Add backend related data
[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"
a9da8ab2 57#include "workqueue.h"
e81ecca3 58#include "lib/mountcheck.h"
40511301 59#include "rate-submit.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 */
e39c0676
JA
147static bool __check_min_rate(struct thread_data *td, struct timeval *now,
148 enum fio_ddir ddir)
2e1df07d
JA
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])
e39c0676 161 return false;
2e1df07d
JA
162
163 /*
164 * allow a 2 second settle period in the beginning
165 */
166 if (mtime_since(&td->start, now) < 2000)
e39c0676 167 return false;
2e1df07d
JA
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)
e39c0676 181 return false;
2e1df07d 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);
e39c0676 190 return true;
2e1df07d 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);
e39c0676 202 return true;
2e1df07d
JA
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);
e39c0676 212 return true;
2e1df07d 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);
e39c0676 224 return true;
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));
e39c0676 233 return false;
2e1df07d
JA
234}
235
e39c0676 236static bool check_min_rate(struct thread_data *td, struct timeval *now)
2e1df07d 237{
e39c0676 238 bool ret = false;
2e1df07d 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 */
e39c0676 289static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
2e1df07d
JA
290{
291 struct io_u *io_u = __get_io_u(td);
292 int ret;
293
294 if (!io_u)
e39c0676 295 return true;
2e1df07d
JA
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);
e39c0676 302 return true;
2e1df07d
JA
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);
e39c0676 310 return true;
2e1df07d 311 } else if (ret == FIO_Q_QUEUED) {
55312f9f 312 if (io_u_queued_complete(td, 1) < 0)
e39c0676 313 return true;
2e1df07d
JA
314 } else if (ret == FIO_Q_COMPLETED) {
315 if (io_u->error) {
316 td_verror(td, io_u->error, "td_io_queue");
e39c0676 317 return true;
2e1df07d
JA
318 }
319
55312f9f 320 if (io_u_sync_complete(td, io_u) < 0)
e39c0676 321 return true;
2e1df07d
JA
322 } else if (ret == FIO_Q_BUSY) {
323 if (td_io_commit(td))
e39c0676 324 return true;
2e1df07d
JA
325 goto requeue;
326 }
327
e39c0676 328 return false;
2e1df07d 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
e39c0676 357static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
2e1df07d
JA
358{
359 if (in_ramp_time(td))
e39c0676 360 return false;
2e1df07d 361 if (!td->o.timeout)
e39c0676 362 return false;
0de5b26f 363 if (utime_since(&td->epoch, t) >= td->o.timeout)
e39c0676 364 return true;
2e1df07d 365
e39c0676 366 return false;
2e1df07d
JA
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
e39c0676
JA
386static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
387 int *retptr)
2e1df07d
JA
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)))
e39c0676 400 return true;
2e1df07d 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;
e39c0676 410 return false;
2e1df07d
JA
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);
e39c0676 418 return true;
2e1df07d
JA
419 } else {
420 /*
421 * Stop the I/O in case of a fatal
422 * error.
423 */
424 update_error_count(td, err);
e39c0676 425 return true;
2e1df07d
JA
426 }
427 }
428
e39c0676 429 return false;
2e1df07d
JA
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
e39c0676 555static inline bool io_in_polling(struct thread_data *td)
82407585
RP
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
e39c0676 716static bool exceeds_number_ios(struct thread_data *td)
3939fe85
JA
717{
718 unsigned long long number_ios;
719
720 if (!td->o.number_ios)
e39c0676 721 return false;
3939fe85 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
e39c0676 729static bool 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
e39c0676 751static bool io_complete_bytes_exceeded(struct thread_data *td)
e28dd2cf
JE
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 930 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
0c5df5f9
JA
931 const unsigned long blen = io_u->xfer_buflen;
932 const enum fio_ddir ddir = acct_ddir(io_u);
933
a9da8ab2
JA
934 if (td->error)
935 break;
0c5df5f9 936
26de50cf
JA
937 workqueue_enqueue(&td->io_wq, &io_u->work);
938 ret = FIO_Q_QUEUED;
50a8ce86 939
26de50cf 940 if (ddir_rw(ddir)) {
0c5df5f9
JA
941 td->io_issues[ddir]++;
942 td->io_issue_bytes[ddir] += blen;
943 td->rate_io_issue_bytes[ddir] += blen;
944 }
945
50a8ce86
D
946 if (should_check_rate(td))
947 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
948
a9da8ab2
JA
949 } else {
950 ret = td_io_queue(td, io_u);
2e1df07d 951
50a8ce86
D
952 if (should_check_rate(td))
953 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
954
fd727d9d 955 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
a9da8ab2 956 break;
2e1df07d 957
a9da8ab2
JA
958 /*
959 * See if we need to complete some commands. Note that
960 * we can get BUSY even without IO queued, if the
961 * system is resource starved.
962 */
3e260a46 963reap:
a9da8ab2
JA
964 full = queue_full(td) ||
965 (ret == FIO_Q_BUSY && td->cur_depth);
82407585 966 if (full || io_in_polling(td))
a9da8ab2
JA
967 ret = wait_for_completions(td, &comp_time);
968 }
2e1df07d
JA
969 if (ret < 0)
970 break;
55312f9f
JA
971 if (!ddir_rw_sum(td->bytes_done) &&
972 !(td->io_ops->flags & FIO_NOIO))
2e1df07d
JA
973 continue;
974
55312f9f
JA
975 if (!in_ramp_time(td) && should_check_rate(td)) {
976 if (check_min_rate(td, &comp_time)) {
f9cafb12 977 if (exitall_on_terminate || td->o.exitall_error)
2e1df07d
JA
978 fio_terminate_threads(td->groupid);
979 td_verror(td, EIO, "check_min_rate");
980 break;
981 }
982 }
3e260a46
JA
983 if (!in_ramp_time(td) && td->o.latency_target)
984 lat_target_check(td);
e155cb64 985
2e1df07d
JA
986 if (td->o.thinktime) {
987 unsigned long long b;
988
342f4be4 989 b = ddir_rw_sum(td->io_blocks);
2e1df07d
JA
990 if (!(b % td->o.thinktime_blocks)) {
991 int left;
992
002e7183
JA
993 io_u_quiesce(td);
994
2e1df07d
JA
995 if (td->o.thinktime_spin)
996 usec_spin(td->o.thinktime_spin);
997
998 left = td->o.thinktime - td->o.thinktime_spin;
999 if (left)
1000 usec_sleep(td, left);
1001 }
1002 }
1003 }
1004
c97f1ad6
JA
1005 check_update_rusage(td);
1006
2e1df07d 1007 if (td->trim_entries)
4e0a8fa2 1008 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
2e1df07d
JA
1009
1010 if (td->o.fill_device && td->error == ENOSPC) {
1011 td->error = 0;
ebea2133 1012 fio_mark_td_terminate(td);
2e1df07d
JA
1013 }
1014 if (!td->error) {
1015 struct fio_file *f;
1016
a9da8ab2
JA
1017 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1018 workqueue_flush(&td->io_wq);
1019 i = 0;
1020 } else
1021 i = td->cur_depth;
1022
2e1df07d 1023 if (i) {
55312f9f 1024 ret = io_u_queued_complete(td, i);
2e1df07d
JA
1025 if (td->o.fill_device && td->error == ENOSPC)
1026 td->error = 0;
1027 }
1028
1029 if (should_fsync(td) && td->o.end_fsync) {
1030 td_set_runstate(td, TD_FSYNCING);
1031
1032 for_each_file(td, f, i) {
61ee0f86 1033 if (!fio_file_fsync(td, f))
2e1df07d 1034 continue;
61ee0f86
JA
1035
1036 log_err("fio: end_fsync failed for file %s\n",
1037 f->file_name);
2e1df07d
JA
1038 }
1039 }
1040 } else
1041 cleanup_pending_aio(td);
1042
1043 /*
1044 * stop job if we failed doing any IO
1045 */
342f4be4 1046 if (!ddir_rw_sum(td->this_io_bytes))
2e1df07d 1047 td->done = 1;
100f49f1 1048
55312f9f 1049 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
2e1df07d
JA
1050}
1051
1052static void cleanup_io_u(struct thread_data *td)
1053{
2e1df07d
JA
1054 struct io_u *io_u;
1055
2ae0b204 1056 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
c73ed246
JA
1057
1058 if (td->io_ops->io_u_free)
1059 td->io_ops->io_u_free(td, io_u);
1060
2e1df07d
JA
1061 fio_memfree(io_u, sizeof(*io_u));
1062 }
1063
1064 free_io_mem(td);
2ae0b204
JA
1065
1066 io_u_rexit(&td->io_u_requeues);
1067 io_u_qexit(&td->io_u_freelist);
1068 io_u_qexit(&td->io_u_all);
ca09be4b
JA
1069
1070 if (td->last_write_comp)
1071 sfree(td->last_write_comp);
2e1df07d
JA
1072}
1073
1074static int init_io_u(struct thread_data *td)
1075{
1076 struct io_u *io_u;
9c42684e 1077 unsigned int max_bs, min_write;
2e1df07d 1078 int cl_align, i, max_units;
2ae0b204 1079 int data_xfer = 1, err;
2e1df07d
JA
1080 char *p;
1081
1082 max_units = td->o.iodepth;
74f4b020 1083 max_bs = td_max_bs(td);
9c42684e 1084 min_write = td->o.min_bs[DDIR_WRITE];
2e1df07d
JA
1085 td->orig_buffer_size = (unsigned long long) max_bs
1086 * (unsigned long long) max_units;
1087
88045e04 1088 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
59d8d0f5
JA
1089 data_xfer = 0;
1090
2ae0b204
JA
1091 err = 0;
1092 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1093 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1094 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1095
1096 if (err) {
1097 log_err("fio: failed setting up IO queues\n");
1098 return 1;
1099 }
1100
fd8a09b8 1101 /*
1102 * if we may later need to do address alignment, then add any
1103 * possible adjustment here so that we don't cause a buffer
1104 * overflow later. this adjustment may be too much if we get
1105 * lucky and the allocator gives us an aligned address.
1106 */
d01612f3
CM
1107 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1108 (td->io_ops->flags & FIO_RAWIO))
fd8a09b8 1109 td->orig_buffer_size += page_mask + td->o.mem_align;
1110
2e1df07d
JA
1111 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1112 unsigned long bs;
1113
1114 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1115 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1116 }
1117
1118 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1119 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1120 return 1;
1121 }
1122
59d8d0f5 1123 if (data_xfer && allocate_io_mem(td))
2e1df07d
JA
1124 return 1;
1125
d01612f3 1126 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
2e1df07d
JA
1127 (td->io_ops->flags & FIO_RAWIO))
1128 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1129 else
1130 p = td->orig_buffer;
1131
1132 cl_align = os_cache_line_size();
1133
1134 for (i = 0; i < max_units; i++) {
1135 void *ptr;
1136
1137 if (td->terminate)
1138 return 1;
1139
1140 ptr = fio_memalign(cl_align, sizeof(*io_u));
1141 if (!ptr) {
1142 log_err("fio: unable to allocate aligned memory\n");
1143 break;
1144 }
1145
1146 io_u = ptr;
1147 memset(io_u, 0, sizeof(*io_u));
2ae0b204 1148 INIT_FLIST_HEAD(&io_u->verify_list);
2e1df07d
JA
1149 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1150
59d8d0f5 1151 if (data_xfer) {
2e1df07d
JA
1152 io_u->buf = p;
1153 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1154
1155 if (td_write(td))
9c42684e 1156 io_u_fill_buffer(td, io_u, min_write, max_bs);
2e1df07d
JA
1157 if (td_write(td) && td->o.verify_pattern_bytes) {
1158 /*
1159 * Fill the buffer with the pattern if we are
1160 * going to be doing writes.
1161 */
ce35b1ec 1162 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
2e1df07d
JA
1163 }
1164 }
1165
1166 io_u->index = i;
1167 io_u->flags = IO_U_F_FREE;
2ae0b204
JA
1168 io_u_qpush(&td->io_u_freelist, io_u);
1169
1170 /*
1171 * io_u never leaves this stack, used for iteration of all
1172 * io_u buffers.
1173 */
1174 io_u_qpush(&td->io_u_all, io_u);
c73ed246
JA
1175
1176 if (td->io_ops->io_u_init) {
1177 int ret = td->io_ops->io_u_init(td, io_u);
1178
1179 if (ret) {
1180 log_err("fio: failed to init engine data: %d\n", ret);
1181 return 1;
1182 }
1183 }
1184
2e1df07d
JA
1185 p += max_bs;
1186 }
1187
ca09be4b
JA
1188 if (td->o.verify != VERIFY_NONE) {
1189 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1190 if (!td->last_write_comp) {
1191 log_err("fio: failed to alloc write comp data\n");
1192 return 1;
1193 }
1194 }
1195
2e1df07d
JA
1196 return 0;
1197}
1198
1199static int switch_ioscheduler(struct thread_data *td)
1200{
1201 char tmp[256], tmp2[128];
1202 FILE *f;
1203 int ret;
1204
1205 if (td->io_ops->flags & FIO_DISKLESSIO)
1206 return 0;
1207
1208 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1209
1210 f = fopen(tmp, "r+");
1211 if (!f) {
1212 if (errno == ENOENT) {
1213 log_err("fio: os or kernel doesn't support IO scheduler"
1214 " switching\n");
1215 return 0;
1216 }
1217 td_verror(td, errno, "fopen iosched");
1218 return 1;
1219 }
1220
1221 /*
1222 * Set io scheduler.
1223 */
1224 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1225 if (ferror(f) || ret != 1) {
1226 td_verror(td, errno, "fwrite");
1227 fclose(f);
1228 return 1;
1229 }
1230
1231 rewind(f);
1232
1233 /*
1234 * Read back and check that the selected scheduler is now the default.
1235 */
b44b9e45 1236 memset(tmp, 0, sizeof(tmp));
49c6f33d 1237 ret = fread(tmp, sizeof(tmp), 1, f);
2e1df07d
JA
1238 if (ferror(f) || ret < 0) {
1239 td_verror(td, errno, "fread");
1240 fclose(f);
1241 return 1;
1242 }
b44b9e45
TK
1243 /*
1244 * either a list of io schedulers or "none\n" is expected.
1245 */
1246 tmp[strlen(tmp) - 1] = '\0';
49c6f33d 1247
2e1df07d
JA
1248
1249 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1250 if (!strstr(tmp, tmp2)) {
1251 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1252 td_verror(td, EINVAL, "iosched_switch");
1253 fclose(f);
1254 return 1;
1255 }
1256
1257 fclose(f);
1258 return 0;
1259}
1260
e39c0676 1261static bool keep_running(struct thread_data *td)
2e1df07d 1262{
77731b29
JA
1263 unsigned long long limit;
1264
2e1df07d 1265 if (td->done)
e39c0676 1266 return false;
2e1df07d 1267 if (td->o.time_based)
e39c0676 1268 return true;
2e1df07d
JA
1269 if (td->o.loops) {
1270 td->o.loops--;
e39c0676 1271 return true;
2e1df07d 1272 }
3939fe85 1273 if (exceeds_number_ios(td))
e39c0676 1274 return false;
26251d8d 1275
77731b29
JA
1276 if (td->o.io_limit)
1277 limit = td->o.io_limit;
1278 else
1279 limit = td->o.size;
1280
1281 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
5bd5f71a
JA
1282 uint64_t diff;
1283
1284 /*
1285 * If the difference is less than the minimum IO size, we
1286 * are done.
1287 */
77731b29 1288 diff = limit - ddir_rw_sum(td->io_bytes);
74f4b020 1289 if (diff < td_max_bs(td))
e39c0676 1290 return false;
5bd5f71a 1291
002fe734 1292 if (fio_files_done(td))
e39c0676 1293 return false;
002fe734 1294
e39c0676 1295 return true;
5bd5f71a 1296 }
2e1df07d 1297
e39c0676 1298 return false;
2e1df07d
JA
1299}
1300
ce486495 1301static int exec_string(struct thread_options *o, const char *string, const char *mode)
2e1df07d 1302{
6c9ce469 1303 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
f491a907 1304 int ret;
2e1df07d
JA
1305 char *str;
1306
1307 str = malloc(newlen);
ce486495 1308 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
2e1df07d 1309
ce486495 1310 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
2e1df07d
JA
1311 ret = system(str);
1312 if (ret == -1)
1313 log_err("fio: exec of cmd <%s> failed\n", str);
1314
1315 free(str);
1316 return ret;
1317}
1318
62167762
JC
1319/*
1320 * Dry run to compute correct state of numberio for verification.
1321 */
1322static uint64_t do_dry_run(struct thread_data *td)
1323{
62167762
JC
1324 td_set_runstate(td, TD_RUNNING);
1325
1326 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
e28dd2cf 1327 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
62167762
JC
1328 struct io_u *io_u;
1329 int ret;
1330
1331 if (td->terminate || td->done)
1332 break;
1333
1334 io_u = get_io_u(td);
1335 if (!io_u)
1336 break;
1337
a9da8ab2 1338 io_u_set(io_u, IO_U_F_FLIGHT);
62167762
JC
1339 io_u->error = 0;
1340 io_u->resid = 0;
1341 if (ddir_rw(acct_ddir(io_u)))
1342 td->io_issues[acct_ddir(io_u)]++;
1343 if (ddir_rw(io_u->ddir)) {
1344 io_u_mark_depth(td, 1);
1345 td->ts.total_io_u[io_u->ddir]++;
1346 }
1347
2e63e96b
PV
1348 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1349 td->o.do_verify &&
1350 td->o.verify != VERIFY_NONE &&
1351 !td->o.experimental_verify)
1352 log_io_piece(td, io_u);
1353
55312f9f 1354 ret = io_u_sync_complete(td, io_u);
62167762
JC
1355 (void) ret;
1356 }
1357
55312f9f 1358 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
62167762
JC
1359}
1360
2e1df07d
JA
1361/*
1362 * Entry point for the thread based jobs. The process based jobs end up
1363 * here as well, after a little setup.
1364 */
1365static void *thread_main(void *data)
1366{
95603b74 1367 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
2e1df07d 1368 struct thread_data *td = data;
4896473e 1369 struct thread_options *o = &td->o;
2e1df07d
JA
1370 pthread_condattr_t attr;
1371 int clear_state;
28727df7 1372 int ret;
2e1df07d 1373
4896473e 1374 if (!o->use_thread) {
2e1df07d
JA
1375 setsid();
1376 td->pid = getpid();
1377 } else
1378 td->pid = gettid();
1379
4896473e 1380 fio_local_clock_init(o->use_thread);
5d879392 1381
2e1df07d
JA
1382 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1383
122c7725
JA
1384 if (is_backend)
1385 fio_server_send_start(td);
1386
2e1df07d
JA
1387 INIT_FLIST_HEAD(&td->io_log_list);
1388 INIT_FLIST_HEAD(&td->io_hist_list);
1389 INIT_FLIST_HEAD(&td->verify_list);
1390 INIT_FLIST_HEAD(&td->trim_list);
1ae83d45 1391 INIT_FLIST_HEAD(&td->next_rand_list);
2e1df07d
JA
1392 pthread_mutex_init(&td->io_u_lock, NULL);
1393 td->io_hist_tree = RB_ROOT;
1394
1395 pthread_condattr_init(&attr);
1396 pthread_cond_init(&td->verify_cond, &attr);
1397 pthread_cond_init(&td->free_cond, &attr);
1398
1399 td_set_runstate(td, TD_INITIALIZED);
1400 dprint(FD_MUTEX, "up startup_mutex\n");
1401 fio_mutex_up(startup_mutex);
1402 dprint(FD_MUTEX, "wait on td->mutex\n");
1403 fio_mutex_down(td->mutex);
1404 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1405
2e1df07d
JA
1406 /*
1407 * A new gid requires privilege, so we need to do this before setting
1408 * the uid.
1409 */
4896473e 1410 if (o->gid != -1U && setgid(o->gid)) {
2e1df07d
JA
1411 td_verror(td, errno, "setgid");
1412 goto err;
1413 }
4896473e 1414 if (o->uid != -1U && setuid(o->uid)) {
2e1df07d
JA
1415 td_verror(td, errno, "setuid");
1416 goto err;
1417 }
1418
1419 /*
1420 * If we have a gettimeofday() thread, make sure we exclude that
1421 * thread from this job
1422 */
4896473e
JA
1423 if (o->gtod_cpu)
1424 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
2e1df07d
JA
1425
1426 /*
1427 * Set affinity first, in case it has an impact on the memory
1428 * allocations.
1429 */
b2a9e649 1430 if (fio_option_is_set(o, cpumask)) {
c2acfbac 1431 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
30cb4c65 1432 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
c2acfbac
JA
1433 if (!ret) {
1434 log_err("fio: no CPUs set\n");
1435 log_err("fio: Try increasing number of available CPUs\n");
1436 td_verror(td, EINVAL, "cpus_split");
1437 goto err;
1438 }
1439 }
28727df7
JA
1440 ret = fio_setaffinity(td->pid, o->cpumask);
1441 if (ret == -1) {
4896473e
JA
1442 td_verror(td, errno, "cpu_set_affinity");
1443 goto err;
1444 }
2e1df07d
JA
1445 }
1446
67bf9823 1447#ifdef CONFIG_LIBNUMA
d0b937ed 1448 /* numa node setup */
b2a9e649
JA
1449 if (fio_option_is_set(o, numa_cpunodes) ||
1450 fio_option_is_set(o, numa_memnodes)) {
43522848 1451 struct bitmask *mask;
d0b937ed
YR
1452
1453 if (numa_available() < 0) {
1454 td_verror(td, errno, "Does not support NUMA API\n");
1455 goto err;
1456 }
1457
b2a9e649 1458 if (fio_option_is_set(o, numa_cpunodes)) {
43522848
DG
1459 mask = numa_parse_nodestring(o->numa_cpunodes);
1460 ret = numa_run_on_node_mask(mask);
1461 numa_free_nodemask(mask);
d0b937ed
YR
1462 if (ret == -1) {
1463 td_verror(td, errno, \
1464 "numa_run_on_node_mask failed\n");
1465 goto err;
1466 }
1467 }
1468
b2a9e649 1469 if (fio_option_is_set(o, numa_memnodes)) {
43522848
DG
1470 mask = NULL;
1471 if (o->numa_memnodes)
1472 mask = numa_parse_nodestring(o->numa_memnodes);
1473
4896473e 1474 switch (o->numa_mem_mode) {
d0b937ed 1475 case MPOL_INTERLEAVE:
43522848 1476 numa_set_interleave_mask(mask);
d0b937ed
YR
1477 break;
1478 case MPOL_BIND:
43522848 1479 numa_set_membind(mask);
d0b937ed
YR
1480 break;
1481 case MPOL_LOCAL:
1482 numa_set_localalloc();
1483 break;
1484 case MPOL_PREFERRED:
4896473e 1485 numa_set_preferred(o->numa_mem_prefer_node);
d0b937ed
YR
1486 break;
1487 case MPOL_DEFAULT:
1488 default:
1489 break;
1490 }
1491
43522848
DG
1492 if (mask)
1493 numa_free_nodemask(mask);
1494
d0b937ed
YR
1495 }
1496 }
1497#endif
1498
9a3f1100
JA
1499 if (fio_pin_memory(td))
1500 goto err;
1501
2e1df07d
JA
1502 /*
1503 * May alter parameters that init_io_u() will use, so we need to
1504 * do this first.
1505 */
1506 if (init_iolog(td))
1507 goto err;
1508
1509 if (init_io_u(td))
1510 goto err;
1511
4896473e 1512 if (o->verify_async && verify_async_init(td))
2e1df07d
JA
1513 goto err;
1514
27d74836
JA
1515 if (fio_option_is_set(o, ioprio) ||
1516 fio_option_is_set(o, ioprio_class)) {
28727df7
JA
1517 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1518 if (ret == -1) {
2e1df07d
JA
1519 td_verror(td, errno, "ioprio_set");
1520 goto err;
1521 }
1522 }
1523
4896473e 1524 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
2e1df07d
JA
1525 goto err;
1526
649c10c5 1527 errno = 0;
4896473e 1528 if (nice(o->nice) == -1 && errno != 0) {
2e1df07d
JA
1529 td_verror(td, errno, "nice");
1530 goto err;
1531 }
1532
4896473e 1533 if (o->ioscheduler && switch_ioscheduler(td))
2e1df07d
JA
1534 goto err;
1535
4896473e 1536 if (!o->create_serialize && setup_files(td))
2e1df07d
JA
1537 goto err;
1538
1539 if (td_io_init(td))
1540 goto err;
1541
1542 if (init_random_map(td))
1543 goto err;
1544
ce486495 1545 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
4896473e 1546 goto err;
2e1df07d 1547
4896473e 1548 if (o->pre_read) {
2e1df07d
JA
1549 if (pre_read_files(td) < 0)
1550 goto err;
1551 }
1552
155f2f02
JA
1553 if (iolog_compress_init(td))
1554 goto err;
aee2ab67 1555
dc5bfbb2
JA
1556 fio_verify_init(td);
1557
103b174e 1558 if (rate_submit_init(td))
a9da8ab2
JA
1559 goto err;
1560
2e1df07d 1561 fio_gettime(&td->epoch, NULL);
44404c5a 1562 fio_getrusage(&td->ru_start);
ac28d905
JA
1563 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1564 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1565
1566 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1567 o->ratemin[DDIR_TRIM]) {
1568 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1569 sizeof(td->bw_sample_time));
1570 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1571 sizeof(td->bw_sample_time));
1572 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1573 sizeof(td->bw_sample_time));
1574 }
1575
2e1df07d
JA
1576 clear_state = 0;
1577 while (keep_running(td)) {
100f49f1
JA
1578 uint64_t verify_bytes;
1579
2e1df07d 1580 fio_gettime(&td->start, NULL);
2e1df07d
JA
1581 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1582
2e1df07d 1583 if (clear_state)
ac28d905 1584 clear_io_state(td, 0);
2e1df07d
JA
1585
1586 prune_io_piece_log(td);
1587
62167762
JC
1588 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1589 verify_bytes = do_dry_run(td);
095196b1 1590 else {
62167762 1591 verify_bytes = do_io(td);
095196b1
JA
1592 if (!verify_bytes)
1593 fio_mark_td_terminate(td);
1594 }
2e1df07d
JA
1595
1596 clear_state = 1;
1597
40c666c8
JA
1598 /*
1599 * Make sure we've successfully updated the rusage stats
1600 * before waiting on the stat mutex. Otherwise we could have
1601 * the stat thread holding stat mutex and waiting for
1602 * the rusage_sem, which would never get upped because
1603 * this thread is waiting for the stat mutex.
1604 */
1605 check_update_rusage(td);
1606
e5437a07 1607 fio_mutex_down(stat_mutex);
95603b74
BF
1608 if (td_read(td) && td->io_bytes[DDIR_READ])
1609 update_runtime(td, elapsed_us, DDIR_READ);
1610 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1611 update_runtime(td, elapsed_us, DDIR_WRITE);
1612 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1613 update_runtime(td, elapsed_us, DDIR_TRIM);
e5437a07
VT
1614 fio_gettime(&td->start, NULL);
1615 fio_mutex_up(stat_mutex);
2e1df07d
JA
1616
1617 if (td->error || td->terminate)
1618 break;
1619
4896473e
JA
1620 if (!o->do_verify ||
1621 o->verify == VERIFY_NONE ||
2e1df07d
JA
1622 (td->io_ops->flags & FIO_UNIDIR))
1623 continue;
1624
ac28d905 1625 clear_io_state(td, 0);
2e1df07d
JA
1626
1627 fio_gettime(&td->start, NULL);
1628
100f49f1 1629 do_verify(td, verify_bytes);
2e1df07d 1630
40c666c8
JA
1631 /*
1632 * See comment further up for why this is done here.
1633 */
1634 check_update_rusage(td);
1635
e5437a07 1636 fio_mutex_down(stat_mutex);
95603b74 1637 update_runtime(td, elapsed_us, DDIR_READ);
e5437a07
VT
1638 fio_gettime(&td->start, NULL);
1639 fio_mutex_up(stat_mutex);
2e1df07d
JA
1640
1641 if (td->error || td->terminate)
1642 break;
1643 }
1644
1645 update_rusage_stat(td);
2e1df07d 1646 td->ts.total_run_time = mtime_since_now(&td->epoch);
6eaf09d6
SL
1647 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1648 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1649 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2e1df07d 1650
ca09be4b 1651 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
d264264a
JA
1652 (td->o.verify != VERIFY_NONE && td_write(td)))
1653 verify_save_state(td->thread_number);
ca09be4b 1654
9a3f1100
JA
1655 fio_unpin_memory(td);
1656
905e3d4f 1657 fio_writeout_logs(td);
2e1df07d 1658
155f2f02 1659 iolog_compress_exit(td);
103b174e 1660 rate_submit_exit(td);
aee2ab67 1661
4896473e 1662 if (o->exec_postrun)
ce486495 1663 exec_string(o, o->exec_postrun, (const char *)"postrun");
2e1df07d 1664
f9cafb12 1665 if (exitall_on_terminate || (o->exitall_error && td->error))
2e1df07d
JA
1666 fio_terminate_threads(td->groupid);
1667
1668err:
1669 if (td->error)
1670 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1671 td->verror);
1672
4896473e 1673 if (o->verify_async)
2e1df07d
JA
1674 verify_async_exit(td);
1675
1676 close_and_free_files(td);
2e1df07d 1677 cleanup_io_u(td);
32dbca2c 1678 close_ioengine(td);
2e1df07d 1679 cgroup_shutdown(td, &cgroup_mnt);
ca09be4b 1680 verify_free_state(td);
2e1df07d 1681
b2a9e649 1682 if (fio_option_is_set(o, cpumask)) {
8a1db9a1
JA
1683 ret = fio_cpuset_exit(&o->cpumask);
1684 if (ret)
1685 td_verror(td, ret, "fio_cpuset_exit");
2e1df07d
JA
1686 }
1687
1688 /*
1689 * do this very late, it will log file closing as well
1690 */
4896473e 1691 if (o->write_iolog_file)
2e1df07d
JA
1692 write_iolog_close(td);
1693
ea66e04f
JA
1694 fio_mutex_remove(td->mutex);
1695 td->mutex = NULL;
1696
2e1df07d 1697 td_set_runstate(td, TD_EXITED);
fda2cfac
JA
1698
1699 /*
1700 * Do this last after setting our runstate to exited, so we
1701 * know that the stat thread is signaled.
1702 */
1703 check_update_rusage(td);
1704
e43606c2 1705 return (void *) (uintptr_t) td->error;
2e1df07d
JA
1706}
1707
1708
1709/*
1710 * We cannot pass the td data into a forked process, so attach the td and
1711 * pass it to the thread worker.
1712 */
1713static int fork_main(int shmid, int offset)
1714{
1715 struct thread_data *td;
1716 void *data, *ret;
1717
c8931876 1718#if !defined(__hpux) && !defined(CONFIG_NO_SHM)
2e1df07d
JA
1719 data = shmat(shmid, NULL, 0);
1720 if (data == (void *) -1) {
1721 int __err = errno;
1722
1723 perror("shmat");
1724 return __err;
1725 }
1726#else
1727 /*
1728 * HP-UX inherits shm mappings?
1729 */
1730 data = threads;
1731#endif
1732
1733 td = data + offset * sizeof(struct thread_data);
1734 ret = thread_main(td);
1735 shmdt(data);
e43606c2 1736 return (int) (uintptr_t) ret;
2e1df07d
JA
1737}
1738
cba5460c
JA
1739static void dump_td_info(struct thread_data *td)
1740{
1741 log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1742 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1743 (unsigned long) time_since_now(&td->terminate_time));
1744}
1745
2e1df07d
JA
1746/*
1747 * Run over the job map and reap the threads that have exited, if any.
1748 */
1749static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1750 unsigned int *m_rate)
1751{
1752 struct thread_data *td;
1753 unsigned int cputhreads, realthreads, pending;
1754 int i, status, ret;
1755
1756 /*
1757 * reap exited threads (TD_EXITED -> TD_REAPED)
1758 */
1759 realthreads = pending = cputhreads = 0;
1760 for_each_td(td, i) {
1761 int flags = 0;
1762
1763 /*
1764 * ->io_ops is NULL for a thread that has closed its
1765 * io engine
1766 */
1767 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1768 cputhreads++;
1769 else
1770 realthreads++;
1771
1772 if (!td->pid) {
1773 pending++;
1774 continue;
1775 }
1776 if (td->runstate == TD_REAPED)
1777 continue;
1778 if (td->o.use_thread) {
1779 if (td->runstate == TD_EXITED) {
1780 td_set_runstate(td, TD_REAPED);
1781 goto reaped;
1782 }
1783 continue;
1784 }
1785
1786 flags = WNOHANG;
1787 if (td->runstate == TD_EXITED)
1788 flags = 0;
1789
1790 /*
1791 * check if someone quit or got killed in an unusual way
1792 */
1793 ret = waitpid(td->pid, &status, flags);
1794 if (ret < 0) {
1795 if (errno == ECHILD) {
1796 log_err("fio: pid=%d disappeared %d\n",
1797 (int) td->pid, td->runstate);
a5e371a6 1798 td->sig = ECHILD;
2e1df07d
JA
1799 td_set_runstate(td, TD_REAPED);
1800 goto reaped;
1801 }
1802 perror("waitpid");
1803 } else if (ret == td->pid) {
1804 if (WIFSIGNALED(status)) {
1805 int sig = WTERMSIG(status);
1806
36d80bc7 1807 if (sig != SIGTERM && sig != SIGUSR2)
2e1df07d
JA
1808 log_err("fio: pid=%d, got signal=%d\n",
1809 (int) td->pid, sig);
a5e371a6 1810 td->sig = sig;
2e1df07d
JA
1811 td_set_runstate(td, TD_REAPED);
1812 goto reaped;
1813 }
1814 if (WIFEXITED(status)) {
1815 if (WEXITSTATUS(status) && !td->error)
1816 td->error = WEXITSTATUS(status);
1817
1818 td_set_runstate(td, TD_REAPED);
1819 goto reaped;
1820 }
1821 }
1822
cba5460c
JA
1823 /*
1824 * If the job is stuck, do a forceful timeout of it and
1825 * move on.
1826 */
1827 if (td->terminate &&
1828 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1829 dump_td_info(td);
1830 td_set_runstate(td, TD_REAPED);
1831 goto reaped;
1832 }
1833
2e1df07d
JA
1834 /*
1835 * thread is not dead, continue
1836 */
1837 pending++;
1838 continue;
1839reaped:
1840 (*nr_running)--;
342f4be4
JA
1841 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1842 (*t_rate) -= ddir_rw_sum(td->o.rate);
2e1df07d
JA
1843 if (!td->pid)
1844 pending--;
1845
1846 if (td->error)
1847 exit_value++;
1848
1849 done_secs += mtime_since_now(&td->epoch) / 1000;
4a88752a 1850 profile_td_exit(td);
2e1df07d
JA
1851 }
1852
1853 if (*nr_running == cputhreads && !pending && realthreads)
1854 fio_terminate_threads(TERMINATE_ALL);
1855}
1856
e39c0676 1857static bool __check_trigger_file(void)
ca09be4b
JA
1858{
1859 struct stat sb;
1860
1861 if (!trigger_file)
e39c0676 1862 return false;
ca09be4b
JA
1863
1864 if (stat(trigger_file, &sb))
e39c0676 1865 return false;
ca09be4b
JA
1866
1867 if (unlink(trigger_file) < 0)
1868 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1869 strerror(errno));
1870
e39c0676 1871 return true;
ca09be4b
JA
1872}
1873
e39c0676 1874static bool trigger_timedout(void)
ca09be4b
JA
1875{
1876 if (trigger_timeout)
1877 return time_since_genesis() >= trigger_timeout;
1878
e39c0676 1879 return false;
ca09be4b
JA
1880}
1881
1882void exec_trigger(const char *cmd)
1883{
1884 int ret;
1885
1886 if (!cmd)
1887 return;
1888
1889 ret = system(cmd);
1890 if (ret == -1)
1891 log_err("fio: failed executing %s trigger\n", cmd);
1892}
1893
1894void check_trigger_file(void)
1895{
1896 if (__check_trigger_file() || trigger_timedout()) {
796fb3ce
JA
1897 if (nr_clients)
1898 fio_clients_send_trigger(trigger_remote_cmd);
1899 else {
d264264a 1900 verify_save_state(IO_LIST_ALL);
ca09be4b
JA
1901 fio_terminate_threads(TERMINATE_ALL);
1902 exec_trigger(trigger_cmd);
1903 }
1904 }
1905}
1906
1907static int fio_verify_load_state(struct thread_data *td)
1908{
1909 int ret;
1910
1911 if (!td->o.verify_state)
1912 return 0;
1913
1914 if (is_backend) {
1915 void *data;
c3546b53 1916 int ver;
ca09be4b
JA
1917
1918 ret = fio_server_get_verify_state(td->o.name,
c3546b53 1919 td->thread_number - 1, &data, &ver);
ca09be4b 1920 if (!ret)
c3546b53 1921 verify_convert_assign_state(td, data, ver);
ca09be4b
JA
1922 } else
1923 ret = verify_load_state(td, "local");
1924
1925 return ret;
1926}
1927
06464907
JA
1928static void do_usleep(unsigned int usecs)
1929{
1930 check_for_running_stats();
ca09be4b 1931 check_trigger_file();
06464907
JA
1932 usleep(usecs);
1933}
1934
e39c0676 1935static bool check_mount_writes(struct thread_data *td)
e81ecca3
JA
1936{
1937 struct fio_file *f;
1938 unsigned int i;
1939
1940 if (!td_write(td) || td->o.allow_mounted_write)
e39c0676 1941 return false;
e81ecca3
JA
1942
1943 for_each_file(td, f, i) {
1944 if (f->filetype != FIO_TYPE_BD)
1945 continue;
1946 if (device_is_mounted(f->file_name))
1947 goto mounted;
1948 }
1949
e39c0676 1950 return false;
e81ecca3
JA
1951mounted:
1952 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
e39c0676 1953 return true;
e81ecca3
JA
1954}
1955
2e1df07d
JA
1956/*
1957 * Main function for kicking off and reaping jobs, as needed.
1958 */
1959static void run_threads(void)
1960{
1961 struct thread_data *td;
2e1df07d 1962 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
0de5b26f 1963 uint64_t spent;
2e1df07d 1964
2e1df07d
JA
1965 if (fio_gtod_offload && fio_start_gtod_thread())
1966 return;
334185e9 1967
f2a2ce0e 1968 fio_idle_prof_init();
2e1df07d
JA
1969
1970 set_sig_handlers();
1971
3a5f6bde
JA
1972 nr_thread = nr_process = 0;
1973 for_each_td(td, i) {
e81ecca3
JA
1974 if (check_mount_writes(td))
1975 return;
3a5f6bde
JA
1976 if (td->o.use_thread)
1977 nr_thread++;
1978 else
1979 nr_process++;
1980 }
1981
01cfefcc 1982 if (output_format & FIO_OUTPUT_NORMAL) {
2e1df07d
JA
1983 log_info("Starting ");
1984 if (nr_thread)
1985 log_info("%d thread%s", nr_thread,
1986 nr_thread > 1 ? "s" : "");
1987 if (nr_process) {
1988 if (nr_thread)
1989 log_info(" and ");
1990 log_info("%d process%s", nr_process,
1991 nr_process > 1 ? "es" : "");
1992 }
1993 log_info("\n");
e411c301 1994 log_info_flush();
2e1df07d
JA
1995 }
1996
1997 todo = thread_number;
1998 nr_running = 0;
1999 nr_started = 0;
2000 m_rate = t_rate = 0;
2001
2002 for_each_td(td, i) {
2003 print_status_init(td->thread_number - 1);
2004
2005 if (!td->o.create_serialize)
2006 continue;
2007
ca09be4b
JA
2008 if (fio_verify_load_state(td))
2009 goto reap;
2010
2e1df07d
JA
2011 /*
2012 * do file setup here so it happens sequentially,
2013 * we don't want X number of threads getting their
2014 * client data interspersed on disk
2015 */
2016 if (setup_files(td)) {
ca09be4b 2017reap:
2e1df07d
JA
2018 exit_value++;
2019 if (td->error)
2020 log_err("fio: pid=%d, err=%d/%s\n",
2021 (int) td->pid, td->error, td->verror);
2022 td_set_runstate(td, TD_REAPED);
2023 todo--;
2024 } else {
2025 struct fio_file *f;
2026 unsigned int j;
2027
2028 /*
2029 * for sharing to work, each job must always open
2030 * its own files. so close them, if we opened them
2031 * for creation
2032 */
2033 for_each_file(td, f, j) {
2034 if (fio_file_open(f))
2035 td_io_close_file(td, f);
2036 }
2037 }
2038 }
2039
f2a2ce0e
HL
2040 /* start idle threads before io threads start to run */
2041 fio_idle_prof_start();
2042
2e1df07d
JA
2043 set_genesis_time();
2044
2045 while (todo) {
2046 struct thread_data *map[REAL_MAX_JOBS];
2047 struct timeval this_start;
2048 int this_jobs = 0, left;
2049
2050 /*
2051 * create threads (TD_NOT_CREATED -> TD_CREATED)
2052 */
2053 for_each_td(td, i) {
2054 if (td->runstate != TD_NOT_CREATED)
2055 continue;
2056
2057 /*
2058 * never got a chance to start, killed by other
2059 * thread for some reason
2060 */
2061 if (td->terminate) {
2062 todo--;
2063 continue;
2064 }
2065
2066 if (td->o.start_delay) {
0de5b26f 2067 spent = utime_since_genesis();
2e1df07d 2068
74454ce4 2069 if (td->o.start_delay > spent)
2e1df07d
JA
2070 continue;
2071 }
2072
2073 if (td->o.stonewall && (nr_started || nr_running)) {
2074 dprint(FD_PROCESS, "%s: stonewall wait\n",
2075 td->o.name);
2076 break;
2077 }
2078
2079 init_disk_util(td);
2080
c97f1ad6
JA
2081 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2082 td->update_rusage = 0;
2083
2e1df07d
JA
2084 /*
2085 * Set state to created. Thread will transition
2086 * to TD_INITIALIZED when it's done setting up.
2087 */
2088 td_set_runstate(td, TD_CREATED);
2089 map[this_jobs++] = td;
2090 nr_started++;
2091
2092 if (td->o.use_thread) {
2093 int ret;
2094
2095 dprint(FD_PROCESS, "will pthread_create\n");
2096 ret = pthread_create(&td->thread, NULL,
2097 thread_main, td);
2098 if (ret) {
2099 log_err("pthread_create: %s\n",
2100 strerror(ret));
2101 nr_started--;
2102 break;
2103 }
2104 ret = pthread_detach(td->thread);
2105 if (ret)
2106 log_err("pthread_detach: %s",
2107 strerror(ret));
2108 } else {
2109 pid_t pid;
2110 dprint(FD_PROCESS, "will fork\n");
2111 pid = fork();
2112 if (!pid) {
2113 int ret = fork_main(shm_id, i);
2114
2115 _exit(ret);
2116 } else if (i == fio_debug_jobno)
2117 *fio_debug_jobp = pid;
2118 }
2119 dprint(FD_MUTEX, "wait on startup_mutex\n");
09400a60 2120 if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2e1df07d
JA
2121 log_err("fio: job startup hung? exiting.\n");
2122 fio_terminate_threads(TERMINATE_ALL);
2123 fio_abort = 1;
2124 nr_started--;
2125 break;
2126 }
2127 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2128 }
2129
2130 /*
2131 * Wait for the started threads to transition to
2132 * TD_INITIALIZED.
2133 */
2134 fio_gettime(&this_start, NULL);
2135 left = this_jobs;
2136 while (left && !fio_abort) {
2137 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2138 break;
2139
06464907 2140 do_usleep(100000);
2e1df07d
JA
2141
2142 for (i = 0; i < this_jobs; i++) {
2143 td = map[i];
2144 if (!td)
2145 continue;
2146 if (td->runstate == TD_INITIALIZED) {
2147 map[i] = NULL;
2148 left--;
2149 } else if (td->runstate >= TD_EXITED) {
2150 map[i] = NULL;
2151 left--;
2152 todo--;
2153 nr_running++; /* work-around... */
2154 }
2155 }
2156 }
2157
2158 if (left) {
4e87c37a
JA
2159 log_err("fio: %d job%s failed to start\n", left,
2160 left > 1 ? "s" : "");
2e1df07d
JA
2161 for (i = 0; i < this_jobs; i++) {
2162 td = map[i];
2163 if (!td)
2164 continue;
2165 kill(td->pid, SIGTERM);
2166 }
2167 break;
2168 }
2169
2170 /*
2171 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2172 */
2173 for_each_td(td, i) {
2174 if (td->runstate != TD_INITIALIZED)
2175 continue;
2176
2177 if (in_ramp_time(td))
2178 td_set_runstate(td, TD_RAMP);
2179 else
2180 td_set_runstate(td, TD_RUNNING);
2181 nr_running++;
2182 nr_started--;
342f4be4
JA
2183 m_rate += ddir_rw_sum(td->o.ratemin);
2184 t_rate += ddir_rw_sum(td->o.rate);
2e1df07d
JA
2185 todo--;
2186 fio_mutex_up(td->mutex);
2187 }
2188
2189 reap_threads(&nr_running, &t_rate, &m_rate);
2190
122c7725 2191 if (todo)
06464907 2192 do_usleep(100000);
2e1df07d
JA
2193 }
2194
2195 while (nr_running) {
2196 reap_threads(&nr_running, &t_rate, &m_rate);
06464907 2197 do_usleep(10000);
2e1df07d
JA
2198 }
2199
f2a2ce0e
HL
2200 fio_idle_prof_stop();
2201
2e1df07d 2202 update_io_ticks();
2e1df07d
JA
2203}
2204
5ddc6707 2205static void wait_for_helper_thread_exit(void)
9ec7779f 2206{
8aab824f
JA
2207 void *ret;
2208
98b4b0a2 2209 helper_exit = 1;
5ddc6707
JA
2210 pthread_cond_signal(&helper_cond);
2211 pthread_join(helper_thread, &ret);
9ec7779f
JA
2212}
2213
27357187
JA
2214static void free_disk_util(void)
2215{
27357187 2216 disk_util_prune_entries();
8aab824f 2217
5ddc6707 2218 pthread_cond_destroy(&helper_cond);
27357187
JA
2219}
2220
5ddc6707 2221static void *helper_thread_main(void *data)
2e1df07d 2222{
2a2804c9 2223 struct backend_data *d = data;
9ec7779f
JA
2224 int ret = 0;
2225
2a2804c9
JA
2226 if (d)
2227 pthread_setspecific(d->key, d->ptr);
2228
2e1df07d
JA
2229 fio_mutex_up(startup_mutex);
2230
8aab824f
JA
2231 while (!ret) {
2232 uint64_t sec = DISK_UTIL_MSEC / 1000;
2233 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2234 struct timespec ts;
2235 struct timeval tv;
2236
2237 gettimeofday(&tv, NULL);
2238 ts.tv_sec = tv.tv_sec + sec;
2239 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
f893b76d 2240
823952af 2241 if (ts.tv_nsec >= 1000000000ULL) {
8aab824f
JA
2242 ts.tv_nsec -= 1000000000ULL;
2243 ts.tv_sec++;
2244 }
2245
5ddc6707 2246 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
8aab824f 2247
9ec7779f 2248 ret = update_io_ticks();
2e1df07d 2249
5ddc6707
JA
2250 if (helper_do_stat) {
2251 helper_do_stat = 0;
2252 __show_running_run_stats();
2253 }
2254
2e1df07d
JA
2255 if (!is_backend)
2256 print_thread_status();
2257 }
2258
2259 return NULL;
2260}
2261
2a2804c9 2262static int create_helper_thread(struct backend_data *data)
2e1df07d
JA
2263{
2264 int ret;
2265
9ec7779f
JA
2266 setup_disk_util();
2267
5ddc6707
JA
2268 pthread_cond_init(&helper_cond, NULL);
2269 pthread_mutex_init(&helper_lock, NULL);
8aab824f 2270
2a2804c9 2271 ret = pthread_create(&helper_thread, NULL, helper_thread_main, data);
2e1df07d 2272 if (ret) {
5ddc6707 2273 log_err("Can't create helper thread: %s\n", strerror(ret));
2e1df07d
JA
2274 return 1;
2275 }
2276
2e1df07d
JA
2277 dprint(FD_MUTEX, "wait on startup_mutex\n");
2278 fio_mutex_down(startup_mutex);
2279 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2280 return 0;
2281}
2282
2a2804c9 2283int fio_backend(struct backend_data *data)
2e1df07d
JA
2284{
2285 struct thread_data *td;
2286 int i;
2287
2288 if (exec_profile) {
2289 if (load_profile(exec_profile))
2290 return 1;
2291 free(exec_profile);
2292 exec_profile = NULL;
2293 }
2294 if (!thread_number)
2295 return 0;
2296
2297 if (write_bw_log) {
aee2ab67
JA
2298 struct log_params p = {
2299 .log_type = IO_LOG_TYPE_BW,
2300 };
2301
2302 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2303 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2304 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2e1df07d
JA
2305 }
2306
521da527 2307 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2e1df07d
JA
2308 if (startup_mutex == NULL)
2309 return 1;
2e1df07d
JA
2310
2311 set_genesis_time();
cef9175e 2312 stat_init();
2a2804c9 2313 create_helper_thread(data);
2e1df07d
JA
2314
2315 cgroup_list = smalloc(sizeof(*cgroup_list));
2316 INIT_FLIST_HEAD(cgroup_list);
2317
2318 run_threads();
2319
5ddc6707 2320 wait_for_helper_thread_exit();
8aab824f 2321
2e1df07d 2322 if (!fio_abort) {
83f7b64e 2323 __show_run_stats();
2e1df07d 2324 if (write_bw_log) {
cb7e0ace
JA
2325 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2326 struct io_log *log = agg_io_log[i];
2327
3a5db920 2328 flush_log(log, 0);
518dac09 2329 free_log(log);
cb7e0ace 2330 }
2e1df07d
JA
2331 }
2332 }
2333
fda2cfac 2334 for_each_td(td, i) {
2e1df07d 2335 fio_options_free(td);
8049adc1
JA
2336 if (td->rusage_sem) {
2337 fio_mutex_remove(td->rusage_sem);
2338 td->rusage_sem = NULL;
2339 }
fda2cfac 2340 }
2e1df07d 2341
a462baef 2342 free_disk_util();
2e1df07d
JA
2343 cgroup_kill(cgroup_list);
2344 sfree(cgroup_list);
2345 sfree(cgroup_mnt);
2346
2347 fio_mutex_remove(startup_mutex);
cef9175e 2348 stat_exit();
2e1df07d
JA
2349 return exit_value;
2350}