mutex: abstract out cond/lock pshared init
[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"
a39fb9ea 60#include "helper_thread.h"
2e1df07d 61
2e1df07d 62static struct fio_mutex *startup_mutex;
2e1df07d
JA
63static struct flist_head *cgroup_list;
64static char *cgroup_mnt;
65static int exit_value;
66static volatile int fio_abort;
3a5f6bde
JA
67static unsigned int nr_process = 0;
68static unsigned int nr_thread = 0;
2e1df07d 69
6eaf09d6 70struct io_log *agg_io_log[DDIR_RWDIR_CNT];
2e1df07d 71
a3efc919
JA
72int groupid = 0;
73unsigned int thread_number = 0;
108fea77 74unsigned int stat_number = 0;
a3efc919
JA
75int shm_id = 0;
76int temp_stall_ts;
77unsigned long done_secs = 0;
a47591e4 78
2e1df07d 79#define PAGE_ALIGN(buf) \
e43606c2 80 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
2e1df07d
JA
81
82#define JOB_START_TIMEOUT (5 * 1000)
83
84static void sig_int(int sig)
85{
86 if (threads) {
87 if (is_backend)
88 fio_server_got_signal(sig);
89 else {
90 log_info("\nfio: terminating on signal %d\n", sig);
e411c301 91 log_info_flush();
2e1df07d
JA
92 exit_value = 128;
93 }
94
95 fio_terminate_threads(TERMINATE_ALL);
96 }
97}
98
d2235e56 99void sig_show_status(int sig)
4c6d91e8
JA
100{
101 show_running_run_stats();
102}
103
2e1df07d
JA
104static void set_sig_handlers(void)
105{
106 struct sigaction act;
107
108 memset(&act, 0, sizeof(act));
109 act.sa_handler = sig_int;
110 act.sa_flags = SA_RESTART;
111 sigaction(SIGINT, &act, NULL);
112
113 memset(&act, 0, sizeof(act));
114 act.sa_handler = sig_int;
115 act.sa_flags = SA_RESTART;
116 sigaction(SIGTERM, &act, NULL);
117
2f694507
BC
118/* Windows uses SIGBREAK as a quit signal from other applications */
119#ifdef WIN32
120 memset(&act, 0, sizeof(act));
121 act.sa_handler = sig_int;
122 act.sa_flags = SA_RESTART;
123 sigaction(SIGBREAK, &act, NULL);
124#endif
125
4c6d91e8
JA
126 memset(&act, 0, sizeof(act));
127 act.sa_handler = sig_show_status;
128 act.sa_flags = SA_RESTART;
129 sigaction(SIGUSR1, &act, NULL);
130
2e1df07d
JA
131 if (is_backend) {
132 memset(&act, 0, sizeof(act));
133 act.sa_handler = sig_int;
134 act.sa_flags = SA_RESTART;
135 sigaction(SIGPIPE, &act, NULL);
136 }
137}
138
139/*
140 * Check if we are above the minimum rate given.
141 */
e39c0676
JA
142static bool __check_min_rate(struct thread_data *td, struct timeval *now,
143 enum fio_ddir ddir)
2e1df07d
JA
144{
145 unsigned long long bytes = 0;
146 unsigned long iops = 0;
147 unsigned long spent;
148 unsigned long rate;
149 unsigned int ratemin = 0;
150 unsigned int rate_iops = 0;
151 unsigned int rate_iops_min = 0;
152
153 assert(ddir_rw(ddir));
154
155 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
e39c0676 156 return false;
2e1df07d
JA
157
158 /*
159 * allow a 2 second settle period in the beginning
160 */
161 if (mtime_since(&td->start, now) < 2000)
e39c0676 162 return false;
2e1df07d
JA
163
164 iops += td->this_io_blocks[ddir];
165 bytes += td->this_io_bytes[ddir];
166 ratemin += td->o.ratemin[ddir];
167 rate_iops += td->o.rate_iops[ddir];
168 rate_iops_min += td->o.rate_iops_min[ddir];
169
170 /*
171 * if rate blocks is set, sample is running
172 */
173 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
174 spent = mtime_since(&td->lastrate[ddir], now);
175 if (spent < td->o.ratecycle)
e39c0676 176 return false;
2e1df07d 177
20e37e4a 178 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
2e1df07d
JA
179 /*
180 * check bandwidth specified rate
181 */
182 if (bytes < td->rate_bytes[ddir]) {
183 log_err("%s: min rate %u not met\n", td->o.name,
184 ratemin);
e39c0676 185 return true;
2e1df07d 186 } else {
49cba9b3
JA
187 if (spent)
188 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
189 else
190 rate = 0;
191
2e1df07d
JA
192 if (rate < ratemin ||
193 bytes < td->rate_bytes[ddir]) {
194 log_err("%s: min rate %u not met, got"
195 " %luKB/sec\n", td->o.name,
196 ratemin, rate);
e39c0676 197 return true;
2e1df07d
JA
198 }
199 }
200 } else {
201 /*
202 * checks iops specified rate
203 */
204 if (iops < rate_iops) {
205 log_err("%s: min iops rate %u not met\n",
206 td->o.name, rate_iops);
e39c0676 207 return true;
2e1df07d 208 } else {
4c707a3b
JA
209 if (spent)
210 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
211 else
212 rate = 0;
213
2e1df07d
JA
214 if (rate < rate_iops_min ||
215 iops < td->rate_blocks[ddir]) {
216 log_err("%s: min iops rate %u not met,"
217 " got %lu\n", td->o.name,
218 rate_iops_min, rate);
e39c0676 219 return true;
2e1df07d
JA
220 }
221 }
222 }
223 }
224
225 td->rate_bytes[ddir] = bytes;
226 td->rate_blocks[ddir] = iops;
227 memcpy(&td->lastrate[ddir], now, sizeof(*now));
e39c0676 228 return false;
2e1df07d
JA
229}
230
e39c0676 231static bool check_min_rate(struct thread_data *td, struct timeval *now)
2e1df07d 232{
e39c0676 233 bool ret = false;
2e1df07d 234
55312f9f 235 if (td->bytes_done[DDIR_READ])
6eaf09d6 236 ret |= __check_min_rate(td, now, DDIR_READ);
55312f9f 237 if (td->bytes_done[DDIR_WRITE])
6eaf09d6 238 ret |= __check_min_rate(td, now, DDIR_WRITE);
55312f9f 239 if (td->bytes_done[DDIR_TRIM])
6eaf09d6 240 ret |= __check_min_rate(td, now, DDIR_TRIM);
2e1df07d
JA
241
242 return ret;
243}
244
245/*
246 * When job exits, we can cancel the in-flight IO if we are using async
247 * io. Attempt to do so.
248 */
249static void cleanup_pending_aio(struct thread_data *td)
250{
2e1df07d
JA
251 int r;
252
253 /*
254 * get immediately available events, if any
255 */
55312f9f 256 r = io_u_queued_complete(td, 0);
2e1df07d
JA
257 if (r < 0)
258 return;
259
260 /*
261 * now cancel remaining active events
262 */
263 if (td->io_ops->cancel) {
2ae0b204
JA
264 struct io_u *io_u;
265 int i;
2e1df07d 266
2ae0b204
JA
267 io_u_qiter(&td->io_u_all, io_u, i) {
268 if (io_u->flags & IO_U_F_FLIGHT) {
2e1df07d
JA
269 r = td->io_ops->cancel(td, io_u);
270 if (!r)
271 put_io_u(td, io_u);
272 }
273 }
274 }
275
276 if (td->cur_depth)
55312f9f 277 r = io_u_queued_complete(td, td->cur_depth);
2e1df07d
JA
278}
279
280/*
281 * Helper to handle the final sync of a file. Works just like the normal
282 * io path, just does everything sync.
283 */
e39c0676 284static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
2e1df07d
JA
285{
286 struct io_u *io_u = __get_io_u(td);
287 int ret;
288
289 if (!io_u)
e39c0676 290 return true;
2e1df07d
JA
291
292 io_u->ddir = DDIR_SYNC;
293 io_u->file = f;
294
295 if (td_io_prep(td, io_u)) {
296 put_io_u(td, io_u);
e39c0676 297 return true;
2e1df07d
JA
298 }
299
300requeue:
301 ret = td_io_queue(td, io_u);
302 if (ret < 0) {
303 td_verror(td, io_u->error, "td_io_queue");
304 put_io_u(td, io_u);
e39c0676 305 return true;
2e1df07d 306 } else if (ret == FIO_Q_QUEUED) {
42ff945c
JA
307 if (td_io_commit(td))
308 return true;
55312f9f 309 if (io_u_queued_complete(td, 1) < 0)
e39c0676 310 return true;
2e1df07d
JA
311 } else if (ret == FIO_Q_COMPLETED) {
312 if (io_u->error) {
313 td_verror(td, io_u->error, "td_io_queue");
e39c0676 314 return true;
2e1df07d
JA
315 }
316
55312f9f 317 if (io_u_sync_complete(td, io_u) < 0)
e39c0676 318 return true;
2e1df07d
JA
319 } else if (ret == FIO_Q_BUSY) {
320 if (td_io_commit(td))
e39c0676 321 return true;
2e1df07d
JA
322 goto requeue;
323 }
324
e39c0676 325 return false;
2e1df07d 326}
a3efc919 327
61ee0f86
JA
328static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
329{
330 int ret;
331
332 if (fio_file_open(f))
333 return fio_io_sync(td, f);
334
335 if (td_io_open_file(td, f))
336 return 1;
337
338 ret = fio_io_sync(td, f);
339 td_io_close_file(td, f);
340 return ret;
341}
342
2e1df07d
JA
343static inline void __update_tv_cache(struct thread_data *td)
344{
345 fio_gettime(&td->tv_cache, NULL);
346}
347
348static inline void update_tv_cache(struct thread_data *td)
349{
350 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
351 __update_tv_cache(td);
352}
353
e39c0676 354static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
2e1df07d
JA
355{
356 if (in_ramp_time(td))
e39c0676 357 return false;
2e1df07d 358 if (!td->o.timeout)
e39c0676 359 return false;
0de5b26f 360 if (utime_since(&td->epoch, t) >= td->o.timeout)
e39c0676 361 return true;
2e1df07d 362
e39c0676 363 return false;
2e1df07d
JA
364}
365
95603b74
BF
366/*
367 * We need to update the runtime consistently in ms, but keep a running
368 * tally of the current elapsed time in microseconds for sub millisecond
369 * updates.
370 */
371static inline void update_runtime(struct thread_data *td,
372 unsigned long long *elapsed_us,
373 const enum fio_ddir ddir)
374{
d5c8ea29
JA
375 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
376 return;
377
95603b74
BF
378 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
379 elapsed_us[ddir] += utime_since_now(&td->start);
380 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
381}
382
e39c0676
JA
383static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
384 int *retptr)
2e1df07d
JA
385{
386 int ret = *retptr;
387
388 if (ret < 0 || td->error) {
8b28bd41
DM
389 int err = td->error;
390 enum error_type_bit eb;
2e1df07d
JA
391
392 if (ret < 0)
393 err = -ret;
2e1df07d 394
8b28bd41
DM
395 eb = td_error_type(ddir, err);
396 if (!(td->o.continue_on_error & (1 << eb)))
e39c0676 397 return true;
2e1df07d 398
8b28bd41 399 if (td_non_fatal_error(td, eb, err)) {
2e1df07d
JA
400 /*
401 * Continue with the I/Os in case of
402 * a non fatal error.
403 */
404 update_error_count(td, err);
405 td_clear_error(td);
406 *retptr = 0;
e39c0676 407 return false;
2e1df07d
JA
408 } else if (td->o.fill_device && err == ENOSPC) {
409 /*
410 * We expect to hit this error if
411 * fill_device option is set.
412 */
413 td_clear_error(td);
ebea2133 414 fio_mark_td_terminate(td);
e39c0676 415 return true;
2e1df07d
JA
416 } else {
417 /*
418 * Stop the I/O in case of a fatal
419 * error.
420 */
421 update_error_count(td, err);
e39c0676 422 return true;
2e1df07d
JA
423 }
424 }
425
e39c0676 426 return false;
2e1df07d
JA
427}
428
c97f1ad6
JA
429static void check_update_rusage(struct thread_data *td)
430{
431 if (td->update_rusage) {
432 td->update_rusage = 0;
433 update_rusage_stat(td);
434 fio_mutex_up(td->rusage_sem);
435 }
436}
437
55312f9f 438static int wait_for_completions(struct thread_data *td, struct timeval *time)
69fea81e
JA
439{
440 const int full = queue_full(td);
441 int min_evts = 0;
442 int ret;
443
1fed2080
JA
444 if (td->flags & TD_F_REGROW_LOGS) {
445 ret = io_u_quiesce(td);
446 regrow_logs(td);
447 return ret;
448 }
449
69fea81e
JA
450 /*
451 * if the queue is full, we MUST reap at least 1 event
452 */
82407585 453 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
4b157ac6 454 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
69fea81e
JA
455 min_evts = 1;
456
66b6c5ef 457 if (time && (__should_check_rate(td, DDIR_READ) ||
69fea81e
JA
458 __should_check_rate(td, DDIR_WRITE) ||
459 __should_check_rate(td, DDIR_TRIM)))
460 fio_gettime(time, NULL);
461
462 do {
55312f9f 463 ret = io_u_queued_complete(td, min_evts);
69fea81e
JA
464 if (ret < 0)
465 break;
466 } while (full && (td->cur_depth > td->o.iodepth_low));
467
468 return ret;
469}
470
e9d512d8
JA
471int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
472 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
473 struct timeval *comp_time)
474{
475 int ret2;
476
477 switch (*ret) {
478 case FIO_Q_COMPLETED:
479 if (io_u->error) {
480 *ret = -io_u->error;
481 clear_io_u(td, io_u);
482 } else if (io_u->resid) {
483 int bytes = io_u->xfer_buflen - io_u->resid;
484 struct fio_file *f = io_u->file;
485
486 if (bytes_issued)
487 *bytes_issued += bytes;
488
489 if (!from_verify)
490 trim_io_piece(td, io_u);
491
492 /*
493 * zero read, fail
494 */
495 if (!bytes) {
496 if (!from_verify)
497 unlog_io_piece(td, io_u);
498 td_verror(td, EIO, "full resid");
499 put_io_u(td, io_u);
500 break;
501 }
502
503 io_u->xfer_buflen = io_u->resid;
504 io_u->xfer_buf += bytes;
505 io_u->offset += bytes;
506
507 if (ddir_rw(io_u->ddir))
508 td->ts.short_io_u[io_u->ddir]++;
509
510 f = io_u->file;
511 if (io_u->offset == f->real_file_size)
512 goto sync_done;
513
514 requeue_io_u(td, &io_u);
515 } else {
516sync_done:
517 if (comp_time && (__should_check_rate(td, DDIR_READ) ||
518 __should_check_rate(td, DDIR_WRITE) ||
519 __should_check_rate(td, DDIR_TRIM)))
520 fio_gettime(comp_time, NULL);
521
522 *ret = io_u_sync_complete(td, io_u);
523 if (*ret < 0)
524 break;
525 }
a0e3e5a6 526
cf2c8d52
JA
527 if (td->flags & TD_F_REGROW_LOGS)
528 regrow_logs(td);
529
a0e3e5a6
EW
530 /*
531 * when doing I/O (not when verifying),
532 * check for any errors that are to be ignored
533 */
534 if (!from_verify)
535 break;
536
e9d512d8
JA
537 return 0;
538 case FIO_Q_QUEUED:
539 /*
540 * if the engine doesn't have a commit hook,
541 * the io_u is really queued. if it does have such
542 * a hook, it has to call io_u_queued() itself.
543 */
544 if (td->io_ops->commit == NULL)
545 io_u_queued(td, io_u);
546 if (bytes_issued)
547 *bytes_issued += io_u->xfer_buflen;
548 break;
549 case FIO_Q_BUSY:
550 if (!from_verify)
551 unlog_io_piece(td, io_u);
552 requeue_io_u(td, &io_u);
553 ret2 = td_io_commit(td);
554 if (ret2 < 0)
555 *ret = ret2;
556 break;
557 default:
b0775325 558 assert(*ret < 0);
e9d512d8
JA
559 td_verror(td, -(*ret), "td_io_queue");
560 break;
561 }
562
563 if (break_on_this_error(td, ddir, ret))
564 return 1;
565
566 return 0;
567}
568
e39c0676 569static inline bool io_in_polling(struct thread_data *td)
82407585
RP
570{
571 return !td->o.iodepth_batch_complete_min &&
572 !td->o.iodepth_batch_complete_max;
573}
574
2e1df07d
JA
575/*
576 * The main verify engine. Runs over the writes we previously submitted,
577 * reads the blocks back in, and checks the crc/md5 of the data.
578 */
100f49f1 579static void do_verify(struct thread_data *td, uint64_t verify_bytes)
2e1df07d
JA
580{
581 struct fio_file *f;
582 struct io_u *io_u;
583 int ret, min_events;
584 unsigned int i;
585
586 dprint(FD_VERIFY, "starting loop\n");
587
588 /*
589 * sync io first and invalidate cache, to make sure we really
590 * read from disk.
591 */
592 for_each_file(td, f, i) {
593 if (!fio_file_open(f))
594 continue;
595 if (fio_io_sync(td, f))
596 break;
597 if (file_invalidate_cache(td, f))
598 break;
599 }
600
c97f1ad6
JA
601 check_update_rusage(td);
602
2e1df07d
JA
603 if (td->error)
604 return;
605
606 td_set_runstate(td, TD_VERIFYING);
607
608 io_u = NULL;
609 while (!td->terminate) {
fbccf46c 610 enum fio_ddir ddir;
e9d512d8 611 int full;
2e1df07d
JA
612
613 update_tv_cache(td);
c97f1ad6 614 check_update_rusage(td);
2e1df07d
JA
615
616 if (runtime_exceeded(td, &td->tv_cache)) {
617 __update_tv_cache(td);
618 if (runtime_exceeded(td, &td->tv_cache)) {
ebea2133 619 fio_mark_td_terminate(td);
2e1df07d
JA
620 break;
621 }
622 }
623
9e684a49
DE
624 if (flow_threshold_exceeded(td))
625 continue;
626
44cbc6da
JA
627 if (!td->o.experimental_verify) {
628 io_u = __get_io_u(td);
629 if (!io_u)
630 break;
2e1df07d 631
44cbc6da
JA
632 if (get_next_verify(td, io_u)) {
633 put_io_u(td, io_u);
634 break;
635 }
2e1df07d 636
44cbc6da
JA
637 if (td_io_prep(td, io_u)) {
638 put_io_u(td, io_u);
639 break;
640 }
641 } else {
55312f9f 642 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
100f49f1
JA
643 break;
644
bcd5abfa 645 while ((io_u = get_io_u(td)) != NULL) {
002fe734
JA
646 if (IS_ERR(io_u)) {
647 io_u = NULL;
648 ret = FIO_Q_BUSY;
649 goto reap;
650 }
651
bcd5abfa
JA
652 /*
653 * We are only interested in the places where
654 * we wrote or trimmed IOs. Turn those into
655 * reads for verification purposes.
656 */
657 if (io_u->ddir == DDIR_READ) {
658 /*
659 * Pretend we issued it for rwmix
660 * accounting
661 */
662 td->io_issues[DDIR_READ]++;
663 put_io_u(td, io_u);
664 continue;
665 } else if (io_u->ddir == DDIR_TRIM) {
666 io_u->ddir = DDIR_READ;
a9da8ab2 667 io_u_set(io_u, IO_U_F_TRIMMED);
bcd5abfa
JA
668 break;
669 } else if (io_u->ddir == DDIR_WRITE) {
670 io_u->ddir = DDIR_READ;
671 break;
672 } else {
673 put_io_u(td, io_u);
674 continue;
675 }
676 }
44cbc6da 677
bcd5abfa 678 if (!io_u)
44cbc6da 679 break;
2e1df07d
JA
680 }
681
ca09be4b
JA
682 if (verify_state_should_stop(td, io_u)) {
683 put_io_u(td, io_u);
684 break;
685 }
686
2e1df07d
JA
687 if (td->o.verify_async)
688 io_u->end_io = verify_io_u_async;
689 else
690 io_u->end_io = verify_io_u;
691
fbccf46c 692 ddir = io_u->ddir;
10adc4a7
GG
693 if (!td->o.disable_slat)
694 fio_gettime(&io_u->start_time, NULL);
fbccf46c 695
2e1df07d 696 ret = td_io_queue(td, io_u);
2e1df07d 697
e9d512d8 698 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
2e1df07d
JA
699 break;
700
701 /*
702 * if we can queue more, do so. but check if there are
703 * completed io_u's first. Note that we can get BUSY even
704 * without IO queued, if the system is resource starved.
705 */
002fe734 706reap:
2e1df07d 707 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
82407585 708 if (full || io_in_polling(td))
55312f9f 709 ret = wait_for_completions(td, NULL);
2e1df07d 710
2e1df07d
JA
711 if (ret < 0)
712 break;
713 }
714
c97f1ad6
JA
715 check_update_rusage(td);
716
2e1df07d
JA
717 if (!td->error) {
718 min_events = td->cur_depth;
719
720 if (min_events)
55312f9f 721 ret = io_u_queued_complete(td, min_events);
2e1df07d
JA
722 } else
723 cleanup_pending_aio(td);
724
725 td_set_runstate(td, TD_RUNNING);
726
727 dprint(FD_VERIFY, "exiting loop\n");
728}
729
e39c0676 730static bool exceeds_number_ios(struct thread_data *td)
3939fe85
JA
731{
732 unsigned long long number_ios;
733
734 if (!td->o.number_ios)
e39c0676 735 return false;
3939fe85 736
cf8a46a7 737 number_ios = ddir_rw_sum(td->io_blocks);
3939fe85
JA
738 number_ios += td->io_u_queued + td->io_u_in_flight;
739
cf8a46a7 740 return number_ios >= (td->o.number_ios * td->loops);
3939fe85
JA
741}
742
e39c0676 743static bool io_issue_bytes_exceeded(struct thread_data *td)
f7078f7b 744{
77731b29 745 unsigned long long bytes, limit;
f7078f7b
JA
746
747 if (td_rw(td))
74d6277f 748 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
f7078f7b 749 else if (td_write(td))
74d6277f 750 bytes = td->io_issue_bytes[DDIR_WRITE];
6eaf09d6 751 else if (td_read(td))
74d6277f 752 bytes = td->io_issue_bytes[DDIR_READ];
f7078f7b 753 else
74d6277f 754 bytes = td->io_issue_bytes[DDIR_TRIM];
f7078f7b 755
77731b29
JA
756 if (td->o.io_limit)
757 limit = td->o.io_limit;
758 else
759 limit = td->o.size;
760
cf8a46a7 761 limit *= td->loops;
77731b29 762 return bytes >= limit || exceeds_number_ios(td);
f7078f7b
JA
763}
764
e39c0676 765static bool io_complete_bytes_exceeded(struct thread_data *td)
e28dd2cf
JE
766{
767 unsigned long long bytes, limit;
768
769 if (td_rw(td))
770 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
771 else if (td_write(td))
772 bytes = td->this_io_bytes[DDIR_WRITE];
773 else if (td_read(td))
774 bytes = td->this_io_bytes[DDIR_READ];
775 else
776 bytes = td->this_io_bytes[DDIR_TRIM];
777
778 if (td->o.io_limit)
779 limit = td->o.io_limit;
780 else
781 limit = td->o.size;
782
cf8a46a7 783 limit *= td->loops;
e28dd2cf
JE
784 return bytes >= limit || exceeds_number_ios(td);
785}
786
50a8ce86
D
787/*
788 * used to calculate the next io time for rate control
789 *
790 */
791static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
792{
ff6bb260 793 uint64_t secs, remainder, bps, bytes, iops;
50a8ce86
D
794
795 assert(!(td->flags & TD_F_CHILD));
796 bytes = td->rate_io_issue_bytes[ddir];
797 bps = td->rate_bps[ddir];
ff6bb260 798
6de65959 799 if (td->o.rate_process == RATE_PROCESS_POISSON) {
a77d139b 800 uint64_t val;
ff6bb260 801 iops = bps / td->o.bs[ddir];
a77d139b
JA
802 val = (int64_t) (1000000 / iops) *
803 -logf(__rand_0_1(&td->poisson_state));
804 if (val) {
805 dprint(FD_RATE, "poisson rate iops=%llu\n",
806 (unsigned long long) 1000000 / val);
807 }
808 td->last_usec += val;
ff6bb260
SL
809 return td->last_usec;
810 } else if (bps) {
50a8ce86
D
811 secs = bytes / bps;
812 remainder = bytes % bps;
813 return remainder * 1000000 / bps + secs * 1000000;
e7b24047
JA
814 }
815
816 return 0;
50a8ce86
D
817}
818
2e1df07d
JA
819/*
820 * Main IO worker function. It retrieves io_u's to process and queues
821 * and reaps them, checking for rate and errors along the way.
100f49f1
JA
822 *
823 * Returns number of bytes written and trimmed.
2e1df07d 824 */
95b03be0 825static void do_io(struct thread_data *td, uint64_t *bytes_done)
2e1df07d
JA
826{
827 unsigned int i;
828 int ret = 0;
c2703bf3 829 uint64_t total_bytes, bytes_issued = 0;
70c68076 830
95b03be0
JA
831 for (i = 0; i < DDIR_RWDIR_CNT; i++)
832 bytes_done[i] = td->bytes_done[i];
2e1df07d
JA
833
834 if (in_ramp_time(td))
835 td_set_runstate(td, TD_RAMP);
836 else
837 td_set_runstate(td, TD_RUNNING);
838
3e260a46
JA
839 lat_target_init(td);
840
1e564979
JE
841 total_bytes = td->o.size;
842 /*
843 * Allow random overwrite workloads to write up to io_limit
844 * before starting verification phase as 'size' doesn't apply.
845 */
846 if (td_write(td) && td_random(td) && td->o.norandommap)
847 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
78a6469c
JA
848 /*
849 * If verify_backlog is enabled, we'll run the verify in this
850 * handler as well. For that case, we may need up to twice the
851 * amount of bytes.
852 */
78a6469c
JA
853 if (td->o.verify != VERIFY_NONE &&
854 (td_write(td) && td->o.verify_backlog))
c2703bf3
JA
855 total_bytes += td->o.size;
856
82a90686 857 /* In trimwrite mode, each byte is trimmed and then written, so
0e4dd95c 858 * allow total_bytes to be twice as big */
82a90686 859 if (td_trimwrite(td))
0e4dd95c
DE
860 total_bytes += td->total_io_size;
861
f7078f7b 862 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
e28dd2cf 863 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
c04e4661 864 td->o.time_based) {
2e1df07d 865 struct timeval comp_time;
2e1df07d 866 struct io_u *io_u;
e9d512d8 867 int full;
2e1df07d
JA
868 enum fio_ddir ddir;
869
c97f1ad6
JA
870 check_update_rusage(td);
871
7d7803fa 872 if (td->terminate || td->done)
2e1df07d
JA
873 break;
874
875 update_tv_cache(td);
876
877 if (runtime_exceeded(td, &td->tv_cache)) {
878 __update_tv_cache(td);
879 if (runtime_exceeded(td, &td->tv_cache)) {
ebea2133 880 fio_mark_td_terminate(td);
2e1df07d
JA
881 break;
882 }
883 }
884
9e684a49
DE
885 if (flow_threshold_exceeded(td))
886 continue;
887
f1a32461
JA
888 /*
889 * Break if we exceeded the bytes. The exception is time
890 * based runs, but we still need to break out of the loop
891 * for those to run verification, if enabled.
892 */
893 if (bytes_issued >= total_bytes &&
894 (!td->o.time_based ||
895 (td->o.time_based && td->o.verify != VERIFY_NONE)))
20876c53
JC
896 break;
897
2e1df07d 898 io_u = get_io_u(td);
002fe734
JA
899 if (IS_ERR_OR_NULL(io_u)) {
900 int err = PTR_ERR(io_u);
901
902 io_u = NULL;
903 if (err == -EBUSY) {
904 ret = FIO_Q_BUSY;
905 goto reap;
906 }
3e260a46
JA
907 if (td->o.latency_target)
908 goto reap;
2e1df07d 909 break;
3e260a46 910 }
2e1df07d
JA
911
912 ddir = io_u->ddir;
913
914 /*
82af2a7c
JA
915 * Add verification end_io handler if:
916 * - Asked to verify (!td_rw(td))
917 * - Or the io_u is from our verify list (mixed write/ver)
2e1df07d
JA
918 */
919 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
82af2a7c 920 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
c4b6117b
PV
921
922 if (!td->o.verify_pattern_bytes) {
d6b72507 923 io_u->rand_seed = __rand(&td->verify_state);
c4b6117b 924 if (sizeof(int) != sizeof(long *))
d6b72507 925 io_u->rand_seed *= __rand(&td->verify_state);
c4b6117b
PV
926 }
927
ca09be4b
JA
928 if (verify_state_should_stop(td, io_u)) {
929 put_io_u(td, io_u);
930 break;
931 }
932
2e1df07d
JA
933 if (td->o.verify_async)
934 io_u->end_io = verify_io_u_async;
935 else
936 io_u->end_io = verify_io_u;
937 td_set_runstate(td, TD_VERIFYING);
938 } else if (in_ramp_time(td))
939 td_set_runstate(td, TD_RAMP);
940 else
941 td_set_runstate(td, TD_RUNNING);
942
9a50c5c5 943 /*
f9401285
JA
944 * Always log IO before it's issued, so we know the specific
945 * order of it. The logged unit will track when the IO has
946 * completed.
9a50c5c5 947 */
c4b6117b
PV
948 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
949 td->o.do_verify &&
950 td->o.verify != VERIFY_NONE &&
f9401285 951 !td->o.experimental_verify)
c4b6117b
PV
952 log_io_piece(td, io_u);
953
a9da8ab2 954 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
0c5df5f9
JA
955 const unsigned long blen = io_u->xfer_buflen;
956 const enum fio_ddir ddir = acct_ddir(io_u);
957
a9da8ab2
JA
958 if (td->error)
959 break;
0c5df5f9 960
26de50cf
JA
961 workqueue_enqueue(&td->io_wq, &io_u->work);
962 ret = FIO_Q_QUEUED;
50a8ce86 963
26de50cf 964 if (ddir_rw(ddir)) {
0c5df5f9
JA
965 td->io_issues[ddir]++;
966 td->io_issue_bytes[ddir] += blen;
967 td->rate_io_issue_bytes[ddir] += blen;
968 }
969
50a8ce86
D
970 if (should_check_rate(td))
971 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
972
a9da8ab2
JA
973 } else {
974 ret = td_io_queue(td, io_u);
2e1df07d 975
50a8ce86
D
976 if (should_check_rate(td))
977 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
978
fd727d9d 979 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
a9da8ab2 980 break;
2e1df07d 981
a9da8ab2
JA
982 /*
983 * See if we need to complete some commands. Note that
984 * we can get BUSY even without IO queued, if the
985 * system is resource starved.
986 */
3e260a46 987reap:
a9da8ab2
JA
988 full = queue_full(td) ||
989 (ret == FIO_Q_BUSY && td->cur_depth);
82407585 990 if (full || io_in_polling(td))
a9da8ab2
JA
991 ret = wait_for_completions(td, &comp_time);
992 }
2e1df07d
JA
993 if (ret < 0)
994 break;
55312f9f
JA
995 if (!ddir_rw_sum(td->bytes_done) &&
996 !(td->io_ops->flags & FIO_NOIO))
2e1df07d
JA
997 continue;
998
55312f9f
JA
999 if (!in_ramp_time(td) && should_check_rate(td)) {
1000 if (check_min_rate(td, &comp_time)) {
f9cafb12 1001 if (exitall_on_terminate || td->o.exitall_error)
2e1df07d
JA
1002 fio_terminate_threads(td->groupid);
1003 td_verror(td, EIO, "check_min_rate");
1004 break;
1005 }
1006 }
3e260a46
JA
1007 if (!in_ramp_time(td) && td->o.latency_target)
1008 lat_target_check(td);
e155cb64 1009
2e1df07d
JA
1010 if (td->o.thinktime) {
1011 unsigned long long b;
1012
342f4be4 1013 b = ddir_rw_sum(td->io_blocks);
2e1df07d
JA
1014 if (!(b % td->o.thinktime_blocks)) {
1015 int left;
1016
002e7183
JA
1017 io_u_quiesce(td);
1018
2e1df07d
JA
1019 if (td->o.thinktime_spin)
1020 usec_spin(td->o.thinktime_spin);
1021
1022 left = td->o.thinktime - td->o.thinktime_spin;
1023 if (left)
1024 usec_sleep(td, left);
1025 }
1026 }
1027 }
1028
c97f1ad6
JA
1029 check_update_rusage(td);
1030
2e1df07d 1031 if (td->trim_entries)
4e0a8fa2 1032 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
2e1df07d
JA
1033
1034 if (td->o.fill_device && td->error == ENOSPC) {
1035 td->error = 0;
ebea2133 1036 fio_mark_td_terminate(td);
2e1df07d
JA
1037 }
1038 if (!td->error) {
1039 struct fio_file *f;
1040
a9da8ab2
JA
1041 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1042 workqueue_flush(&td->io_wq);
1043 i = 0;
1044 } else
1045 i = td->cur_depth;
1046
2e1df07d 1047 if (i) {
55312f9f 1048 ret = io_u_queued_complete(td, i);
2e1df07d
JA
1049 if (td->o.fill_device && td->error == ENOSPC)
1050 td->error = 0;
1051 }
1052
1053 if (should_fsync(td) && td->o.end_fsync) {
1054 td_set_runstate(td, TD_FSYNCING);
1055
1056 for_each_file(td, f, i) {
61ee0f86 1057 if (!fio_file_fsync(td, f))
2e1df07d 1058 continue;
61ee0f86
JA
1059
1060 log_err("fio: end_fsync failed for file %s\n",
1061 f->file_name);
2e1df07d
JA
1062 }
1063 }
1064 } else
1065 cleanup_pending_aio(td);
1066
1067 /*
1068 * stop job if we failed doing any IO
1069 */
342f4be4 1070 if (!ddir_rw_sum(td->this_io_bytes))
2e1df07d 1071 td->done = 1;
100f49f1 1072
95b03be0
JA
1073 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1074 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
2e1df07d
JA
1075}
1076
94a6e1bb
JA
1077static void free_file_completion_logging(struct thread_data *td)
1078{
1079 struct fio_file *f;
1080 unsigned int i;
1081
1082 for_each_file(td, f, i) {
1083 if (!f->last_write_comp)
1084 break;
1085 sfree(f->last_write_comp);
1086 }
1087}
1088
1089static int init_file_completion_logging(struct thread_data *td,
1090 unsigned int depth)
1091{
1092 struct fio_file *f;
1093 unsigned int i;
1094
1095 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1096 return 0;
1097
1098 for_each_file(td, f, i) {
1099 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1100 if (!f->last_write_comp)
1101 goto cleanup;
1102 }
1103
1104 return 0;
1105
1106cleanup:
1107 free_file_completion_logging(td);
1108 log_err("fio: failed to alloc write comp data\n");
1109 return 1;
1110}
1111
2e1df07d
JA
1112static void cleanup_io_u(struct thread_data *td)
1113{
2e1df07d
JA
1114 struct io_u *io_u;
1115
2ae0b204 1116 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
c73ed246
JA
1117
1118 if (td->io_ops->io_u_free)
1119 td->io_ops->io_u_free(td, io_u);
1120
2e1df07d
JA
1121 fio_memfree(io_u, sizeof(*io_u));
1122 }
1123
1124 free_io_mem(td);
2ae0b204
JA
1125
1126 io_u_rexit(&td->io_u_requeues);
1127 io_u_qexit(&td->io_u_freelist);
1128 io_u_qexit(&td->io_u_all);
ca09be4b 1129
94a6e1bb 1130 free_file_completion_logging(td);
2e1df07d
JA
1131}
1132
1133static int init_io_u(struct thread_data *td)
1134{
1135 struct io_u *io_u;
9c42684e 1136 unsigned int max_bs, min_write;
2e1df07d 1137 int cl_align, i, max_units;
2ae0b204 1138 int data_xfer = 1, err;
2e1df07d
JA
1139 char *p;
1140
1141 max_units = td->o.iodepth;
74f4b020 1142 max_bs = td_max_bs(td);
9c42684e 1143 min_write = td->o.min_bs[DDIR_WRITE];
2e1df07d
JA
1144 td->orig_buffer_size = (unsigned long long) max_bs
1145 * (unsigned long long) max_units;
1146
88045e04 1147 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
59d8d0f5
JA
1148 data_xfer = 0;
1149
2ae0b204
JA
1150 err = 0;
1151 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1152 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1153 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1154
1155 if (err) {
1156 log_err("fio: failed setting up IO queues\n");
1157 return 1;
1158 }
1159
fd8a09b8 1160 /*
1161 * if we may later need to do address alignment, then add any
1162 * possible adjustment here so that we don't cause a buffer
1163 * overflow later. this adjustment may be too much if we get
1164 * lucky and the allocator gives us an aligned address.
1165 */
d01612f3
CM
1166 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1167 (td->io_ops->flags & FIO_RAWIO))
fd8a09b8 1168 td->orig_buffer_size += page_mask + td->o.mem_align;
1169
2e1df07d
JA
1170 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1171 unsigned long bs;
1172
1173 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1174 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1175 }
1176
1177 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1178 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1179 return 1;
1180 }
1181
59d8d0f5 1182 if (data_xfer && allocate_io_mem(td))
2e1df07d
JA
1183 return 1;
1184
d01612f3 1185 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
2e1df07d
JA
1186 (td->io_ops->flags & FIO_RAWIO))
1187 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1188 else
1189 p = td->orig_buffer;
1190
1191 cl_align = os_cache_line_size();
1192
1193 for (i = 0; i < max_units; i++) {
1194 void *ptr;
1195
1196 if (td->terminate)
1197 return 1;
1198
1199 ptr = fio_memalign(cl_align, sizeof(*io_u));
1200 if (!ptr) {
1201 log_err("fio: unable to allocate aligned memory\n");
1202 break;
1203 }
1204
1205 io_u = ptr;
1206 memset(io_u, 0, sizeof(*io_u));
2ae0b204 1207 INIT_FLIST_HEAD(&io_u->verify_list);
2e1df07d
JA
1208 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1209
59d8d0f5 1210 if (data_xfer) {
2e1df07d
JA
1211 io_u->buf = p;
1212 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1213
1214 if (td_write(td))
9c42684e 1215 io_u_fill_buffer(td, io_u, min_write, max_bs);
2e1df07d
JA
1216 if (td_write(td) && td->o.verify_pattern_bytes) {
1217 /*
1218 * Fill the buffer with the pattern if we are
1219 * going to be doing writes.
1220 */
ce35b1ec 1221 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
2e1df07d
JA
1222 }
1223 }
1224
1225 io_u->index = i;
1226 io_u->flags = IO_U_F_FREE;
2ae0b204
JA
1227 io_u_qpush(&td->io_u_freelist, io_u);
1228
1229 /*
1230 * io_u never leaves this stack, used for iteration of all
1231 * io_u buffers.
1232 */
1233 io_u_qpush(&td->io_u_all, io_u);
c73ed246
JA
1234
1235 if (td->io_ops->io_u_init) {
1236 int ret = td->io_ops->io_u_init(td, io_u);
1237
1238 if (ret) {
1239 log_err("fio: failed to init engine data: %d\n", ret);
1240 return 1;
1241 }
1242 }
1243
2e1df07d
JA
1244 p += max_bs;
1245 }
1246
94a6e1bb
JA
1247 if (init_file_completion_logging(td, max_units))
1248 return 1;
ca09be4b 1249
2e1df07d
JA
1250 return 0;
1251}
1252
1253static int switch_ioscheduler(struct thread_data *td)
1254{
1255 char tmp[256], tmp2[128];
1256 FILE *f;
1257 int ret;
1258
1259 if (td->io_ops->flags & FIO_DISKLESSIO)
1260 return 0;
1261
1262 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1263
1264 f = fopen(tmp, "r+");
1265 if (!f) {
1266 if (errno == ENOENT) {
1267 log_err("fio: os or kernel doesn't support IO scheduler"
1268 " switching\n");
1269 return 0;
1270 }
1271 td_verror(td, errno, "fopen iosched");
1272 return 1;
1273 }
1274
1275 /*
1276 * Set io scheduler.
1277 */
1278 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1279 if (ferror(f) || ret != 1) {
1280 td_verror(td, errno, "fwrite");
1281 fclose(f);
1282 return 1;
1283 }
1284
1285 rewind(f);
1286
1287 /*
1288 * Read back and check that the selected scheduler is now the default.
1289 */
b44b9e45 1290 memset(tmp, 0, sizeof(tmp));
49c6f33d 1291 ret = fread(tmp, sizeof(tmp), 1, f);
2e1df07d
JA
1292 if (ferror(f) || ret < 0) {
1293 td_verror(td, errno, "fread");
1294 fclose(f);
1295 return 1;
1296 }
b44b9e45
TK
1297 /*
1298 * either a list of io schedulers or "none\n" is expected.
1299 */
1300 tmp[strlen(tmp) - 1] = '\0';
49c6f33d 1301
2e1df07d
JA
1302
1303 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1304 if (!strstr(tmp, tmp2)) {
1305 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1306 td_verror(td, EINVAL, "iosched_switch");
1307 fclose(f);
1308 return 1;
1309 }
1310
1311 fclose(f);
1312 return 0;
1313}
1314
e39c0676 1315static bool keep_running(struct thread_data *td)
2e1df07d 1316{
77731b29
JA
1317 unsigned long long limit;
1318
2e1df07d 1319 if (td->done)
e39c0676 1320 return false;
2e1df07d 1321 if (td->o.time_based)
e39c0676 1322 return true;
2e1df07d
JA
1323 if (td->o.loops) {
1324 td->o.loops--;
e39c0676 1325 return true;
2e1df07d 1326 }
3939fe85 1327 if (exceeds_number_ios(td))
e39c0676 1328 return false;
26251d8d 1329
77731b29
JA
1330 if (td->o.io_limit)
1331 limit = td->o.io_limit;
1332 else
1333 limit = td->o.size;
1334
1335 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
5bd5f71a
JA
1336 uint64_t diff;
1337
1338 /*
1339 * If the difference is less than the minimum IO size, we
1340 * are done.
1341 */
77731b29 1342 diff = limit - ddir_rw_sum(td->io_bytes);
74f4b020 1343 if (diff < td_max_bs(td))
e39c0676 1344 return false;
5bd5f71a 1345
543e2e9d 1346 if (fio_files_done(td) && !td->o.io_limit)
e39c0676 1347 return false;
002fe734 1348
e39c0676 1349 return true;
5bd5f71a 1350 }
2e1df07d 1351
e39c0676 1352 return false;
2e1df07d
JA
1353}
1354
ce486495 1355static int exec_string(struct thread_options *o, const char *string, const char *mode)
2e1df07d 1356{
6c9ce469 1357 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
f491a907 1358 int ret;
2e1df07d
JA
1359 char *str;
1360
1361 str = malloc(newlen);
ce486495 1362 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
2e1df07d 1363
ce486495 1364 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
2e1df07d
JA
1365 ret = system(str);
1366 if (ret == -1)
1367 log_err("fio: exec of cmd <%s> failed\n", str);
1368
1369 free(str);
1370 return ret;
1371}
1372
62167762
JC
1373/*
1374 * Dry run to compute correct state of numberio for verification.
1375 */
1376static uint64_t do_dry_run(struct thread_data *td)
1377{
62167762
JC
1378 td_set_runstate(td, TD_RUNNING);
1379
1380 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
e28dd2cf 1381 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
62167762
JC
1382 struct io_u *io_u;
1383 int ret;
1384
1385 if (td->terminate || td->done)
1386 break;
1387
1388 io_u = get_io_u(td);
1389 if (!io_u)
1390 break;
1391
a9da8ab2 1392 io_u_set(io_u, IO_U_F_FLIGHT);
62167762
JC
1393 io_u->error = 0;
1394 io_u->resid = 0;
1395 if (ddir_rw(acct_ddir(io_u)))
1396 td->io_issues[acct_ddir(io_u)]++;
1397 if (ddir_rw(io_u->ddir)) {
1398 io_u_mark_depth(td, 1);
1399 td->ts.total_io_u[io_u->ddir]++;
1400 }
1401
2e63e96b
PV
1402 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1403 td->o.do_verify &&
1404 td->o.verify != VERIFY_NONE &&
1405 !td->o.experimental_verify)
1406 log_io_piece(td, io_u);
1407
55312f9f 1408 ret = io_u_sync_complete(td, io_u);
62167762
JC
1409 (void) ret;
1410 }
1411
55312f9f 1412 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
62167762
JC
1413}
1414
24660963
JA
1415struct fork_data {
1416 struct thread_data *td;
1417 struct sk_out *sk_out;
1418};
1419
2e1df07d
JA
1420/*
1421 * Entry point for the thread based jobs. The process based jobs end up
1422 * here as well, after a little setup.
1423 */
1424static void *thread_main(void *data)
1425{
24660963 1426 struct fork_data *fd = data;
95603b74 1427 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
24660963 1428 struct thread_data *td = fd->td;
4896473e 1429 struct thread_options *o = &td->o;
24660963 1430 struct sk_out *sk_out = fd->sk_out;
2e1df07d 1431 int clear_state;
28727df7 1432 int ret;
2e1df07d 1433
24660963
JA
1434 sk_out_assign(sk_out);
1435 free(fd);
1436
4896473e 1437 if (!o->use_thread) {
2e1df07d
JA
1438 setsid();
1439 td->pid = getpid();
1440 } else
1441 td->pid = gettid();
1442
4896473e 1443 fio_local_clock_init(o->use_thread);
5d879392 1444
2e1df07d
JA
1445 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1446
122c7725
JA
1447 if (is_backend)
1448 fio_server_send_start(td);
1449
2e1df07d
JA
1450 INIT_FLIST_HEAD(&td->io_log_list);
1451 INIT_FLIST_HEAD(&td->io_hist_list);
1452 INIT_FLIST_HEAD(&td->verify_list);
1453 INIT_FLIST_HEAD(&td->trim_list);
1ae83d45 1454 INIT_FLIST_HEAD(&td->next_rand_list);
2e1df07d
JA
1455 td->io_hist_tree = RB_ROOT;
1456
34febb23 1457 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
f9e5b5ee 1458 if (ret) {
34febb23 1459 td_verror(td, ret, "mutex_cond_init_pshared");
f9e5b5ee
JK
1460 goto err;
1461 }
34febb23 1462 ret = cond_init_pshared(&td->verify_cond);
f9e5b5ee 1463 if (ret) {
34febb23 1464 td_verror(td, ret, "mutex_cond_pshared");
f9e5b5ee
JK
1465 goto err;
1466 }
2e1df07d
JA
1467
1468 td_set_runstate(td, TD_INITIALIZED);
1469 dprint(FD_MUTEX, "up startup_mutex\n");
1470 fio_mutex_up(startup_mutex);
1471 dprint(FD_MUTEX, "wait on td->mutex\n");
1472 fio_mutex_down(td->mutex);
1473 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1474
2e1df07d
JA
1475 /*
1476 * A new gid requires privilege, so we need to do this before setting
1477 * the uid.
1478 */
4896473e 1479 if (o->gid != -1U && setgid(o->gid)) {
2e1df07d
JA
1480 td_verror(td, errno, "setgid");
1481 goto err;
1482 }
4896473e 1483 if (o->uid != -1U && setuid(o->uid)) {
2e1df07d
JA
1484 td_verror(td, errno, "setuid");
1485 goto err;
1486 }
1487
2354d810
JA
1488 /*
1489 * Do this early, we don't want the compress threads to be limited
1490 * to the same CPUs as the IO workers. So do this before we set
1491 * any potential CPU affinity
1492 */
1493 if (iolog_compress_init(td, sk_out))
1494 goto err;
1495
2e1df07d
JA
1496 /*
1497 * If we have a gettimeofday() thread, make sure we exclude that
1498 * thread from this job
1499 */
4896473e
JA
1500 if (o->gtod_cpu)
1501 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
2e1df07d
JA
1502
1503 /*
1504 * Set affinity first, in case it has an impact on the memory
1505 * allocations.
1506 */
b2a9e649 1507 if (fio_option_is_set(o, cpumask)) {
c2acfbac 1508 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
30cb4c65 1509 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
c2acfbac
JA
1510 if (!ret) {
1511 log_err("fio: no CPUs set\n");
1512 log_err("fio: Try increasing number of available CPUs\n");
1513 td_verror(td, EINVAL, "cpus_split");
1514 goto err;
1515 }
1516 }
28727df7
JA
1517 ret = fio_setaffinity(td->pid, o->cpumask);
1518 if (ret == -1) {
4896473e
JA
1519 td_verror(td, errno, "cpu_set_affinity");
1520 goto err;
1521 }
2e1df07d
JA
1522 }
1523
67bf9823 1524#ifdef CONFIG_LIBNUMA
d0b937ed 1525 /* numa node setup */
b2a9e649
JA
1526 if (fio_option_is_set(o, numa_cpunodes) ||
1527 fio_option_is_set(o, numa_memnodes)) {
43522848 1528 struct bitmask *mask;
d0b937ed
YR
1529
1530 if (numa_available() < 0) {
1531 td_verror(td, errno, "Does not support NUMA API\n");
1532 goto err;
1533 }
1534
b2a9e649 1535 if (fio_option_is_set(o, numa_cpunodes)) {
43522848
DG
1536 mask = numa_parse_nodestring(o->numa_cpunodes);
1537 ret = numa_run_on_node_mask(mask);
1538 numa_free_nodemask(mask);
d0b937ed
YR
1539 if (ret == -1) {
1540 td_verror(td, errno, \
1541 "numa_run_on_node_mask failed\n");
1542 goto err;
1543 }
1544 }
1545
b2a9e649 1546 if (fio_option_is_set(o, numa_memnodes)) {
43522848
DG
1547 mask = NULL;
1548 if (o->numa_memnodes)
1549 mask = numa_parse_nodestring(o->numa_memnodes);
1550
4896473e 1551 switch (o->numa_mem_mode) {
d0b937ed 1552 case MPOL_INTERLEAVE:
43522848 1553 numa_set_interleave_mask(mask);
d0b937ed
YR
1554 break;
1555 case MPOL_BIND:
43522848 1556 numa_set_membind(mask);
d0b937ed
YR
1557 break;
1558 case MPOL_LOCAL:
1559 numa_set_localalloc();
1560 break;
1561 case MPOL_PREFERRED:
4896473e 1562 numa_set_preferred(o->numa_mem_prefer_node);
d0b937ed
YR
1563 break;
1564 case MPOL_DEFAULT:
1565 default:
1566 break;
1567 }
1568
43522848
DG
1569 if (mask)
1570 numa_free_nodemask(mask);
1571
d0b937ed
YR
1572 }
1573 }
1574#endif
1575
9a3f1100
JA
1576 if (fio_pin_memory(td))
1577 goto err;
1578
2e1df07d
JA
1579 /*
1580 * May alter parameters that init_io_u() will use, so we need to
1581 * do this first.
1582 */
1583 if (init_iolog(td))
1584 goto err;
1585
1586 if (init_io_u(td))
1587 goto err;
1588
4896473e 1589 if (o->verify_async && verify_async_init(td))
2e1df07d
JA
1590 goto err;
1591
27d74836
JA
1592 if (fio_option_is_set(o, ioprio) ||
1593 fio_option_is_set(o, ioprio_class)) {
28727df7
JA
1594 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1595 if (ret == -1) {
2e1df07d
JA
1596 td_verror(td, errno, "ioprio_set");
1597 goto err;
1598 }
1599 }
1600
4896473e 1601 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
2e1df07d
JA
1602 goto err;
1603
649c10c5 1604 errno = 0;
4896473e 1605 if (nice(o->nice) == -1 && errno != 0) {
2e1df07d
JA
1606 td_verror(td, errno, "nice");
1607 goto err;
1608 }
1609
4896473e 1610 if (o->ioscheduler && switch_ioscheduler(td))
2e1df07d
JA
1611 goto err;
1612
4896473e 1613 if (!o->create_serialize && setup_files(td))
2e1df07d
JA
1614 goto err;
1615
1616 if (td_io_init(td))
1617 goto err;
1618
1619 if (init_random_map(td))
1620 goto err;
1621
ce486495 1622 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
4896473e 1623 goto err;
2e1df07d 1624
4896473e 1625 if (o->pre_read) {
2e1df07d
JA
1626 if (pre_read_files(td) < 0)
1627 goto err;
1628 }
1629
dc5bfbb2
JA
1630 fio_verify_init(td);
1631
24660963 1632 if (rate_submit_init(td, sk_out))
a9da8ab2
JA
1633 goto err;
1634
2e1df07d 1635 fio_gettime(&td->epoch, NULL);
44404c5a 1636 fio_getrusage(&td->ru_start);
ac28d905
JA
1637 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1638 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1639
1640 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1641 o->ratemin[DDIR_TRIM]) {
1642 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1643 sizeof(td->bw_sample_time));
1644 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1645 sizeof(td->bw_sample_time));
1646 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1647 sizeof(td->bw_sample_time));
1648 }
1649
2e1df07d
JA
1650 clear_state = 0;
1651 while (keep_running(td)) {
100f49f1
JA
1652 uint64_t verify_bytes;
1653
2e1df07d 1654 fio_gettime(&td->start, NULL);
2e1df07d
JA
1655 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1656
2e1df07d 1657 if (clear_state)
ac28d905 1658 clear_io_state(td, 0);
2e1df07d
JA
1659
1660 prune_io_piece_log(td);
1661
62167762
JC
1662 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1663 verify_bytes = do_dry_run(td);
095196b1 1664 else {
95b03be0
JA
1665 uint64_t bytes_done[DDIR_RWDIR_CNT];
1666
1667 do_io(td, bytes_done);
1668
1669 if (!ddir_rw_sum(bytes_done)) {
095196b1 1670 fio_mark_td_terminate(td);
95b03be0
JA
1671 verify_bytes = 0;
1672 } else {
1673 verify_bytes = bytes_done[DDIR_WRITE] +
1674 bytes_done[DDIR_TRIM];
1675 }
095196b1 1676 }
2e1df07d
JA
1677
1678 clear_state = 1;
1679
40c666c8
JA
1680 /*
1681 * Make sure we've successfully updated the rusage stats
1682 * before waiting on the stat mutex. Otherwise we could have
1683 * the stat thread holding stat mutex and waiting for
1684 * the rusage_sem, which would never get upped because
1685 * this thread is waiting for the stat mutex.
1686 */
1687 check_update_rusage(td);
1688
e5437a07 1689 fio_mutex_down(stat_mutex);
95603b74
BF
1690 if (td_read(td) && td->io_bytes[DDIR_READ])
1691 update_runtime(td, elapsed_us, DDIR_READ);
1692 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1693 update_runtime(td, elapsed_us, DDIR_WRITE);
1694 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1695 update_runtime(td, elapsed_us, DDIR_TRIM);
e5437a07
VT
1696 fio_gettime(&td->start, NULL);
1697 fio_mutex_up(stat_mutex);
2e1df07d
JA
1698
1699 if (td->error || td->terminate)
1700 break;
1701
4896473e
JA
1702 if (!o->do_verify ||
1703 o->verify == VERIFY_NONE ||
2e1df07d
JA
1704 (td->io_ops->flags & FIO_UNIDIR))
1705 continue;
1706
ac28d905 1707 clear_io_state(td, 0);
2e1df07d
JA
1708
1709 fio_gettime(&td->start, NULL);
1710
100f49f1 1711 do_verify(td, verify_bytes);
2e1df07d 1712
40c666c8
JA
1713 /*
1714 * See comment further up for why this is done here.
1715 */
1716 check_update_rusage(td);
1717
e5437a07 1718 fio_mutex_down(stat_mutex);
95603b74 1719 update_runtime(td, elapsed_us, DDIR_READ);
e5437a07
VT
1720 fio_gettime(&td->start, NULL);
1721 fio_mutex_up(stat_mutex);
2e1df07d
JA
1722
1723 if (td->error || td->terminate)
1724 break;
1725 }
1726
2cea0b4c
JA
1727 td_set_runstate(td, TD_FINISHING);
1728
2e1df07d 1729 update_rusage_stat(td);
2e1df07d 1730 td->ts.total_run_time = mtime_since_now(&td->epoch);
6eaf09d6
SL
1731 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1732 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1733 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2e1df07d 1734
ca09be4b 1735 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
d264264a
JA
1736 (td->o.verify != VERIFY_NONE && td_write(td)))
1737 verify_save_state(td->thread_number);
ca09be4b 1738
9a3f1100
JA
1739 fio_unpin_memory(td);
1740
a47591e4 1741 td_writeout_logs(td, true);
2e1df07d 1742
155f2f02 1743 iolog_compress_exit(td);
103b174e 1744 rate_submit_exit(td);
aee2ab67 1745
4896473e 1746 if (o->exec_postrun)
ce486495 1747 exec_string(o, o->exec_postrun, (const char *)"postrun");
2e1df07d 1748
f9cafb12 1749 if (exitall_on_terminate || (o->exitall_error && td->error))
2e1df07d
JA
1750 fio_terminate_threads(td->groupid);
1751
1752err:
1753 if (td->error)
1754 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1755 td->verror);
1756
4896473e 1757 if (o->verify_async)
2e1df07d
JA
1758 verify_async_exit(td);
1759
1760 close_and_free_files(td);
2e1df07d 1761 cleanup_io_u(td);
32dbca2c 1762 close_ioengine(td);
2e1df07d 1763 cgroup_shutdown(td, &cgroup_mnt);
ca09be4b 1764 verify_free_state(td);
2e1df07d 1765
7c4d0fb7
JA
1766 if (td->zone_state_index) {
1767 int i;
1768
1769 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1770 free(td->zone_state_index[i]);
1771 free(td->zone_state_index);
1772 td->zone_state_index = NULL;
1773 }
1774
b2a9e649 1775 if (fio_option_is_set(o, cpumask)) {
8a1db9a1
JA
1776 ret = fio_cpuset_exit(&o->cpumask);
1777 if (ret)
1778 td_verror(td, ret, "fio_cpuset_exit");
2e1df07d
JA
1779 }
1780
1781 /*
1782 * do this very late, it will log file closing as well
1783 */
4896473e 1784 if (o->write_iolog_file)
2e1df07d
JA
1785 write_iolog_close(td);
1786
ea66e04f
JA
1787 fio_mutex_remove(td->mutex);
1788 td->mutex = NULL;
1789
2e1df07d 1790 td_set_runstate(td, TD_EXITED);
fda2cfac
JA
1791
1792 /*
1793 * Do this last after setting our runstate to exited, so we
1794 * know that the stat thread is signaled.
1795 */
1796 check_update_rusage(td);
1797
24660963 1798 sk_out_drop();
e43606c2 1799 return (void *) (uintptr_t) td->error;
2e1df07d
JA
1800}
1801
cba5460c
JA
1802static void dump_td_info(struct thread_data *td)
1803{
a507505d
JA
1804 log_err("fio: job '%s' (state=%d) hasn't exited in %lu seconds, it "
1805 "appears to be stuck. Doing forceful exit of this job.\n",
1806 td->o.name, td->runstate,
cba5460c
JA
1807 (unsigned long) time_since_now(&td->terminate_time));
1808}
1809
2e1df07d
JA
1810/*
1811 * Run over the job map and reap the threads that have exited, if any.
1812 */
1813static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1814 unsigned int *m_rate)
1815{
1816 struct thread_data *td;
1817 unsigned int cputhreads, realthreads, pending;
1818 int i, status, ret;
1819
1820 /*
1821 * reap exited threads (TD_EXITED -> TD_REAPED)
1822 */
1823 realthreads = pending = cputhreads = 0;
1824 for_each_td(td, i) {
1825 int flags = 0;
1826
1827 /*
1828 * ->io_ops is NULL for a thread that has closed its
1829 * io engine
1830 */
1831 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1832 cputhreads++;
1833 else
1834 realthreads++;
1835
1836 if (!td->pid) {
1837 pending++;
1838 continue;
1839 }
1840 if (td->runstate == TD_REAPED)
1841 continue;
1842 if (td->o.use_thread) {
1843 if (td->runstate == TD_EXITED) {
1844 td_set_runstate(td, TD_REAPED);
1845 goto reaped;
1846 }
1847 continue;
1848 }
1849
1850 flags = WNOHANG;
1851 if (td->runstate == TD_EXITED)
1852 flags = 0;
1853
1854 /*
1855 * check if someone quit or got killed in an unusual way
1856 */
1857 ret = waitpid(td->pid, &status, flags);
1858 if (ret < 0) {
1859 if (errno == ECHILD) {
1860 log_err("fio: pid=%d disappeared %d\n",
1861 (int) td->pid, td->runstate);
a5e371a6 1862 td->sig = ECHILD;
2e1df07d
JA
1863 td_set_runstate(td, TD_REAPED);
1864 goto reaped;
1865 }
1866 perror("waitpid");
1867 } else if (ret == td->pid) {
1868 if (WIFSIGNALED(status)) {
1869 int sig = WTERMSIG(status);
1870
36d80bc7 1871 if (sig != SIGTERM && sig != SIGUSR2)
2e1df07d
JA
1872 log_err("fio: pid=%d, got signal=%d\n",
1873 (int) td->pid, sig);
a5e371a6 1874 td->sig = sig;
2e1df07d
JA
1875 td_set_runstate(td, TD_REAPED);
1876 goto reaped;
1877 }
1878 if (WIFEXITED(status)) {
1879 if (WEXITSTATUS(status) && !td->error)
1880 td->error = WEXITSTATUS(status);
1881
1882 td_set_runstate(td, TD_REAPED);
1883 goto reaped;
1884 }
1885 }
1886
cba5460c
JA
1887 /*
1888 * If the job is stuck, do a forceful timeout of it and
1889 * move on.
1890 */
1891 if (td->terminate &&
2c45a4ac 1892 td->runstate < TD_FSYNCING &&
cba5460c
JA
1893 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1894 dump_td_info(td);
1895 td_set_runstate(td, TD_REAPED);
1896 goto reaped;
1897 }
1898
2e1df07d
JA
1899 /*
1900 * thread is not dead, continue
1901 */
1902 pending++;
1903 continue;
1904reaped:
1905 (*nr_running)--;
342f4be4
JA
1906 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1907 (*t_rate) -= ddir_rw_sum(td->o.rate);
2e1df07d
JA
1908 if (!td->pid)
1909 pending--;
1910
1911 if (td->error)
1912 exit_value++;
1913
1914 done_secs += mtime_since_now(&td->epoch) / 1000;
4a88752a 1915 profile_td_exit(td);
2e1df07d
JA
1916 }
1917
1918 if (*nr_running == cputhreads && !pending && realthreads)
1919 fio_terminate_threads(TERMINATE_ALL);
1920}
1921
e39c0676 1922static bool __check_trigger_file(void)
ca09be4b
JA
1923{
1924 struct stat sb;
1925
1926 if (!trigger_file)
e39c0676 1927 return false;
ca09be4b
JA
1928
1929 if (stat(trigger_file, &sb))
e39c0676 1930 return false;
ca09be4b
JA
1931
1932 if (unlink(trigger_file) < 0)
1933 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1934 strerror(errno));
1935
e39c0676 1936 return true;
ca09be4b
JA
1937}
1938
e39c0676 1939static bool trigger_timedout(void)
ca09be4b
JA
1940{
1941 if (trigger_timeout)
1942 return time_since_genesis() >= trigger_timeout;
1943
e39c0676 1944 return false;
ca09be4b
JA
1945}
1946
1947void exec_trigger(const char *cmd)
1948{
1949 int ret;
1950
1951 if (!cmd)
1952 return;
1953
1954 ret = system(cmd);
1955 if (ret == -1)
1956 log_err("fio: failed executing %s trigger\n", cmd);
1957}
1958
1959void check_trigger_file(void)
1960{
1961 if (__check_trigger_file() || trigger_timedout()) {
796fb3ce
JA
1962 if (nr_clients)
1963 fio_clients_send_trigger(trigger_remote_cmd);
1964 else {
d264264a 1965 verify_save_state(IO_LIST_ALL);
ca09be4b
JA
1966 fio_terminate_threads(TERMINATE_ALL);
1967 exec_trigger(trigger_cmd);
1968 }
1969 }
1970}
1971
1972static int fio_verify_load_state(struct thread_data *td)
1973{
1974 int ret;
1975
1976 if (!td->o.verify_state)
1977 return 0;
1978
1979 if (is_backend) {
1980 void *data;
1981
1982 ret = fio_server_get_verify_state(td->o.name,
94a6e1bb 1983 td->thread_number - 1, &data);
ca09be4b 1984 if (!ret)
94a6e1bb 1985 verify_assign_state(td, data);
ca09be4b
JA
1986 } else
1987 ret = verify_load_state(td, "local");
1988
1989 return ret;
1990}
1991
06464907
JA
1992static void do_usleep(unsigned int usecs)
1993{
1994 check_for_running_stats();
ca09be4b 1995 check_trigger_file();
06464907
JA
1996 usleep(usecs);
1997}
1998
e39c0676 1999static bool check_mount_writes(struct thread_data *td)
e81ecca3
JA
2000{
2001 struct fio_file *f;
2002 unsigned int i;
2003
2004 if (!td_write(td) || td->o.allow_mounted_write)
e39c0676 2005 return false;
e81ecca3
JA
2006
2007 for_each_file(td, f, i) {
2008 if (f->filetype != FIO_TYPE_BD)
2009 continue;
2010 if (device_is_mounted(f->file_name))
2011 goto mounted;
2012 }
2013
e39c0676 2014 return false;
e81ecca3
JA
2015mounted:
2016 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
e39c0676 2017 return true;
e81ecca3
JA
2018}
2019
9cc8cb91
AK
2020static bool waitee_running(struct thread_data *me)
2021{
2022 const char *waitee = me->o.wait_for;
2023 const char *self = me->o.name;
2024 struct thread_data *td;
2025 int i;
2026
2027 if (!waitee)
2028 return false;
2029
2030 for_each_td(td, i) {
2031 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2032 continue;
2033
2034 if (td->runstate < TD_EXITED) {
2035 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2036 self, td->o.name,
2037 runstate_to_name(td->runstate));
2038 return true;
2039 }
2040 }
2041
2042 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2043 return false;
2044}
2045
2e1df07d
JA
2046/*
2047 * Main function for kicking off and reaping jobs, as needed.
2048 */
24660963 2049static void run_threads(struct sk_out *sk_out)
2e1df07d
JA
2050{
2051 struct thread_data *td;
2e1df07d 2052 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
0de5b26f 2053 uint64_t spent;
2e1df07d 2054
2e1df07d
JA
2055 if (fio_gtod_offload && fio_start_gtod_thread())
2056 return;
334185e9 2057
f2a2ce0e 2058 fio_idle_prof_init();
2e1df07d
JA
2059
2060 set_sig_handlers();
2061
3a5f6bde
JA
2062 nr_thread = nr_process = 0;
2063 for_each_td(td, i) {
e81ecca3
JA
2064 if (check_mount_writes(td))
2065 return;
3a5f6bde
JA
2066 if (td->o.use_thread)
2067 nr_thread++;
2068 else
2069 nr_process++;
2070 }
2071
01cfefcc 2072 if (output_format & FIO_OUTPUT_NORMAL) {
2e1df07d
JA
2073 log_info("Starting ");
2074 if (nr_thread)
2075 log_info("%d thread%s", nr_thread,
2076 nr_thread > 1 ? "s" : "");
2077 if (nr_process) {
2078 if (nr_thread)
2079 log_info(" and ");
2080 log_info("%d process%s", nr_process,
2081 nr_process > 1 ? "es" : "");
2082 }
2083 log_info("\n");
e411c301 2084 log_info_flush();
2e1df07d
JA
2085 }
2086
2087 todo = thread_number;
2088 nr_running = 0;
2089 nr_started = 0;
2090 m_rate = t_rate = 0;
2091
2092 for_each_td(td, i) {
2093 print_status_init(td->thread_number - 1);
2094
2095 if (!td->o.create_serialize)
2096 continue;
2097
ca09be4b
JA
2098 if (fio_verify_load_state(td))
2099 goto reap;
2100
2e1df07d
JA
2101 /*
2102 * do file setup here so it happens sequentially,
2103 * we don't want X number of threads getting their
2104 * client data interspersed on disk
2105 */
2106 if (setup_files(td)) {
ca09be4b 2107reap:
2e1df07d
JA
2108 exit_value++;
2109 if (td->error)
2110 log_err("fio: pid=%d, err=%d/%s\n",
2111 (int) td->pid, td->error, td->verror);
2112 td_set_runstate(td, TD_REAPED);
2113 todo--;
2114 } else {
2115 struct fio_file *f;
2116 unsigned int j;
2117
2118 /*
2119 * for sharing to work, each job must always open
2120 * its own files. so close them, if we opened them
2121 * for creation
2122 */
2123 for_each_file(td, f, j) {
2124 if (fio_file_open(f))
2125 td_io_close_file(td, f);
2126 }
2127 }
2128 }
2129
f2a2ce0e
HL
2130 /* start idle threads before io threads start to run */
2131 fio_idle_prof_start();
2132
2e1df07d
JA
2133 set_genesis_time();
2134
2135 while (todo) {
2136 struct thread_data *map[REAL_MAX_JOBS];
2137 struct timeval this_start;
2138 int this_jobs = 0, left;
d7982dd0 2139 struct fork_data *fd;
2e1df07d
JA
2140
2141 /*
2142 * create threads (TD_NOT_CREATED -> TD_CREATED)
2143 */
2144 for_each_td(td, i) {
2145 if (td->runstate != TD_NOT_CREATED)
2146 continue;
2147
2148 /*
2149 * never got a chance to start, killed by other
2150 * thread for some reason
2151 */
2152 if (td->terminate) {
2153 todo--;
2154 continue;
2155 }
2156
2157 if (td->o.start_delay) {
0de5b26f 2158 spent = utime_since_genesis();
2e1df07d 2159
74454ce4 2160 if (td->o.start_delay > spent)
2e1df07d
JA
2161 continue;
2162 }
2163
2164 if (td->o.stonewall && (nr_started || nr_running)) {
2165 dprint(FD_PROCESS, "%s: stonewall wait\n",
2166 td->o.name);
2167 break;
2168 }
2169
9cc8cb91
AK
2170 if (waitee_running(td)) {
2171 dprint(FD_PROCESS, "%s: waiting for %s\n",
2172 td->o.name, td->o.wait_for);
5c74fc76 2173 continue;
9cc8cb91
AK
2174 }
2175
2e1df07d
JA
2176 init_disk_util(td);
2177
c97f1ad6
JA
2178 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2179 td->update_rusage = 0;
2180
2e1df07d
JA
2181 /*
2182 * Set state to created. Thread will transition
2183 * to TD_INITIALIZED when it's done setting up.
2184 */
2185 td_set_runstate(td, TD_CREATED);
2186 map[this_jobs++] = td;
2187 nr_started++;
2188
d7982dd0
JK
2189 fd = calloc(1, sizeof(*fd));
2190 fd->td = td;
2191 fd->sk_out = sk_out;
2192
2e1df07d
JA
2193 if (td->o.use_thread) {
2194 int ret;
2195
2196 dprint(FD_PROCESS, "will pthread_create\n");
2197 ret = pthread_create(&td->thread, NULL,
24660963 2198 thread_main, fd);
2e1df07d
JA
2199 if (ret) {
2200 log_err("pthread_create: %s\n",
2201 strerror(ret));
24660963 2202 free(fd);
2e1df07d
JA
2203 nr_started--;
2204 break;
2205 }
2206 ret = pthread_detach(td->thread);
2207 if (ret)
2208 log_err("pthread_detach: %s",
2209 strerror(ret));
2210 } else {
2211 pid_t pid;
2212 dprint(FD_PROCESS, "will fork\n");
2213 pid = fork();
2214 if (!pid) {
d7982dd0 2215 int ret;
2e1df07d 2216
d7982dd0 2217 ret = (int)(uintptr_t)thread_main(fd);
2e1df07d
JA
2218 _exit(ret);
2219 } else if (i == fio_debug_jobno)
2220 *fio_debug_jobp = pid;
2221 }
2222 dprint(FD_MUTEX, "wait on startup_mutex\n");
09400a60 2223 if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2e1df07d
JA
2224 log_err("fio: job startup hung? exiting.\n");
2225 fio_terminate_threads(TERMINATE_ALL);
2226 fio_abort = 1;
2227 nr_started--;
2228 break;
2229 }
2230 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2231 }
2232
2233 /*
2234 * Wait for the started threads to transition to
2235 * TD_INITIALIZED.
2236 */
2237 fio_gettime(&this_start, NULL);
2238 left = this_jobs;
2239 while (left && !fio_abort) {
2240 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2241 break;
2242
06464907 2243 do_usleep(100000);
2e1df07d
JA
2244
2245 for (i = 0; i < this_jobs; i++) {
2246 td = map[i];
2247 if (!td)
2248 continue;
2249 if (td->runstate == TD_INITIALIZED) {
2250 map[i] = NULL;
2251 left--;
2252 } else if (td->runstate >= TD_EXITED) {
2253 map[i] = NULL;
2254 left--;
2255 todo--;
2256 nr_running++; /* work-around... */
2257 }
2258 }
2259 }
2260
2261 if (left) {
4e87c37a
JA
2262 log_err("fio: %d job%s failed to start\n", left,
2263 left > 1 ? "s" : "");
2e1df07d
JA
2264 for (i = 0; i < this_jobs; i++) {
2265 td = map[i];
2266 if (!td)
2267 continue;
2268 kill(td->pid, SIGTERM);
2269 }
2270 break;
2271 }
2272
2273 /*
2274 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2275 */
2276 for_each_td(td, i) {
2277 if (td->runstate != TD_INITIALIZED)
2278 continue;
2279
2280 if (in_ramp_time(td))
2281 td_set_runstate(td, TD_RAMP);
2282 else
2283 td_set_runstate(td, TD_RUNNING);
2284 nr_running++;
2285 nr_started--;
342f4be4
JA
2286 m_rate += ddir_rw_sum(td->o.ratemin);
2287 t_rate += ddir_rw_sum(td->o.rate);
2e1df07d
JA
2288 todo--;
2289 fio_mutex_up(td->mutex);
2290 }
2291
2292 reap_threads(&nr_running, &t_rate, &m_rate);
2293
122c7725 2294 if (todo)
06464907 2295 do_usleep(100000);
2e1df07d
JA
2296 }
2297
2298 while (nr_running) {
2299 reap_threads(&nr_running, &t_rate, &m_rate);
06464907 2300 do_usleep(10000);
2e1df07d
JA
2301 }
2302
f2a2ce0e
HL
2303 fio_idle_prof_stop();
2304
2e1df07d 2305 update_io_ticks();
2e1df07d
JA
2306}
2307
27357187
JA
2308static void free_disk_util(void)
2309{
27357187 2310 disk_util_prune_entries();
a39fb9ea 2311 helper_thread_destroy();
2e1df07d
JA
2312}
2313
24660963 2314int fio_backend(struct sk_out *sk_out)
2e1df07d
JA
2315{
2316 struct thread_data *td;
2317 int i;
2318
2319 if (exec_profile) {
2320 if (load_profile(exec_profile))
2321 return 1;
2322 free(exec_profile);
2323 exec_profile = NULL;
2324 }
2325 if (!thread_number)
2326 return 0;
2327
2328 if (write_bw_log) {
aee2ab67
JA
2329 struct log_params p = {
2330 .log_type = IO_LOG_TYPE_BW,
2331 };
2332
2333 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2334 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2335 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2e1df07d
JA
2336 }
2337
521da527 2338 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2e1df07d
JA
2339 if (startup_mutex == NULL)
2340 return 1;
2e1df07d
JA
2341
2342 set_genesis_time();
cef9175e 2343 stat_init();
a39fb9ea 2344 helper_thread_create(startup_mutex, sk_out);
2e1df07d
JA
2345
2346 cgroup_list = smalloc(sizeof(*cgroup_list));
2347 INIT_FLIST_HEAD(cgroup_list);
2348
24660963 2349 run_threads(sk_out);
2e1df07d 2350
a39fb9ea 2351 helper_thread_exit();
8aab824f 2352
2e1df07d 2353 if (!fio_abort) {
83f7b64e 2354 __show_run_stats();
2e1df07d 2355 if (write_bw_log) {
cb7e0ace
JA
2356 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2357 struct io_log *log = agg_io_log[i];
2358
3a5db920 2359 flush_log(log, 0);
518dac09 2360 free_log(log);
cb7e0ace 2361 }
2e1df07d
JA
2362 }
2363 }
2364
fda2cfac 2365 for_each_td(td, i) {
2e1df07d 2366 fio_options_free(td);
8049adc1
JA
2367 if (td->rusage_sem) {
2368 fio_mutex_remove(td->rusage_sem);
2369 td->rusage_sem = NULL;
2370 }
fda2cfac 2371 }
2e1df07d 2372
a462baef 2373 free_disk_util();
2e1df07d
JA
2374 cgroup_kill(cgroup_list);
2375 sfree(cgroup_list);
2376 sfree(cgroup_mnt);
2377
2378 fio_mutex_remove(startup_mutex);
cef9175e 2379 stat_exit();
2e1df07d
JA
2380 return exit_value;
2381}