, used for iteration of all
1161 * io_u buffers.
1162 */
1163 io_u_qpush(&td->io_u_all, io_u);
c73ed246
JA
1164
1165 if (td->io_ops->io_u_init) {
1166 int ret = td->io_ops->io_u_init(td, io_u);
1167
1168 if (ret) {
1169 log_err("fio: failed to init engine data: %d\n", ret);
1170 return 1;
1171 }
1172 }
1173
2e1df07d
JA
1174 p += max_bs;
1175 }
1176
ca09be4b
JA
1177 if (td->o.verify != VERIFY_NONE) {
1178 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1179 if (!td->last_write_comp) {
1180 log_err("fio: failed to alloc write comp data\n");
1181 return 1;
1182 }
1183 }
1184
2e1df07d
JA
1185 return 0;
1186}
1187
1188static int switch_ioscheduler(struct thread_data *td)
1189{
1190 char tmp[256], tmp2[128];
1191 FILE *f;
1192 int ret;
1193
1194 if (td->io_ops->flags & FIO_DISKLESSIO)
1195 return 0;
1196
1197 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1198
1199 f = fopen(tmp, "r+");
1200 if (!f) {
1201 if (errno == ENOENT) {
1202 log_err("fio: os or kernel doesn't support IO scheduler"
1203 " switching\n");
1204 return 0;
1205 }
1206 td_verror(td, errno, "fopen iosched");
1207 return 1;
1208 }
1209
1210 /*
1211 * Set io scheduler.
1212 */
1213 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1214 if (ferror(f) || ret != 1) {
1215 td_verror(td, errno, "fwrite");
1216 fclose(f);
1217 return 1;
1218 }
1219
1220 rewind(f);
1221
1222 /*
1223 * Read back and check that the selected scheduler is now the default.
1224 */
b44b9e45 1225 memset(tmp, 0, sizeof(tmp));
49c6f33d 1226 ret = fread(tmp, sizeof(tmp), 1, f);
2e1df07d
JA
1227 if (ferror(f) || ret < 0) {
1228 td_verror(td, errno, "fread");
1229 fclose(f);
1230 return 1;
1231 }
b44b9e45
TK
1232 /*
1233 * either a list of io schedulers or "none\n" is expected.
1234 */
1235 tmp[strlen(tmp) - 1] = '\0';
49c6f33d 1236
2e1df07d
JA
1237
1238 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1239 if (!strstr(tmp, tmp2)) {
1240 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1241 td_verror(td, EINVAL, "iosched_switch");
1242 fclose(f);
1243 return 1;
1244 }
1245
1246 fclose(f);
1247 return 0;
1248}
1249
1250static int keep_running(struct thread_data *td)
1251{
77731b29
JA
1252 unsigned long long limit;
1253
2e1df07d
JA
1254 if (td->done)
1255 return 0;
1256 if (td->o.time_based)
1257 return 1;
1258 if (td->o.loops) {
1259 td->o.loops--;
1260 return 1;
1261 }
3939fe85
JA
1262 if (exceeds_number_ios(td))
1263 return 0;
26251d8d 1264
77731b29
JA
1265 if (td->o.io_limit)
1266 limit = td->o.io_limit;
1267 else
1268 limit = td->o.size;
1269
1270 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
5bd5f71a
JA
1271 uint64_t diff;
1272
1273 /*
1274 * If the difference is less than the minimum IO size, we
1275 * are done.
1276 */
77731b29 1277 diff = limit - ddir_rw_sum(td->io_bytes);
74f4b020 1278 if (diff < td_max_bs(td))
5bd5f71a
JA
1279 return 0;
1280
002fe734
JA
1281 if (fio_files_done(td))
1282 return 0;
1283
2e1df07d 1284 return 1;
5bd5f71a 1285 }
2e1df07d
JA
1286
1287 return 0;
1288}
1289
ce486495 1290static int exec_string(struct thread_options *o, const char *string, const char *mode)
2e1df07d 1291{
6c9ce469 1292 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
f491a907 1293 int ret;
2e1df07d
JA
1294 char *str;
1295
1296 str = malloc(newlen);
ce486495 1297 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
2e1df07d 1298
ce486495 1299 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
2e1df07d
JA
1300 ret = system(str);
1301 if (ret == -1)
1302 log_err("fio: exec of cmd <%s> failed\n", str);
1303
1304 free(str);
1305 return ret;
1306}
1307
62167762
JC
1308/*
1309 * Dry run to compute correct state of numberio for verification.
1310 */
1311static uint64_t do_dry_run(struct thread_data *td)
1312{
62167762
JC
1313 td_set_runstate(td, TD_RUNNING);
1314
1315 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
e28dd2cf 1316 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
62167762
JC
1317 struct io_u *io_u;
1318 int ret;
1319
1320 if (td->terminate || td->done)
1321 break;
1322
1323 io_u = get_io_u(td);
1324 if (!io_u)
1325 break;
1326
a9da8ab2 1327 io_u_set(io_u, IO_U_F_FLIGHT);
62167762
JC
1328 io_u->error = 0;
1329 io_u->resid = 0;
1330 if (ddir_rw(acct_ddir(io_u)))
1331 td->io_issues[acct_ddir(io_u)]++;
1332 if (ddir_rw(io_u->ddir)) {
1333 io_u_mark_depth(td, 1);
1334 td->ts.total_io_u[io_u->ddir]++;
1335 }
1336
2e63e96b
PV
1337 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1338 td->o.do_verify &&
1339 td->o.verify != VERIFY_NONE &&
1340 !td->o.experimental_verify)
1341 log_io_piece(td, io_u);
1342
55312f9f 1343 ret = io_u_sync_complete(td, io_u);
62167762
JC
1344 (void) ret;
1345 }
1346
55312f9f 1347 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
62167762
JC
1348}
1349
a9da8ab2
JA
1350static void io_workqueue_fn(struct thread_data *td, struct io_u *io_u)
1351{
1352 const enum fio_ddir ddir = io_u->ddir;
1353 int ret;
1354
1355 dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
1356
1357 io_u_set(io_u, IO_U_F_NO_FILE_PUT);
1358
1359 td->cur_depth++;
1360
1361 ret = td_io_queue(td, io_u);
1362
1363 dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
1364
1365 io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
1366
1367 if (ret == FIO_Q_QUEUED)
1368 ret = io_u_queued_complete(td, 1);
1369
1370 td->cur_depth--;
1371}
1372
2e1df07d
JA
1373/*
1374 * Entry point for the thread based jobs. The process based jobs end up
1375 * here as well, after a little setup.
1376 */
1377static void *thread_main(void *data)
1378{
95603b74 1379 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
2e1df07d 1380 struct thread_data *td = data;
4896473e 1381 struct thread_options *o = &td->o;
2e1df07d
JA
1382 pthread_condattr_t attr;
1383 int clear_state;
28727df7 1384 int ret;
2e1df07d 1385
4896473e 1386 if (!o->use_thread) {
2e1df07d
JA
1387 setsid();
1388 td->pid = getpid();
1389 } else
1390 td->pid = gettid();
1391
4896473e 1392 fio_local_clock_init(o->use_thread);
5d879392 1393
2e1df07d
JA
1394 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1395
122c7725
JA
1396 if (is_backend)
1397 fio_server_send_start(td);
1398
2e1df07d
JA
1399 INIT_FLIST_HEAD(&td->io_log_list);
1400 INIT_FLIST_HEAD(&td->io_hist_list);
1401 INIT_FLIST_HEAD(&td->verify_list);
1402 INIT_FLIST_HEAD(&td->trim_list);
1ae83d45 1403 INIT_FLIST_HEAD(&td->next_rand_list);
2e1df07d
JA
1404 pthread_mutex_init(&td->io_u_lock, NULL);
1405 td->io_hist_tree = RB_ROOT;
1406
1407 pthread_condattr_init(&attr);
1408 pthread_cond_init(&td->verify_cond, &attr);
1409 pthread_cond_init(&td->free_cond, &attr);
1410
1411 td_set_runstate(td, TD_INITIALIZED);
1412 dprint(FD_MUTEX, "up startup_mutex\n");
1413 fio_mutex_up(startup_mutex);
1414 dprint(FD_MUTEX, "wait on td->mutex\n");
1415 fio_mutex_down(td->mutex);
1416 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1417
2e1df07d
JA
1418 /*
1419 * A new gid requires privilege, so we need to do this before setting
1420 * the uid.
1421 */
4896473e 1422 if (o->gid != -1U && setgid(o->gid)) {
2e1df07d
JA
1423 td_verror(td, errno, "setgid");
1424 goto err;
1425 }
4896473e 1426 if (o->uid != -1U && setuid(o->uid)) {
2e1df07d
JA
1427 td_verror(td, errno, "setuid");
1428 goto err;
1429 }
1430
1431 /*
1432 * If we have a gettimeofday() thread, make sure we exclude that
1433 * thread from this job
1434 */
4896473e
JA
1435 if (o->gtod_cpu)
1436 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
2e1df07d
JA
1437
1438 /*
1439 * Set affinity first, in case it has an impact on the memory
1440 * allocations.
1441 */
b2a9e649 1442 if (fio_option_is_set(o, cpumask)) {
c2acfbac 1443 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
30cb4c65 1444 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
c2acfbac
JA
1445 if (!ret) {
1446 log_err("fio: no CPUs set\n");
1447 log_err("fio: Try increasing number of available CPUs\n");
1448 td_verror(td, EINVAL, "cpus_split");
1449 goto err;
1450 }
1451 }
28727df7
JA
1452 ret = fio_setaffinity(td->pid, o->cpumask);
1453 if (ret == -1) {
4896473e
JA
1454 td_verror(td, errno, "cpu_set_affinity");
1455 goto err;
1456 }
2e1df07d
JA
1457 }
1458
67bf9823 1459#ifdef CONFIG_LIBNUMA
d0b937ed 1460 /* numa node setup */
b2a9e649
JA
1461 if (fio_option_is_set(o, numa_cpunodes) ||
1462 fio_option_is_set(o, numa_memnodes)) {
43522848 1463 struct bitmask *mask;
d0b937ed
YR
1464
1465 if (numa_available() < 0) {
1466 td_verror(td, errno, "Does not support NUMA API\n");
1467 goto err;
1468 }
1469
b2a9e649 1470 if (fio_option_is_set(o, numa_cpunodes)) {
43522848
DG
1471 mask = numa_parse_nodestring(o->numa_cpunodes);
1472 ret = numa_run_on_node_mask(mask);
1473 numa_free_nodemask(mask);
d0b937ed
YR
1474 if (ret == -1) {
1475 td_verror(td, errno, \
1476 "numa_run_on_node_mask failed\n");
1477 goto err;
1478 }
1479 }
1480
b2a9e649 1481 if (fio_option_is_set(o, numa_memnodes)) {
43522848
DG
1482 mask = NULL;
1483 if (o->numa_memnodes)
1484 mask = numa_parse_nodestring(o->numa_memnodes);
1485
4896473e 1486 switch (o->numa_mem_mode) {
d0b937ed 1487 case MPOL_INTERLEAVE:
43522848 1488 numa_set_interleave_mask(mask);
d0b937ed
YR
1489 break;
1490 case MPOL_BIND:
43522848 1491 numa_set_membind(mask);
d0b937ed
YR
1492 break;
1493 case MPOL_LOCAL:
1494 numa_set_localalloc();
1495 break;
1496 case MPOL_PREFERRED:
4896473e 1497 numa_set_preferred(o->numa_mem_prefer_node);
d0b937ed
YR
1498 break;
1499 case MPOL_DEFAULT:
1500 default:
1501 break;
1502 }
1503
43522848
DG
1504 if (mask)
1505 numa_free_nodemask(mask);
1506
d0b937ed
YR
1507 }
1508 }
1509#endif
1510
9a3f1100
JA
1511 if (fio_pin_memory(td))
1512 goto err;
1513
2e1df07d
JA
1514 /*
1515 * May alter parameters that init_io_u() will use, so we need to
1516 * do this first.
1517 */
1518 if (init_iolog(td))
1519 goto err;
1520
1521 if (init_io_u(td))
1522 goto err;
1523
4896473e 1524 if (o->verify_async && verify_async_init(td))
2e1df07d
JA
1525 goto err;
1526
27d74836
JA
1527 if (fio_option_is_set(o, ioprio) ||
1528 fio_option_is_set(o, ioprio_class)) {
28727df7
JA
1529 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1530 if (ret == -1) {
2e1df07d
JA
1531 td_verror(td, errno, "ioprio_set");
1532 goto err;
1533 }
1534 }
1535
4896473e 1536 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
2e1df07d
JA
1537 goto err;
1538
649c10c5 1539 errno = 0;
4896473e 1540 if (nice(o->nice) == -1 && errno != 0) {
2e1df07d
JA
1541 td_verror(td, errno, "nice");
1542 goto err;
1543 }
1544
4896473e 1545 if (o->ioscheduler && switch_ioscheduler(td))
2e1df07d
JA
1546 goto err;
1547
4896473e 1548 if (!o->create_serialize && setup_files(td))
2e1df07d
JA
1549 goto err;
1550
1551 if (td_io_init(td))
1552 goto err;
1553
1554 if (init_random_map(td))
1555 goto err;
1556
ce486495 1557 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
4896473e 1558 goto err;
2e1df07d 1559
4896473e 1560 if (o->pre_read) {
2e1df07d
JA
1561 if (pre_read_files(td) < 0)
1562 goto err;
1563 }
1564
aee2ab67
JA
1565 if (td->flags & TD_F_COMPRESS_LOG)
1566 tp_init(&td->tp_data);
1567
dc5bfbb2
JA
1568 fio_verify_init(td);
1569
a9da8ab2
JA
1570 if ((o->io_submit_mode == IO_MODE_OFFLOAD) &&
1571 workqueue_init(td, &td->io_wq, io_workqueue_fn, td->o.iodepth))
1572 goto err;
1573
2e1df07d 1574 fio_gettime(&td->epoch, NULL);
44404c5a 1575 fio_getrusage(&td->ru_start);
ac28d905
JA
1576 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1577 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1578
1579 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1580 o->ratemin[DDIR_TRIM]) {
1581 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1582 sizeof(td->bw_sample_time));
1583 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1584 sizeof(td->bw_sample_time));
1585 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1586 sizeof(td->bw_sample_time));
1587 }
1588
2e1df07d
JA
1589 clear_state = 0;
1590 while (keep_running(td)) {
100f49f1
JA
1591 uint64_t verify_bytes;
1592
2e1df07d 1593 fio_gettime(&td->start, NULL);
2e1df07d
JA
1594 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1595
2e1df07d 1596 if (clear_state)
ac28d905 1597 clear_io_state(td, 0);
2e1df07d
JA
1598
1599 prune_io_piece_log(td);
1600
62167762
JC
1601 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1602 verify_bytes = do_dry_run(td);
1603 else
1604 verify_bytes = do_io(td);
2e1df07d
JA
1605
1606 clear_state = 1;
1607
40c666c8
JA
1608 /*
1609 * Make sure we've successfully updated the rusage stats
1610 * before waiting on the stat mutex. Otherwise we could have
1611 * the stat thread holding stat mutex and waiting for
1612 * the rusage_sem, which would never get upped because
1613 * this thread is waiting for the stat mutex.
1614 */
1615 check_update_rusage(td);
1616
e5437a07 1617 fio_mutex_down(stat_mutex);
95603b74
BF
1618 if (td_read(td) && td->io_bytes[DDIR_READ])
1619 update_runtime(td, elapsed_us, DDIR_READ);
1620 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1621 update_runtime(td, elapsed_us, DDIR_WRITE);
1622 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1623 update_runtime(td, elapsed_us, DDIR_TRIM);
e5437a07
VT
1624 fio_gettime(&td->start, NULL);
1625 fio_mutex_up(stat_mutex);
2e1df07d
JA
1626
1627 if (td->error || td->terminate)
1628 break;
1629
4896473e
JA
1630 if (!o->do_verify ||
1631 o->verify == VERIFY_NONE ||
2e1df07d
JA
1632 (td->io_ops->flags & FIO_UNIDIR))
1633 continue;
1634
ac28d905 1635 clear_io_state(td, 0);
2e1df07d
JA
1636
1637 fio_gettime(&td->start, NULL);
1638
100f49f1 1639 do_verify(td, verify_bytes);
2e1df07d 1640
40c666c8
JA
1641 /*
1642 * See comment further up for why this is done here.
1643 */
1644 check_update_rusage(td);
1645
e5437a07 1646 fio_mutex_down(stat_mutex);
95603b74 1647 update_runtime(td, elapsed_us, DDIR_READ);
e5437a07
VT
1648 fio_gettime(&td->start, NULL);
1649 fio_mutex_up(stat_mutex);
2e1df07d
JA
1650
1651 if (td->error || td->terminate)
1652 break;
1653 }
1654
1655 update_rusage_stat(td);
2e1df07d 1656 td->ts.total_run_time = mtime_since_now(&td->epoch);
6eaf09d6
SL
1657 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1658 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1659 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2e1df07d 1660
ca09be4b 1661 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
d264264a
JA
1662 (td->o.verify != VERIFY_NONE && td_write(td)))
1663 verify_save_state(td->thread_number);
ca09be4b 1664
9a3f1100
JA
1665 fio_unpin_memory(td);
1666
905e3d4f 1667 fio_writeout_logs(td);
2e1df07d 1668
a9da8ab2
JA
1669 if (o->io_submit_mode == IO_MODE_OFFLOAD)
1670 workqueue_exit(&td->io_wq);
1671
aee2ab67
JA
1672 if (td->flags & TD_F_COMPRESS_LOG)
1673 tp_exit(&td->tp_data);
1674
4896473e 1675 if (o->exec_postrun)
ce486495 1676 exec_string(o, o->exec_postrun, (const char *)"postrun");
2e1df07d
JA
1677
1678 if (exitall_on_terminate)
1679 fio_terminate_threads(td->groupid);
1680
1681err:
1682 if (td->error)
1683 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1684 td->verror);
1685
4896473e 1686 if (o->verify_async)
2e1df07d
JA
1687 verify_async_exit(td);
1688
1689 close_and_free_files(td);
2e1df07d 1690 cleanup_io_u(td);
32dbca2c 1691 close_ioengine(td);
2e1df07d 1692 cgroup_shutdown(td, &cgroup_mnt);
ca09be4b 1693 verify_free_state(td);
2e1df07d 1694
b2a9e649 1695 if (fio_option_is_set(o, cpumask)) {
8a1db9a1
JA
1696 ret = fio_cpuset_exit(&o->cpumask);
1697 if (ret)
1698 td_verror(td, ret, "fio_cpuset_exit");
2e1df07d
JA
1699 }
1700
1701 /*
1702 * do this very late, it will log file closing as well
1703 */
4896473e 1704 if (o->write_iolog_file)
2e1df07d
JA
1705 write_iolog_close(td);
1706
ea66e04f
JA
1707 fio_mutex_remove(td->mutex);
1708 td->mutex = NULL;
1709
2e1df07d 1710 td_set_runstate(td, TD_EXITED);
fda2cfac
JA
1711
1712 /*
1713 * Do this last after setting our runstate to exited, so we
1714 * know that the stat thread is signaled.
1715 */
1716 check_update_rusage(td);
1717
e43606c2 1718 return (void *) (uintptr_t) td->error;
2e1df07d
JA
1719}
1720
1721
1722/*
1723 * We cannot pass the td data into a forked process, so attach the td and
1724 * pass it to the thread worker.
1725 */
1726static int fork_main(int shmid, int offset)
1727{
1728 struct thread_data *td;
1729 void *data, *ret;
1730
c8931876 1731#if !defined(__hpux) && !defined(CONFIG_NO_SHM)
2e1df07d
JA
1732 data = shmat(shmid, NULL, 0);
1733 if (data == (void *) -1) {
1734 int __err = errno;
1735
1736 perror("shmat");
1737 return __err;
1738 }
1739#else
1740 /*
1741 * HP-UX inherits shm mappings?
1742 */
1743 data = threads;
1744#endif
1745
1746 td = data + offset * sizeof(struct thread_data);
1747 ret = thread_main(td);
1748 shmdt(data);
e43606c2 1749 return (int) (uintptr_t) ret;
2e1df07d
JA
1750}
1751
cba5460c
JA
1752static void dump_td_info(struct thread_data *td)
1753{
1754 log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1755 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1756 (unsigned long) time_since_now(&td->terminate_time));
1757}
1758
2e1df07d
JA
1759/*
1760 * Run over the job map and reap the threads that have exited, if any.
1761 */
1762static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1763 unsigned int *m_rate)
1764{
1765 struct thread_data *td;
1766 unsigned int cputhreads, realthreads, pending;
1767 int i, status, ret;
1768
1769 /*
1770 * reap exited threads (TD_EXITED -> TD_REAPED)
1771 */
1772 realthreads = pending = cputhreads = 0;
1773 for_each_td(td, i) {
1774 int flags = 0;
1775
1776 /*
1777 * ->io_ops is NULL for a thread that has closed its
1778 * io engine
1779 */
1780 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1781 cputhreads++;
1782 else
1783 realthreads++;
1784
1785 if (!td->pid) {
1786 pending++;
1787 continue;
1788 }
1789 if (td->runstate == TD_REAPED)
1790 continue;
1791 if (td->o.use_thread) {
1792 if (td->runstate == TD_EXITED) {
1793 td_set_runstate(td, TD_REAPED);
1794 goto reaped;
1795 }
1796 continue;
1797 }
1798
1799 flags = WNOHANG;
1800 if (td->runstate == TD_EXITED)
1801 flags = 0;
1802
1803 /*
1804 * check if someone quit or got killed in an unusual way
1805 */
1806 ret = waitpid(td->pid, &status, flags);
1807 if (ret < 0) {
1808 if (errno == ECHILD) {
1809 log_err("fio: pid=%d disappeared %d\n",
1810 (int) td->pid, td->runstate);
a5e371a6 1811 td->sig = ECHILD;
2e1df07d
JA
1812 td_set_runstate(td, TD_REAPED);
1813 goto reaped;
1814 }
1815 perror("waitpid");
1816 } else if (ret == td->pid) {
1817 if (WIFSIGNALED(status)) {
1818 int sig = WTERMSIG(status);
1819
36d80bc7 1820 if (sig != SIGTERM && sig != SIGUSR2)
2e1df07d
JA
1821 log_err("fio: pid=%d, got signal=%d\n",
1822 (int) td->pid, sig);
a5e371a6 1823 td->sig = sig;
2e1df07d
JA
1824 td_set_runstate(td, TD_REAPED);
1825 goto reaped;
1826 }
1827 if (WIFEXITED(status)) {
1828 if (WEXITSTATUS(status) && !td->error)
1829 td->error = WEXITSTATUS(status);
1830
1831 td_set_runstate(td, TD_REAPED);
1832 goto reaped;
1833 }
1834 }
1835
cba5460c
JA
1836 /*
1837 * If the job is stuck, do a forceful timeout of it and
1838 * move on.
1839 */
1840 if (td->terminate &&
1841 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1842 dump_td_info(td);
1843 td_set_runstate(td, TD_REAPED);
1844 goto reaped;
1845 }
1846
2e1df07d
JA
1847 /*
1848 * thread is not dead, continue
1849 */
1850 pending++;
1851 continue;
1852reaped:
1853 (*nr_running)--;
342f4be4
JA
1854 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1855 (*t_rate) -= ddir_rw_sum(td->o.rate);
2e1df07d
JA
1856 if (!td->pid)
1857 pending--;
1858
1859 if (td->error)
1860 exit_value++;
1861
1862 done_secs += mtime_since_now(&td->epoch) / 1000;
4a88752a 1863 profile_td_exit(td);
2e1df07d
JA
1864 }
1865
1866 if (*nr_running == cputhreads && !pending && realthreads)
1867 fio_terminate_threads(TERMINATE_ALL);
1868}
1869
ca09be4b
JA
1870static int __check_trigger_file(void)
1871{
1872 struct stat sb;
1873
1874 if (!trigger_file)
1875 return 0;
1876
1877 if (stat(trigger_file, &sb))
1878 return 0;
1879
1880 if (unlink(trigger_file) < 0)
1881 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1882 strerror(errno));
1883
1884 return 1;
1885}
1886
1887static int trigger_timedout(void)
1888{
1889 if (trigger_timeout)
1890 return time_since_genesis() >= trigger_timeout;
1891
1892 return 0;
1893}
1894
1895void exec_trigger(const char *cmd)
1896{
1897 int ret;
1898
1899 if (!cmd)
1900 return;
1901
1902 ret = system(cmd);
1903 if (ret == -1)
1904 log_err("fio: failed executing %s trigger\n", cmd);
1905}
1906
1907void check_trigger_file(void)
1908{
1909 if (__check_trigger_file() || trigger_timedout()) {
796fb3ce
JA
1910 if (nr_clients)
1911 fio_clients_send_trigger(trigger_remote_cmd);
1912 else {
d264264a 1913 verify_save_state(IO_LIST_ALL);
ca09be4b
JA
1914 fio_terminate_threads(TERMINATE_ALL);
1915 exec_trigger(trigger_cmd);
1916 }
1917 }
1918}
1919
1920static int fio_verify_load_state(struct thread_data *td)
1921{
1922 int ret;
1923
1924 if (!td->o.verify_state)
1925 return 0;
1926
1927 if (is_backend) {
1928 void *data;
c3546b53 1929 int ver;
ca09be4b
JA
1930
1931 ret = fio_server_get_verify_state(td->o.name,
c3546b53 1932 td->thread_number - 1, &data, &ver);
ca09be4b 1933 if (!ret)
c3546b53 1934 verify_convert_assign_state(td, data, ver);
ca09be4b
JA
1935 } else
1936 ret = verify_load_state(td, "local");
1937
1938 return ret;
1939}
1940
06464907
JA
1941static void do_usleep(unsigned int usecs)
1942{
1943 check_for_running_stats();
ca09be4b 1944 check_trigger_file();
06464907
JA
1945 usleep(usecs);
1946}
1947
e81ecca3
JA
1948static int check_mount_writes(struct thread_data *td)
1949{
1950 struct fio_file *f;
1951 unsigned int i;
1952
1953 if (!td_write(td) || td->o.allow_mounted_write)
1954 return 0;
1955
1956 for_each_file(td, f, i) {
1957 if (f->filetype != FIO_TYPE_BD)
1958 continue;
1959 if (device_is_mounted(f->file_name))
1960 goto mounted;
1961 }
1962
1963 return 0;
1964mounted:
1965 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
1966 return 1;
1967}
1968
2e1df07d
JA
1969/*
1970 * Main function for kicking off and reaping jobs, as needed.
1971 */
1972static void run_threads(void)
1973{
1974 struct thread_data *td;
2e1df07d 1975 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
0de5b26f 1976 uint64_t spent;
2e1df07d 1977
2e1df07d
JA
1978 if (fio_gtod_offload && fio_start_gtod_thread())
1979 return;
334185e9 1980
f2a2ce0e 1981 fio_idle_prof_init();
2e1df07d
JA
1982
1983 set_sig_handlers();
1984
3a5f6bde
JA
1985 nr_thread = nr_process = 0;
1986 for_each_td(td, i) {
e81ecca3
JA
1987 if (check_mount_writes(td))
1988 return;
3a5f6bde
JA
1989 if (td->o.use_thread)
1990 nr_thread++;
1991 else
1992 nr_process++;
1993 }
1994
01cfefcc 1995 if (output_format & FIO_OUTPUT_NORMAL) {
2e1df07d
JA
1996 log_info("Starting ");
1997 if (nr_thread)
1998 log_info("%d thread%s", nr_thread,
1999 nr_thread > 1 ? "s" : "");
2000 if (nr_process) {
2001 if (nr_thread)
2002 log_info(" and ");
2003 log_info("%d process%s", nr_process,
2004 nr_process > 1 ? "es" : "");
2005 }
2006 log_info("\n");
e411c301 2007 log_info_flush();
2e1df07d
JA
2008 }
2009
2010 todo = thread_number;
2011 nr_running = 0;
2012 nr_started = 0;
2013 m_rate = t_rate = 0;
2014
2015 for_each_td(td, i) {
2016 print_status_init(td->thread_number - 1);
2017
2018 if (!td->o.create_serialize)
2019 continue;
2020
ca09be4b
JA
2021 if (fio_verify_load_state(td))
2022 goto reap;
2023
2e1df07d
JA
2024 /*
2025 * do file setup here so it happens sequentially,
2026 * we don't want X number of threads getting their
2027 * client data interspersed on disk
2028 */
2029 if (setup_files(td)) {
ca09be4b 2030reap:
2e1df07d
JA
2031 exit_value++;
2032 if (td->error)
2033 log_err("fio: pid=%d, err=%d/%s\n",
2034 (int) td->pid, td->error, td->verror);
2035 td_set_runstate(td, TD_REAPED);
2036 todo--;
2037 } else {
2038 struct fio_file *f;
2039 unsigned int j;
2040
2041 /*
2042 * for sharing to work, each job must always open
2043 * its own files. so close them, if we opened them
2044 * for creation
2045 */
2046 for_each_file(td, f, j) {
2047 if (fio_file_open(f))
2048 td_io_close_file(td, f);
2049 }
2050 }
2051 }
2052
f2a2ce0e
HL
2053 /* start idle threads before io threads start to run */
2054 fio_idle_prof_start();
2055
2e1df07d
JA
2056 set_genesis_time();
2057
2058 while (todo) {
2059 struct thread_data *map[REAL_MAX_JOBS];
2060 struct timeval this_start;
2061 int this_jobs = 0, left;
2062
2063 /*
2064 * create threads (TD_NOT_CREATED -> TD_CREATED)
2065 */
2066 for_each_td(td, i) {
2067 if (td->runstate != TD_NOT_CREATED)
2068 continue;
2069
2070 /*
2071 * never got a chance to start, killed by other
2072 * thread for some reason
2073 */
2074 if (td->terminate) {
2075 todo--;
2076 continue;
2077 }
2078
2079 if (td->o.start_delay) {
0de5b26f 2080 spent = utime_since_genesis();
2e1df07d 2081
74454ce4 2082 if (td->o.start_delay > spent)
2e1df07d
JA
2083 continue;
2084 }
2085
2086 if (td->o.stonewall && (nr_started || nr_running)) {
2087 dprint(FD_PROCESS, "%s: stonewall wait\n",
2088 td->o.name);
2089 break;
2090 }
2091
2092 init_disk_util(td);
2093
c97f1ad6
JA
2094 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2095 td->update_rusage = 0;
2096
2e1df07d
JA
2097 /*
2098 * Set state to created. Thread will transition
2099 * to TD_INITIALIZED when it's done setting up.
2100 */
2101 td_set_runstate(td, TD_CREATED);
2102 map[this_jobs++] = td;
2103 nr_started++;
2104
2105 if (td->o.use_thread) {
2106 int ret;
2107
2108 dprint(FD_PROCESS, "will pthread_create\n");
2109 ret = pthread_create(&td->thread, NULL,
2110 thread_main, td);
2111 if (ret) {
2112 log_err("pthread_create: %s\n",
2113 strerror(ret));
2114 nr_started--;
2115 break;
2116 }
2117 ret = pthread_detach(td->thread);
2118 if (ret)
2119 log_err("pthread_detach: %s",
2120 strerror(ret));
2121 } else {
2122 pid_t pid;
2123 dprint(FD_PROCESS, "will fork\n");
2124 pid = fork();
2125 if (!pid) {
2126 int ret = fork_main(shm_id, i);
2127
2128 _exit(ret);
2129 } else if (i == fio_debug_jobno)
2130 *fio_debug_jobp = pid;
2131 }
2132 dprint(FD_MUTEX, "wait on startup_mutex\n");
2133 if (fio_mutex_down_timeout(startup_mutex, 10)) {
2134 log_err("fio: job startup hung? exiting.\n");
2135 fio_terminate_threads(TERMINATE_ALL);
2136 fio_abort = 1;
2137 nr_started--;
2138 break;
2139 }
2140 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2141 }
2142
2143 /*
2144 * Wait for the started threads to transition to
2145 * TD_INITIALIZED.
2146 */
2147 fio_gettime(&this_start, NULL);
2148 left = this_jobs;
2149 while (left && !fio_abort) {
2150 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2151 break;
2152
06464907 2153 do_usleep(100000);
2e1df07d
JA
2154
2155 for (i = 0; i < this_jobs; i++) {
2156 td = map[i];
2157 if (!td)
2158 continue;
2159 if (td->runstate == TD_INITIALIZED) {
2160 map[i] = NULL;
2161 left--;
2162 } else if (td->runstate >= TD_EXITED) {
2163 map[i] = NULL;
2164 left--;
2165 todo--;
2166 nr_running++; /* work-around... */
2167 }
2168 }
2169 }
2170
2171 if (left) {
4e87c37a
JA
2172 log_err("fio: %d job%s failed to start\n", left,
2173 left > 1 ? "s" : "");
2e1df07d
JA
2174 for (i = 0; i < this_jobs; i++) {
2175 td = map[i];
2176 if (!td)
2177 continue;
2178 kill(td->pid, SIGTERM);
2179 }
2180 break;
2181 }
2182
2183 /*
2184 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2185 */
2186 for_each_td(td, i) {
2187 if (td->runstate != TD_INITIALIZED)
2188 continue;
2189
2190 if (in_ramp_time(td))
2191 td_set_runstate(td, TD_RAMP);
2192 else
2193 td_set_runstate(td, TD_RUNNING);
2194 nr_running++;
2195 nr_started--;
342f4be4
JA
2196 m_rate += ddir_rw_sum(td->o.ratemin);
2197 t_rate += ddir_rw_sum(td->o.rate);
2e1df07d
JA
2198 todo--;
2199 fio_mutex_up(td->mutex);
2200 }
2201
2202 reap_threads(&nr_running, &t_rate, &m_rate);
2203
122c7725 2204 if (todo)
06464907 2205 do_usleep(100000);
2e1df07d
JA
2206 }
2207
2208 while (nr_running) {
2209 reap_threads(&nr_running, &t_rate, &m_rate);
06464907 2210 do_usleep(10000);
2e1df07d
JA
2211 }
2212
f2a2ce0e
HL
2213 fio_idle_prof_stop();
2214
2e1df07d 2215 update_io_ticks();
2e1df07d
JA
2216}
2217
5ddc6707 2218static void wait_for_helper_thread_exit(void)
9ec7779f 2219{
8aab824f
JA
2220 void *ret;
2221
98b4b0a2 2222 helper_exit = 1;
5ddc6707
JA
2223 pthread_cond_signal(&helper_cond);
2224 pthread_join(helper_thread, &ret);
9ec7779f
JA
2225}
2226
27357187
JA
2227static void free_disk_util(void)
2228{
27357187 2229 disk_util_prune_entries();
8aab824f 2230
5ddc6707 2231 pthread_cond_destroy(&helper_cond);
27357187
JA
2232}
2233
5ddc6707 2234static void *helper_thread_main(void *data)
2e1df07d 2235{
9ec7779f
JA
2236 int ret = 0;
2237
2e1df07d
JA
2238 fio_mutex_up(startup_mutex);
2239
8aab824f
JA
2240 while (!ret) {
2241 uint64_t sec = DISK_UTIL_MSEC / 1000;
2242 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2243 struct timespec ts;
2244 struct timeval tv;
2245
2246 gettimeofday(&tv, NULL);
2247 ts.tv_sec = tv.tv_sec + sec;
2248 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
f893b76d 2249
823952af 2250 if (ts.tv_nsec >= 1000000000ULL) {
8aab824f
JA
2251 ts.tv_nsec -= 1000000000ULL;
2252 ts.tv_sec++;
2253 }
2254
5ddc6707 2255 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
8aab824f 2256
9ec7779f 2257 ret = update_io_ticks();
2e1df07d 2258
5ddc6707
JA
2259 if (helper_do_stat) {
2260 helper_do_stat = 0;
2261 __show_running_run_stats();
2262 }
2263
2e1df07d
JA
2264 if (!is_backend)
2265 print_thread_status();
2266 }
2267
2268 return NULL;
2269}
2270
5ddc6707 2271static int create_helper_thread(void)
2e1df07d
JA
2272{
2273 int ret;
2274
9ec7779f
JA
2275 setup_disk_util();
2276
5ddc6707
JA
2277 pthread_cond_init(&helper_cond, NULL);
2278 pthread_mutex_init(&helper_lock, NULL);
8aab824f 2279
5ddc6707 2280 ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL);
2e1df07d 2281 if (ret) {
5ddc6707 2282 log_err("Can't create helper thread: %s\n", strerror(ret));
2e1df07d
JA
2283 return 1;
2284 }
2285
2e1df07d
JA
2286 dprint(FD_MUTEX, "wait on startup_mutex\n");
2287 fio_mutex_down(startup_mutex);
2288 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2289 return 0;
2290}
2291
2e1df07d
JA
2292int fio_backend(void)
2293{
2294 struct thread_data *td;
2295 int i;
2296
2297 if (exec_profile) {
2298 if (load_profile(exec_profile))
2299 return 1;
2300 free(exec_profile);
2301 exec_profile = NULL;
2302 }
2303 if (!thread_number)
2304 return 0;
2305
2306 if (write_bw_log) {
aee2ab67
JA
2307 struct log_params p = {
2308 .log_type = IO_LOG_TYPE_BW,
2309 };
2310
2311 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2312 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2313 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2e1df07d
JA
2314 }
2315
521da527 2316 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2e1df07d
JA
2317 if (startup_mutex == NULL)
2318 return 1;
2e1df07d
JA
2319
2320 set_genesis_time();
cef9175e 2321 stat_init();
5ddc6707 2322 create_helper_thread();
2e1df07d
JA
2323
2324 cgroup_list = smalloc(sizeof(*cgroup_list));
2325 INIT_FLIST_HEAD(cgroup_list);
2326
2327 run_threads();
2328
5ddc6707 2329 wait_for_helper_thread_exit();
8aab824f 2330
2e1df07d 2331 if (!fio_abort) {
83f7b64e 2332 __show_run_stats();
2e1df07d 2333 if (write_bw_log) {
cb7e0ace
JA
2334 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2335 struct io_log *log = agg_io_log[i];
2336
3a5db920 2337 flush_log(log, 0);
518dac09 2338 free_log(log);
cb7e0ace 2339 }
2e1df07d
JA
2340 }
2341 }
2342
fda2cfac 2343 for_each_td(td, i) {
2e1df07d 2344 fio_options_free(td);
8049adc1
JA
2345 if (td->rusage_sem) {
2346 fio_mutex_remove(td->rusage_sem);
2347 td->rusage_sem = NULL;
2348 }
fda2cfac 2349 }
2e1df07d 2350
a462baef 2351 free_disk_util();
2e1df07d
JA
2352 cgroup_kill(cgroup_list);
2353 sfree(cgroup_list);
2354 sfree(cgroup_mnt);
2355
2356 fio_mutex_remove(startup_mutex);
cef9175e 2357 stat_exit();
2e1df07d
JA
2358 return exit_value;
2359}