idletime: fix another missing unlock on error
[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"
2e1df07d
JA
56
57static pthread_t disk_util_thread;
9ec7779f 58static struct fio_mutex *disk_thread_mutex;
2e1df07d 59static struct fio_mutex *startup_mutex;
2e1df07d
JA
60static struct flist_head *cgroup_list;
61static char *cgroup_mnt;
62static int exit_value;
63static volatile int fio_abort;
3a5f6bde
JA
64static unsigned int nr_process = 0;
65static unsigned int nr_thread = 0;
2e1df07d 66
6eaf09d6 67struct io_log *agg_io_log[DDIR_RWDIR_CNT];
2e1df07d 68
a3efc919
JA
69int groupid = 0;
70unsigned int thread_number = 0;
108fea77 71unsigned int stat_number = 0;
a3efc919
JA
72int shm_id = 0;
73int temp_stall_ts;
74unsigned long done_secs = 0;
27357187 75volatile int disk_util_exit = 0;
a3efc919 76
2e1df07d 77#define PAGE_ALIGN(buf) \
e43606c2 78 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
2e1df07d
JA
79
80#define JOB_START_TIMEOUT (5 * 1000)
81
82static void sig_int(int sig)
83{
84 if (threads) {
85 if (is_backend)
86 fio_server_got_signal(sig);
87 else {
88 log_info("\nfio: terminating on signal %d\n", sig);
89 fflush(stdout);
90 exit_value = 128;
91 }
92
93 fio_terminate_threads(TERMINATE_ALL);
94 }
95}
96
4c6d91e8
JA
97static void sig_show_status(int sig)
98{
99 show_running_run_stats();
100}
101
2e1df07d
JA
102static void set_sig_handlers(void)
103{
104 struct sigaction act;
105
106 memset(&act, 0, sizeof(act));
107 act.sa_handler = sig_int;
108 act.sa_flags = SA_RESTART;
109 sigaction(SIGINT, &act, NULL);
110
111 memset(&act, 0, sizeof(act));
112 act.sa_handler = sig_int;
113 act.sa_flags = SA_RESTART;
114 sigaction(SIGTERM, &act, NULL);
115
2f694507
BC
116/* Windows uses SIGBREAK as a quit signal from other applications */
117#ifdef WIN32
118 memset(&act, 0, sizeof(act));
119 act.sa_handler = sig_int;
120 act.sa_flags = SA_RESTART;
121 sigaction(SIGBREAK, &act, NULL);
122#endif
123
4c6d91e8
JA
124 memset(&act, 0, sizeof(act));
125 act.sa_handler = sig_show_status;
126 act.sa_flags = SA_RESTART;
127 sigaction(SIGUSR1, &act, NULL);
128
2e1df07d
JA
129 if (is_backend) {
130 memset(&act, 0, sizeof(act));
131 act.sa_handler = sig_int;
132 act.sa_flags = SA_RESTART;
133 sigaction(SIGPIPE, &act, NULL);
134 }
135}
136
137/*
138 * Check if we are above the minimum rate given.
139 */
140static int __check_min_rate(struct thread_data *td, struct timeval *now,
141 enum fio_ddir ddir)
142{
143 unsigned long long bytes = 0;
144 unsigned long iops = 0;
145 unsigned long spent;
146 unsigned long rate;
147 unsigned int ratemin = 0;
148 unsigned int rate_iops = 0;
149 unsigned int rate_iops_min = 0;
150
151 assert(ddir_rw(ddir));
152
153 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
154 return 0;
155
156 /*
157 * allow a 2 second settle period in the beginning
158 */
159 if (mtime_since(&td->start, now) < 2000)
160 return 0;
161
162 iops += td->this_io_blocks[ddir];
163 bytes += td->this_io_bytes[ddir];
164 ratemin += td->o.ratemin[ddir];
165 rate_iops += td->o.rate_iops[ddir];
166 rate_iops_min += td->o.rate_iops_min[ddir];
167
168 /*
169 * if rate blocks is set, sample is running
170 */
171 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
172 spent = mtime_since(&td->lastrate[ddir], now);
173 if (spent < td->o.ratecycle)
174 return 0;
175
176 if (td->o.rate[ddir]) {
177 /*
178 * check bandwidth specified rate
179 */
180 if (bytes < td->rate_bytes[ddir]) {
181 log_err("%s: min rate %u not met\n", td->o.name,
182 ratemin);
183 return 1;
184 } else {
49cba9b3
JA
185 if (spent)
186 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
187 else
188 rate = 0;
189
2e1df07d
JA
190 if (rate < ratemin ||
191 bytes < td->rate_bytes[ddir]) {
192 log_err("%s: min rate %u not met, got"
193 " %luKB/sec\n", td->o.name,
194 ratemin, rate);
195 return 1;
196 }
197 }
198 } else {
199 /*
200 * checks iops specified rate
201 */
202 if (iops < rate_iops) {
203 log_err("%s: min iops rate %u not met\n",
204 td->o.name, rate_iops);
205 return 1;
206 } else {
207 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
208 if (rate < rate_iops_min ||
209 iops < td->rate_blocks[ddir]) {
210 log_err("%s: min iops rate %u not met,"
211 " got %lu\n", td->o.name,
212 rate_iops_min, rate);
213 }
214 }
215 }
216 }
217
218 td->rate_bytes[ddir] = bytes;
219 td->rate_blocks[ddir] = iops;
220 memcpy(&td->lastrate[ddir], now, sizeof(*now));
221 return 0;
222}
223
224static int check_min_rate(struct thread_data *td, struct timeval *now,
100f49f1 225 uint64_t *bytes_done)
2e1df07d
JA
226{
227 int ret = 0;
228
6eaf09d6
SL
229 if (bytes_done[DDIR_READ])
230 ret |= __check_min_rate(td, now, DDIR_READ);
231 if (bytes_done[DDIR_WRITE])
232 ret |= __check_min_rate(td, now, DDIR_WRITE);
233 if (bytes_done[DDIR_TRIM])
234 ret |= __check_min_rate(td, now, DDIR_TRIM);
2e1df07d
JA
235
236 return ret;
237}
238
239/*
240 * When job exits, we can cancel the in-flight IO if we are using async
241 * io. Attempt to do so.
242 */
243static void cleanup_pending_aio(struct thread_data *td)
244{
2e1df07d
JA
245 int r;
246
247 /*
248 * get immediately available events, if any
249 */
250 r = io_u_queued_complete(td, 0, NULL);
251 if (r < 0)
252 return;
253
254 /*
255 * now cancel remaining active events
256 */
257 if (td->io_ops->cancel) {
2ae0b204
JA
258 struct io_u *io_u;
259 int i;
2e1df07d 260
2ae0b204
JA
261 io_u_qiter(&td->io_u_all, io_u, i) {
262 if (io_u->flags & IO_U_F_FLIGHT) {
2e1df07d
JA
263 r = td->io_ops->cancel(td, io_u);
264 if (!r)
265 put_io_u(td, io_u);
266 }
267 }
268 }
269
270 if (td->cur_depth)
271 r = io_u_queued_complete(td, td->cur_depth, NULL);
272}
273
274/*
275 * Helper to handle the final sync of a file. Works just like the normal
276 * io path, just does everything sync.
277 */
278static int fio_io_sync(struct thread_data *td, struct fio_file *f)
279{
280 struct io_u *io_u = __get_io_u(td);
281 int ret;
282
283 if (!io_u)
284 return 1;
285
286 io_u->ddir = DDIR_SYNC;
287 io_u->file = f;
288
289 if (td_io_prep(td, io_u)) {
290 put_io_u(td, io_u);
291 return 1;
292 }
293
294requeue:
295 ret = td_io_queue(td, io_u);
296 if (ret < 0) {
297 td_verror(td, io_u->error, "td_io_queue");
298 put_io_u(td, io_u);
299 return 1;
300 } else if (ret == FIO_Q_QUEUED) {
301 if (io_u_queued_complete(td, 1, NULL) < 0)
302 return 1;
303 } else if (ret == FIO_Q_COMPLETED) {
304 if (io_u->error) {
305 td_verror(td, io_u->error, "td_io_queue");
306 return 1;
307 }
308
309 if (io_u_sync_complete(td, io_u, NULL) < 0)
310 return 1;
311 } else if (ret == FIO_Q_BUSY) {
312 if (td_io_commit(td))
313 return 1;
314 goto requeue;
315 }
316
317 return 0;
318}
a3efc919 319
61ee0f86
JA
320static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
321{
322 int ret;
323
324 if (fio_file_open(f))
325 return fio_io_sync(td, f);
326
327 if (td_io_open_file(td, f))
328 return 1;
329
330 ret = fio_io_sync(td, f);
331 td_io_close_file(td, f);
332 return ret;
333}
334
2e1df07d
JA
335static inline void __update_tv_cache(struct thread_data *td)
336{
337 fio_gettime(&td->tv_cache, NULL);
338}
339
340static inline void update_tv_cache(struct thread_data *td)
341{
342 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
343 __update_tv_cache(td);
344}
345
346static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
347{
348 if (in_ramp_time(td))
349 return 0;
350 if (!td->o.timeout)
351 return 0;
0de5b26f 352 if (utime_since(&td->epoch, t) >= td->o.timeout)
2e1df07d
JA
353 return 1;
354
355 return 0;
356}
357
358static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
359 int *retptr)
360{
361 int ret = *retptr;
362
363 if (ret < 0 || td->error) {
8b28bd41
DM
364 int err = td->error;
365 enum error_type_bit eb;
2e1df07d
JA
366
367 if (ret < 0)
368 err = -ret;
2e1df07d 369
8b28bd41
DM
370 eb = td_error_type(ddir, err);
371 if (!(td->o.continue_on_error & (1 << eb)))
2e1df07d
JA
372 return 1;
373
8b28bd41 374 if (td_non_fatal_error(td, eb, err)) {
2e1df07d
JA
375 /*
376 * Continue with the I/Os in case of
377 * a non fatal error.
378 */
379 update_error_count(td, err);
380 td_clear_error(td);
381 *retptr = 0;
382 return 0;
383 } else if (td->o.fill_device && err == ENOSPC) {
384 /*
385 * We expect to hit this error if
386 * fill_device option is set.
387 */
388 td_clear_error(td);
389 td->terminate = 1;
390 return 1;
391 } else {
392 /*
393 * Stop the I/O in case of a fatal
394 * error.
395 */
396 update_error_count(td, err);
397 return 1;
398 }
399 }
400
401 return 0;
402}
403
c97f1ad6
JA
404static void check_update_rusage(struct thread_data *td)
405{
406 if (td->update_rusage) {
407 td->update_rusage = 0;
408 update_rusage_stat(td);
409 fio_mutex_up(td->rusage_sem);
410 }
411}
412
2e1df07d
JA
413/*
414 * The main verify engine. Runs over the writes we previously submitted,
415 * reads the blocks back in, and checks the crc/md5 of the data.
416 */
100f49f1 417static void do_verify(struct thread_data *td, uint64_t verify_bytes)
2e1df07d 418{
100f49f1 419 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
2e1df07d
JA
420 struct fio_file *f;
421 struct io_u *io_u;
422 int ret, min_events;
423 unsigned int i;
424
425 dprint(FD_VERIFY, "starting loop\n");
426
427 /*
428 * sync io first and invalidate cache, to make sure we really
429 * read from disk.
430 */
431 for_each_file(td, f, i) {
432 if (!fio_file_open(f))
433 continue;
434 if (fio_io_sync(td, f))
435 break;
436 if (file_invalidate_cache(td, f))
437 break;
438 }
439
c97f1ad6
JA
440 check_update_rusage(td);
441
2e1df07d
JA
442 if (td->error)
443 return;
444
445 td_set_runstate(td, TD_VERIFYING);
446
447 io_u = NULL;
448 while (!td->terminate) {
fbccf46c 449 enum fio_ddir ddir;
2e1df07d
JA
450 int ret2, full;
451
452 update_tv_cache(td);
c97f1ad6 453 check_update_rusage(td);
2e1df07d
JA
454
455 if (runtime_exceeded(td, &td->tv_cache)) {
456 __update_tv_cache(td);
457 if (runtime_exceeded(td, &td->tv_cache)) {
458 td->terminate = 1;
459 break;
460 }
461 }
462
9e684a49
DE
463 if (flow_threshold_exceeded(td))
464 continue;
465
44cbc6da
JA
466 if (!td->o.experimental_verify) {
467 io_u = __get_io_u(td);
468 if (!io_u)
469 break;
2e1df07d 470
44cbc6da
JA
471 if (get_next_verify(td, io_u)) {
472 put_io_u(td, io_u);
473 break;
474 }
2e1df07d 475
44cbc6da
JA
476 if (td_io_prep(td, io_u)) {
477 put_io_u(td, io_u);
478 break;
479 }
480 } else {
100f49f1
JA
481 if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
482 break;
483
bcd5abfa 484 while ((io_u = get_io_u(td)) != NULL) {
002fe734
JA
485 if (IS_ERR(io_u)) {
486 io_u = NULL;
487 ret = FIO_Q_BUSY;
488 goto reap;
489 }
490
bcd5abfa
JA
491 /*
492 * We are only interested in the places where
493 * we wrote or trimmed IOs. Turn those into
494 * reads for verification purposes.
495 */
496 if (io_u->ddir == DDIR_READ) {
497 /*
498 * Pretend we issued it for rwmix
499 * accounting
500 */
501 td->io_issues[DDIR_READ]++;
502 put_io_u(td, io_u);
503 continue;
504 } else if (io_u->ddir == DDIR_TRIM) {
505 io_u->ddir = DDIR_READ;
506 io_u->flags |= IO_U_F_TRIMMED;
507 break;
508 } else if (io_u->ddir == DDIR_WRITE) {
509 io_u->ddir = DDIR_READ;
510 break;
511 } else {
512 put_io_u(td, io_u);
513 continue;
514 }
515 }
44cbc6da 516
bcd5abfa 517 if (!io_u)
44cbc6da 518 break;
2e1df07d
JA
519 }
520
521 if (td->o.verify_async)
522 io_u->end_io = verify_io_u_async;
523 else
524 io_u->end_io = verify_io_u;
525
fbccf46c
JA
526 ddir = io_u->ddir;
527
2e1df07d
JA
528 ret = td_io_queue(td, io_u);
529 switch (ret) {
530 case FIO_Q_COMPLETED:
531 if (io_u->error) {
532 ret = -io_u->error;
533 clear_io_u(td, io_u);
534 } else if (io_u->resid) {
535 int bytes = io_u->xfer_buflen - io_u->resid;
536
537 /*
538 * zero read, fail
539 */
540 if (!bytes) {
541 td_verror(td, EIO, "full resid");
542 put_io_u(td, io_u);
543 break;
544 }
545
546 io_u->xfer_buflen = io_u->resid;
547 io_u->xfer_buf += bytes;
548 io_u->offset += bytes;
549
550 if (ddir_rw(io_u->ddir))
551 td->ts.short_io_u[io_u->ddir]++;
552
553 f = io_u->file;
554 if (io_u->offset == f->real_file_size)
555 goto sync_done;
556
557 requeue_io_u(td, &io_u);
558 } else {
559sync_done:
100f49f1 560 ret = io_u_sync_complete(td, io_u, bytes_done);
2e1df07d
JA
561 if (ret < 0)
562 break;
563 }
564 continue;
565 case FIO_Q_QUEUED:
566 break;
567 case FIO_Q_BUSY:
568 requeue_io_u(td, &io_u);
569 ret2 = td_io_commit(td);
570 if (ret2 < 0)
571 ret = ret2;
572 break;
573 default:
574 assert(ret < 0);
575 td_verror(td, -ret, "td_io_queue");
576 break;
577 }
578
fbccf46c 579 if (break_on_this_error(td, ddir, &ret))
2e1df07d
JA
580 break;
581
582 /*
583 * if we can queue more, do so. but check if there are
584 * completed io_u's first. Note that we can get BUSY even
585 * without IO queued, if the system is resource starved.
586 */
002fe734 587reap:
2e1df07d
JA
588 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
589 if (full || !td->o.iodepth_batch_complete) {
590 min_events = min(td->o.iodepth_batch_complete,
591 td->cur_depth);
8a74b56d
JA
592 /*
593 * if the queue is full, we MUST reap at least 1 event
594 */
595 if (full && !min_events)
2e1df07d
JA
596 min_events = 1;
597
598 do {
599 /*
600 * Reap required number of io units, if any,
601 * and do the verification on them through
602 * the callback handler
603 */
100f49f1 604 if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
2e1df07d
JA
605 ret = -1;
606 break;
607 }
608 } while (full && (td->cur_depth > td->o.iodepth_low));
609 }
610 if (ret < 0)
611 break;
612 }
613
c97f1ad6
JA
614 check_update_rusage(td);
615
2e1df07d
JA
616 if (!td->error) {
617 min_events = td->cur_depth;
618
619 if (min_events)
620 ret = io_u_queued_complete(td, min_events, NULL);
621 } else
622 cleanup_pending_aio(td);
623
624 td_set_runstate(td, TD_RUNNING);
625
626 dprint(FD_VERIFY, "exiting loop\n");
627}
628
3939fe85
JA
629static unsigned int exceeds_number_ios(struct thread_data *td)
630{
631 unsigned long long number_ios;
632
633 if (!td->o.number_ios)
634 return 0;
635
636 number_ios = ddir_rw_sum(td->this_io_blocks);
637 number_ios += td->io_u_queued + td->io_u_in_flight;
638
639 return number_ios >= td->o.number_ios;
640}
641
f7078f7b
JA
642static int io_bytes_exceeded(struct thread_data *td)
643{
644 unsigned long long bytes;
645
646 if (td_rw(td))
6eaf09d6 647 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
f7078f7b 648 else if (td_write(td))
6eaf09d6
SL
649 bytes = td->this_io_bytes[DDIR_WRITE];
650 else if (td_read(td))
651 bytes = td->this_io_bytes[DDIR_READ];
f7078f7b 652 else
6eaf09d6 653 bytes = td->this_io_bytes[DDIR_TRIM];
f7078f7b 654
3939fe85 655 return bytes >= td->o.size || exceeds_number_ios(td);
f7078f7b
JA
656}
657
2e1df07d
JA
658/*
659 * Main IO worker function. It retrieves io_u's to process and queues
660 * and reaps them, checking for rate and errors along the way.
100f49f1
JA
661 *
662 * Returns number of bytes written and trimmed.
2e1df07d 663 */
100f49f1 664static uint64_t do_io(struct thread_data *td)
2e1df07d 665{
100f49f1 666 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
2e1df07d
JA
667 unsigned int i;
668 int ret = 0;
c2703bf3 669 uint64_t total_bytes, bytes_issued = 0;
2e1df07d
JA
670
671 if (in_ramp_time(td))
672 td_set_runstate(td, TD_RAMP);
673 else
674 td_set_runstate(td, TD_RUNNING);
675
3e260a46
JA
676 lat_target_init(td);
677
78a6469c
JA
678 /*
679 * If verify_backlog is enabled, we'll run the verify in this
680 * handler as well. For that case, we may need up to twice the
681 * amount of bytes.
682 */
c2703bf3 683 total_bytes = td->o.size;
78a6469c
JA
684 if (td->o.verify != VERIFY_NONE &&
685 (td_write(td) && td->o.verify_backlog))
c2703bf3
JA
686 total_bytes += td->o.size;
687
f7078f7b 688 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
c04e4661
DE
689 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
690 td->o.time_based) {
2e1df07d 691 struct timeval comp_time;
2e1df07d
JA
692 int min_evts = 0;
693 struct io_u *io_u;
694 int ret2, full;
695 enum fio_ddir ddir;
696
c97f1ad6
JA
697 check_update_rusage(td);
698
7d7803fa 699 if (td->terminate || td->done)
2e1df07d
JA
700 break;
701
702 update_tv_cache(td);
703
704 if (runtime_exceeded(td, &td->tv_cache)) {
705 __update_tv_cache(td);
706 if (runtime_exceeded(td, &td->tv_cache)) {
707 td->terminate = 1;
708 break;
709 }
710 }
711
9e684a49
DE
712 if (flow_threshold_exceeded(td))
713 continue;
714
c2703bf3 715 if (bytes_issued >= total_bytes)
20876c53
JC
716 break;
717
2e1df07d 718 io_u = get_io_u(td);
002fe734
JA
719 if (IS_ERR_OR_NULL(io_u)) {
720 int err = PTR_ERR(io_u);
721
722 io_u = NULL;
723 if (err == -EBUSY) {
724 ret = FIO_Q_BUSY;
725 goto reap;
726 }
3e260a46
JA
727 if (td->o.latency_target)
728 goto reap;
2e1df07d 729 break;
3e260a46 730 }
2e1df07d
JA
731
732 ddir = io_u->ddir;
733
734 /*
82af2a7c
JA
735 * Add verification end_io handler if:
736 * - Asked to verify (!td_rw(td))
737 * - Or the io_u is from our verify list (mixed write/ver)
2e1df07d
JA
738 */
739 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
82af2a7c 740 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
c4b6117b
PV
741
742 if (!td->o.verify_pattern_bytes) {
743 io_u->rand_seed = __rand(&td->__verify_state);
744 if (sizeof(int) != sizeof(long *))
745 io_u->rand_seed *= __rand(&td->__verify_state);
746 }
747
2e1df07d
JA
748 if (td->o.verify_async)
749 io_u->end_io = verify_io_u_async;
750 else
751 io_u->end_io = verify_io_u;
752 td_set_runstate(td, TD_VERIFYING);
753 } else if (in_ramp_time(td))
754 td_set_runstate(td, TD_RAMP);
755 else
756 td_set_runstate(td, TD_RUNNING);
757
9a50c5c5 758 /*
f9401285
JA
759 * Always log IO before it's issued, so we know the specific
760 * order of it. The logged unit will track when the IO has
761 * completed.
9a50c5c5 762 */
c4b6117b
PV
763 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
764 td->o.do_verify &&
765 td->o.verify != VERIFY_NONE &&
f9401285 766 !td->o.experimental_verify)
c4b6117b
PV
767 log_io_piece(td, io_u);
768
2e1df07d
JA
769 ret = td_io_queue(td, io_u);
770 switch (ret) {
771 case FIO_Q_COMPLETED:
772 if (io_u->error) {
773 ret = -io_u->error;
774 clear_io_u(td, io_u);
775 } else if (io_u->resid) {
776 int bytes = io_u->xfer_buflen - io_u->resid;
777 struct fio_file *f = io_u->file;
778
20876c53 779 bytes_issued += bytes;
2e1df07d
JA
780 /*
781 * zero read, fail
782 */
783 if (!bytes) {
784 td_verror(td, EIO, "full resid");
785 put_io_u(td, io_u);
786 break;
787 }
788
789 io_u->xfer_buflen = io_u->resid;
790 io_u->xfer_buf += bytes;
791 io_u->offset += bytes;
792
793 if (ddir_rw(io_u->ddir))
794 td->ts.short_io_u[io_u->ddir]++;
795
796 if (io_u->offset == f->real_file_size)
797 goto sync_done;
798
799 requeue_io_u(td, &io_u);
800 } else {
801sync_done:
6eaf09d6
SL
802 if (__should_check_rate(td, DDIR_READ) ||
803 __should_check_rate(td, DDIR_WRITE) ||
804 __should_check_rate(td, DDIR_TRIM))
2e1df07d
JA
805 fio_gettime(&comp_time, NULL);
806
807 ret = io_u_sync_complete(td, io_u, bytes_done);
808 if (ret < 0)
809 break;
20876c53 810 bytes_issued += io_u->xfer_buflen;
2e1df07d
JA
811 }
812 break;
813 case FIO_Q_QUEUED:
814 /*
815 * if the engine doesn't have a commit hook,
816 * the io_u is really queued. if it does have such
817 * a hook, it has to call io_u_queued() itself.
818 */
819 if (td->io_ops->commit == NULL)
820 io_u_queued(td, io_u);
20876c53 821 bytes_issued += io_u->xfer_buflen;
2e1df07d
JA
822 break;
823 case FIO_Q_BUSY:
824 requeue_io_u(td, &io_u);
825 ret2 = td_io_commit(td);
826 if (ret2 < 0)
827 ret = ret2;
828 break;
829 default:
830 assert(ret < 0);
831 put_io_u(td, io_u);
832 break;
833 }
834
835 if (break_on_this_error(td, ddir, &ret))
836 break;
837
838 /*
839 * See if we need to complete some commands. Note that we
840 * can get BUSY even without IO queued, if the system is
841 * resource starved.
842 */
3e260a46 843reap:
2e1df07d
JA
844 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
845 if (full || !td->o.iodepth_batch_complete) {
846 min_evts = min(td->o.iodepth_batch_complete,
847 td->cur_depth);
8a74b56d
JA
848 /*
849 * if the queue is full, we MUST reap at least 1 event
850 */
851 if (full && !min_evts)
2e1df07d
JA
852 min_evts = 1;
853
6eaf09d6
SL
854 if (__should_check_rate(td, DDIR_READ) ||
855 __should_check_rate(td, DDIR_WRITE) ||
856 __should_check_rate(td, DDIR_TRIM))
2e1df07d
JA
857 fio_gettime(&comp_time, NULL);
858
859 do {
860 ret = io_u_queued_complete(td, min_evts, bytes_done);
861 if (ret < 0)
862 break;
863
864 } while (full && (td->cur_depth > td->o.iodepth_low));
865 }
866
867 if (ret < 0)
868 break;
d5abee06 869 if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
2e1df07d
JA
870 continue;
871
872 if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
873 if (check_min_rate(td, &comp_time, bytes_done)) {
874 if (exitall_on_terminate)
875 fio_terminate_threads(td->groupid);
876 td_verror(td, EIO, "check_min_rate");
877 break;
878 }
879 }
3e260a46
JA
880 if (!in_ramp_time(td) && td->o.latency_target)
881 lat_target_check(td);
2e1df07d
JA
882
883 if (td->o.thinktime) {
884 unsigned long long b;
885
342f4be4 886 b = ddir_rw_sum(td->io_blocks);
2e1df07d
JA
887 if (!(b % td->o.thinktime_blocks)) {
888 int left;
889
002e7183
JA
890 io_u_quiesce(td);
891
2e1df07d
JA
892 if (td->o.thinktime_spin)
893 usec_spin(td->o.thinktime_spin);
894
895 left = td->o.thinktime - td->o.thinktime_spin;
896 if (left)
897 usec_sleep(td, left);
898 }
899 }
900 }
901
c97f1ad6
JA
902 check_update_rusage(td);
903
2e1df07d 904 if (td->trim_entries)
4e0a8fa2 905 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
2e1df07d
JA
906
907 if (td->o.fill_device && td->error == ENOSPC) {
908 td->error = 0;
909 td->terminate = 1;
910 }
911 if (!td->error) {
912 struct fio_file *f;
913
914 i = td->cur_depth;
915 if (i) {
5bd5f71a 916 ret = io_u_queued_complete(td, i, bytes_done);
2e1df07d
JA
917 if (td->o.fill_device && td->error == ENOSPC)
918 td->error = 0;
919 }
920
921 if (should_fsync(td) && td->o.end_fsync) {
922 td_set_runstate(td, TD_FSYNCING);
923
924 for_each_file(td, f, i) {
61ee0f86 925 if (!fio_file_fsync(td, f))
2e1df07d 926 continue;
61ee0f86
JA
927
928 log_err("fio: end_fsync failed for file %s\n",
929 f->file_name);
2e1df07d
JA
930 }
931 }
932 } else
933 cleanup_pending_aio(td);
934
935 /*
936 * stop job if we failed doing any IO
937 */
342f4be4 938 if (!ddir_rw_sum(td->this_io_bytes))
2e1df07d 939 td->done = 1;
100f49f1
JA
940
941 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
2e1df07d
JA
942}
943
944static void cleanup_io_u(struct thread_data *td)
945{
2e1df07d
JA
946 struct io_u *io_u;
947
2ae0b204 948 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
c73ed246
JA
949
950 if (td->io_ops->io_u_free)
951 td->io_ops->io_u_free(td, io_u);
952
2e1df07d
JA
953 fio_memfree(io_u, sizeof(*io_u));
954 }
955
956 free_io_mem(td);
2ae0b204
JA
957
958 io_u_rexit(&td->io_u_requeues);
959 io_u_qexit(&td->io_u_freelist);
960 io_u_qexit(&td->io_u_all);
2e1df07d
JA
961}
962
963static int init_io_u(struct thread_data *td)
964{
965 struct io_u *io_u;
9c42684e 966 unsigned int max_bs, min_write;
2e1df07d 967 int cl_align, i, max_units;
2ae0b204 968 int data_xfer = 1, err;
2e1df07d
JA
969 char *p;
970
971 max_units = td->o.iodepth;
74f4b020 972 max_bs = td_max_bs(td);
9c42684e 973 min_write = td->o.min_bs[DDIR_WRITE];
2e1df07d
JA
974 td->orig_buffer_size = (unsigned long long) max_bs
975 * (unsigned long long) max_units;
976
88045e04 977 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
59d8d0f5
JA
978 data_xfer = 0;
979
2ae0b204
JA
980 err = 0;
981 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
982 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
983 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
984
985 if (err) {
986 log_err("fio: failed setting up IO queues\n");
987 return 1;
988 }
989
fd8a09b8 990 /*
991 * if we may later need to do address alignment, then add any
992 * possible adjustment here so that we don't cause a buffer
993 * overflow later. this adjustment may be too much if we get
994 * lucky and the allocator gives us an aligned address.
995 */
d01612f3
CM
996 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
997 (td->io_ops->flags & FIO_RAWIO))
fd8a09b8 998 td->orig_buffer_size += page_mask + td->o.mem_align;
999
2e1df07d
JA
1000 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1001 unsigned long bs;
1002
1003 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1004 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1005 }
1006
1007 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1008 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1009 return 1;
1010 }
1011
59d8d0f5 1012 if (data_xfer && allocate_io_mem(td))
2e1df07d
JA
1013 return 1;
1014
d01612f3 1015 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
2e1df07d
JA
1016 (td->io_ops->flags & FIO_RAWIO))
1017 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1018 else
1019 p = td->orig_buffer;
1020
1021 cl_align = os_cache_line_size();
1022
1023 for (i = 0; i < max_units; i++) {
1024 void *ptr;
1025
1026 if (td->terminate)
1027 return 1;
1028
1029 ptr = fio_memalign(cl_align, sizeof(*io_u));
1030 if (!ptr) {
1031 log_err("fio: unable to allocate aligned memory\n");
1032 break;
1033 }
1034
1035 io_u = ptr;
1036 memset(io_u, 0, sizeof(*io_u));
2ae0b204 1037 INIT_FLIST_HEAD(&io_u->verify_list);
2e1df07d
JA
1038 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1039
59d8d0f5 1040 if (data_xfer) {
2e1df07d
JA
1041 io_u->buf = p;
1042 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1043
1044 if (td_write(td))
9c42684e 1045 io_u_fill_buffer(td, io_u, min_write, max_bs);
2e1df07d
JA
1046 if (td_write(td) && td->o.verify_pattern_bytes) {
1047 /*
1048 * Fill the buffer with the pattern if we are
1049 * going to be doing writes.
1050 */
ce35b1ec 1051 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
2e1df07d
JA
1052 }
1053 }
1054
1055 io_u->index = i;
1056 io_u->flags = IO_U_F_FREE;
2ae0b204
JA
1057 io_u_qpush(&td->io_u_freelist, io_u);
1058
1059 /*
1060 * io_u never leaves this stack, used for iteration of all
1061 * io_u buffers.
1062 */
1063 io_u_qpush(&td->io_u_all, io_u);
c73ed246
JA
1064
1065 if (td->io_ops->io_u_init) {
1066 int ret = td->io_ops->io_u_init(td, io_u);
1067
1068 if (ret) {
1069 log_err("fio: failed to init engine data: %d\n", ret);
1070 return 1;
1071 }
1072 }
1073
2e1df07d
JA
1074 p += max_bs;
1075 }
1076
1077 return 0;
1078}
1079
1080static int switch_ioscheduler(struct thread_data *td)
1081{
1082 char tmp[256], tmp2[128];
1083 FILE *f;
1084 int ret;
1085
1086 if (td->io_ops->flags & FIO_DISKLESSIO)
1087 return 0;
1088
1089 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1090
1091 f = fopen(tmp, "r+");
1092 if (!f) {
1093 if (errno == ENOENT) {
1094 log_err("fio: os or kernel doesn't support IO scheduler"
1095 " switching\n");
1096 return 0;
1097 }
1098 td_verror(td, errno, "fopen iosched");
1099 return 1;
1100 }
1101
1102 /*
1103 * Set io scheduler.
1104 */
1105 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1106 if (ferror(f) || ret != 1) {
1107 td_verror(td, errno, "fwrite");
1108 fclose(f);
1109 return 1;
1110 }
1111
1112 rewind(f);
1113
1114 /*
1115 * Read back and check that the selected scheduler is now the default.
1116 */
49c6f33d 1117 ret = fread(tmp, sizeof(tmp), 1, f);
2e1df07d
JA
1118 if (ferror(f) || ret < 0) {
1119 td_verror(td, errno, "fread");
1120 fclose(f);
1121 return 1;
1122 }
49c6f33d
JA
1123 tmp[sizeof(tmp) - 1] = '\0';
1124
2e1df07d
JA
1125
1126 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1127 if (!strstr(tmp, tmp2)) {
1128 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1129 td_verror(td, EINVAL, "iosched_switch");
1130 fclose(f);
1131 return 1;
1132 }
1133
1134 fclose(f);
1135 return 0;
1136}
1137
1138static int keep_running(struct thread_data *td)
1139{
2e1df07d
JA
1140 if (td->done)
1141 return 0;
1142 if (td->o.time_based)
1143 return 1;
1144 if (td->o.loops) {
1145 td->o.loops--;
1146 return 1;
1147 }
3939fe85
JA
1148 if (exceeds_number_ios(td))
1149 return 0;
26251d8d 1150
ee98e234 1151 if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
5bd5f71a
JA
1152 uint64_t diff;
1153
1154 /*
1155 * If the difference is less than the minimum IO size, we
1156 * are done.
1157 */
1158 diff = td->o.size - ddir_rw_sum(td->io_bytes);
74f4b020 1159 if (diff < td_max_bs(td))
5bd5f71a
JA
1160 return 0;
1161
002fe734
JA
1162 if (fio_files_done(td))
1163 return 0;
1164
2e1df07d 1165 return 1;
5bd5f71a 1166 }
2e1df07d
JA
1167
1168 return 0;
1169}
1170
ce486495 1171static int exec_string(struct thread_options *o, const char *string, const char *mode)
2e1df07d 1172{
ce486495 1173 int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
2e1df07d
JA
1174 char *str;
1175
1176 str = malloc(newlen);
ce486495 1177 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
2e1df07d 1178
ce486495 1179 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
2e1df07d
JA
1180 ret = system(str);
1181 if (ret == -1)
1182 log_err("fio: exec of cmd <%s> failed\n", str);
1183
1184 free(str);
1185 return ret;
1186}
1187
62167762
JC
1188/*
1189 * Dry run to compute correct state of numberio for verification.
1190 */
1191static uint64_t do_dry_run(struct thread_data *td)
1192{
1193 uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
1194
1195 td_set_runstate(td, TD_RUNNING);
1196
1197 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
1198 (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
1199 struct io_u *io_u;
1200 int ret;
1201
1202 if (td->terminate || td->done)
1203 break;
1204
1205 io_u = get_io_u(td);
1206 if (!io_u)
1207 break;
1208
1209 io_u->flags |= IO_U_F_FLIGHT;
1210 io_u->error = 0;
1211 io_u->resid = 0;
1212 if (ddir_rw(acct_ddir(io_u)))
1213 td->io_issues[acct_ddir(io_u)]++;
1214 if (ddir_rw(io_u->ddir)) {
1215 io_u_mark_depth(td, 1);
1216 td->ts.total_io_u[io_u->ddir]++;
1217 }
1218
2e63e96b
PV
1219 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1220 td->o.do_verify &&
1221 td->o.verify != VERIFY_NONE &&
1222 !td->o.experimental_verify)
1223 log_io_piece(td, io_u);
1224
62167762
JC
1225 ret = io_u_sync_complete(td, io_u, bytes_done);
1226 (void) ret;
1227 }
1228
1229 return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
1230}
1231
2e1df07d
JA
1232/*
1233 * Entry point for the thread based jobs. The process based jobs end up
1234 * here as well, after a little setup.
1235 */
1236static void *thread_main(void *data)
1237{
1238 unsigned long long elapsed;
1239 struct thread_data *td = data;
4896473e 1240 struct thread_options *o = &td->o;
2e1df07d
JA
1241 pthread_condattr_t attr;
1242 int clear_state;
28727df7 1243 int ret;
2e1df07d 1244
4896473e 1245 if (!o->use_thread) {
2e1df07d
JA
1246 setsid();
1247 td->pid = getpid();
1248 } else
1249 td->pid = gettid();
1250
334185e9
BC
1251 /*
1252 * fio_time_init() may not have been called yet if running as a server
1253 */
1254 fio_time_init();
1255
4896473e 1256 fio_local_clock_init(o->use_thread);
5d879392 1257
2e1df07d
JA
1258 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1259
122c7725
JA
1260 if (is_backend)
1261 fio_server_send_start(td);
1262
2e1df07d
JA
1263 INIT_FLIST_HEAD(&td->io_log_list);
1264 INIT_FLIST_HEAD(&td->io_hist_list);
1265 INIT_FLIST_HEAD(&td->verify_list);
1266 INIT_FLIST_HEAD(&td->trim_list);
1ae83d45 1267 INIT_FLIST_HEAD(&td->next_rand_list);
2e1df07d
JA
1268 pthread_mutex_init(&td->io_u_lock, NULL);
1269 td->io_hist_tree = RB_ROOT;
1270
1271 pthread_condattr_init(&attr);
1272 pthread_cond_init(&td->verify_cond, &attr);
1273 pthread_cond_init(&td->free_cond, &attr);
1274
1275 td_set_runstate(td, TD_INITIALIZED);
1276 dprint(FD_MUTEX, "up startup_mutex\n");
1277 fio_mutex_up(startup_mutex);
1278 dprint(FD_MUTEX, "wait on td->mutex\n");
1279 fio_mutex_down(td->mutex);
1280 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1281
2e1df07d
JA
1282 /*
1283 * A new gid requires privilege, so we need to do this before setting
1284 * the uid.
1285 */
4896473e 1286 if (o->gid != -1U && setgid(o->gid)) {
2e1df07d
JA
1287 td_verror(td, errno, "setgid");
1288 goto err;
1289 }
4896473e 1290 if (o->uid != -1U && setuid(o->uid)) {
2e1df07d
JA
1291 td_verror(td, errno, "setuid");
1292 goto err;
1293 }
1294
1295 /*
1296 * If we have a gettimeofday() thread, make sure we exclude that
1297 * thread from this job
1298 */
4896473e
JA
1299 if (o->gtod_cpu)
1300 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
2e1df07d
JA
1301
1302 /*
1303 * Set affinity first, in case it has an impact on the memory
1304 * allocations.
1305 */
4896473e 1306 if (o->cpumask_set) {
c2acfbac 1307 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
30cb4c65 1308 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
c2acfbac
JA
1309 if (!ret) {
1310 log_err("fio: no CPUs set\n");
1311 log_err("fio: Try increasing number of available CPUs\n");
1312 td_verror(td, EINVAL, "cpus_split");
1313 goto err;
1314 }
1315 }
28727df7
JA
1316 ret = fio_setaffinity(td->pid, o->cpumask);
1317 if (ret == -1) {
4896473e
JA
1318 td_verror(td, errno, "cpu_set_affinity");
1319 goto err;
1320 }
2e1df07d
JA
1321 }
1322
67bf9823 1323#ifdef CONFIG_LIBNUMA
d0b937ed 1324 /* numa node setup */
4896473e 1325 if (o->numa_cpumask_set || o->numa_memmask_set) {
d0b937ed
YR
1326 int ret;
1327
1328 if (numa_available() < 0) {
1329 td_verror(td, errno, "Does not support NUMA API\n");
1330 goto err;
1331 }
1332
4896473e
JA
1333 if (o->numa_cpumask_set) {
1334 ret = numa_run_on_node_mask(o->numa_cpunodesmask);
d0b937ed
YR
1335 if (ret == -1) {
1336 td_verror(td, errno, \
1337 "numa_run_on_node_mask failed\n");
1338 goto err;
1339 }
1340 }
1341
4896473e 1342 if (o->numa_memmask_set) {
d0b937ed 1343
4896473e 1344 switch (o->numa_mem_mode) {
d0b937ed 1345 case MPOL_INTERLEAVE:
4896473e 1346 numa_set_interleave_mask(o->numa_memnodesmask);
d0b937ed
YR
1347 break;
1348 case MPOL_BIND:
4896473e 1349 numa_set_membind(o->numa_memnodesmask);
d0b937ed
YR
1350 break;
1351 case MPOL_LOCAL:
1352 numa_set_localalloc();
1353 break;
1354 case MPOL_PREFERRED:
4896473e 1355 numa_set_preferred(o->numa_mem_prefer_node);
d0b937ed
YR
1356 break;
1357 case MPOL_DEFAULT:
1358 default:
1359 break;
1360 }
1361
1362 }
1363 }
1364#endif
1365
9a3f1100
JA
1366 if (fio_pin_memory(td))
1367 goto err;
1368
2e1df07d
JA
1369 /*
1370 * May alter parameters that init_io_u() will use, so we need to
1371 * do this first.
1372 */
1373 if (init_iolog(td))
1374 goto err;
1375
1376 if (init_io_u(td))
1377 goto err;
1378
4896473e 1379 if (o->verify_async && verify_async_init(td))
2e1df07d
JA
1380 goto err;
1381
28727df7
JA
1382 if (o->ioprio) {
1383 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1384 if (ret == -1) {
2e1df07d
JA
1385 td_verror(td, errno, "ioprio_set");
1386 goto err;
1387 }
1388 }
1389
4896473e 1390 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
2e1df07d
JA
1391 goto err;
1392
649c10c5 1393 errno = 0;
4896473e 1394 if (nice(o->nice) == -1 && errno != 0) {
2e1df07d
JA
1395 td_verror(td, errno, "nice");
1396 goto err;
1397 }
1398
4896473e 1399 if (o->ioscheduler && switch_ioscheduler(td))
2e1df07d
JA
1400 goto err;
1401
4896473e 1402 if (!o->create_serialize && setup_files(td))
2e1df07d
JA
1403 goto err;
1404
1405 if (td_io_init(td))
1406 goto err;
1407
1408 if (init_random_map(td))
1409 goto err;
1410
ce486495 1411 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
4896473e 1412 goto err;
2e1df07d 1413
4896473e 1414 if (o->pre_read) {
2e1df07d
JA
1415 if (pre_read_files(td) < 0)
1416 goto err;
1417 }
1418
dc5bfbb2
JA
1419 fio_verify_init(td);
1420
2e1df07d 1421 fio_gettime(&td->epoch, NULL);
44404c5a 1422 fio_getrusage(&td->ru_start);
2e1df07d
JA
1423 clear_state = 0;
1424 while (keep_running(td)) {
100f49f1
JA
1425 uint64_t verify_bytes;
1426
2e1df07d
JA
1427 fio_gettime(&td->start, NULL);
1428 memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
1429 memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
1430 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1431
4896473e
JA
1432 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1433 o->ratemin[DDIR_TRIM]) {
6eaf09d6 1434 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
2e1df07d 1435 sizeof(td->bw_sample_time));
6eaf09d6
SL
1436 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1437 sizeof(td->bw_sample_time));
1438 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
2e1df07d
JA
1439 sizeof(td->bw_sample_time));
1440 }
1441
1442 if (clear_state)
1443 clear_io_state(td);
1444
1445 prune_io_piece_log(td);
1446
62167762
JC
1447 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1448 verify_bytes = do_dry_run(td);
1449 else
1450 verify_bytes = do_io(td);
2e1df07d
JA
1451
1452 clear_state = 1;
1453
1454 if (td_read(td) && td->io_bytes[DDIR_READ]) {
1455 elapsed = utime_since_now(&td->start);
1456 td->ts.runtime[DDIR_READ] += elapsed;
1457 }
1458 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
1459 elapsed = utime_since_now(&td->start);
1460 td->ts.runtime[DDIR_WRITE] += elapsed;
1461 }
6eaf09d6
SL
1462 if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
1463 elapsed = utime_since_now(&td->start);
1464 td->ts.runtime[DDIR_TRIM] += elapsed;
1465 }
2e1df07d
JA
1466
1467 if (td->error || td->terminate)
1468 break;
1469
4896473e
JA
1470 if (!o->do_verify ||
1471 o->verify == VERIFY_NONE ||
2e1df07d
JA
1472 (td->io_ops->flags & FIO_UNIDIR))
1473 continue;
1474
1475 clear_io_state(td);
1476
1477 fio_gettime(&td->start, NULL);
1478
100f49f1 1479 do_verify(td, verify_bytes);
2e1df07d
JA
1480
1481 td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
1482
1483 if (td->error || td->terminate)
1484 break;
1485 }
1486
1487 update_rusage_stat(td);
6eaf09d6
SL
1488 td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
1489 td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
1490 td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000;
2e1df07d 1491 td->ts.total_run_time = mtime_since_now(&td->epoch);
6eaf09d6
SL
1492 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1493 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1494 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2e1df07d 1495
9a3f1100
JA
1496 fio_unpin_memory(td);
1497
905e3d4f 1498 fio_writeout_logs(td);
2e1df07d 1499
4896473e 1500 if (o->exec_postrun)
ce486495 1501 exec_string(o, o->exec_postrun, (const char *)"postrun");
2e1df07d
JA
1502
1503 if (exitall_on_terminate)
1504 fio_terminate_threads(td->groupid);
1505
1506err:
1507 if (td->error)
1508 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1509 td->verror);
1510
4896473e 1511 if (o->verify_async)
2e1df07d
JA
1512 verify_async_exit(td);
1513
1514 close_and_free_files(td);
2e1df07d 1515 cleanup_io_u(td);
32dbca2c 1516 close_ioengine(td);
2e1df07d
JA
1517 cgroup_shutdown(td, &cgroup_mnt);
1518
4896473e
JA
1519 if (o->cpumask_set) {
1520 int ret = fio_cpuset_exit(&o->cpumask);
2e1df07d
JA
1521
1522 td_verror(td, ret, "fio_cpuset_exit");
1523 }
1524
1525 /*
1526 * do this very late, it will log file closing as well
1527 */
4896473e 1528 if (o->write_iolog_file)
2e1df07d
JA
1529 write_iolog_close(td);
1530
c97f1ad6
JA
1531 fio_mutex_remove(td->rusage_sem);
1532 td->rusage_sem = NULL;
1533
ea66e04f
JA
1534 fio_mutex_remove(td->mutex);
1535 td->mutex = NULL;
1536
2e1df07d 1537 td_set_runstate(td, TD_EXITED);
e43606c2 1538 return (void *) (uintptr_t) td->error;
2e1df07d
JA
1539}
1540
1541
1542/*
1543 * We cannot pass the td data into a forked process, so attach the td and
1544 * pass it to the thread worker.
1545 */
1546static int fork_main(int shmid, int offset)
1547{
1548 struct thread_data *td;
1549 void *data, *ret;
1550
1551#ifndef __hpux
1552 data = shmat(shmid, NULL, 0);
1553 if (data == (void *) -1) {
1554 int __err = errno;
1555
1556 perror("shmat");
1557 return __err;
1558 }
1559#else
1560 /*
1561 * HP-UX inherits shm mappings?
1562 */
1563 data = threads;
1564#endif
1565
1566 td = data + offset * sizeof(struct thread_data);
1567 ret = thread_main(td);
1568 shmdt(data);
e43606c2 1569 return (int) (uintptr_t) ret;
2e1df07d
JA
1570}
1571
1572/*
1573 * Run over the job map and reap the threads that have exited, if any.
1574 */
1575static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1576 unsigned int *m_rate)
1577{
1578 struct thread_data *td;
1579 unsigned int cputhreads, realthreads, pending;
1580 int i, status, ret;
1581
1582 /*
1583 * reap exited threads (TD_EXITED -> TD_REAPED)
1584 */
1585 realthreads = pending = cputhreads = 0;
1586 for_each_td(td, i) {
1587 int flags = 0;
1588
1589 /*
1590 * ->io_ops is NULL for a thread that has closed its
1591 * io engine
1592 */
1593 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1594 cputhreads++;
1595 else
1596 realthreads++;
1597
1598 if (!td->pid) {
1599 pending++;
1600 continue;
1601 }
1602 if (td->runstate == TD_REAPED)
1603 continue;
1604 if (td->o.use_thread) {
1605 if (td->runstate == TD_EXITED) {
1606 td_set_runstate(td, TD_REAPED);
1607 goto reaped;
1608 }
1609 continue;
1610 }
1611
1612 flags = WNOHANG;
1613 if (td->runstate == TD_EXITED)
1614 flags = 0;
1615
1616 /*
1617 * check if someone quit or got killed in an unusual way
1618 */
1619 ret = waitpid(td->pid, &status, flags);
1620 if (ret < 0) {
1621 if (errno == ECHILD) {
1622 log_err("fio: pid=%d disappeared %d\n",
1623 (int) td->pid, td->runstate);
a5e371a6 1624 td->sig = ECHILD;
2e1df07d
JA
1625 td_set_runstate(td, TD_REAPED);
1626 goto reaped;
1627 }
1628 perror("waitpid");
1629 } else if (ret == td->pid) {
1630 if (WIFSIGNALED(status)) {
1631 int sig = WTERMSIG(status);
1632
36d80bc7 1633 if (sig != SIGTERM && sig != SIGUSR2)
2e1df07d
JA
1634 log_err("fio: pid=%d, got signal=%d\n",
1635 (int) td->pid, sig);
a5e371a6 1636 td->sig = sig;
2e1df07d
JA
1637 td_set_runstate(td, TD_REAPED);
1638 goto reaped;
1639 }
1640 if (WIFEXITED(status)) {
1641 if (WEXITSTATUS(status) && !td->error)
1642 td->error = WEXITSTATUS(status);
1643
1644 td_set_runstate(td, TD_REAPED);
1645 goto reaped;
1646 }
1647 }
1648
1649 /*
1650 * thread is not dead, continue
1651 */
1652 pending++;
1653 continue;
1654reaped:
1655 (*nr_running)--;
342f4be4
JA
1656 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1657 (*t_rate) -= ddir_rw_sum(td->o.rate);
2e1df07d
JA
1658 if (!td->pid)
1659 pending--;
1660
1661 if (td->error)
1662 exit_value++;
1663
1664 done_secs += mtime_since_now(&td->epoch) / 1000;
4a88752a 1665 profile_td_exit(td);
2e1df07d
JA
1666 }
1667
1668 if (*nr_running == cputhreads && !pending && realthreads)
1669 fio_terminate_threads(TERMINATE_ALL);
1670}
1671
06464907
JA
1672static void do_usleep(unsigned int usecs)
1673{
1674 check_for_running_stats();
1675 usleep(usecs);
1676}
1677
2e1df07d
JA
1678/*
1679 * Main function for kicking off and reaping jobs, as needed.
1680 */
1681static void run_threads(void)
1682{
1683 struct thread_data *td;
2e1df07d 1684 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
0de5b26f 1685 uint64_t spent;
2e1df07d 1686
2e1df07d
JA
1687 if (fio_gtod_offload && fio_start_gtod_thread())
1688 return;
334185e9 1689
f2a2ce0e 1690 fio_idle_prof_init();
2e1df07d
JA
1691
1692 set_sig_handlers();
1693
3a5f6bde
JA
1694 nr_thread = nr_process = 0;
1695 for_each_td(td, i) {
1696 if (td->o.use_thread)
1697 nr_thread++;
1698 else
1699 nr_process++;
1700 }
1701
f3afa57e 1702 if (output_format == FIO_OUTPUT_NORMAL) {
2e1df07d
JA
1703 log_info("Starting ");
1704 if (nr_thread)
1705 log_info("%d thread%s", nr_thread,
1706 nr_thread > 1 ? "s" : "");
1707 if (nr_process) {
1708 if (nr_thread)
1709 log_info(" and ");
1710 log_info("%d process%s", nr_process,
1711 nr_process > 1 ? "es" : "");
1712 }
1713 log_info("\n");
1714 fflush(stdout);
1715 }
1716
1717 todo = thread_number;
1718 nr_running = 0;
1719 nr_started = 0;
1720 m_rate = t_rate = 0;
1721
1722 for_each_td(td, i) {
1723 print_status_init(td->thread_number - 1);
1724
1725 if (!td->o.create_serialize)
1726 continue;
1727
1728 /*
1729 * do file setup here so it happens sequentially,
1730 * we don't want X number of threads getting their
1731 * client data interspersed on disk
1732 */
1733 if (setup_files(td)) {
1734 exit_value++;
1735 if (td->error)
1736 log_err("fio: pid=%d, err=%d/%s\n",
1737 (int) td->pid, td->error, td->verror);
1738 td_set_runstate(td, TD_REAPED);
1739 todo--;
1740 } else {
1741 struct fio_file *f;
1742 unsigned int j;
1743
1744 /*
1745 * for sharing to work, each job must always open
1746 * its own files. so close them, if we opened them
1747 * for creation
1748 */
1749 for_each_file(td, f, j) {
1750 if (fio_file_open(f))
1751 td_io_close_file(td, f);
1752 }
1753 }
1754 }
1755
f2a2ce0e
HL
1756 /* start idle threads before io threads start to run */
1757 fio_idle_prof_start();
1758
2e1df07d
JA
1759 set_genesis_time();
1760
1761 while (todo) {
1762 struct thread_data *map[REAL_MAX_JOBS];
1763 struct timeval this_start;
1764 int this_jobs = 0, left;
1765
1766 /*
1767 * create threads (TD_NOT_CREATED -> TD_CREATED)
1768 */
1769 for_each_td(td, i) {
1770 if (td->runstate != TD_NOT_CREATED)
1771 continue;
1772
1773 /*
1774 * never got a chance to start, killed by other
1775 * thread for some reason
1776 */
1777 if (td->terminate) {
1778 todo--;
1779 continue;
1780 }
1781
1782 if (td->o.start_delay) {
0de5b26f 1783 spent = utime_since_genesis();
2e1df07d 1784
74454ce4 1785 if (td->o.start_delay > spent)
2e1df07d
JA
1786 continue;
1787 }
1788
1789 if (td->o.stonewall && (nr_started || nr_running)) {
1790 dprint(FD_PROCESS, "%s: stonewall wait\n",
1791 td->o.name);
1792 break;
1793 }
1794
1795 init_disk_util(td);
1796
c97f1ad6
JA
1797 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
1798 td->update_rusage = 0;
1799
2e1df07d
JA
1800 /*
1801 * Set state to created. Thread will transition
1802 * to TD_INITIALIZED when it's done setting up.
1803 */
1804 td_set_runstate(td, TD_CREATED);
1805 map[this_jobs++] = td;
1806 nr_started++;
1807
1808 if (td->o.use_thread) {
1809 int ret;
1810
1811 dprint(FD_PROCESS, "will pthread_create\n");
1812 ret = pthread_create(&td->thread, NULL,
1813 thread_main, td);
1814 if (ret) {
1815 log_err("pthread_create: %s\n",
1816 strerror(ret));
1817 nr_started--;
1818 break;
1819 }
1820 ret = pthread_detach(td->thread);
1821 if (ret)
1822 log_err("pthread_detach: %s",
1823 strerror(ret));
1824 } else {
1825 pid_t pid;
1826 dprint(FD_PROCESS, "will fork\n");
1827 pid = fork();
1828 if (!pid) {
1829 int ret = fork_main(shm_id, i);
1830
1831 _exit(ret);
1832 } else if (i == fio_debug_jobno)
1833 *fio_debug_jobp = pid;
1834 }
1835 dprint(FD_MUTEX, "wait on startup_mutex\n");
1836 if (fio_mutex_down_timeout(startup_mutex, 10)) {
1837 log_err("fio: job startup hung? exiting.\n");
1838 fio_terminate_threads(TERMINATE_ALL);
1839 fio_abort = 1;
1840 nr_started--;
1841 break;
1842 }
1843 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1844 }
1845
1846 /*
1847 * Wait for the started threads to transition to
1848 * TD_INITIALIZED.
1849 */
1850 fio_gettime(&this_start, NULL);
1851 left = this_jobs;
1852 while (left && !fio_abort) {
1853 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1854 break;
1855
06464907 1856 do_usleep(100000);
2e1df07d
JA
1857
1858 for (i = 0; i < this_jobs; i++) {
1859 td = map[i];
1860 if (!td)
1861 continue;
1862 if (td->runstate == TD_INITIALIZED) {
1863 map[i] = NULL;
1864 left--;
1865 } else if (td->runstate >= TD_EXITED) {
1866 map[i] = NULL;
1867 left--;
1868 todo--;
1869 nr_running++; /* work-around... */
1870 }
1871 }
1872 }
1873
1874 if (left) {
4e87c37a
JA
1875 log_err("fio: %d job%s failed to start\n", left,
1876 left > 1 ? "s" : "");
2e1df07d
JA
1877 for (i = 0; i < this_jobs; i++) {
1878 td = map[i];
1879 if (!td)
1880 continue;
1881 kill(td->pid, SIGTERM);
1882 }
1883 break;
1884 }
1885
1886 /*
1887 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1888 */
1889 for_each_td(td, i) {
1890 if (td->runstate != TD_INITIALIZED)
1891 continue;
1892
1893 if (in_ramp_time(td))
1894 td_set_runstate(td, TD_RAMP);
1895 else
1896 td_set_runstate(td, TD_RUNNING);
1897 nr_running++;
1898 nr_started--;
342f4be4
JA
1899 m_rate += ddir_rw_sum(td->o.ratemin);
1900 t_rate += ddir_rw_sum(td->o.rate);
2e1df07d
JA
1901 todo--;
1902 fio_mutex_up(td->mutex);
1903 }
1904
1905 reap_threads(&nr_running, &t_rate, &m_rate);
1906
122c7725 1907 if (todo)
06464907 1908 do_usleep(100000);
2e1df07d
JA
1909 }
1910
1911 while (nr_running) {
1912 reap_threads(&nr_running, &t_rate, &m_rate);
06464907 1913 do_usleep(10000);
2e1df07d
JA
1914 }
1915
f2a2ce0e
HL
1916 fio_idle_prof_stop();
1917
2e1df07d 1918 update_io_ticks();
2e1df07d
JA
1919}
1920
9ec7779f
JA
1921void wait_for_disk_thread_exit(void)
1922{
1923 fio_mutex_down(disk_thread_mutex);
1924}
1925
27357187
JA
1926static void free_disk_util(void)
1927{
1928 disk_util_start_exit();
1929 wait_for_disk_thread_exit();
1930 disk_util_prune_entries();
1931}
1932
2e1df07d
JA
1933static void *disk_thread_main(void *data)
1934{
9ec7779f
JA
1935 int ret = 0;
1936
2e1df07d
JA
1937 fio_mutex_up(startup_mutex);
1938
9ec7779f 1939 while (threads && !ret) {
2e1df07d
JA
1940 usleep(DISK_UTIL_MSEC * 1000);
1941 if (!threads)
1942 break;
9ec7779f 1943 ret = update_io_ticks();
2e1df07d
JA
1944
1945 if (!is_backend)
1946 print_thread_status();
1947 }
1948
9ec7779f 1949 fio_mutex_up(disk_thread_mutex);
2e1df07d
JA
1950 return NULL;
1951}
1952
1953static int create_disk_util_thread(void)
1954{
1955 int ret;
1956
9ec7779f
JA
1957 setup_disk_util();
1958
521da527 1959 disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
9ec7779f 1960
2e1df07d
JA
1961 ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
1962 if (ret) {
9ec7779f 1963 fio_mutex_remove(disk_thread_mutex);
2e1df07d
JA
1964 log_err("Can't create disk util thread: %s\n", strerror(ret));
1965 return 1;
1966 }
1967
1968 ret = pthread_detach(disk_util_thread);
1969 if (ret) {
9ec7779f 1970 fio_mutex_remove(disk_thread_mutex);
2e1df07d
JA
1971 log_err("Can't detatch disk util thread: %s\n", strerror(ret));
1972 return 1;
1973 }
1974
1975 dprint(FD_MUTEX, "wait on startup_mutex\n");
1976 fio_mutex_down(startup_mutex);
1977 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
1978 return 0;
1979}
1980
2e1df07d
JA
1981int fio_backend(void)
1982{
1983 struct thread_data *td;
1984 int i;
1985
1986 if (exec_profile) {
1987 if (load_profile(exec_profile))
1988 return 1;
1989 free(exec_profile);
1990 exec_profile = NULL;
1991 }
1992 if (!thread_number)
1993 return 0;
1994
1995 if (write_bw_log) {
5a812f9b
JA
1996 setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
1997 setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
1998 setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
2e1df07d
JA
1999 }
2000
521da527 2001 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2e1df07d
JA
2002 if (startup_mutex == NULL)
2003 return 1;
2e1df07d
JA
2004
2005 set_genesis_time();
cef9175e 2006 stat_init();
2e1df07d
JA
2007 create_disk_util_thread();
2008
2009 cgroup_list = smalloc(sizeof(*cgroup_list));
2010 INIT_FLIST_HEAD(cgroup_list);
2011
2012 run_threads();
2013
2014 if (!fio_abort) {
2015 show_run_stats();
2016 if (write_bw_log) {
2017 __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
2018 __finish_log(agg_io_log[DDIR_WRITE],
2019 "agg-write_bw.log");
6eaf09d6
SL
2020 __finish_log(agg_io_log[DDIR_TRIM],
2021 "agg-write_bw.log");
2e1df07d
JA
2022 }
2023 }
2024
2025 for_each_td(td, i)
2026 fio_options_free(td);
2027
a462baef 2028 free_disk_util();
2e1df07d
JA
2029 cgroup_kill(cgroup_list);
2030 sfree(cgroup_list);
2031 sfree(cgroup_mnt);
2032
2033 fio_mutex_remove(startup_mutex);
9ec7779f 2034 fio_mutex_remove(disk_thread_mutex);
cef9175e 2035 stat_exit();
2e1df07d
JA
2036 return exit_value;
2037}