docs: move rate_cycle description
[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
fa07eaa6 21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
2e1df07d
JA
22 *
23 */
24#include <unistd.h>
2e1df07d 25#include <string.h>
2e1df07d 26#include <signal.h>
2e1df07d 27#include <assert.h>
e43606c2 28#include <inttypes.h>
2e1df07d
JA
29#include <sys/stat.h>
30#include <sys/wait.h>
ff6bb260 31#include <math.h>
c06379a6 32#include <pthread.h>
2e1df07d
JA
33
34#include "fio.h"
2e1df07d
JA
35#include "smalloc.h"
36#include "verify.h"
2e1df07d
JA
37#include "diskutil.h"
38#include "cgroup.h"
39#include "profile.h"
40#include "lib/rand.h"
f7690c4a 41#include "lib/memalign.h"
2e1df07d 42#include "server.h"
44404c5a 43#include "lib/getrusage.h"
f2a2ce0e 44#include "idletime.h"
002fe734 45#include "err.h"
a9da8ab2 46#include "workqueue.h"
e81ecca3 47#include "lib/mountcheck.h"
40511301 48#include "rate-submit.h"
a39fb9ea 49#include "helper_thread.h"
ae626d4e 50#include "pshared.h"
07776127 51#include "zone-dist.h"
2e1df07d 52
971caeb1 53static struct fio_sem *startup_sem;
2e1df07d 54static struct flist_head *cgroup_list;
c3dc516a 55static struct cgroup_mnt *cgroup_mnt;
2e1df07d 56static int exit_value;
f1867a7f 57static volatile bool fio_abort;
3a5f6bde
JA
58static unsigned int nr_process = 0;
59static unsigned int nr_thread = 0;
2e1df07d 60
6eaf09d6 61struct io_log *agg_io_log[DDIR_RWDIR_CNT];
2e1df07d 62
a3efc919
JA
63int groupid = 0;
64unsigned int thread_number = 0;
2fbe1e19
JA
65unsigned int nr_segments = 0;
66unsigned int cur_segment = 0;
108fea77 67unsigned int stat_number = 0;
a3efc919
JA
68int temp_stall_ts;
69unsigned long done_secs = 0;
a38c70a8
BVA
70#ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP
71pthread_mutex_t overlap_check = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
72#else
c06379a6 73pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER;
a38c70a8 74#endif
a47591e4 75
2e1df07d
JA
76#define JOB_START_TIMEOUT (5 * 1000)
77
78static void sig_int(int sig)
79{
2fbe1e19 80 if (nr_segments) {
2e1df07d
JA
81 if (is_backend)
82 fio_server_got_signal(sig);
83 else {
84 log_info("\nfio: terminating on signal %d\n", sig);
e411c301 85 log_info_flush();
2e1df07d
JA
86 exit_value = 128;
87 }
88
64402a8a 89 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2e1df07d
JA
90 }
91}
92
0dea1f55
BP
93#ifdef WIN32
94static void sig_break(int sig)
95{
0dea1f55
BP
96 sig_int(sig);
97
98 /**
99 * Windows terminates all job processes on SIGBREAK after the handler
100 * returns, so give them time to wrap-up and give stats
101 */
da8f124f 102 for_each_td(td) {
0dea1f55
BP
103 while (td->runstate < TD_EXITED)
104 sleep(1);
da8f124f 105 } end_for_each();
0dea1f55
BP
106}
107#endif
108
d2235e56 109void sig_show_status(int sig)
4c6d91e8
JA
110{
111 show_running_run_stats();
112}
113
2e1df07d
JA
114static void set_sig_handlers(void)
115{
116 struct sigaction act;
117
118 memset(&act, 0, sizeof(act));
119 act.sa_handler = sig_int;
120 act.sa_flags = SA_RESTART;
121 sigaction(SIGINT, &act, NULL);
122
123 memset(&act, 0, sizeof(act));
124 act.sa_handler = sig_int;
125 act.sa_flags = SA_RESTART;
126 sigaction(SIGTERM, &act, NULL);
127
2f694507
BC
128/* Windows uses SIGBREAK as a quit signal from other applications */
129#ifdef WIN32
130 memset(&act, 0, sizeof(act));
0dea1f55 131 act.sa_handler = sig_break;
2f694507
BC
132 act.sa_flags = SA_RESTART;
133 sigaction(SIGBREAK, &act, NULL);
134#endif
135
4c6d91e8
JA
136 memset(&act, 0, sizeof(act));
137 act.sa_handler = sig_show_status;
138 act.sa_flags = SA_RESTART;
139 sigaction(SIGUSR1, &act, NULL);
140
2e1df07d
JA
141 if (is_backend) {
142 memset(&act, 0, sizeof(act));
143 act.sa_handler = sig_int;
144 act.sa_flags = SA_RESTART;
145 sigaction(SIGPIPE, &act, NULL);
146 }
147}
148
149/*
150 * Check if we are above the minimum rate given.
151 */
8b6a404c 152static bool __check_min_rate(struct thread_data *td, struct timespec *now,
e39c0676 153 enum fio_ddir ddir)
2e1df07d 154{
f5204b8a 155 unsigned long long current_rate_check_bytes = td->this_io_bytes[ddir];
156 unsigned long current_rate_check_blocks = td->this_io_blocks[ddir];
157 unsigned long long option_rate_bytes_min = td->o.ratemin[ddir];
158 unsigned int option_rate_iops_min = td->o.rate_iops_min[ddir];
2e1df07d
JA
159
160 assert(ddir_rw(ddir));
161
162 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
e39c0676 163 return false;
2e1df07d
JA
164
165 /*
166 * allow a 2 second settle period in the beginning
167 */
168 if (mtime_since(&td->start, now) < 2000)
e39c0676 169 return false;
2e1df07d 170
2e1df07d 171 /*
f5204b8a 172 * if last_rate_check_blocks or last_rate_check_bytes is set,
173 * we can compute a rate per ratecycle
2e1df07d 174 */
f5204b8a 175 if (td->last_rate_check_bytes[ddir] || td->last_rate_check_blocks[ddir]) {
176 unsigned long spent = mtime_since(&td->last_rate_check_time[ddir], now);
177 if (spent < td->o.ratecycle || spent==0)
e39c0676 178 return false;
2e1df07d 179
f5204b8a 180 if (td->o.ratemin[ddir]) {
2e1df07d
JA
181 /*
182 * check bandwidth specified rate
183 */
f5204b8a 184 unsigned long long current_rate_bytes =
185 ((current_rate_check_bytes - td->last_rate_check_bytes[ddir]) * 1000) / spent;
186 if (current_rate_bytes < option_rate_bytes_min) {
187 log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n",
188 td->o.name, option_rate_bytes_min, current_rate_bytes);
e39c0676 189 return true;
2e1df07d
JA
190 }
191 } else {
192 /*
193 * checks iops specified rate
194 */
f5204b8a 195 unsigned long long current_rate_iops =
196 ((current_rate_check_blocks - td->last_rate_check_blocks[ddir]) * 1000) / spent;
197
198 if (current_rate_iops < option_rate_iops_min) {
199 log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n",
200 td->o.name, option_rate_iops_min, current_rate_iops);
e39c0676 201 return true;
2e1df07d
JA
202 }
203 }
204 }
205
f5204b8a 206 td->last_rate_check_bytes[ddir] = current_rate_check_bytes;
207 td->last_rate_check_blocks[ddir] = current_rate_check_blocks;
208 memcpy(&td->last_rate_check_time[ddir], now, sizeof(*now));
e39c0676 209 return false;
2e1df07d
JA
210}
211
8b6a404c 212static bool check_min_rate(struct thread_data *td, struct timespec *now)
2e1df07d 213{
e39c0676 214 bool ret = false;
2e1df07d 215
dde39f8c
AD
216 for_each_rw_ddir(ddir) {
217 if (td->bytes_done[ddir])
218 ret |= __check_min_rate(td, now, ddir);
219 }
2e1df07d
JA
220
221 return ret;
222}
223
224/*
225 * When job exits, we can cancel the in-flight IO if we are using async
226 * io. Attempt to do so.
227 */
228static void cleanup_pending_aio(struct thread_data *td)
229{
2e1df07d
JA
230 int r;
231
232 /*
233 * get immediately available events, if any
234 */
55312f9f 235 r = io_u_queued_complete(td, 0);
2e1df07d
JA
236
237 /*
238 * now cancel remaining active events
239 */
240 if (td->io_ops->cancel) {
2ae0b204
JA
241 struct io_u *io_u;
242 int i;
2e1df07d 243
2ae0b204
JA
244 io_u_qiter(&td->io_u_all, io_u, i) {
245 if (io_u->flags & IO_U_F_FLIGHT) {
2e1df07d
JA
246 r = td->io_ops->cancel(td, io_u);
247 if (!r)
248 put_io_u(td, io_u);
249 }
250 }
251 }
252
253 if (td->cur_depth)
55312f9f 254 r = io_u_queued_complete(td, td->cur_depth);
2e1df07d
JA
255}
256
257/*
258 * Helper to handle the final sync of a file. Works just like the normal
259 * io path, just does everything sync.
260 */
e39c0676 261static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
2e1df07d
JA
262{
263 struct io_u *io_u = __get_io_u(td);
d3b07186 264 enum fio_q_status ret;
2e1df07d
JA
265
266 if (!io_u)
e39c0676 267 return true;
2e1df07d
JA
268
269 io_u->ddir = DDIR_SYNC;
270 io_u->file = f;
00615bfb 271 io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
2e1df07d
JA
272
273 if (td_io_prep(td, io_u)) {
274 put_io_u(td, io_u);
e39c0676 275 return true;
2e1df07d
JA
276 }
277
278requeue:
279 ret = td_io_queue(td, io_u);
18f6165e
BVA
280 switch (ret) {
281 case FIO_Q_QUEUED:
a80cb54b 282 td_io_commit(td);
55312f9f 283 if (io_u_queued_complete(td, 1) < 0)
e39c0676 284 return true;
18f6165e
BVA
285 break;
286 case FIO_Q_COMPLETED:
2e1df07d
JA
287 if (io_u->error) {
288 td_verror(td, io_u->error, "td_io_queue");
e39c0676 289 return true;
2e1df07d
JA
290 }
291
55312f9f 292 if (io_u_sync_complete(td, io_u) < 0)
e39c0676 293 return true;
18f6165e
BVA
294 break;
295 case FIO_Q_BUSY:
a80cb54b 296 td_io_commit(td);
2e1df07d
JA
297 goto requeue;
298 }
299
e39c0676 300 return false;
2e1df07d 301}
a3efc919 302
61ee0f86
JA
303static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
304{
00615bfb 305 int ret, ret2;
61ee0f86
JA
306
307 if (fio_file_open(f))
308 return fio_io_sync(td, f);
309
310 if (td_io_open_file(td, f))
311 return 1;
312
313 ret = fio_io_sync(td, f);
00615bfb
SW
314 ret2 = 0;
315 if (fio_file_open(f))
316 ret2 = td_io_close_file(td, f);
317 return (ret || ret2);
61ee0f86
JA
318}
319
8b6a404c 320static inline void __update_ts_cache(struct thread_data *td)
2e1df07d 321{
8b6a404c 322 fio_gettime(&td->ts_cache, NULL);
2e1df07d
JA
323}
324
8b6a404c 325static inline void update_ts_cache(struct thread_data *td)
2e1df07d 326{
8b6a404c
VF
327 if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
328 __update_ts_cache(td);
2e1df07d
JA
329}
330
8b6a404c 331static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
2e1df07d
JA
332{
333 if (in_ramp_time(td))
e39c0676 334 return false;
2e1df07d 335 if (!td->o.timeout)
e39c0676 336 return false;
0de5b26f 337 if (utime_since(&td->epoch, t) >= td->o.timeout)
e39c0676 338 return true;
2e1df07d 339
e39c0676 340 return false;
2e1df07d
JA
341}
342
95603b74
BF
343/*
344 * We need to update the runtime consistently in ms, but keep a running
345 * tally of the current elapsed time in microseconds for sub millisecond
346 * updates.
347 */
348static inline void update_runtime(struct thread_data *td,
349 unsigned long long *elapsed_us,
350 const enum fio_ddir ddir)
351{
d5c8ea29
JA
352 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
353 return;
354
95603b74
BF
355 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
356 elapsed_us[ddir] += utime_since_now(&td->start);
357 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
358}
359
e39c0676
JA
360static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
361 int *retptr)
2e1df07d
JA
362{
363 int ret = *retptr;
364
365 if (ret < 0 || td->error) {
8b28bd41
DM
366 int err = td->error;
367 enum error_type_bit eb;
2e1df07d
JA
368
369 if (ret < 0)
370 err = -ret;
2e1df07d 371
8b28bd41
DM
372 eb = td_error_type(ddir, err);
373 if (!(td->o.continue_on_error & (1 << eb)))
e39c0676 374 return true;
2e1df07d 375
8b28bd41 376 if (td_non_fatal_error(td, eb, err)) {
2e1df07d
JA
377 /*
378 * Continue with the I/Os in case of
379 * a non fatal error.
380 */
381 update_error_count(td, err);
382 td_clear_error(td);
383 *retptr = 0;
e39c0676 384 return false;
418f5399 385 } else if (td->o.fill_device && (err == ENOSPC || err == EDQUOT)) {
2e1df07d
JA
386 /*
387 * We expect to hit this error if
388 * fill_device option is set.
389 */
390 td_clear_error(td);
ebea2133 391 fio_mark_td_terminate(td);
e39c0676 392 return true;
2e1df07d
JA
393 } else {
394 /*
395 * Stop the I/O in case of a fatal
396 * error.
397 */
398 update_error_count(td, err);
e39c0676 399 return true;
2e1df07d
JA
400 }
401 }
402
e39c0676 403 return false;
2e1df07d
JA
404}
405
c97f1ad6
JA
406static void check_update_rusage(struct thread_data *td)
407{
408 if (td->update_rusage) {
409 td->update_rusage = 0;
410 update_rusage_stat(td);
971caeb1 411 fio_sem_up(td->rusage_sem);
c97f1ad6
JA
412 }
413}
414
8b6a404c 415static int wait_for_completions(struct thread_data *td, struct timespec *time)
69fea81e
JA
416{
417 const int full = queue_full(td);
418 int min_evts = 0;
419 int ret;
420
6be06c46
JA
421 if (td->flags & TD_F_REGROW_LOGS)
422 return io_u_quiesce(td);
1fed2080 423
69fea81e
JA
424 /*
425 * if the queue is full, we MUST reap at least 1 event
426 */
82407585 427 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
4b157ac6 428 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
69fea81e
JA
429 min_evts = 1;
430
90711edc 431 if (time && should_check_rate(td))
69fea81e
JA
432 fio_gettime(time, NULL);
433
434 do {
55312f9f 435 ret = io_u_queued_complete(td, min_evts);
69fea81e
JA
436 if (ret < 0)
437 break;
438 } while (full && (td->cur_depth > td->o.iodepth_low));
439
440 return ret;
441}
442
e9d512d8
JA
443int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
444 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
8b6a404c 445 struct timespec *comp_time)
e9d512d8 446{
e9d512d8
JA
447 switch (*ret) {
448 case FIO_Q_COMPLETED:
449 if (io_u->error) {
450 *ret = -io_u->error;
451 clear_io_u(td, io_u);
452 } else if (io_u->resid) {
5fff9543 453 long long bytes = io_u->xfer_buflen - io_u->resid;
e9d512d8
JA
454 struct fio_file *f = io_u->file;
455
456 if (bytes_issued)
457 *bytes_issued += bytes;
458
459 if (!from_verify)
94a78d52 460 trim_io_piece(io_u);
e9d512d8
JA
461
462 /*
463 * zero read, fail
464 */
465 if (!bytes) {
466 if (!from_verify)
467 unlog_io_piece(td, io_u);
468 td_verror(td, EIO, "full resid");
469 put_io_u(td, io_u);
470 break;
471 }
472
473 io_u->xfer_buflen = io_u->resid;
474 io_u->xfer_buf += bytes;
475 io_u->offset += bytes;
476
477 if (ddir_rw(io_u->ddir))
478 td->ts.short_io_u[io_u->ddir]++;
479
e9d512d8
JA
480 if (io_u->offset == f->real_file_size)
481 goto sync_done;
482
483 requeue_io_u(td, &io_u);
484 } else {
485sync_done:
90711edc 486 if (comp_time && should_check_rate(td))
e9d512d8
JA
487 fio_gettime(comp_time, NULL);
488
489 *ret = io_u_sync_complete(td, io_u);
490 if (*ret < 0)
491 break;
492 }
a0e3e5a6 493
cf2c8d52
JA
494 if (td->flags & TD_F_REGROW_LOGS)
495 regrow_logs(td);
496
a0e3e5a6
EW
497 /*
498 * when doing I/O (not when verifying),
499 * check for any errors that are to be ignored
500 */
501 if (!from_verify)
502 break;
503
e9d512d8
JA
504 return 0;
505 case FIO_Q_QUEUED:
506 /*
507 * if the engine doesn't have a commit hook,
508 * the io_u is really queued. if it does have such
509 * a hook, it has to call io_u_queued() itself.
510 */
511 if (td->io_ops->commit == NULL)
512 io_u_queued(td, io_u);
513 if (bytes_issued)
514 *bytes_issued += io_u->xfer_buflen;
515 break;
516 case FIO_Q_BUSY:
517 if (!from_verify)
518 unlog_io_piece(td, io_u);
519 requeue_io_u(td, &io_u);
a80cb54b 520 td_io_commit(td);
e9d512d8
JA
521 break;
522 default:
b0775325 523 assert(*ret < 0);
e9d512d8
JA
524 td_verror(td, -(*ret), "td_io_queue");
525 break;
526 }
527
528 if (break_on_this_error(td, ddir, ret))
529 return 1;
530
531 return 0;
532}
533
e39c0676 534static inline bool io_in_polling(struct thread_data *td)
82407585
RP
535{
536 return !td->o.iodepth_batch_complete_min &&
537 !td->o.iodepth_batch_complete_max;
538}
39c1c323 539/*
540 * Unlinks files from thread data fio_file structure
541 */
542static int unlink_all_files(struct thread_data *td)
543{
544 struct fio_file *f;
545 unsigned int i;
546 int ret = 0;
547
548 for_each_file(td, f, i) {
549 if (f->filetype != FIO_TYPE_FILE)
550 continue;
551 ret = td_io_unlink_file(td, f);
552 if (ret)
553 break;
554 }
555
556 if (ret)
557 td_verror(td, ret, "unlink_all_files");
558
559 return ret;
560}
82407585 561
997b5680
SW
562/*
563 * Check if io_u will overlap an in-flight IO in the queue
564 */
c06379a6 565bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
997b5680
SW
566{
567 bool overlap;
568 struct io_u *check_io_u;
569 unsigned long long x1, x2, y1, y2;
570 int i;
571
572 x1 = io_u->offset;
573 x2 = io_u->offset + io_u->buflen;
574 overlap = false;
575 io_u_qiter(q, check_io_u, i) {
576 if (check_io_u->flags & IO_U_F_FLIGHT) {
577 y1 = check_io_u->offset;
578 y2 = check_io_u->offset + check_io_u->buflen;
579
580 if (x1 < y2 && y1 < x2) {
581 overlap = true;
5fff9543 582 dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n",
997b5680
SW
583 x1, io_u->buflen,
584 y1, check_io_u->buflen);
585 break;
586 }
587 }
588 }
589
590 return overlap;
591}
592
d3b07186 593static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
08b2b386
JA
594{
595 /*
596 * Check for overlap if the user asked us to, and we have
597 * at least one IO in flight besides this one.
598 */
599 if (td->o.serialize_overlap && td->cur_depth > 1 &&
600 in_flight_overlap(&td->io_u_all, io_u))
601 return FIO_Q_BUSY;
602
603 return td_io_queue(td, io_u);
604}
605
2e1df07d
JA
606/*
607 * The main verify engine. Runs over the writes we previously submitted,
608 * reads the blocks back in, and checks the crc/md5 of the data.
609 */
100f49f1 610static void do_verify(struct thread_data *td, uint64_t verify_bytes)
2e1df07d
JA
611{
612 struct fio_file *f;
613 struct io_u *io_u;
614 int ret, min_events;
615 unsigned int i;
616
617 dprint(FD_VERIFY, "starting loop\n");
618
619 /*
620 * sync io first and invalidate cache, to make sure we really
621 * read from disk.
622 */
623 for_each_file(td, f, i) {
624 if (!fio_file_open(f))
625 continue;
626 if (fio_io_sync(td, f))
627 break;
628 if (file_invalidate_cache(td, f))
629 break;
630 }
631
c97f1ad6
JA
632 check_update_rusage(td);
633
2e1df07d
JA
634 if (td->error)
635 return;
636
637 td_set_runstate(td, TD_VERIFYING);
638
639 io_u = NULL;
640 while (!td->terminate) {
fbccf46c 641 enum fio_ddir ddir;
e9d512d8 642 int full;
2e1df07d 643
8b6a404c 644 update_ts_cache(td);
c97f1ad6 645 check_update_rusage(td);
2e1df07d 646
8b6a404c
VF
647 if (runtime_exceeded(td, &td->ts_cache)) {
648 __update_ts_cache(td);
649 if (runtime_exceeded(td, &td->ts_cache)) {
ebea2133 650 fio_mark_td_terminate(td);
2e1df07d
JA
651 break;
652 }
653 }
654
9e684a49
DE
655 if (flow_threshold_exceeded(td))
656 continue;
657
44cbc6da
JA
658 if (!td->o.experimental_verify) {
659 io_u = __get_io_u(td);
660 if (!io_u)
661 break;
2e1df07d 662
44cbc6da
JA
663 if (get_next_verify(td, io_u)) {
664 put_io_u(td, io_u);
665 break;
666 }
2e1df07d 667
44cbc6da
JA
668 if (td_io_prep(td, io_u)) {
669 put_io_u(td, io_u);
670 break;
671 }
672 } else {
191d6634 673 if (td->bytes_verified + td->o.rw_min_bs > verify_bytes)
100f49f1
JA
674 break;
675
bcd5abfa 676 while ((io_u = get_io_u(td)) != NULL) {
059b61f2 677 if (IS_ERR_OR_NULL(io_u)) {
002fe734
JA
678 io_u = NULL;
679 ret = FIO_Q_BUSY;
680 goto reap;
681 }
682
bcd5abfa
JA
683 /*
684 * We are only interested in the places where
685 * we wrote or trimmed IOs. Turn those into
686 * reads for verification purposes.
687 */
688 if (io_u->ddir == DDIR_READ) {
689 /*
690 * Pretend we issued it for rwmix
691 * accounting
692 */
693 td->io_issues[DDIR_READ]++;
694 put_io_u(td, io_u);
695 continue;
696 } else if (io_u->ddir == DDIR_TRIM) {
697 io_u->ddir = DDIR_READ;
1651e431 698 io_u_set(td, io_u, IO_U_F_TRIMMED);
bcd5abfa
JA
699 break;
700 } else if (io_u->ddir == DDIR_WRITE) {
701 io_u->ddir = DDIR_READ;
28921bc0
SK
702 io_u->numberio = td->verify_read_issues;
703 td->verify_read_issues++;
5c5c33c1 704 populate_verify_io_u(td, io_u);
bcd5abfa
JA
705 break;
706 } else {
707 put_io_u(td, io_u);
708 continue;
709 }
710 }
44cbc6da 711
bcd5abfa 712 if (!io_u)
44cbc6da 713 break;
2e1df07d
JA
714 }
715
ca09be4b
JA
716 if (verify_state_should_stop(td, io_u)) {
717 put_io_u(td, io_u);
718 break;
719 }
720
2e1df07d
JA
721 if (td->o.verify_async)
722 io_u->end_io = verify_io_u_async;
723 else
724 io_u->end_io = verify_io_u;
725
fbccf46c 726 ddir = io_u->ddir;
10adc4a7
GG
727 if (!td->o.disable_slat)
728 fio_gettime(&io_u->start_time, NULL);
fbccf46c 729
08b2b386 730 ret = io_u_submit(td, io_u);
2e1df07d 731
e9d512d8 732 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
2e1df07d
JA
733 break;
734
735 /*
736 * if we can queue more, do so. but check if there are
737 * completed io_u's first. Note that we can get BUSY even
738 * without IO queued, if the system is resource starved.
739 */
002fe734 740reap:
2e1df07d 741 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
82407585 742 if (full || io_in_polling(td))
55312f9f 743 ret = wait_for_completions(td, NULL);
2e1df07d 744
2e1df07d
JA
745 if (ret < 0)
746 break;
747 }
748
c97f1ad6
JA
749 check_update_rusage(td);
750
2e1df07d
JA
751 if (!td->error) {
752 min_events = td->cur_depth;
753
754 if (min_events)
55312f9f 755 ret = io_u_queued_complete(td, min_events);
2e1df07d
JA
756 } else
757 cleanup_pending_aio(td);
758
759 td_set_runstate(td, TD_RUNNING);
760
761 dprint(FD_VERIFY, "exiting loop\n");
762}
763
e39c0676 764static bool exceeds_number_ios(struct thread_data *td)
3939fe85
JA
765{
766 unsigned long long number_ios;
767
768 if (!td->o.number_ios)
e39c0676 769 return false;
3939fe85 770
cf8a46a7 771 number_ios = ddir_rw_sum(td->io_blocks);
3939fe85
JA
772 number_ios += td->io_u_queued + td->io_u_in_flight;
773
cf8a46a7 774 return number_ios >= (td->o.number_ios * td->loops);
3939fe85
JA
775}
776
926098e0 777static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
f7078f7b 778{
77731b29 779 unsigned long long bytes, limit;
f7078f7b
JA
780
781 if (td_rw(td))
926098e0 782 bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE];
f7078f7b 783 else if (td_write(td))
926098e0 784 bytes = this_bytes[DDIR_WRITE];
6eaf09d6 785 else if (td_read(td))
926098e0 786 bytes = this_bytes[DDIR_READ];
f7078f7b 787 else
926098e0 788 bytes = this_bytes[DDIR_TRIM];
f7078f7b 789
5be9bf09
TK
790 if (td->o.io_size)
791 limit = td->o.io_size;
77731b29
JA
792 else
793 limit = td->o.size;
794
cf8a46a7 795 limit *= td->loops;
77731b29 796 return bytes >= limit || exceeds_number_ios(td);
f7078f7b
JA
797}
798
926098e0 799static bool io_issue_bytes_exceeded(struct thread_data *td)
e28dd2cf 800{
926098e0
JA
801 return io_bytes_exceeded(td, td->io_issue_bytes);
802}
e28dd2cf 803
926098e0
JA
804static bool io_complete_bytes_exceeded(struct thread_data *td)
805{
806 return io_bytes_exceeded(td, td->this_io_bytes);
e28dd2cf
JE
807}
808
50a8ce86
D
809/*
810 * used to calculate the next io time for rate control
811 *
812 */
813static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
814{
1a9bf814 815 uint64_t bps = td->rate_bps[ddir];
50a8ce86
D
816
817 assert(!(td->flags & TD_F_CHILD));
ff6bb260 818
6de65959 819 if (td->o.rate_process == RATE_PROCESS_POISSON) {
1a9bf814
JA
820 uint64_t val, iops;
821
d6299558 822 iops = bps / td->o.min_bs[ddir];
a77d139b 823 val = (int64_t) (1000000 / iops) *
6c8611c7 824 -logf(__rand_0_1(&td->poisson_state[ddir]));
a77d139b 825 if (val) {
98e51d7c
JA
826 dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
827 (unsigned long long) 1000000 / val,
828 ddir);
a77d139b 829 }
6c8611c7
JA
830 td->last_usec[ddir] += val;
831 return td->last_usec[ddir];
ff6bb260 832 } else if (bps) {
1a9bf814
JA
833 uint64_t bytes = td->rate_io_issue_bytes[ddir];
834 uint64_t secs = bytes / bps;
835 uint64_t remainder = bytes % bps;
836
50a8ce86 837 return remainder * 1000000 / bps + secs * 1000000;
e7b24047
JA
838 }
839
840 return 0;
50a8ce86
D
841}
842
f7942acd
SK
843static void init_thinktime(struct thread_data *td)
844{
845 if (td->o.thinktime_blocks_type == THINKTIME_BLOCKS_TYPE_COMPLETE)
846 td->thinktime_blocks_counter = td->io_blocks;
847 else
848 td->thinktime_blocks_counter = td->io_issues;
849 td->last_thinktime = td->epoch;
850 td->last_thinktime_blocks = 0;
851}
852
7216b664
HQ
853static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir,
854 struct timespec *time)
1a9bf814
JA
855{
856 unsigned long long b;
f9fc7a27 857 unsigned long long runtime_left;
1a9bf814
JA
858 uint64_t total;
859 int left;
f7942acd
SK
860 struct timespec now;
861 bool stall = false;
862
863 if (td->o.thinktime_iotime) {
864 fio_gettime(&now, NULL);
865 if (utime_since(&td->last_thinktime, &now)
f9fc7a27 866 >= td->o.thinktime_iotime) {
f7942acd
SK
867 stall = true;
868 } else if (!fio_option_is_set(&td->o, thinktime_blocks)) {
869 /*
870 * When thinktime_iotime is set and thinktime_blocks is
871 * not set, skip the thinktime_blocks check, since
872 * thinktime_blocks default value 1 does not work
873 * together with thinktime_iotime.
874 */
875 return;
876 }
877
878 }
1a9bf814 879
33f42c20 880 b = ddir_rw_sum(td->thinktime_blocks_counter);
f7942acd
SK
881 if (b >= td->last_thinktime_blocks + td->o.thinktime_blocks)
882 stall = true;
883
884 if (!stall)
1a9bf814
JA
885 return;
886
887 io_u_quiesce(td);
888
f9fc7a27
AK
889 left = td->o.thinktime_spin;
890 if (td->o.timeout) {
891 runtime_left = td->o.timeout - utime_since_now(&td->epoch);
892 if (runtime_left < (unsigned long long)left)
893 left = runtime_left;
894 }
895
1a9bf814 896 total = 0;
f9fc7a27
AK
897 if (left)
898 total = usec_spin(left);
1a9bf814
JA
899
900 left = td->o.thinktime - total;
f9fc7a27
AK
901 if (td->o.timeout) {
902 runtime_left = td->o.timeout - utime_since_now(&td->epoch);
903 if (runtime_left < (unsigned long long)left)
904 left = runtime_left;
905 }
906
1a9bf814
JA
907 if (left)
908 total += usec_sleep(td, left);
909
910 /*
911 * If we're ignoring thinktime for the rate, add the number of bytes
1aa39b0c
JA
912 * we would have done while sleeping, minus one block to ensure we
913 * start issuing immediately after the sleep.
1a9bf814 914 */
1aa39b0c 915 if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
f2aedb17 916 uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
62b858f6
JA
917 uint64_t bs = td->o.min_bs[ddir];
918 uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
f2aedb17
JA
919 uint64_t over;
920
62b858f6
JA
921 if (usperop <= total)
922 over = bs;
f2aedb17 923 else
62b858f6 924 over = (usperop - total) / usperop * -bs;
f2aedb17
JA
925
926 td->rate_io_issue_bytes[ddir] += (missed - over);
182cab24
JA
927 /* adjust for rate_process=poisson */
928 td->last_usec[ddir] += total;
1aa39b0c 929 }
7216b664
HQ
930
931 if (time && should_check_rate(td))
932 fio_gettime(time, NULL);
f7942acd
SK
933
934 td->last_thinktime_blocks = b;
f9fc7a27
AK
935 if (td->o.thinktime_iotime) {
936 fio_gettime(&now, NULL);
f7942acd 937 td->last_thinktime = now;
f9fc7a27 938 }
1a9bf814
JA
939}
940
2e1df07d
JA
941/*
942 * Main IO worker function. It retrieves io_u's to process and queues
943 * and reaps them, checking for rate and errors along the way.
100f49f1
JA
944 *
945 * Returns number of bytes written and trimmed.
2e1df07d 946 */
95b03be0 947static void do_io(struct thread_data *td, uint64_t *bytes_done)
2e1df07d
JA
948{
949 unsigned int i;
950 int ret = 0;
c2703bf3 951 uint64_t total_bytes, bytes_issued = 0;
70c68076 952
95b03be0
JA
953 for (i = 0; i < DDIR_RWDIR_CNT; i++)
954 bytes_done[i] = td->bytes_done[i];
2e1df07d
JA
955
956 if (in_ramp_time(td))
957 td_set_runstate(td, TD_RAMP);
958 else
959 td_set_runstate(td, TD_RUNNING);
960
3e260a46
JA
961 lat_target_init(td);
962
1e564979
JE
963 total_bytes = td->o.size;
964 /*
5be9bf09 965 * Allow random overwrite workloads to write up to io_size
1e564979
JE
966 * before starting verification phase as 'size' doesn't apply.
967 */
968 if (td_write(td) && td_random(td) && td->o.norandommap)
5be9bf09 969 total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
78a6469c
JA
970 /*
971 * If verify_backlog is enabled, we'll run the verify in this
972 * handler as well. For that case, we may need up to twice the
973 * amount of bytes.
974 */
78a6469c
JA
975 if (td->o.verify != VERIFY_NONE &&
976 (td_write(td) && td->o.verify_backlog))
c2703bf3
JA
977 total_bytes += td->o.size;
978
82a90686 979 /* In trimwrite mode, each byte is trimmed and then written, so
08996af4
AK
980 * allow total_bytes or number of ios to be twice as big */
981 if (td_trimwrite(td)) {
0e4dd95c 982 total_bytes += td->total_io_size;
08996af4
AK
983 td->o.number_ios *= 2;
984 }
0e4dd95c 985
f7078f7b 986 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
e28dd2cf 987 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
c04e4661 988 td->o.time_based) {
8b6a404c 989 struct timespec comp_time;
2e1df07d 990 struct io_u *io_u;
e9d512d8 991 int full;
2e1df07d
JA
992 enum fio_ddir ddir;
993
c97f1ad6
JA
994 check_update_rusage(td);
995
7d7803fa 996 if (td->terminate || td->done)
2e1df07d
JA
997 break;
998
8b6a404c 999 update_ts_cache(td);
2e1df07d 1000
8b6a404c
VF
1001 if (runtime_exceeded(td, &td->ts_cache)) {
1002 __update_ts_cache(td);
1003 if (runtime_exceeded(td, &td->ts_cache)) {
ebea2133 1004 fio_mark_td_terminate(td);
2e1df07d
JA
1005 break;
1006 }
1007 }
1008
9e684a49
DE
1009 if (flow_threshold_exceeded(td))
1010 continue;
1011
f1a32461
JA
1012 /*
1013 * Break if we exceeded the bytes. The exception is time
1014 * based runs, but we still need to break out of the loop
1015 * for those to run verification, if enabled.
98e7161c 1016 * Jobs read from iolog do not use this stop condition.
f1a32461
JA
1017 */
1018 if (bytes_issued >= total_bytes &&
98e7161c 1019 !td->o.read_iolog_file &&
f1a32461
JA
1020 (!td->o.time_based ||
1021 (td->o.time_based && td->o.verify != VERIFY_NONE)))
20876c53
JC
1022 break;
1023
2e1df07d 1024 io_u = get_io_u(td);
002fe734
JA
1025 if (IS_ERR_OR_NULL(io_u)) {
1026 int err = PTR_ERR(io_u);
1027
1028 io_u = NULL;
1a9bf814 1029 ddir = DDIR_INVAL;
002fe734
JA
1030 if (err == -EBUSY) {
1031 ret = FIO_Q_BUSY;
1032 goto reap;
1033 }
3e260a46
JA
1034 if (td->o.latency_target)
1035 goto reap;
2e1df07d 1036 break;
3e260a46 1037 }
2e1df07d 1038
28921bc0 1039 if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY) {
ad21b61d
H
1040 if (!(io_u->flags & IO_U_F_PATTERN_DONE)) {
1041 io_u_set(td, io_u, IO_U_F_PATTERN_DONE);
1042 io_u->numberio = td->io_issues[io_u->ddir];
1043 populate_verify_io_u(td, io_u);
1044 }
28921bc0 1045 }
5c5c33c1 1046
2e1df07d
JA
1047 ddir = io_u->ddir;
1048
1049 /*
82af2a7c
JA
1050 * Add verification end_io handler if:
1051 * - Asked to verify (!td_rw(td))
1052 * - Or the io_u is from our verify list (mixed write/ver)
2e1df07d
JA
1053 */
1054 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
82af2a7c 1055 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
c4b6117b 1056
ca09be4b
JA
1057 if (verify_state_should_stop(td, io_u)) {
1058 put_io_u(td, io_u);
1059 break;
1060 }
1061
2e1df07d
JA
1062 if (td->o.verify_async)
1063 io_u->end_io = verify_io_u_async;
1064 else
1065 io_u->end_io = verify_io_u;
1066 td_set_runstate(td, TD_VERIFYING);
1067 } else if (in_ramp_time(td))
1068 td_set_runstate(td, TD_RAMP);
1069 else
1070 td_set_runstate(td, TD_RUNNING);
1071
9a50c5c5 1072 /*
f9401285
JA
1073 * Always log IO before it's issued, so we know the specific
1074 * order of it. The logged unit will track when the IO has
1075 * completed.
9a50c5c5 1076 */
c4b6117b
PV
1077 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1078 td->o.do_verify &&
1079 td->o.verify != VERIFY_NONE &&
f9401285 1080 !td->o.experimental_verify)
c4b6117b
PV
1081 log_io_piece(td, io_u);
1082
a9da8ab2 1083 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
5fff9543 1084 const unsigned long long blen = io_u->xfer_buflen;
a43f4461 1085 const enum fio_ddir __ddir = acct_ddir(io_u);
0c5df5f9 1086
a9da8ab2
JA
1087 if (td->error)
1088 break;
0c5df5f9 1089
26de50cf
JA
1090 workqueue_enqueue(&td->io_wq, &io_u->work);
1091 ret = FIO_Q_QUEUED;
50a8ce86 1092
a43f4461
JA
1093 if (ddir_rw(__ddir)) {
1094 td->io_issues[__ddir]++;
1095 td->io_issue_bytes[__ddir] += blen;
1096 td->rate_io_issue_bytes[__ddir] += blen;
0c5df5f9
JA
1097 }
1098
5b28b75b 1099 if (should_check_rate(td)) {
a43f4461 1100 td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir);
5b28b75b 1101 fio_gettime(&comp_time, NULL);
1102 }
50a8ce86 1103
a9da8ab2 1104 } else {
08b2b386 1105 ret = io_u_submit(td, io_u);
2e1df07d 1106
50a8ce86
D
1107 if (should_check_rate(td))
1108 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
1109
fd727d9d 1110 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
a9da8ab2 1111 break;
2e1df07d 1112
a9da8ab2
JA
1113 /*
1114 * See if we need to complete some commands. Note that
1115 * we can get BUSY even without IO queued, if the
1116 * system is resource starved.
1117 */
3e260a46 1118reap:
a9da8ab2
JA
1119 full = queue_full(td) ||
1120 (ret == FIO_Q_BUSY && td->cur_depth);
82407585 1121 if (full || io_in_polling(td))
a9da8ab2
JA
1122 ret = wait_for_completions(td, &comp_time);
1123 }
2e1df07d
JA
1124 if (ret < 0)
1125 break;
33f42c20
HQ
1126
1127 if (ddir_rw(ddir) && td->o.thinktime)
7216b664 1128 handle_thinktime(td, ddir, &comp_time);
33f42c20 1129
55312f9f 1130 if (!ddir_rw_sum(td->bytes_done) &&
9b87f09b 1131 !td_ioengine_flagged(td, FIO_NOIO))
2e1df07d
JA
1132 continue;
1133
55312f9f
JA
1134 if (!in_ramp_time(td) && should_check_rate(td)) {
1135 if (check_min_rate(td, &comp_time)) {
f9cafb12 1136 if (exitall_on_terminate || td->o.exitall_error)
64402a8a 1137 fio_terminate_threads(td->groupid, td->o.exit_what);
2e1df07d
JA
1138 td_verror(td, EIO, "check_min_rate");
1139 break;
1140 }
1141 }
3e260a46
JA
1142 if (!in_ramp_time(td) && td->o.latency_target)
1143 lat_target_check(td);
2e1df07d
JA
1144 }
1145
c97f1ad6
JA
1146 check_update_rusage(td);
1147
2e1df07d 1148 if (td->trim_entries)
4e0a8fa2 1149 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
2e1df07d 1150
418f5399 1151 if (td->o.fill_device && (td->error == ENOSPC || td->error == EDQUOT)) {
2e1df07d 1152 td->error = 0;
ebea2133 1153 fio_mark_td_terminate(td);
2e1df07d
JA
1154 }
1155 if (!td->error) {
1156 struct fio_file *f;
1157
a9da8ab2
JA
1158 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1159 workqueue_flush(&td->io_wq);
1160 i = 0;
1161 } else
1162 i = td->cur_depth;
1163
2e1df07d 1164 if (i) {
55312f9f 1165 ret = io_u_queued_complete(td, i);
418f5399
MB
1166 if (td->o.fill_device &&
1167 (td->error == ENOSPC || td->error == EDQUOT))
2e1df07d
JA
1168 td->error = 0;
1169 }
1170
00615bfb 1171 if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) {
2e1df07d
JA
1172 td_set_runstate(td, TD_FSYNCING);
1173
1174 for_each_file(td, f, i) {
61ee0f86 1175 if (!fio_file_fsync(td, f))
2e1df07d 1176 continue;
61ee0f86
JA
1177
1178 log_err("fio: end_fsync failed for file %s\n",
1179 f->file_name);
2e1df07d
JA
1180 }
1181 }
5b28b75b 1182 } else {
1183 if (td->o.io_submit_mode == IO_MODE_OFFLOAD)
1184 workqueue_flush(&td->io_wq);
2e1df07d 1185 cleanup_pending_aio(td);
5b28b75b 1186 }
2e1df07d
JA
1187
1188 /*
1189 * stop job if we failed doing any IO
1190 */
342f4be4 1191 if (!ddir_rw_sum(td->this_io_bytes))
2e1df07d 1192 td->done = 1;
100f49f1 1193
95b03be0
JA
1194 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1195 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
2e1df07d
JA
1196}
1197
94a6e1bb
JA
1198static void free_file_completion_logging(struct thread_data *td)
1199{
1200 struct fio_file *f;
1201 unsigned int i;
1202
1203 for_each_file(td, f, i) {
1204 if (!f->last_write_comp)
1205 break;
1206 sfree(f->last_write_comp);
1207 }
1208}
1209
1210static int init_file_completion_logging(struct thread_data *td,
1211 unsigned int depth)
1212{
1213 struct fio_file *f;
1214 unsigned int i;
1215
1216 if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save)
1217 return 0;
1218
1219 for_each_file(td, f, i) {
1220 f->last_write_comp = scalloc(depth, sizeof(uint64_t));
1221 if (!f->last_write_comp)
1222 goto cleanup;
1223 }
1224
1225 return 0;
1226
1227cleanup:
1228 free_file_completion_logging(td);
1229 log_err("fio: failed to alloc write comp data\n");
1230 return 1;
1231}
1232
2e1df07d
JA
1233static void cleanup_io_u(struct thread_data *td)
1234{
2e1df07d
JA
1235 struct io_u *io_u;
1236
2ae0b204 1237 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
c73ed246
JA
1238
1239 if (td->io_ops->io_u_free)
1240 td->io_ops->io_u_free(td, io_u);
1241
3114b675 1242 fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td));
2e1df07d
JA
1243 }
1244
1245 free_io_mem(td);
2ae0b204
JA
1246
1247 io_u_rexit(&td->io_u_requeues);
3114b675
VF
1248 io_u_qexit(&td->io_u_freelist, false);
1249 io_u_qexit(&td->io_u_all, td_offload_overlap(td));
ca09be4b 1250
94a6e1bb 1251 free_file_completion_logging(td);
2e1df07d
JA
1252}
1253
1254static int init_io_u(struct thread_data *td)
1255{
1256 struct io_u *io_u;
2e1df07d 1257 int cl_align, i, max_units;
71e6e5a2 1258 int err;
2e1df07d
JA
1259
1260 max_units = td->o.iodepth;
59d8d0f5 1261
2ae0b204 1262 err = 0;
34851ad5 1263 err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth);
3114b675
VF
1264 err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false);
1265 err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td));
2ae0b204
JA
1266
1267 if (err) {
1268 log_err("fio: failed setting up IO queues\n");
1269 return 1;
1270 }
1271
71e6e5a2
AK
1272 cl_align = os_cache_line_size();
1273
1274 for (i = 0; i < max_units; i++) {
1275 void *ptr;
1276
1277 if (td->terminate)
1278 return 1;
1279
3114b675 1280 ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td));
71e6e5a2
AK
1281 if (!ptr) {
1282 log_err("fio: unable to allocate aligned memory\n");
ab43b8cb 1283 return 1;
71e6e5a2
AK
1284 }
1285
1286 io_u = ptr;
1287 memset(io_u, 0, sizeof(*io_u));
1288 INIT_FLIST_HEAD(&io_u->verify_list);
1289 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1290
1291 io_u->index = i;
1292 io_u->flags = IO_U_F_FREE;
1293 io_u_qpush(&td->io_u_freelist, io_u);
1294
1295 /*
1296 * io_u never leaves this stack, used for iteration of all
1297 * io_u buffers.
1298 */
1299 io_u_qpush(&td->io_u_all, io_u);
1300
1301 if (td->io_ops->io_u_init) {
1302 int ret = td->io_ops->io_u_init(td, io_u);
1303
1304 if (ret) {
1305 log_err("fio: failed to init engine data: %d\n", ret);
1306 return 1;
1307 }
1308 }
1309 }
1310
6d8fe6e8
SK
1311 if (init_io_u_buffers(td))
1312 return 1;
71e6e5a2
AK
1313
1314 if (init_file_completion_logging(td, max_units))
1315 return 1;
1316
1317 return 0;
1318}
1319
1320int init_io_u_buffers(struct thread_data *td)
1321{
1322 struct io_u *io_u;
1323 unsigned long long max_bs, min_write;
1324 int i, max_units;
1325 int data_xfer = 1;
1326 char *p;
1327
1328 max_units = td->o.iodepth;
1329 max_bs = td_max_bs(td);
1330 min_write = td->o.min_bs[DDIR_WRITE];
1331 td->orig_buffer_size = (unsigned long long) max_bs
1332 * (unsigned long long) max_units;
1333
1334 if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td)))
1335 data_xfer = 0;
1336
fd8a09b8 1337 /*
1338 * if we may later need to do address alignment, then add any
1339 * possible adjustment here so that we don't cause a buffer
1340 * overflow later. this adjustment may be too much if we get
1341 * lucky and the allocator gives us an aligned address.
1342 */
a25ba6c6 1343 if (td->o.odirect || td->o.mem_align ||
9b87f09b 1344 td_ioengine_flagged(td, FIO_RAWIO))
fd8a09b8 1345 td->orig_buffer_size += page_mask + td->o.mem_align;
1346
2e1df07d 1347 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
5fff9543 1348 unsigned long long bs;
2e1df07d
JA
1349
1350 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1351 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1352 }
1353
1354 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1355 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1356 return 1;
1357 }
1358
59d8d0f5 1359 if (data_xfer && allocate_io_mem(td))
2e1df07d
JA
1360 return 1;
1361
a25ba6c6 1362 if (td->o.odirect || td->o.mem_align ||
9b87f09b 1363 td_ioengine_flagged(td, FIO_RAWIO))
248c9436 1364 p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
2e1df07d
JA
1365 else
1366 p = td->orig_buffer;
1367
2e1df07d 1368 for (i = 0; i < max_units; i++) {
71e6e5a2 1369 io_u = td->io_u_all.io_us[i];
2e1df07d
JA
1370 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1371
59d8d0f5 1372 if (data_xfer) {
2e1df07d
JA
1373 io_u->buf = p;
1374 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1375
1376 if (td_write(td))
9c42684e 1377 io_u_fill_buffer(td, io_u, min_write, max_bs);
2e1df07d
JA
1378 if (td_write(td) && td->o.verify_pattern_bytes) {
1379 /*
1380 * Fill the buffer with the pattern if we are
1381 * going to be doing writes.
1382 */
ce35b1ec 1383 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
2e1df07d
JA
1384 }
1385 }
2e1df07d
JA
1386 p += max_bs;
1387 }
1388
1389 return 0;
1390}
1391
5592e992 1392#ifdef FIO_HAVE_IOSCHED_SWITCH
bd4d9bdc 1393/*
5592e992 1394 * These functions are Linux specific.
bd4d9bdc
TK
1395 * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
1396 */
5592e992 1397static int set_ioscheduler(struct thread_data *td, struct fio_file *file)
2e1df07d 1398{
178ef27a 1399 char tmp[256], tmp2[128], *p;
2e1df07d
JA
1400 FILE *f;
1401 int ret;
1402
5592e992
DLM
1403 assert(file->du && file->du->sysfs_root);
1404 sprintf(tmp, "%s/queue/scheduler", file->du->sysfs_root);
2e1df07d
JA
1405
1406 f = fopen(tmp, "r+");
1407 if (!f) {
1408 if (errno == ENOENT) {
1409 log_err("fio: os or kernel doesn't support IO scheduler"
1410 " switching\n");
1411 return 0;
1412 }
1413 td_verror(td, errno, "fopen iosched");
1414 return 1;
1415 }
1416
1417 /*
1418 * Set io scheduler.
1419 */
1420 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1421 if (ferror(f) || ret != 1) {
1422 td_verror(td, errno, "fwrite");
1423 fclose(f);
1424 return 1;
1425 }
1426
1427 rewind(f);
1428
1429 /*
1430 * Read back and check that the selected scheduler is now the default.
1431 */
0e23d456 1432 ret = fread(tmp, 1, sizeof(tmp) - 1, f);
2e1df07d
JA
1433 if (ferror(f) || ret < 0) {
1434 td_verror(td, errno, "fread");
1435 fclose(f);
1436 return 1;
1437 }
0e23d456 1438 tmp[ret] = '\0';
b44b9e45 1439 /*
178ef27a
BVA
1440 * either a list of io schedulers or "none\n" is expected. Strip the
1441 * trailing newline.
b44b9e45 1442 */
178ef27a
BVA
1443 p = tmp;
1444 strsep(&p, "\n");
49c6f33d 1445
67f28a8c
TK
1446 /*
1447 * Write to "none" entry doesn't fail, so check the result here.
1448 */
1449 if (!strcmp(tmp, "none")) {
1450 log_err("fio: io scheduler is not tunable\n");
1451 fclose(f);
1452 return 0;
1453 }
2e1df07d
JA
1454
1455 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1456 if (!strstr(tmp, tmp2)) {
da44720c 1457 log_err("fio: unable to set io scheduler to %s\n", td->o.ioscheduler);
2e1df07d
JA
1458 td_verror(td, EINVAL, "iosched_switch");
1459 fclose(f);
1460 return 1;
1461 }
1462
1463 fclose(f);
1464 return 0;
5592e992
DLM
1465}
1466
1467static int switch_ioscheduler(struct thread_data *td)
1468{
1469 struct fio_file *f;
1470 unsigned int i;
1471 int ret = 0;
1472
1473 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
1474 return 0;
1475
1476 assert(td->files && td->files[0]);
1477
1478 for_each_file(td, f, i) {
1479
1480 /* Only consider regular files and block device files */
1481 switch (f->filetype) {
1482 case FIO_TYPE_FILE:
1483 case FIO_TYPE_BLOCK:
1484 /*
1485 * Make sure that the device hosting the file could
1486 * be determined.
1487 */
1488 if (!f->du)
1489 continue;
1490 break;
1491 case FIO_TYPE_CHAR:
1492 case FIO_TYPE_PIPE:
1493 default:
1494 continue;
1495 }
1496
1497 ret = set_ioscheduler(td, f);
1498 if (ret)
1499 return ret;
1500 }
1501
1502 return 0;
1503}
1504
db815fb3 1505#else
5592e992
DLM
1506
1507static int switch_ioscheduler(struct thread_data *td)
1508{
db815fb3 1509 return 0;
2e1df07d
JA
1510}
1511
5592e992
DLM
1512#endif /* FIO_HAVE_IOSCHED_SWITCH */
1513
e39c0676 1514static bool keep_running(struct thread_data *td)
2e1df07d 1515{
77731b29
JA
1516 unsigned long long limit;
1517
2e1df07d 1518 if (td->done)
e39c0676 1519 return false;
98413be7
SW
1520 if (td->terminate)
1521 return false;
2e1df07d 1522 if (td->o.time_based)
e39c0676 1523 return true;
2e1df07d
JA
1524 if (td->o.loops) {
1525 td->o.loops--;
e39c0676 1526 return true;
2e1df07d 1527 }
3939fe85 1528 if (exceeds_number_ios(td))
e39c0676 1529 return false;
26251d8d 1530
5be9bf09
TK
1531 if (td->o.io_size)
1532 limit = td->o.io_size;
77731b29
JA
1533 else
1534 limit = td->o.size;
1535
1536 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
5bd5f71a
JA
1537 uint64_t diff;
1538
1539 /*
589d5d4f 1540 * If the difference is less than the maximum IO size, we
5bd5f71a
JA
1541 * are done.
1542 */
77731b29 1543 diff = limit - ddir_rw_sum(td->io_bytes);
74f4b020 1544 if (diff < td_max_bs(td))
e39c0676 1545 return false;
5bd5f71a 1546
5be9bf09 1547 if (fio_files_done(td) && !td->o.io_size)
e39c0676 1548 return false;
002fe734 1549
e39c0676 1550 return true;
5bd5f71a 1551 }
2e1df07d 1552
e39c0676 1553 return false;
2e1df07d
JA
1554}
1555
b7e4a53c
BVA
1556static int exec_string(struct thread_options *o, const char *string,
1557 const char *mode)
2e1df07d 1558{
f491a907 1559 int ret;
2e1df07d
JA
1560 char *str;
1561
b7e4a53c
BVA
1562 if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0)
1563 return -1;
2e1df07d 1564
b7e4a53c
BVA
1565 log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode,
1566 o->name, mode);
2e1df07d
JA
1567 ret = system(str);
1568 if (ret == -1)
1569 log_err("fio: exec of cmd <%s> failed\n", str);
1570
1571 free(str);
1572 return ret;
1573}
1574
62167762
JC
1575/*
1576 * Dry run to compute correct state of numberio for verification.
1577 */
1578static uint64_t do_dry_run(struct thread_data *td)
1579{
62167762
JC
1580 td_set_runstate(td, TD_RUNNING);
1581
1582 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
e28dd2cf 1583 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
62167762
JC
1584 struct io_u *io_u;
1585 int ret;
1586
1587 if (td->terminate || td->done)
1588 break;
1589
1590 io_u = get_io_u(td);
99955d3d 1591 if (IS_ERR_OR_NULL(io_u))
62167762
JC
1592 break;
1593
1651e431 1594 io_u_set(td, io_u, IO_U_F_FLIGHT);
62167762
JC
1595 io_u->error = 0;
1596 io_u->resid = 0;
1597 if (ddir_rw(acct_ddir(io_u)))
1598 td->io_issues[acct_ddir(io_u)]++;
1599 if (ddir_rw(io_u->ddir)) {
1600 io_u_mark_depth(td, 1);
1601 td->ts.total_io_u[io_u->ddir]++;
1602 }
1603
2e63e96b
PV
1604 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1605 td->o.do_verify &&
1606 td->o.verify != VERIFY_NONE &&
1607 !td->o.experimental_verify)
1608 log_io_piece(td, io_u);
1609
55312f9f 1610 ret = io_u_sync_complete(td, io_u);
62167762
JC
1611 (void) ret;
1612 }
1613
55312f9f 1614 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
62167762
JC
1615}
1616
24660963
JA
1617struct fork_data {
1618 struct thread_data *td;
1619 struct sk_out *sk_out;
1620};
1621
2e1df07d
JA
1622/*
1623 * Entry point for the thread based jobs. The process based jobs end up
1624 * here as well, after a little setup.
1625 */
1626static void *thread_main(void *data)
1627{
24660963 1628 struct fork_data *fd = data;
95603b74 1629 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
24660963 1630 struct thread_data *td = fd->td;
4896473e 1631 struct thread_options *o = &td->o;
24660963 1632 struct sk_out *sk_out = fd->sk_out;
48366f3a 1633 uint64_t bytes_done[DDIR_RWDIR_CNT];
ca0122d8 1634 int deadlock_loop_cnt;
ff3aa922 1635 bool clear_state;
a38c70a8 1636 int res, ret;
2e1df07d 1637
24660963
JA
1638 sk_out_assign(sk_out);
1639 free(fd);
1640
4896473e 1641 if (!o->use_thread) {
2e1df07d
JA
1642 setsid();
1643 td->pid = getpid();
1644 } else
1645 td->pid = gettid();
1646
a0eba820 1647 fio_local_clock_init();
5d879392 1648
2e1df07d
JA
1649 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1650
122c7725
JA
1651 if (is_backend)
1652 fio_server_send_start(td);
1653
2e1df07d
JA
1654 INIT_FLIST_HEAD(&td->io_log_list);
1655 INIT_FLIST_HEAD(&td->io_hist_list);
1656 INIT_FLIST_HEAD(&td->verify_list);
1657 INIT_FLIST_HEAD(&td->trim_list);
2e1df07d
JA
1658 td->io_hist_tree = RB_ROOT;
1659
34febb23 1660 ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
f9e5b5ee 1661 if (ret) {
34febb23 1662 td_verror(td, ret, "mutex_cond_init_pshared");
f9e5b5ee
JK
1663 goto err;
1664 }
34febb23 1665 ret = cond_init_pshared(&td->verify_cond);
f9e5b5ee 1666 if (ret) {
34febb23 1667 td_verror(td, ret, "mutex_cond_pshared");
f9e5b5ee
JK
1668 goto err;
1669 }
2e1df07d
JA
1670
1671 td_set_runstate(td, TD_INITIALIZED);
971caeb1
BVA
1672 dprint(FD_MUTEX, "up startup_sem\n");
1673 fio_sem_up(startup_sem);
1674 dprint(FD_MUTEX, "wait on td->sem\n");
1675 fio_sem_down(td->sem);
1676 dprint(FD_MUTEX, "done waiting on td->sem\n");
2e1df07d 1677
2e1df07d
JA
1678 /*
1679 * A new gid requires privilege, so we need to do this before setting
1680 * the uid.
1681 */
4896473e 1682 if (o->gid != -1U && setgid(o->gid)) {
2e1df07d
JA
1683 td_verror(td, errno, "setgid");
1684 goto err;
1685 }
4896473e 1686 if (o->uid != -1U && setuid(o->uid)) {
2e1df07d
JA
1687 td_verror(td, errno, "setuid");
1688 goto err;
1689 }
1690
07776127
JA
1691 td_zone_gen_index(td);
1692
2354d810
JA
1693 /*
1694 * Do this early, we don't want the compress threads to be limited
1695 * to the same CPUs as the IO workers. So do this before we set
1696 * any potential CPU affinity
1697 */
1698 if (iolog_compress_init(td, sk_out))
1699 goto err;
1700
2e1df07d
JA
1701 /*
1702 * If we have a gettimeofday() thread, make sure we exclude that
1703 * thread from this job
1704 */
4896473e
JA
1705 if (o->gtod_cpu)
1706 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
2e1df07d
JA
1707
1708 /*
1709 * Set affinity first, in case it has an impact on the memory
1710 * allocations.
1711 */
b2a9e649 1712 if (fio_option_is_set(o, cpumask)) {
c2acfbac 1713 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
30cb4c65 1714 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
c2acfbac
JA
1715 if (!ret) {
1716 log_err("fio: no CPUs set\n");
1717 log_err("fio: Try increasing number of available CPUs\n");
1718 td_verror(td, EINVAL, "cpus_split");
1719 goto err;
1720 }
1721 }
28727df7
JA
1722 ret = fio_setaffinity(td->pid, o->cpumask);
1723 if (ret == -1) {
4896473e
JA
1724 td_verror(td, errno, "cpu_set_affinity");
1725 goto err;
1726 }
2e1df07d
JA
1727 }
1728
67bf9823 1729#ifdef CONFIG_LIBNUMA
d0b937ed 1730 /* numa node setup */
b2a9e649
JA
1731 if (fio_option_is_set(o, numa_cpunodes) ||
1732 fio_option_is_set(o, numa_memnodes)) {
43522848 1733 struct bitmask *mask;
d0b937ed
YR
1734
1735 if (numa_available() < 0) {
1736 td_verror(td, errno, "Does not support NUMA API\n");
1737 goto err;
1738 }
1739
b2a9e649 1740 if (fio_option_is_set(o, numa_cpunodes)) {
43522848
DG
1741 mask = numa_parse_nodestring(o->numa_cpunodes);
1742 ret = numa_run_on_node_mask(mask);
1743 numa_free_nodemask(mask);
d0b937ed
YR
1744 if (ret == -1) {
1745 td_verror(td, errno, \
1746 "numa_run_on_node_mask failed\n");
1747 goto err;
1748 }
1749 }
1750
b2a9e649 1751 if (fio_option_is_set(o, numa_memnodes)) {
43522848
DG
1752 mask = NULL;
1753 if (o->numa_memnodes)
1754 mask = numa_parse_nodestring(o->numa_memnodes);
1755
4896473e 1756 switch (o->numa_mem_mode) {
d0b937ed 1757 case MPOL_INTERLEAVE:
43522848 1758 numa_set_interleave_mask(mask);
d0b937ed
YR
1759 break;
1760 case MPOL_BIND:
43522848 1761 numa_set_membind(mask);
d0b937ed
YR
1762 break;
1763 case MPOL_LOCAL:
1764 numa_set_localalloc();
1765 break;
1766 case MPOL_PREFERRED:
4896473e 1767 numa_set_preferred(o->numa_mem_prefer_node);
d0b937ed
YR
1768 break;
1769 case MPOL_DEFAULT:
1770 default:
1771 break;
1772 }
1773
43522848
DG
1774 if (mask)
1775 numa_free_nodemask(mask);
1776
d0b937ed
YR
1777 }
1778 }
1779#endif
1780
9a3f1100
JA
1781 if (fio_pin_memory(td))
1782 goto err;
1783
2e1df07d
JA
1784 /*
1785 * May alter parameters that init_io_u() will use, so we need to
1786 * do this first.
1787 */
b153f94a 1788 if (!init_iolog(td))
2e1df07d
JA
1789 goto err;
1790
61b20c5c
NC
1791 /* ioprio_set() has to be done before td_io_init() */
1792 if (fio_option_is_set(o, ioprio) ||
1793 fio_option_is_set(o, ioprio_class)) {
1794 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1795 if (ret == -1) {
1796 td_verror(td, errno, "ioprio_set");
1797 goto err;
1798 }
1799 td->ioprio = ioprio_value(o->ioprio_class, o->ioprio);
2df7f081 1800 td->ts.ioprio = td->ioprio;
61b20c5c
NC
1801 }
1802
08dc3bd5
JA
1803 if (td_io_init(td))
1804 goto err;
1805
b8d07a36 1806 if (td_ioengine_flagged(td, FIO_SYNCIO) && td->o.iodepth > 1 && td->o.io_submit_mode != IO_MODE_OFFLOAD) {
349d5f6c
AD
1807 log_info("note: both iodepth >= 1 and synchronous I/O engine "
1808 "are selected, queue depth will be capped at 1\n");
1809 }
1810
2e1df07d
JA
1811 if (init_io_u(td))
1812 goto err;
1813
2041bd34
JA
1814 if (td->io_ops->post_init && td->io_ops->post_init(td))
1815 goto err;
1816
4896473e 1817 if (o->verify_async && verify_async_init(td))
2e1df07d
JA
1818 goto err;
1819
4896473e 1820 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
2e1df07d
JA
1821 goto err;
1822
649c10c5 1823 errno = 0;
4896473e 1824 if (nice(o->nice) == -1 && errno != 0) {
2e1df07d
JA
1825 td_verror(td, errno, "nice");
1826 goto err;
1827 }
1828
4896473e 1829 if (o->ioscheduler && switch_ioscheduler(td))
2e1df07d
JA
1830 goto err;
1831
4896473e 1832 if (!o->create_serialize && setup_files(td))
2e1df07d
JA
1833 goto err;
1834
4d832322 1835 if (!init_random_map(td))
2e1df07d
JA
1836 goto err;
1837
0518f09f 1838 if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun"))
4896473e 1839 goto err;
2e1df07d 1840
c139f3b4
JA
1841 if (o->pre_read && !pre_read_files(td))
1842 goto err;
2e1df07d 1843
dc5bfbb2
JA
1844 fio_verify_init(td);
1845
24660963 1846 if (rate_submit_init(td, sk_out))
a9da8ab2
JA
1847 goto err;
1848
d5b3cfd4 1849 set_epoch_time(td, o->log_unix_epoch | o->log_alternate_epoch, o->log_alternate_epoch_clock_id);
44404c5a 1850 fio_getrusage(&td->ru_start);
ac28d905
JA
1851 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1852 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
16e56d25 1853 memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
ac28d905 1854
f7942acd
SK
1855 init_thinktime(td);
1856
ac28d905
JA
1857 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1858 o->ratemin[DDIR_TRIM]) {
f5204b8a 1859 memcpy(&td->last_rate_check_time[DDIR_READ], &td->bw_sample_time,
ac28d905 1860 sizeof(td->bw_sample_time));
f5204b8a 1861 memcpy(&td->last_rate_check_time[DDIR_WRITE], &td->bw_sample_time,
ac28d905 1862 sizeof(td->bw_sample_time));
f5204b8a 1863 memcpy(&td->last_rate_check_time[DDIR_TRIM], &td->bw_sample_time,
ac28d905
JA
1864 sizeof(td->bw_sample_time));
1865 }
1866
48366f3a 1867 memset(bytes_done, 0, sizeof(bytes_done));
c16035aa 1868 clear_state = false;
48366f3a 1869
2e1df07d 1870 while (keep_running(td)) {
100f49f1
JA
1871 uint64_t verify_bytes;
1872
2e1df07d 1873 fio_gettime(&td->start, NULL);
8b6a404c 1874 memcpy(&td->ts_cache, &td->start, sizeof(td->start));
2e1df07d 1875
39c1c323 1876 if (clear_state) {
ac28d905 1877 clear_io_state(td, 0);
2e1df07d 1878
39c1c323 1879 if (o->unlink_each_loop && unlink_all_files(td))
1880 break;
1881 }
1882
2e1df07d
JA
1883 prune_io_piece_log(td);
1884
dcf9844e 1885 if (td->o.verify_only && td_write(td))
62167762 1886 verify_bytes = do_dry_run(td);
095196b1 1887 else {
50eaefa5
H
1888 if (!td->o.rand_repeatable)
1889 /* save verify rand state to replay hdr seeds later at verify */
1890 frand_copy(&td->verify_state_last_do_io, &td->verify_state);
95b03be0 1891 do_io(td, bytes_done);
50eaefa5
H
1892 if (!td->o.rand_repeatable)
1893 frand_copy(&td->verify_state, &td->verify_state_last_do_io);
95b03be0 1894 if (!ddir_rw_sum(bytes_done)) {
095196b1 1895 fio_mark_td_terminate(td);
95b03be0
JA
1896 verify_bytes = 0;
1897 } else {
1898 verify_bytes = bytes_done[DDIR_WRITE] +
1899 bytes_done[DDIR_TRIM];
1900 }
095196b1 1901 }
2e1df07d 1902
49f0b8e0
JA
1903 /*
1904 * If we took too long to shut down, the main thread could
1905 * already consider us reaped/exited. If that happens, break
1906 * out and clean up.
1907 */
1908 if (td->runstate >= TD_EXITED)
1909 break;
1910
c16035aa 1911 clear_state = true;
2e1df07d 1912
40c666c8
JA
1913 /*
1914 * Make sure we've successfully updated the rusage stats
1915 * before waiting on the stat mutex. Otherwise we could have
1916 * the stat thread holding stat mutex and waiting for
1917 * the rusage_sem, which would never get upped because
1918 * this thread is waiting for the stat mutex.
1919 */
ca0122d8 1920 deadlock_loop_cnt = 0;
39d13e67
JA
1921 do {
1922 check_update_rusage(td);
971caeb1 1923 if (!fio_sem_down_trylock(stat_sem))
39d13e67
JA
1924 break;
1925 usleep(1000);
ca0122d8 1926 if (deadlock_loop_cnt++ > 5000) {
971caeb1 1927 log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
43f248c4 1928 td->error = EDEADLK;
ca0122d8
TT
1929 goto err;
1930 }
39d13e67 1931 } while (1);
40c666c8 1932
615c794c
H
1933 if (td->io_bytes[DDIR_READ] && (td_read(td) ||
1934 ((td->flags & TD_F_VER_BACKLOG) && td_write(td))))
95603b74
BF
1935 update_runtime(td, elapsed_us, DDIR_READ);
1936 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1937 update_runtime(td, elapsed_us, DDIR_WRITE);
1938 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1939 update_runtime(td, elapsed_us, DDIR_TRIM);
e5437a07 1940 fio_gettime(&td->start, NULL);
971caeb1 1941 fio_sem_up(stat_sem);
2e1df07d
JA
1942
1943 if (td->error || td->terminate)
1944 break;
1945
4896473e
JA
1946 if (!o->do_verify ||
1947 o->verify == VERIFY_NONE ||
9b87f09b 1948 td_ioengine_flagged(td, FIO_UNIDIR))
2e1df07d
JA
1949 continue;
1950
2d44f2c6
SK
1951 clear_io_state(td, 0);
1952
2e1df07d
JA
1953 fio_gettime(&td->start, NULL);
1954
100f49f1 1955 do_verify(td, verify_bytes);
2e1df07d 1956
40c666c8
JA
1957 /*
1958 * See comment further up for why this is done here.
1959 */
1960 check_update_rusage(td);
1961
971caeb1 1962 fio_sem_down(stat_sem);
95603b74 1963 update_runtime(td, elapsed_us, DDIR_READ);
e5437a07 1964 fio_gettime(&td->start, NULL);
971caeb1 1965 fio_sem_up(stat_sem);
2e1df07d
JA
1966
1967 if (td->error || td->terminate)
1968 break;
1969 }
1970
817ae977
VF
1971 /*
1972 * Acquire this lock if we were doing overlap checking in
1973 * offload mode so that we don't clean up this job while
1974 * another thread is checking its io_u's for overlap
1975 */
a38c70a8
BVA
1976 if (td_offload_overlap(td)) {
1977 int res = pthread_mutex_lock(&overlap_check);
1978 assert(res == 0);
1979 }
2cea0b4c 1980 td_set_runstate(td, TD_FINISHING);
a38c70a8
BVA
1981 if (td_offload_overlap(td)) {
1982 res = pthread_mutex_unlock(&overlap_check);
1983 assert(res == 0);
1984 }
2cea0b4c 1985
2e1df07d 1986 update_rusage_stat(td);
2e1df07d 1987 td->ts.total_run_time = mtime_since_now(&td->epoch);
dde39f8c
AD
1988 for_each_rw_ddir(ddir) {
1989 td->ts.io_bytes[ddir] = td->io_bytes[ddir];
1990 }
2e1df07d 1991
ca09be4b 1992 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
d264264a
JA
1993 (td->o.verify != VERIFY_NONE && td_write(td)))
1994 verify_save_state(td->thread_number);
ca09be4b 1995
9a3f1100
JA
1996 fio_unpin_memory(td);
1997
a47591e4 1998 td_writeout_logs(td, true);
2e1df07d 1999
155f2f02 2000 iolog_compress_exit(td);
103b174e 2001 rate_submit_exit(td);
aee2ab67 2002
4896473e 2003 if (o->exec_postrun)
0518f09f 2004 exec_string(o, o->exec_postrun, "postrun");
2e1df07d 2005
f9cafb12 2006 if (exitall_on_terminate || (o->exitall_error && td->error))
64402a8a 2007 fio_terminate_threads(td->groupid, td->o.exit_what);
2e1df07d
JA
2008
2009err:
2010 if (td->error)
2011 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
2012 td->verror);
2013
4896473e 2014 if (o->verify_async)
2e1df07d
JA
2015 verify_async_exit(td);
2016
2017 close_and_free_files(td);
2e1df07d 2018 cleanup_io_u(td);
32dbca2c 2019 close_ioengine(td);
c3dc516a 2020 cgroup_shutdown(td, cgroup_mnt);
ca09be4b 2021 verify_free_state(td);
07776127 2022 td_zone_free_index(td);
7c4d0fb7 2023
b2a9e649 2024 if (fio_option_is_set(o, cpumask)) {
8a1db9a1
JA
2025 ret = fio_cpuset_exit(&o->cpumask);
2026 if (ret)
2027 td_verror(td, ret, "fio_cpuset_exit");
2e1df07d
JA
2028 }
2029
2030 /*
2031 * do this very late, it will log file closing as well
2032 */
4896473e 2033 if (o->write_iolog_file)
2e1df07d 2034 write_iolog_close(td);
98e7161c
AK
2035 if (td->io_log_rfile)
2036 fclose(td->io_log_rfile);
2e1df07d
JA
2037
2038 td_set_runstate(td, TD_EXITED);
fda2cfac
JA
2039
2040 /*
2041 * Do this last after setting our runstate to exited, so we
2042 * know that the stat thread is signaled.
2043 */
2044 check_update_rusage(td);
2045
24660963 2046 sk_out_drop();
e43606c2 2047 return (void *) (uintptr_t) td->error;
2e1df07d
JA
2048}
2049
2e1df07d
JA
2050/*
2051 * Run over the job map and reap the threads that have exited, if any.
2052 */
90eff1c9
SW
2053static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
2054 uint64_t *m_rate)
2e1df07d 2055{
2e1df07d 2056 unsigned int cputhreads, realthreads, pending;
da8f124f 2057 int status, ret;
2e1df07d
JA
2058
2059 /*
2060 * reap exited threads (TD_EXITED -> TD_REAPED)
2061 */
2062 realthreads = pending = cputhreads = 0;
da8f124f 2063 for_each_td(td) {
2e1df07d
JA
2064 int flags = 0;
2065
74ee1904 2066 if (!strcmp(td->o.ioengine, "cpuio"))
2e1df07d
JA
2067 cputhreads++;
2068 else
2069 realthreads++;
2070
2071 if (!td->pid) {
2072 pending++;
2073 continue;
2074 }
2075 if (td->runstate == TD_REAPED)
2076 continue;
2077 if (td->o.use_thread) {
2078 if (td->runstate == TD_EXITED) {
2079 td_set_runstate(td, TD_REAPED);
2080 goto reaped;
2081 }
2082 continue;
2083 }
2084
2085 flags = WNOHANG;
2086 if (td->runstate == TD_EXITED)
2087 flags = 0;
2088
2089 /*
2090 * check if someone quit or got killed in an unusual way
2091 */
2092 ret = waitpid(td->pid, &status, flags);
2093 if (ret < 0) {
2094 if (errno == ECHILD) {
2095 log_err("fio: pid=%d disappeared %d\n",
2096 (int) td->pid, td->runstate);
a5e371a6 2097 td->sig = ECHILD;
2e1df07d
JA
2098 td_set_runstate(td, TD_REAPED);
2099 goto reaped;
2100 }
2101 perror("waitpid");
2102 } else if (ret == td->pid) {
2103 if (WIFSIGNALED(status)) {
2104 int sig = WTERMSIG(status);
2105
36d80bc7 2106 if (sig != SIGTERM && sig != SIGUSR2)
2e1df07d
JA
2107 log_err("fio: pid=%d, got signal=%d\n",
2108 (int) td->pid, sig);
a5e371a6 2109 td->sig = sig;
2e1df07d
JA
2110 td_set_runstate(td, TD_REAPED);
2111 goto reaped;
2112 }
2113 if (WIFEXITED(status)) {
2114 if (WEXITSTATUS(status) && !td->error)
2115 td->error = WEXITSTATUS(status);
2116
2117 td_set_runstate(td, TD_REAPED);
2118 goto reaped;
2119 }
2120 }
2121
cba5460c
JA
2122 /*
2123 * If the job is stuck, do a forceful timeout of it and
2124 * move on.
2125 */
2126 if (td->terminate &&
2c45a4ac 2127 td->runstate < TD_FSYNCING &&
cba5460c 2128 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
86310a1f
TK
2129 log_err("fio: job '%s' (state=%d) hasn't exited in "
2130 "%lu seconds, it appears to be stuck. Doing "
2131 "forceful exit of this job.\n",
2132 td->o.name, td->runstate,
2133 (unsigned long) time_since_now(&td->terminate_time));
cba5460c
JA
2134 td_set_runstate(td, TD_REAPED);
2135 goto reaped;
2136 }
2137
2e1df07d
JA
2138 /*
2139 * thread is not dead, continue
2140 */
2141 pending++;
2142 continue;
2143reaped:
2144 (*nr_running)--;
342f4be4
JA
2145 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
2146 (*t_rate) -= ddir_rw_sum(td->o.rate);
2e1df07d
JA
2147 if (!td->pid)
2148 pending--;
2149
2150 if (td->error)
2151 exit_value++;
2152
2153 done_secs += mtime_since_now(&td->epoch) / 1000;
4a88752a 2154 profile_td_exit(td);
42e26e94 2155 flow_exit_job(td);
da8f124f 2156 } end_for_each();
2e1df07d
JA
2157
2158 if (*nr_running == cputhreads && !pending && realthreads)
64402a8a 2159 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
2e1df07d
JA
2160}
2161
e39c0676 2162static bool __check_trigger_file(void)
ca09be4b
JA
2163{
2164 struct stat sb;
2165
2166 if (!trigger_file)
e39c0676 2167 return false;
ca09be4b
JA
2168
2169 if (stat(trigger_file, &sb))
e39c0676 2170 return false;
ca09be4b
JA
2171
2172 if (unlink(trigger_file) < 0)
2173 log_err("fio: failed to unlink %s: %s\n", trigger_file,
2174 strerror(errno));
2175
e39c0676 2176 return true;
ca09be4b
JA
2177}
2178
e39c0676 2179static bool trigger_timedout(void)
ca09be4b
JA
2180{
2181 if (trigger_timeout)
98413be7
SW
2182 if (time_since_genesis() >= trigger_timeout) {
2183 trigger_timeout = 0;
2184 return true;
2185 }
ca09be4b 2186
e39c0676 2187 return false;
ca09be4b
JA
2188}
2189
2190void exec_trigger(const char *cmd)
2191{
2192 int ret;
2193
98413be7 2194 if (!cmd || cmd[0] == '\0')
ca09be4b
JA
2195 return;
2196
2197 ret = system(cmd);
2198 if (ret == -1)
2199 log_err("fio: failed executing %s trigger\n", cmd);
2200}
2201
2202void check_trigger_file(void)
2203{
2204 if (__check_trigger_file() || trigger_timedout()) {
796fb3ce
JA
2205 if (nr_clients)
2206 fio_clients_send_trigger(trigger_remote_cmd);
2207 else {
d264264a 2208 verify_save_state(IO_LIST_ALL);
64402a8a 2209 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
ca09be4b
JA
2210 exec_trigger(trigger_cmd);
2211 }
2212 }
2213}
2214
2215static int fio_verify_load_state(struct thread_data *td)
2216{
2217 int ret;
2218
2219 if (!td->o.verify_state)
2220 return 0;
2221
2222 if (is_backend) {
2223 void *data;
2224
2225 ret = fio_server_get_verify_state(td->o.name,
94a6e1bb 2226 td->thread_number - 1, &data);
ca09be4b 2227 if (!ret)
94a6e1bb 2228 verify_assign_state(td, data);
67710881
AK
2229 } else {
2230 char prefix[PATH_MAX];
2231
2232 if (aux_path)
2233 sprintf(prefix, "%s%clocal", aux_path,
2234 FIO_OS_PATH_SEPARATOR);
2235 else
2236 strcpy(prefix, "local");
2237 ret = verify_load_state(td, prefix);
2238 }
ca09be4b
JA
2239
2240 return ret;
2241}
2242
06464907
JA
2243static void do_usleep(unsigned int usecs)
2244{
2245 check_for_running_stats();
ca09be4b 2246 check_trigger_file();
06464907
JA
2247 usleep(usecs);
2248}
2249
e39c0676 2250static bool check_mount_writes(struct thread_data *td)
e81ecca3
JA
2251{
2252 struct fio_file *f;
2253 unsigned int i;
2254
2255 if (!td_write(td) || td->o.allow_mounted_write)
e39c0676 2256 return false;
e81ecca3 2257
5e81f9c1
TK
2258 /*
2259 * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
2260 * are mkfs'd and mounted.
2261 */
e81ecca3 2262 for_each_file(td, f, i) {
5e81f9c1
TK
2263#ifdef FIO_HAVE_CHARDEV_SIZE
2264 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
2265#else
686fbd31 2266 if (f->filetype != FIO_TYPE_BLOCK)
5e81f9c1 2267#endif
e81ecca3
JA
2268 continue;
2269 if (device_is_mounted(f->file_name))
2270 goto mounted;
2271 }
2272
e39c0676 2273 return false;
e81ecca3 2274mounted:
c5839011 2275 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
e39c0676 2276 return true;
e81ecca3
JA
2277}
2278
9cc8cb91
AK
2279static bool waitee_running(struct thread_data *me)
2280{
2281 const char *waitee = me->o.wait_for;
2282 const char *self = me->o.name;
9cc8cb91
AK
2283
2284 if (!waitee)
2285 return false;
2286
da8f124f 2287 for_each_td(td) {
9cc8cb91
AK
2288 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2289 continue;
2290
2291 if (td->runstate < TD_EXITED) {
2292 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2293 self, td->o.name,
2294 runstate_to_name(td->runstate));
2295 return true;
2296 }
da8f124f 2297 } end_for_each();
9cc8cb91
AK
2298
2299 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2300 return false;
2301}
2302
2e1df07d
JA
2303/*
2304 * Main function for kicking off and reaping jobs, as needed.
2305 */
24660963 2306static void run_threads(struct sk_out *sk_out)
2e1df07d
JA
2307{
2308 struct thread_data *td;
90eff1c9
SW
2309 unsigned int i, todo, nr_running, nr_started;
2310 uint64_t m_rate, t_rate;
0de5b26f 2311 uint64_t spent;
2e1df07d 2312
2e1df07d
JA
2313 if (fio_gtod_offload && fio_start_gtod_thread())
2314 return;
334185e9 2315
f2a2ce0e 2316 fio_idle_prof_init();
2e1df07d
JA
2317
2318 set_sig_handlers();
2319
3a5f6bde 2320 nr_thread = nr_process = 0;
da8f124f 2321 for_each_td(td) {
e81ecca3
JA
2322 if (check_mount_writes(td))
2323 return;
3a5f6bde
JA
2324 if (td->o.use_thread)
2325 nr_thread++;
2326 else
2327 nr_process++;
da8f124f 2328 } end_for_each();
3a5f6bde 2329
01cfefcc 2330 if (output_format & FIO_OUTPUT_NORMAL) {
a41fc529
JA
2331 struct buf_output out;
2332
2333 buf_output_init(&out);
c2fcb6bc 2334 __log_buf(&out, "Starting ");
2e1df07d 2335 if (nr_thread)
c2fcb6bc 2336 __log_buf(&out, "%d thread%s", nr_thread,
2e1df07d
JA
2337 nr_thread > 1 ? "s" : "");
2338 if (nr_process) {
2339 if (nr_thread)
c2fcb6bc
JA
2340 __log_buf(&out, " and ");
2341 __log_buf(&out, "%d process%s", nr_process,
2e1df07d
JA
2342 nr_process > 1 ? "es" : "");
2343 }
c2fcb6bc 2344 __log_buf(&out, "\n");
a41fc529 2345 log_info_buf(out.buf, out.buflen);
d5bb0a89 2346 buf_output_free(&out);
2e1df07d
JA
2347 }
2348
2349 todo = thread_number;
2350 nr_running = 0;
2351 nr_started = 0;
2352 m_rate = t_rate = 0;
2353
da8f124f 2354 for_each_td(td) {
2e1df07d
JA
2355 print_status_init(td->thread_number - 1);
2356
d5e3f7cb 2357 if (!td->o.create_serialize)
2e1df07d
JA
2358 continue;
2359
ca09be4b
JA
2360 if (fio_verify_load_state(td))
2361 goto reap;
2362
2e1df07d
JA
2363 /*
2364 * do file setup here so it happens sequentially,
2365 * we don't want X number of threads getting their
2366 * client data interspersed on disk
2367 */
2368 if (setup_files(td)) {
ca09be4b 2369reap:
2e1df07d
JA
2370 exit_value++;
2371 if (td->error)
2372 log_err("fio: pid=%d, err=%d/%s\n",
2373 (int) td->pid, td->error, td->verror);
2374 td_set_runstate(td, TD_REAPED);
2375 todo--;
2376 } else {
2377 struct fio_file *f;
2378 unsigned int j;
2379
2380 /*
2381 * for sharing to work, each job must always open
2382 * its own files. so close them, if we opened them
2383 * for creation
2384 */
2385 for_each_file(td, f, j) {
2386 if (fio_file_open(f))
2387 td_io_close_file(td, f);
2388 }
2389 }
da8f124f 2390 } end_for_each();
2e1df07d 2391
f2a2ce0e
HL
2392 /* start idle threads before io threads start to run */
2393 fio_idle_prof_start();
2394
2e1df07d
JA
2395 set_genesis_time();
2396
2397 while (todo) {
2398 struct thread_data *map[REAL_MAX_JOBS];
8b6a404c 2399 struct timespec this_start;
2e1df07d 2400 int this_jobs = 0, left;
ce8b6139 2401 struct fork_data *fd;
2e1df07d
JA
2402
2403 /*
2404 * create threads (TD_NOT_CREATED -> TD_CREATED)
2405 */
da8f124f 2406 for_each_td(td) {
2e1df07d
JA
2407 if (td->runstate != TD_NOT_CREATED)
2408 continue;
2409
2410 /*
2411 * never got a chance to start, killed by other
2412 * thread for some reason
2413 */
2414 if (td->terminate) {
2415 todo--;
2416 continue;
2417 }
2418
2419 if (td->o.start_delay) {
0de5b26f 2420 spent = utime_since_genesis();
2e1df07d 2421
74454ce4 2422 if (td->o.start_delay > spent)
2e1df07d
JA
2423 continue;
2424 }
2425
2426 if (td->o.stonewall && (nr_started || nr_running)) {
2427 dprint(FD_PROCESS, "%s: stonewall wait\n",
2428 td->o.name);
2429 break;
2430 }
2431
9cc8cb91
AK
2432 if (waitee_running(td)) {
2433 dprint(FD_PROCESS, "%s: waiting for %s\n",
2434 td->o.name, td->o.wait_for);
5c74fc76 2435 continue;
9cc8cb91
AK
2436 }
2437
2e1df07d
JA
2438 init_disk_util(td);
2439
971caeb1 2440 td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
c97f1ad6
JA
2441 td->update_rusage = 0;
2442
2e1df07d
JA
2443 /*
2444 * Set state to created. Thread will transition
2445 * to TD_INITIALIZED when it's done setting up.
2446 */
2447 td_set_runstate(td, TD_CREATED);
2448 map[this_jobs++] = td;
2449 nr_started++;
2450
ce8b6139
JA
2451 fd = calloc(1, sizeof(*fd));
2452 fd->td = td;
2453 fd->sk_out = sk_out;
2454
2e1df07d
JA
2455 if (td->o.use_thread) {
2456 int ret;
2457
2458 dprint(FD_PROCESS, "will pthread_create\n");
2459 ret = pthread_create(&td->thread, NULL,
24660963 2460 thread_main, fd);
2e1df07d
JA
2461 if (ret) {
2462 log_err("pthread_create: %s\n",
2463 strerror(ret));
24660963 2464 free(fd);
2e1df07d
JA
2465 nr_started--;
2466 break;
2467 }
493dd6c1 2468 fd = NULL;
2e1df07d
JA
2469 ret = pthread_detach(td->thread);
2470 if (ret)
2471 log_err("pthread_detach: %s",
2472 strerror(ret));
2473 } else {
2474 pid_t pid;
6203763e 2475 void *eo;
2e1df07d 2476 dprint(FD_PROCESS, "will fork\n");
6203763e 2477 eo = td->eo;
807473c3 2478 read_barrier();
2e1df07d
JA
2479 pid = fork();
2480 if (!pid) {
ce8b6139 2481 int ret;
2e1df07d 2482
ce8b6139 2483 ret = (int)(uintptr_t)thread_main(fd);
2e1df07d 2484 _exit(ret);
da8f124f 2485 } else if (__td_index == fio_debug_jobno)
2e1df07d 2486 *fio_debug_jobp = pid;
6203763e 2487 free(eo);
36aac34e
DP
2488 free(fd);
2489 fd = NULL;
2e1df07d 2490 }
971caeb1
BVA
2491 dprint(FD_MUTEX, "wait on startup_sem\n");
2492 if (fio_sem_down_timeout(startup_sem, 10000)) {
2e1df07d 2493 log_err("fio: job startup hung? exiting.\n");
64402a8a 2494 fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL);
f1867a7f 2495 fio_abort = true;
2e1df07d 2496 nr_started--;
6541f4e0 2497 free(fd);
2e1df07d
JA
2498 break;
2499 }
971caeb1 2500 dprint(FD_MUTEX, "done waiting on startup_sem\n");
da8f124f 2501 } end_for_each();
2e1df07d
JA
2502
2503 /*
2504 * Wait for the started threads to transition to
2505 * TD_INITIALIZED.
2506 */
2507 fio_gettime(&this_start, NULL);
2508 left = this_jobs;
2509 while (left && !fio_abort) {
2510 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2511 break;
2512
06464907 2513 do_usleep(100000);
2e1df07d
JA
2514
2515 for (i = 0; i < this_jobs; i++) {
2516 td = map[i];
2517 if (!td)
2518 continue;
2519 if (td->runstate == TD_INITIALIZED) {
2520 map[i] = NULL;
2521 left--;
2522 } else if (td->runstate >= TD_EXITED) {
2523 map[i] = NULL;
2524 left--;
2525 todo--;
2526 nr_running++; /* work-around... */
2527 }
2528 }
2529 }
2530
2531 if (left) {
4e87c37a
JA
2532 log_err("fio: %d job%s failed to start\n", left,
2533 left > 1 ? "s" : "");
2e1df07d
JA
2534 for (i = 0; i < this_jobs; i++) {
2535 td = map[i];
2536 if (!td)
2537 continue;
2538 kill(td->pid, SIGTERM);
2539 }
2540 break;
2541 }
2542
2543 /*
2544 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2545 */
da8f124f 2546 for_each_td(td) {
2e1df07d
JA
2547 if (td->runstate != TD_INITIALIZED)
2548 continue;
2549
2550 if (in_ramp_time(td))
2551 td_set_runstate(td, TD_RAMP);
2552 else
2553 td_set_runstate(td, TD_RUNNING);
2554 nr_running++;
2555 nr_started--;
342f4be4
JA
2556 m_rate += ddir_rw_sum(td->o.ratemin);
2557 t_rate += ddir_rw_sum(td->o.rate);
2e1df07d 2558 todo--;
971caeb1 2559 fio_sem_up(td->sem);
da8f124f 2560 } end_for_each();
2e1df07d
JA
2561
2562 reap_threads(&nr_running, &t_rate, &m_rate);
2563
122c7725 2564 if (todo)
06464907 2565 do_usleep(100000);
2e1df07d
JA
2566 }
2567
2568 while (nr_running) {
2569 reap_threads(&nr_running, &t_rate, &m_rate);
06464907 2570 do_usleep(10000);
2e1df07d
JA
2571 }
2572
f2a2ce0e
HL
2573 fio_idle_prof_stop();
2574
2e1df07d 2575 update_io_ticks();
2e1df07d
JA
2576}
2577
27357187
JA
2578static void free_disk_util(void)
2579{
27357187 2580 disk_util_prune_entries();
a39fb9ea 2581 helper_thread_destroy();
2e1df07d
JA
2582}
2583
24660963 2584int fio_backend(struct sk_out *sk_out)
2e1df07d 2585{
2e1df07d 2586 int i;
2e1df07d
JA
2587 if (exec_profile) {
2588 if (load_profile(exec_profile))
2589 return 1;
2590 free(exec_profile);
2591 exec_profile = NULL;
2592 }
2593 if (!thread_number)
2594 return 0;
2595
2596 if (write_bw_log) {
aee2ab67
JA
2597 struct log_params p = {
2598 .log_type = IO_LOG_TYPE_BW,
2599 };
2600
2601 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2602 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2603 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2e1df07d
JA
2604 }
2605
c49cfc76
BD
2606 if (init_global_dedupe_working_set_seeds()) {
2607 log_err("fio: failed to initialize global dedupe working set\n");
2608 return 1;
2609 }
2610
971caeb1 2611 startup_sem = fio_sem_init(FIO_SEM_LOCKED);
24a97130
LC
2612 if (!sk_out)
2613 is_local_backend = true;
971caeb1 2614 if (startup_sem == NULL)
2e1df07d 2615 return 1;
2e1df07d
JA
2616
2617 set_genesis_time();
cef9175e 2618 stat_init();
998e9ebb
BVA
2619 if (helper_thread_create(startup_sem, sk_out))
2620 log_err("fio: failed to create helper thread\n");
2e1df07d
JA
2621
2622 cgroup_list = smalloc(sizeof(*cgroup_list));
b2e6494c
JA
2623 if (cgroup_list)
2624 INIT_FLIST_HEAD(cgroup_list);
2e1df07d 2625
24660963 2626 run_threads(sk_out);
2e1df07d 2627
a39fb9ea 2628 helper_thread_exit();
8aab824f 2629
2e1df07d 2630 if (!fio_abort) {
83f7b64e 2631 __show_run_stats();
2e1df07d 2632 if (write_bw_log) {
cb7e0ace
JA
2633 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2634 struct io_log *log = agg_io_log[i];
2635
60a25727 2636 flush_log(log, false);
518dac09 2637 free_log(log);
cb7e0ace 2638 }
2e1df07d
JA
2639 }
2640 }
2641
da8f124f 2642 for_each_td(td) {
f0547200
NC
2643 struct thread_stat *ts = &td->ts;
2644
2645 free_clat_prio_stats(ts);
c27cc65f 2646 steadystate_free(td);
2e1df07d 2647 fio_options_free(td);
8009f543 2648 fio_dump_options_free(td);
8049adc1 2649 if (td->rusage_sem) {
971caeb1 2650 fio_sem_remove(td->rusage_sem);
8049adc1
JA
2651 td->rusage_sem = NULL;
2652 }
971caeb1
BVA
2653 fio_sem_remove(td->sem);
2654 td->sem = NULL;
da8f124f 2655 } end_for_each();
2e1df07d 2656
a462baef 2657 free_disk_util();
b2e6494c
JA
2658 if (cgroup_list) {
2659 cgroup_kill(cgroup_list);
2660 sfree(cgroup_list);
2661 }
2e1df07d 2662
971caeb1 2663 fio_sem_remove(startup_sem);
cef9175e 2664 stat_exit();
2e1df07d
JA
2665 return exit_value;
2666}