Modify RDMA engine to print strerror messages.
[fio.git] / backend.c
CommitLineData
2e1df07d
JA
1/*
2 * fio - the flexible io tester
3 *
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk>
6 *
7 * The license below covers all files distributed with fio unless otherwise
8 * noted in the file itself.
9 *
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License version 2 as
12 * published by the Free Software Foundation.
13 *
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
18 *
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
22 *
23 */
24#include <unistd.h>
25#include <fcntl.h>
26#include <string.h>
27#include <limits.h>
28#include <signal.h>
29#include <time.h>
30#include <locale.h>
31#include <assert.h>
32#include <time.h>
e43606c2 33#include <inttypes.h>
2e1df07d
JA
34#include <sys/stat.h>
35#include <sys/wait.h>
36#include <sys/ipc.h>
2e1df07d 37#include <sys/mman.h>
ff6bb260 38#include <math.h>
2e1df07d
JA
39
40#include "fio.h"
a5e0ee11
O
41#ifndef FIO_NO_HAVE_SHM_H
42#include <sys/shm.h>
43#endif
2e1df07d
JA
44#include "hash.h"
45#include "smalloc.h"
46#include "verify.h"
47#include "trim.h"
48#include "diskutil.h"
49#include "cgroup.h"
50#include "profile.h"
51#include "lib/rand.h"
f7690c4a 52#include "lib/memalign.h"
2e1df07d 53#include "server.h"
44404c5a 54#include "lib/getrusage.h"
f2a2ce0e 55#include "idletime.h"
002fe734 56#include "err.h"
a9da8ab2 57#include "workqueue.h"
e81ecca3 58#include "lib/mountcheck.h"
40511301 59#include "rate-submit.h"
2e1df07d 60
5ddc6707
JA
61static pthread_t helper_thread;
62static pthread_mutex_t helper_lock;
63pthread_cond_t helper_cond;
64int helper_do_stat = 0;
8aab824f 65
2e1df07d 66static struct fio_mutex *startup_mutex;
2e1df07d
JA
67static struct flist_head *cgroup_list;
68static char *cgroup_mnt;
69static int exit_value;
70static volatile int fio_abort;
3a5f6bde
JA
71static unsigned int nr_process = 0;
72static unsigned int nr_thread = 0;
2e1df07d 73
6eaf09d6 74struct io_log *agg_io_log[DDIR_RWDIR_CNT];
2e1df07d 75
a3efc919
JA
76int groupid = 0;
77unsigned int thread_number = 0;
108fea77 78unsigned int stat_number = 0;
a3efc919
JA
79int shm_id = 0;
80int temp_stall_ts;
81unsigned long done_secs = 0;
5ddc6707 82volatile int helper_exit = 0;
a3efc919 83
2e1df07d 84#define PAGE_ALIGN(buf) \
e43606c2 85 (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
2e1df07d
JA
86
87#define JOB_START_TIMEOUT (5 * 1000)
88
89static void sig_int(int sig)
90{
91 if (threads) {
92 if (is_backend)
93 fio_server_got_signal(sig);
94 else {
95 log_info("\nfio: terminating on signal %d\n", sig);
e411c301 96 log_info_flush();
2e1df07d
JA
97 exit_value = 128;
98 }
99
100 fio_terminate_threads(TERMINATE_ALL);
101 }
102}
103
d2235e56 104void sig_show_status(int sig)
4c6d91e8
JA
105{
106 show_running_run_stats();
107}
108
2e1df07d
JA
109static void set_sig_handlers(void)
110{
111 struct sigaction act;
112
113 memset(&act, 0, sizeof(act));
114 act.sa_handler = sig_int;
115 act.sa_flags = SA_RESTART;
116 sigaction(SIGINT, &act, NULL);
117
118 memset(&act, 0, sizeof(act));
119 act.sa_handler = sig_int;
120 act.sa_flags = SA_RESTART;
121 sigaction(SIGTERM, &act, NULL);
122
2f694507
BC
123/* Windows uses SIGBREAK as a quit signal from other applications */
124#ifdef WIN32
125 memset(&act, 0, sizeof(act));
126 act.sa_handler = sig_int;
127 act.sa_flags = SA_RESTART;
128 sigaction(SIGBREAK, &act, NULL);
129#endif
130
4c6d91e8
JA
131 memset(&act, 0, sizeof(act));
132 act.sa_handler = sig_show_status;
133 act.sa_flags = SA_RESTART;
134 sigaction(SIGUSR1, &act, NULL);
135
2e1df07d
JA
136 if (is_backend) {
137 memset(&act, 0, sizeof(act));
138 act.sa_handler = sig_int;
139 act.sa_flags = SA_RESTART;
140 sigaction(SIGPIPE, &act, NULL);
141 }
142}
143
144/*
145 * Check if we are above the minimum rate given.
146 */
e39c0676
JA
147static bool __check_min_rate(struct thread_data *td, struct timeval *now,
148 enum fio_ddir ddir)
2e1df07d
JA
149{
150 unsigned long long bytes = 0;
151 unsigned long iops = 0;
152 unsigned long spent;
153 unsigned long rate;
154 unsigned int ratemin = 0;
155 unsigned int rate_iops = 0;
156 unsigned int rate_iops_min = 0;
157
158 assert(ddir_rw(ddir));
159
160 if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
e39c0676 161 return false;
2e1df07d
JA
162
163 /*
164 * allow a 2 second settle period in the beginning
165 */
166 if (mtime_since(&td->start, now) < 2000)
e39c0676 167 return false;
2e1df07d
JA
168
169 iops += td->this_io_blocks[ddir];
170 bytes += td->this_io_bytes[ddir];
171 ratemin += td->o.ratemin[ddir];
172 rate_iops += td->o.rate_iops[ddir];
173 rate_iops_min += td->o.rate_iops_min[ddir];
174
175 /*
176 * if rate blocks is set, sample is running
177 */
178 if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
179 spent = mtime_since(&td->lastrate[ddir], now);
180 if (spent < td->o.ratecycle)
e39c0676 181 return false;
2e1df07d 182
20e37e4a 183 if (td->o.rate[ddir] || td->o.ratemin[ddir]) {
2e1df07d
JA
184 /*
185 * check bandwidth specified rate
186 */
187 if (bytes < td->rate_bytes[ddir]) {
188 log_err("%s: min rate %u not met\n", td->o.name,
189 ratemin);
e39c0676 190 return true;
2e1df07d 191 } else {
49cba9b3
JA
192 if (spent)
193 rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
194 else
195 rate = 0;
196
2e1df07d
JA
197 if (rate < ratemin ||
198 bytes < td->rate_bytes[ddir]) {
199 log_err("%s: min rate %u not met, got"
200 " %luKB/sec\n", td->o.name,
201 ratemin, rate);
e39c0676 202 return true;
2e1df07d
JA
203 }
204 }
205 } else {
206 /*
207 * checks iops specified rate
208 */
209 if (iops < rate_iops) {
210 log_err("%s: min iops rate %u not met\n",
211 td->o.name, rate_iops);
e39c0676 212 return true;
2e1df07d 213 } else {
4c707a3b
JA
214 if (spent)
215 rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
216 else
217 rate = 0;
218
2e1df07d
JA
219 if (rate < rate_iops_min ||
220 iops < td->rate_blocks[ddir]) {
221 log_err("%s: min iops rate %u not met,"
222 " got %lu\n", td->o.name,
223 rate_iops_min, rate);
e39c0676 224 return true;
2e1df07d
JA
225 }
226 }
227 }
228 }
229
230 td->rate_bytes[ddir] = bytes;
231 td->rate_blocks[ddir] = iops;
232 memcpy(&td->lastrate[ddir], now, sizeof(*now));
e39c0676 233 return false;
2e1df07d
JA
234}
235
e39c0676 236static bool check_min_rate(struct thread_data *td, struct timeval *now)
2e1df07d 237{
e39c0676 238 bool ret = false;
2e1df07d 239
55312f9f 240 if (td->bytes_done[DDIR_READ])
6eaf09d6 241 ret |= __check_min_rate(td, now, DDIR_READ);
55312f9f 242 if (td->bytes_done[DDIR_WRITE])
6eaf09d6 243 ret |= __check_min_rate(td, now, DDIR_WRITE);
55312f9f 244 if (td->bytes_done[DDIR_TRIM])
6eaf09d6 245 ret |= __check_min_rate(td, now, DDIR_TRIM);
2e1df07d
JA
246
247 return ret;
248}
249
250/*
251 * When job exits, we can cancel the in-flight IO if we are using async
252 * io. Attempt to do so.
253 */
254static void cleanup_pending_aio(struct thread_data *td)
255{
2e1df07d
JA
256 int r;
257
258 /*
259 * get immediately available events, if any
260 */
55312f9f 261 r = io_u_queued_complete(td, 0);
2e1df07d
JA
262 if (r < 0)
263 return;
264
265 /*
266 * now cancel remaining active events
267 */
268 if (td->io_ops->cancel) {
2ae0b204
JA
269 struct io_u *io_u;
270 int i;
2e1df07d 271
2ae0b204
JA
272 io_u_qiter(&td->io_u_all, io_u, i) {
273 if (io_u->flags & IO_U_F_FLIGHT) {
2e1df07d
JA
274 r = td->io_ops->cancel(td, io_u);
275 if (!r)
276 put_io_u(td, io_u);
277 }
278 }
279 }
280
281 if (td->cur_depth)
55312f9f 282 r = io_u_queued_complete(td, td->cur_depth);
2e1df07d
JA
283}
284
285/*
286 * Helper to handle the final sync of a file. Works just like the normal
287 * io path, just does everything sync.
288 */
e39c0676 289static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
2e1df07d
JA
290{
291 struct io_u *io_u = __get_io_u(td);
292 int ret;
293
294 if (!io_u)
e39c0676 295 return true;
2e1df07d
JA
296
297 io_u->ddir = DDIR_SYNC;
298 io_u->file = f;
299
300 if (td_io_prep(td, io_u)) {
301 put_io_u(td, io_u);
e39c0676 302 return true;
2e1df07d
JA
303 }
304
305requeue:
306 ret = td_io_queue(td, io_u);
307 if (ret < 0) {
308 td_verror(td, io_u->error, "td_io_queue");
309 put_io_u(td, io_u);
e39c0676 310 return true;
2e1df07d 311 } else if (ret == FIO_Q_QUEUED) {
42ff945c
JA
312 if (td_io_commit(td))
313 return true;
55312f9f 314 if (io_u_queued_complete(td, 1) < 0)
e39c0676 315 return true;
2e1df07d
JA
316 } else if (ret == FIO_Q_COMPLETED) {
317 if (io_u->error) {
318 td_verror(td, io_u->error, "td_io_queue");
e39c0676 319 return true;
2e1df07d
JA
320 }
321
55312f9f 322 if (io_u_sync_complete(td, io_u) < 0)
e39c0676 323 return true;
2e1df07d
JA
324 } else if (ret == FIO_Q_BUSY) {
325 if (td_io_commit(td))
e39c0676 326 return true;
2e1df07d
JA
327 goto requeue;
328 }
329
e39c0676 330 return false;
2e1df07d 331}
a3efc919 332
61ee0f86
JA
333static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
334{
335 int ret;
336
337 if (fio_file_open(f))
338 return fio_io_sync(td, f);
339
340 if (td_io_open_file(td, f))
341 return 1;
342
343 ret = fio_io_sync(td, f);
344 td_io_close_file(td, f);
345 return ret;
346}
347
2e1df07d
JA
348static inline void __update_tv_cache(struct thread_data *td)
349{
350 fio_gettime(&td->tv_cache, NULL);
351}
352
353static inline void update_tv_cache(struct thread_data *td)
354{
355 if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
356 __update_tv_cache(td);
357}
358
e39c0676 359static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
2e1df07d
JA
360{
361 if (in_ramp_time(td))
e39c0676 362 return false;
2e1df07d 363 if (!td->o.timeout)
e39c0676 364 return false;
0de5b26f 365 if (utime_since(&td->epoch, t) >= td->o.timeout)
e39c0676 366 return true;
2e1df07d 367
e39c0676 368 return false;
2e1df07d
JA
369}
370
95603b74
BF
371/*
372 * We need to update the runtime consistently in ms, but keep a running
373 * tally of the current elapsed time in microseconds for sub millisecond
374 * updates.
375 */
376static inline void update_runtime(struct thread_data *td,
377 unsigned long long *elapsed_us,
378 const enum fio_ddir ddir)
379{
d5c8ea29
JA
380 if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only)
381 return;
382
95603b74
BF
383 td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000;
384 elapsed_us[ddir] += utime_since_now(&td->start);
385 td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000;
386}
387
e39c0676
JA
388static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
389 int *retptr)
2e1df07d
JA
390{
391 int ret = *retptr;
392
393 if (ret < 0 || td->error) {
8b28bd41
DM
394 int err = td->error;
395 enum error_type_bit eb;
2e1df07d
JA
396
397 if (ret < 0)
398 err = -ret;
2e1df07d 399
8b28bd41
DM
400 eb = td_error_type(ddir, err);
401 if (!(td->o.continue_on_error & (1 << eb)))
e39c0676 402 return true;
2e1df07d 403
8b28bd41 404 if (td_non_fatal_error(td, eb, err)) {
2e1df07d
JA
405 /*
406 * Continue with the I/Os in case of
407 * a non fatal error.
408 */
409 update_error_count(td, err);
410 td_clear_error(td);
411 *retptr = 0;
e39c0676 412 return false;
2e1df07d
JA
413 } else if (td->o.fill_device && err == ENOSPC) {
414 /*
415 * We expect to hit this error if
416 * fill_device option is set.
417 */
418 td_clear_error(td);
ebea2133 419 fio_mark_td_terminate(td);
e39c0676 420 return true;
2e1df07d
JA
421 } else {
422 /*
423 * Stop the I/O in case of a fatal
424 * error.
425 */
426 update_error_count(td, err);
e39c0676 427 return true;
2e1df07d
JA
428 }
429 }
430
e39c0676 431 return false;
2e1df07d
JA
432}
433
c97f1ad6
JA
434static void check_update_rusage(struct thread_data *td)
435{
436 if (td->update_rusage) {
437 td->update_rusage = 0;
438 update_rusage_stat(td);
439 fio_mutex_up(td->rusage_sem);
440 }
441}
442
55312f9f 443static int wait_for_completions(struct thread_data *td, struct timeval *time)
69fea81e
JA
444{
445 const int full = queue_full(td);
446 int min_evts = 0;
447 int ret;
448
449 /*
450 * if the queue is full, we MUST reap at least 1 event
451 */
82407585 452 min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
4b157ac6 453 if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
69fea81e
JA
454 min_evts = 1;
455
66b6c5ef 456 if (time && (__should_check_rate(td, DDIR_READ) ||
69fea81e
JA
457 __should_check_rate(td, DDIR_WRITE) ||
458 __should_check_rate(td, DDIR_TRIM)))
459 fio_gettime(time, NULL);
460
461 do {
55312f9f 462 ret = io_u_queued_complete(td, min_evts);
69fea81e
JA
463 if (ret < 0)
464 break;
465 } while (full && (td->cur_depth > td->o.iodepth_low));
466
467 return ret;
468}
469
e9d512d8
JA
470int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
471 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
472 struct timeval *comp_time)
473{
474 int ret2;
475
476 switch (*ret) {
477 case FIO_Q_COMPLETED:
478 if (io_u->error) {
479 *ret = -io_u->error;
480 clear_io_u(td, io_u);
481 } else if (io_u->resid) {
482 int bytes = io_u->xfer_buflen - io_u->resid;
483 struct fio_file *f = io_u->file;
484
485 if (bytes_issued)
486 *bytes_issued += bytes;
487
488 if (!from_verify)
489 trim_io_piece(td, io_u);
490
491 /*
492 * zero read, fail
493 */
494 if (!bytes) {
495 if (!from_verify)
496 unlog_io_piece(td, io_u);
497 td_verror(td, EIO, "full resid");
498 put_io_u(td, io_u);
499 break;
500 }
501
502 io_u->xfer_buflen = io_u->resid;
503 io_u->xfer_buf += bytes;
504 io_u->offset += bytes;
505
506 if (ddir_rw(io_u->ddir))
507 td->ts.short_io_u[io_u->ddir]++;
508
509 f = io_u->file;
510 if (io_u->offset == f->real_file_size)
511 goto sync_done;
512
513 requeue_io_u(td, &io_u);
514 } else {
515sync_done:
516 if (comp_time && (__should_check_rate(td, DDIR_READ) ||
517 __should_check_rate(td, DDIR_WRITE) ||
518 __should_check_rate(td, DDIR_TRIM)))
519 fio_gettime(comp_time, NULL);
520
521 *ret = io_u_sync_complete(td, io_u);
522 if (*ret < 0)
523 break;
524 }
a0e3e5a6
EW
525
526 /*
527 * when doing I/O (not when verifying),
528 * check for any errors that are to be ignored
529 */
530 if (!from_verify)
531 break;
532
e9d512d8
JA
533 return 0;
534 case FIO_Q_QUEUED:
535 /*
536 * if the engine doesn't have a commit hook,
537 * the io_u is really queued. if it does have such
538 * a hook, it has to call io_u_queued() itself.
539 */
540 if (td->io_ops->commit == NULL)
541 io_u_queued(td, io_u);
542 if (bytes_issued)
543 *bytes_issued += io_u->xfer_buflen;
544 break;
545 case FIO_Q_BUSY:
546 if (!from_verify)
547 unlog_io_piece(td, io_u);
548 requeue_io_u(td, &io_u);
549 ret2 = td_io_commit(td);
550 if (ret2 < 0)
551 *ret = ret2;
552 break;
553 default:
b0775325 554 assert(*ret < 0);
e9d512d8
JA
555 td_verror(td, -(*ret), "td_io_queue");
556 break;
557 }
558
559 if (break_on_this_error(td, ddir, ret))
560 return 1;
561
562 return 0;
563}
564
e39c0676 565static inline bool io_in_polling(struct thread_data *td)
82407585
RP
566{
567 return !td->o.iodepth_batch_complete_min &&
568 !td->o.iodepth_batch_complete_max;
569}
570
2e1df07d
JA
571/*
572 * The main verify engine. Runs over the writes we previously submitted,
573 * reads the blocks back in, and checks the crc/md5 of the data.
574 */
100f49f1 575static void do_verify(struct thread_data *td, uint64_t verify_bytes)
2e1df07d
JA
576{
577 struct fio_file *f;
578 struct io_u *io_u;
579 int ret, min_events;
580 unsigned int i;
581
582 dprint(FD_VERIFY, "starting loop\n");
583
584 /*
585 * sync io first and invalidate cache, to make sure we really
586 * read from disk.
587 */
588 for_each_file(td, f, i) {
589 if (!fio_file_open(f))
590 continue;
591 if (fio_io_sync(td, f))
592 break;
593 if (file_invalidate_cache(td, f))
594 break;
595 }
596
c97f1ad6
JA
597 check_update_rusage(td);
598
2e1df07d
JA
599 if (td->error)
600 return;
601
602 td_set_runstate(td, TD_VERIFYING);
603
604 io_u = NULL;
605 while (!td->terminate) {
fbccf46c 606 enum fio_ddir ddir;
e9d512d8 607 int full;
2e1df07d
JA
608
609 update_tv_cache(td);
c97f1ad6 610 check_update_rusage(td);
2e1df07d
JA
611
612 if (runtime_exceeded(td, &td->tv_cache)) {
613 __update_tv_cache(td);
614 if (runtime_exceeded(td, &td->tv_cache)) {
ebea2133 615 fio_mark_td_terminate(td);
2e1df07d
JA
616 break;
617 }
618 }
619
9e684a49
DE
620 if (flow_threshold_exceeded(td))
621 continue;
622
44cbc6da
JA
623 if (!td->o.experimental_verify) {
624 io_u = __get_io_u(td);
625 if (!io_u)
626 break;
2e1df07d 627
44cbc6da
JA
628 if (get_next_verify(td, io_u)) {
629 put_io_u(td, io_u);
630 break;
631 }
2e1df07d 632
44cbc6da
JA
633 if (td_io_prep(td, io_u)) {
634 put_io_u(td, io_u);
635 break;
636 }
637 } else {
55312f9f 638 if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes)
100f49f1
JA
639 break;
640
bcd5abfa 641 while ((io_u = get_io_u(td)) != NULL) {
002fe734
JA
642 if (IS_ERR(io_u)) {
643 io_u = NULL;
644 ret = FIO_Q_BUSY;
645 goto reap;
646 }
647
bcd5abfa
JA
648 /*
649 * We are only interested in the places where
650 * we wrote or trimmed IOs. Turn those into
651 * reads for verification purposes.
652 */
653 if (io_u->ddir == DDIR_READ) {
654 /*
655 * Pretend we issued it for rwmix
656 * accounting
657 */
658 td->io_issues[DDIR_READ]++;
659 put_io_u(td, io_u);
660 continue;
661 } else if (io_u->ddir == DDIR_TRIM) {
662 io_u->ddir = DDIR_READ;
a9da8ab2 663 io_u_set(io_u, IO_U_F_TRIMMED);
bcd5abfa
JA
664 break;
665 } else if (io_u->ddir == DDIR_WRITE) {
666 io_u->ddir = DDIR_READ;
667 break;
668 } else {
669 put_io_u(td, io_u);
670 continue;
671 }
672 }
44cbc6da 673
bcd5abfa 674 if (!io_u)
44cbc6da 675 break;
2e1df07d
JA
676 }
677
ca09be4b
JA
678 if (verify_state_should_stop(td, io_u)) {
679 put_io_u(td, io_u);
680 break;
681 }
682
2e1df07d
JA
683 if (td->o.verify_async)
684 io_u->end_io = verify_io_u_async;
685 else
686 io_u->end_io = verify_io_u;
687
fbccf46c 688 ddir = io_u->ddir;
10adc4a7
GG
689 if (!td->o.disable_slat)
690 fio_gettime(&io_u->start_time, NULL);
fbccf46c 691
2e1df07d 692 ret = td_io_queue(td, io_u);
2e1df07d 693
e9d512d8 694 if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
2e1df07d
JA
695 break;
696
697 /*
698 * if we can queue more, do so. but check if there are
699 * completed io_u's first. Note that we can get BUSY even
700 * without IO queued, if the system is resource starved.
701 */
002fe734 702reap:
2e1df07d 703 full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
82407585 704 if (full || io_in_polling(td))
55312f9f 705 ret = wait_for_completions(td, NULL);
2e1df07d 706
2e1df07d
JA
707 if (ret < 0)
708 break;
709 }
710
c97f1ad6
JA
711 check_update_rusage(td);
712
2e1df07d
JA
713 if (!td->error) {
714 min_events = td->cur_depth;
715
716 if (min_events)
55312f9f 717 ret = io_u_queued_complete(td, min_events);
2e1df07d
JA
718 } else
719 cleanup_pending_aio(td);
720
721 td_set_runstate(td, TD_RUNNING);
722
723 dprint(FD_VERIFY, "exiting loop\n");
724}
725
e39c0676 726static bool exceeds_number_ios(struct thread_data *td)
3939fe85
JA
727{
728 unsigned long long number_ios;
729
730 if (!td->o.number_ios)
e39c0676 731 return false;
3939fe85 732
cf8a46a7 733 number_ios = ddir_rw_sum(td->io_blocks);
3939fe85
JA
734 number_ios += td->io_u_queued + td->io_u_in_flight;
735
cf8a46a7 736 return number_ios >= (td->o.number_ios * td->loops);
3939fe85
JA
737}
738
e39c0676 739static bool io_issue_bytes_exceeded(struct thread_data *td)
f7078f7b 740{
77731b29 741 unsigned long long bytes, limit;
f7078f7b
JA
742
743 if (td_rw(td))
74d6277f 744 bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
f7078f7b 745 else if (td_write(td))
74d6277f 746 bytes = td->io_issue_bytes[DDIR_WRITE];
6eaf09d6 747 else if (td_read(td))
74d6277f 748 bytes = td->io_issue_bytes[DDIR_READ];
f7078f7b 749 else
74d6277f 750 bytes = td->io_issue_bytes[DDIR_TRIM];
f7078f7b 751
77731b29
JA
752 if (td->o.io_limit)
753 limit = td->o.io_limit;
754 else
755 limit = td->o.size;
756
cf8a46a7 757 limit *= td->loops;
77731b29 758 return bytes >= limit || exceeds_number_ios(td);
f7078f7b
JA
759}
760
e39c0676 761static bool io_complete_bytes_exceeded(struct thread_data *td)
e28dd2cf
JE
762{
763 unsigned long long bytes, limit;
764
765 if (td_rw(td))
766 bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
767 else if (td_write(td))
768 bytes = td->this_io_bytes[DDIR_WRITE];
769 else if (td_read(td))
770 bytes = td->this_io_bytes[DDIR_READ];
771 else
772 bytes = td->this_io_bytes[DDIR_TRIM];
773
774 if (td->o.io_limit)
775 limit = td->o.io_limit;
776 else
777 limit = td->o.size;
778
cf8a46a7 779 limit *= td->loops;
e28dd2cf
JE
780 return bytes >= limit || exceeds_number_ios(td);
781}
782
50a8ce86
D
783/*
784 * used to calculate the next io time for rate control
785 *
786 */
787static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
788{
ff6bb260 789 uint64_t secs, remainder, bps, bytes, iops;
50a8ce86
D
790
791 assert(!(td->flags & TD_F_CHILD));
792 bytes = td->rate_io_issue_bytes[ddir];
793 bps = td->rate_bps[ddir];
ff6bb260 794
6de65959 795 if (td->o.rate_process == RATE_PROCESS_POISSON) {
a77d139b 796 uint64_t val;
ff6bb260 797 iops = bps / td->o.bs[ddir];
a77d139b
JA
798 val = (int64_t) (1000000 / iops) *
799 -logf(__rand_0_1(&td->poisson_state));
800 if (val) {
801 dprint(FD_RATE, "poisson rate iops=%llu\n",
802 (unsigned long long) 1000000 / val);
803 }
804 td->last_usec += val;
ff6bb260
SL
805 return td->last_usec;
806 } else if (bps) {
50a8ce86
D
807 secs = bytes / bps;
808 remainder = bytes % bps;
809 return remainder * 1000000 / bps + secs * 1000000;
e7b24047
JA
810 }
811
812 return 0;
50a8ce86
D
813}
814
2e1df07d
JA
815/*
816 * Main IO worker function. It retrieves io_u's to process and queues
817 * and reaps them, checking for rate and errors along the way.
100f49f1
JA
818 *
819 * Returns number of bytes written and trimmed.
2e1df07d 820 */
95b03be0 821static void do_io(struct thread_data *td, uint64_t *bytes_done)
2e1df07d
JA
822{
823 unsigned int i;
824 int ret = 0;
c2703bf3 825 uint64_t total_bytes, bytes_issued = 0;
70c68076 826
95b03be0
JA
827 for (i = 0; i < DDIR_RWDIR_CNT; i++)
828 bytes_done[i] = td->bytes_done[i];
2e1df07d
JA
829
830 if (in_ramp_time(td))
831 td_set_runstate(td, TD_RAMP);
832 else
833 td_set_runstate(td, TD_RUNNING);
834
3e260a46
JA
835 lat_target_init(td);
836
1e564979
JE
837 total_bytes = td->o.size;
838 /*
839 * Allow random overwrite workloads to write up to io_limit
840 * before starting verification phase as 'size' doesn't apply.
841 */
842 if (td_write(td) && td_random(td) && td->o.norandommap)
843 total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
78a6469c
JA
844 /*
845 * If verify_backlog is enabled, we'll run the verify in this
846 * handler as well. For that case, we may need up to twice the
847 * amount of bytes.
848 */
78a6469c
JA
849 if (td->o.verify != VERIFY_NONE &&
850 (td_write(td) && td->o.verify_backlog))
c2703bf3
JA
851 total_bytes += td->o.size;
852
82a90686 853 /* In trimwrite mode, each byte is trimmed and then written, so
0e4dd95c 854 * allow total_bytes to be twice as big */
82a90686 855 if (td_trimwrite(td))
0e4dd95c
DE
856 total_bytes += td->total_io_size;
857
f7078f7b 858 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
e28dd2cf 859 (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
c04e4661 860 td->o.time_based) {
2e1df07d 861 struct timeval comp_time;
2e1df07d 862 struct io_u *io_u;
e9d512d8 863 int full;
2e1df07d
JA
864 enum fio_ddir ddir;
865
c97f1ad6
JA
866 check_update_rusage(td);
867
7d7803fa 868 if (td->terminate || td->done)
2e1df07d
JA
869 break;
870
871 update_tv_cache(td);
872
873 if (runtime_exceeded(td, &td->tv_cache)) {
874 __update_tv_cache(td);
875 if (runtime_exceeded(td, &td->tv_cache)) {
ebea2133 876 fio_mark_td_terminate(td);
2e1df07d
JA
877 break;
878 }
879 }
880
9e684a49
DE
881 if (flow_threshold_exceeded(td))
882 continue;
883
f1a32461
JA
884 /*
885 * Break if we exceeded the bytes. The exception is time
886 * based runs, but we still need to break out of the loop
887 * for those to run verification, if enabled.
888 */
889 if (bytes_issued >= total_bytes &&
890 (!td->o.time_based ||
891 (td->o.time_based && td->o.verify != VERIFY_NONE)))
20876c53
JC
892 break;
893
2e1df07d 894 io_u = get_io_u(td);
002fe734
JA
895 if (IS_ERR_OR_NULL(io_u)) {
896 int err = PTR_ERR(io_u);
897
898 io_u = NULL;
899 if (err == -EBUSY) {
900 ret = FIO_Q_BUSY;
901 goto reap;
902 }
3e260a46
JA
903 if (td->o.latency_target)
904 goto reap;
2e1df07d 905 break;
3e260a46 906 }
2e1df07d
JA
907
908 ddir = io_u->ddir;
909
910 /*
82af2a7c
JA
911 * Add verification end_io handler if:
912 * - Asked to verify (!td_rw(td))
913 * - Or the io_u is from our verify list (mixed write/ver)
2e1df07d
JA
914 */
915 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
82af2a7c 916 ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
c4b6117b
PV
917
918 if (!td->o.verify_pattern_bytes) {
d6b72507 919 io_u->rand_seed = __rand(&td->verify_state);
c4b6117b 920 if (sizeof(int) != sizeof(long *))
d6b72507 921 io_u->rand_seed *= __rand(&td->verify_state);
c4b6117b
PV
922 }
923
ca09be4b
JA
924 if (verify_state_should_stop(td, io_u)) {
925 put_io_u(td, io_u);
926 break;
927 }
928
2e1df07d
JA
929 if (td->o.verify_async)
930 io_u->end_io = verify_io_u_async;
931 else
932 io_u->end_io = verify_io_u;
933 td_set_runstate(td, TD_VERIFYING);
934 } else if (in_ramp_time(td))
935 td_set_runstate(td, TD_RAMP);
936 else
937 td_set_runstate(td, TD_RUNNING);
938
9a50c5c5 939 /*
f9401285
JA
940 * Always log IO before it's issued, so we know the specific
941 * order of it. The logged unit will track when the IO has
942 * completed.
9a50c5c5 943 */
c4b6117b
PV
944 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
945 td->o.do_verify &&
946 td->o.verify != VERIFY_NONE &&
f9401285 947 !td->o.experimental_verify)
c4b6117b
PV
948 log_io_piece(td, io_u);
949
a9da8ab2 950 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
0c5df5f9
JA
951 const unsigned long blen = io_u->xfer_buflen;
952 const enum fio_ddir ddir = acct_ddir(io_u);
953
a9da8ab2
JA
954 if (td->error)
955 break;
0c5df5f9 956
26de50cf
JA
957 workqueue_enqueue(&td->io_wq, &io_u->work);
958 ret = FIO_Q_QUEUED;
50a8ce86 959
26de50cf 960 if (ddir_rw(ddir)) {
0c5df5f9
JA
961 td->io_issues[ddir]++;
962 td->io_issue_bytes[ddir] += blen;
963 td->rate_io_issue_bytes[ddir] += blen;
964 }
965
50a8ce86
D
966 if (should_check_rate(td))
967 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
968
a9da8ab2
JA
969 } else {
970 ret = td_io_queue(td, io_u);
2e1df07d 971
50a8ce86
D
972 if (should_check_rate(td))
973 td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
974
fd727d9d 975 if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
a9da8ab2 976 break;
2e1df07d 977
a9da8ab2
JA
978 /*
979 * See if we need to complete some commands. Note that
980 * we can get BUSY even without IO queued, if the
981 * system is resource starved.
982 */
3e260a46 983reap:
a9da8ab2
JA
984 full = queue_full(td) ||
985 (ret == FIO_Q_BUSY && td->cur_depth);
82407585 986 if (full || io_in_polling(td))
a9da8ab2
JA
987 ret = wait_for_completions(td, &comp_time);
988 }
2e1df07d
JA
989 if (ret < 0)
990 break;
55312f9f
JA
991 if (!ddir_rw_sum(td->bytes_done) &&
992 !(td->io_ops->flags & FIO_NOIO))
2e1df07d
JA
993 continue;
994
55312f9f
JA
995 if (!in_ramp_time(td) && should_check_rate(td)) {
996 if (check_min_rate(td, &comp_time)) {
f9cafb12 997 if (exitall_on_terminate || td->o.exitall_error)
2e1df07d
JA
998 fio_terminate_threads(td->groupid);
999 td_verror(td, EIO, "check_min_rate");
1000 break;
1001 }
1002 }
3e260a46
JA
1003 if (!in_ramp_time(td) && td->o.latency_target)
1004 lat_target_check(td);
e155cb64 1005
2e1df07d
JA
1006 if (td->o.thinktime) {
1007 unsigned long long b;
1008
342f4be4 1009 b = ddir_rw_sum(td->io_blocks);
2e1df07d
JA
1010 if (!(b % td->o.thinktime_blocks)) {
1011 int left;
1012
002e7183
JA
1013 io_u_quiesce(td);
1014
2e1df07d
JA
1015 if (td->o.thinktime_spin)
1016 usec_spin(td->o.thinktime_spin);
1017
1018 left = td->o.thinktime - td->o.thinktime_spin;
1019 if (left)
1020 usec_sleep(td, left);
1021 }
1022 }
1023 }
1024
c97f1ad6
JA
1025 check_update_rusage(td);
1026
2e1df07d 1027 if (td->trim_entries)
4e0a8fa2 1028 log_err("fio: %lu trim entries leaked?\n", td->trim_entries);
2e1df07d
JA
1029
1030 if (td->o.fill_device && td->error == ENOSPC) {
1031 td->error = 0;
ebea2133 1032 fio_mark_td_terminate(td);
2e1df07d
JA
1033 }
1034 if (!td->error) {
1035 struct fio_file *f;
1036
a9da8ab2
JA
1037 if (td->o.io_submit_mode == IO_MODE_OFFLOAD) {
1038 workqueue_flush(&td->io_wq);
1039 i = 0;
1040 } else
1041 i = td->cur_depth;
1042
2e1df07d 1043 if (i) {
55312f9f 1044 ret = io_u_queued_complete(td, i);
2e1df07d
JA
1045 if (td->o.fill_device && td->error == ENOSPC)
1046 td->error = 0;
1047 }
1048
1049 if (should_fsync(td) && td->o.end_fsync) {
1050 td_set_runstate(td, TD_FSYNCING);
1051
1052 for_each_file(td, f, i) {
61ee0f86 1053 if (!fio_file_fsync(td, f))
2e1df07d 1054 continue;
61ee0f86
JA
1055
1056 log_err("fio: end_fsync failed for file %s\n",
1057 f->file_name);
2e1df07d
JA
1058 }
1059 }
1060 } else
1061 cleanup_pending_aio(td);
1062
1063 /*
1064 * stop job if we failed doing any IO
1065 */
342f4be4 1066 if (!ddir_rw_sum(td->this_io_bytes))
2e1df07d 1067 td->done = 1;
100f49f1 1068
95b03be0
JA
1069 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1070 bytes_done[i] = td->bytes_done[i] - bytes_done[i];
2e1df07d
JA
1071}
1072
1073static void cleanup_io_u(struct thread_data *td)
1074{
2e1df07d
JA
1075 struct io_u *io_u;
1076
2ae0b204 1077 while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
c73ed246
JA
1078
1079 if (td->io_ops->io_u_free)
1080 td->io_ops->io_u_free(td, io_u);
1081
2e1df07d
JA
1082 fio_memfree(io_u, sizeof(*io_u));
1083 }
1084
1085 free_io_mem(td);
2ae0b204
JA
1086
1087 io_u_rexit(&td->io_u_requeues);
1088 io_u_qexit(&td->io_u_freelist);
1089 io_u_qexit(&td->io_u_all);
ca09be4b
JA
1090
1091 if (td->last_write_comp)
1092 sfree(td->last_write_comp);
2e1df07d
JA
1093}
1094
1095static int init_io_u(struct thread_data *td)
1096{
1097 struct io_u *io_u;
9c42684e 1098 unsigned int max_bs, min_write;
2e1df07d 1099 int cl_align, i, max_units;
2ae0b204 1100 int data_xfer = 1, err;
2e1df07d
JA
1101 char *p;
1102
1103 max_units = td->o.iodepth;
74f4b020 1104 max_bs = td_max_bs(td);
9c42684e 1105 min_write = td->o.min_bs[DDIR_WRITE];
2e1df07d
JA
1106 td->orig_buffer_size = (unsigned long long) max_bs
1107 * (unsigned long long) max_units;
1108
88045e04 1109 if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
59d8d0f5
JA
1110 data_xfer = 0;
1111
2ae0b204
JA
1112 err = 0;
1113 err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
1114 err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
1115 err += io_u_qinit(&td->io_u_all, td->o.iodepth);
1116
1117 if (err) {
1118 log_err("fio: failed setting up IO queues\n");
1119 return 1;
1120 }
1121
fd8a09b8 1122 /*
1123 * if we may later need to do address alignment, then add any
1124 * possible adjustment here so that we don't cause a buffer
1125 * overflow later. this adjustment may be too much if we get
1126 * lucky and the allocator gives us an aligned address.
1127 */
d01612f3
CM
1128 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
1129 (td->io_ops->flags & FIO_RAWIO))
fd8a09b8 1130 td->orig_buffer_size += page_mask + td->o.mem_align;
1131
2e1df07d
JA
1132 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
1133 unsigned long bs;
1134
1135 bs = td->orig_buffer_size + td->o.hugepage_size - 1;
1136 td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
1137 }
1138
1139 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
1140 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
1141 return 1;
1142 }
1143
59d8d0f5 1144 if (data_xfer && allocate_io_mem(td))
2e1df07d
JA
1145 return 1;
1146
d01612f3 1147 if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
2e1df07d
JA
1148 (td->io_ops->flags & FIO_RAWIO))
1149 p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
1150 else
1151 p = td->orig_buffer;
1152
1153 cl_align = os_cache_line_size();
1154
1155 for (i = 0; i < max_units; i++) {
1156 void *ptr;
1157
1158 if (td->terminate)
1159 return 1;
1160
1161 ptr = fio_memalign(cl_align, sizeof(*io_u));
1162 if (!ptr) {
1163 log_err("fio: unable to allocate aligned memory\n");
1164 break;
1165 }
1166
1167 io_u = ptr;
1168 memset(io_u, 0, sizeof(*io_u));
2ae0b204 1169 INIT_FLIST_HEAD(&io_u->verify_list);
2e1df07d
JA
1170 dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
1171
59d8d0f5 1172 if (data_xfer) {
2e1df07d
JA
1173 io_u->buf = p;
1174 dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
1175
1176 if (td_write(td))
9c42684e 1177 io_u_fill_buffer(td, io_u, min_write, max_bs);
2e1df07d
JA
1178 if (td_write(td) && td->o.verify_pattern_bytes) {
1179 /*
1180 * Fill the buffer with the pattern if we are
1181 * going to be doing writes.
1182 */
ce35b1ec 1183 fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
2e1df07d
JA
1184 }
1185 }
1186
1187 io_u->index = i;
1188 io_u->flags = IO_U_F_FREE;
2ae0b204
JA
1189 io_u_qpush(&td->io_u_freelist, io_u);
1190
1191 /*
1192 * io_u never leaves this stack, used for iteration of all
1193 * io_u buffers.
1194 */
1195 io_u_qpush(&td->io_u_all, io_u);
c73ed246
JA
1196
1197 if (td->io_ops->io_u_init) {
1198 int ret = td->io_ops->io_u_init(td, io_u);
1199
1200 if (ret) {
1201 log_err("fio: failed to init engine data: %d\n", ret);
1202 return 1;
1203 }
1204 }
1205
2e1df07d
JA
1206 p += max_bs;
1207 }
1208
ca09be4b
JA
1209 if (td->o.verify != VERIFY_NONE) {
1210 td->last_write_comp = scalloc(max_units, sizeof(uint64_t));
1211 if (!td->last_write_comp) {
1212 log_err("fio: failed to alloc write comp data\n");
1213 return 1;
1214 }
1215 }
1216
2e1df07d
JA
1217 return 0;
1218}
1219
1220static int switch_ioscheduler(struct thread_data *td)
1221{
1222 char tmp[256], tmp2[128];
1223 FILE *f;
1224 int ret;
1225
1226 if (td->io_ops->flags & FIO_DISKLESSIO)
1227 return 0;
1228
1229 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1230
1231 f = fopen(tmp, "r+");
1232 if (!f) {
1233 if (errno == ENOENT) {
1234 log_err("fio: os or kernel doesn't support IO scheduler"
1235 " switching\n");
1236 return 0;
1237 }
1238 td_verror(td, errno, "fopen iosched");
1239 return 1;
1240 }
1241
1242 /*
1243 * Set io scheduler.
1244 */
1245 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
1246 if (ferror(f) || ret != 1) {
1247 td_verror(td, errno, "fwrite");
1248 fclose(f);
1249 return 1;
1250 }
1251
1252 rewind(f);
1253
1254 /*
1255 * Read back and check that the selected scheduler is now the default.
1256 */
b44b9e45 1257 memset(tmp, 0, sizeof(tmp));
49c6f33d 1258 ret = fread(tmp, sizeof(tmp), 1, f);
2e1df07d
JA
1259 if (ferror(f) || ret < 0) {
1260 td_verror(td, errno, "fread");
1261 fclose(f);
1262 return 1;
1263 }
b44b9e45
TK
1264 /*
1265 * either a list of io schedulers or "none\n" is expected.
1266 */
1267 tmp[strlen(tmp) - 1] = '\0';
49c6f33d 1268
2e1df07d
JA
1269
1270 sprintf(tmp2, "[%s]", td->o.ioscheduler);
1271 if (!strstr(tmp, tmp2)) {
1272 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
1273 td_verror(td, EINVAL, "iosched_switch");
1274 fclose(f);
1275 return 1;
1276 }
1277
1278 fclose(f);
1279 return 0;
1280}
1281
e39c0676 1282static bool keep_running(struct thread_data *td)
2e1df07d 1283{
77731b29
JA
1284 unsigned long long limit;
1285
2e1df07d 1286 if (td->done)
e39c0676 1287 return false;
2e1df07d 1288 if (td->o.time_based)
e39c0676 1289 return true;
2e1df07d
JA
1290 if (td->o.loops) {
1291 td->o.loops--;
e39c0676 1292 return true;
2e1df07d 1293 }
3939fe85 1294 if (exceeds_number_ios(td))
e39c0676 1295 return false;
26251d8d 1296
77731b29
JA
1297 if (td->o.io_limit)
1298 limit = td->o.io_limit;
1299 else
1300 limit = td->o.size;
1301
1302 if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
5bd5f71a
JA
1303 uint64_t diff;
1304
1305 /*
1306 * If the difference is less than the minimum IO size, we
1307 * are done.
1308 */
77731b29 1309 diff = limit - ddir_rw_sum(td->io_bytes);
74f4b020 1310 if (diff < td_max_bs(td))
e39c0676 1311 return false;
5bd5f71a 1312
543e2e9d 1313 if (fio_files_done(td) && !td->o.io_limit)
e39c0676 1314 return false;
002fe734 1315
e39c0676 1316 return true;
5bd5f71a 1317 }
2e1df07d 1318
e39c0676 1319 return false;
2e1df07d
JA
1320}
1321
ce486495 1322static int exec_string(struct thread_options *o, const char *string, const char *mode)
2e1df07d 1323{
6c9ce469 1324 size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
f491a907 1325 int ret;
2e1df07d
JA
1326 char *str;
1327
1328 str = malloc(newlen);
ce486495 1329 sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
2e1df07d 1330
ce486495 1331 log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
2e1df07d
JA
1332 ret = system(str);
1333 if (ret == -1)
1334 log_err("fio: exec of cmd <%s> failed\n", str);
1335
1336 free(str);
1337 return ret;
1338}
1339
62167762
JC
1340/*
1341 * Dry run to compute correct state of numberio for verification.
1342 */
1343static uint64_t do_dry_run(struct thread_data *td)
1344{
62167762
JC
1345 td_set_runstate(td, TD_RUNNING);
1346
1347 while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
e28dd2cf 1348 (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
62167762
JC
1349 struct io_u *io_u;
1350 int ret;
1351
1352 if (td->terminate || td->done)
1353 break;
1354
1355 io_u = get_io_u(td);
1356 if (!io_u)
1357 break;
1358
a9da8ab2 1359 io_u_set(io_u, IO_U_F_FLIGHT);
62167762
JC
1360 io_u->error = 0;
1361 io_u->resid = 0;
1362 if (ddir_rw(acct_ddir(io_u)))
1363 td->io_issues[acct_ddir(io_u)]++;
1364 if (ddir_rw(io_u->ddir)) {
1365 io_u_mark_depth(td, 1);
1366 td->ts.total_io_u[io_u->ddir]++;
1367 }
1368
2e63e96b
PV
1369 if (td_write(td) && io_u->ddir == DDIR_WRITE &&
1370 td->o.do_verify &&
1371 td->o.verify != VERIFY_NONE &&
1372 !td->o.experimental_verify)
1373 log_io_piece(td, io_u);
1374
55312f9f 1375 ret = io_u_sync_complete(td, io_u);
62167762
JC
1376 (void) ret;
1377 }
1378
55312f9f 1379 return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
62167762
JC
1380}
1381
24660963
JA
1382struct fork_data {
1383 struct thread_data *td;
1384 struct sk_out *sk_out;
1385};
1386
2e1df07d
JA
1387/*
1388 * Entry point for the thread based jobs. The process based jobs end up
1389 * here as well, after a little setup.
1390 */
1391static void *thread_main(void *data)
1392{
24660963 1393 struct fork_data *fd = data;
95603b74 1394 unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
24660963 1395 struct thread_data *td = fd->td;
4896473e 1396 struct thread_options *o = &td->o;
24660963 1397 struct sk_out *sk_out = fd->sk_out;
2e1df07d
JA
1398 pthread_condattr_t attr;
1399 int clear_state;
28727df7 1400 int ret;
2e1df07d 1401
24660963
JA
1402 sk_out_assign(sk_out);
1403 free(fd);
1404
4896473e 1405 if (!o->use_thread) {
2e1df07d
JA
1406 setsid();
1407 td->pid = getpid();
1408 } else
1409 td->pid = gettid();
1410
4896473e 1411 fio_local_clock_init(o->use_thread);
5d879392 1412
2e1df07d
JA
1413 dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
1414
122c7725
JA
1415 if (is_backend)
1416 fio_server_send_start(td);
1417
2e1df07d
JA
1418 INIT_FLIST_HEAD(&td->io_log_list);
1419 INIT_FLIST_HEAD(&td->io_hist_list);
1420 INIT_FLIST_HEAD(&td->verify_list);
1421 INIT_FLIST_HEAD(&td->trim_list);
1ae83d45 1422 INIT_FLIST_HEAD(&td->next_rand_list);
2e1df07d
JA
1423 pthread_mutex_init(&td->io_u_lock, NULL);
1424 td->io_hist_tree = RB_ROOT;
1425
1426 pthread_condattr_init(&attr);
1427 pthread_cond_init(&td->verify_cond, &attr);
1428 pthread_cond_init(&td->free_cond, &attr);
1429
1430 td_set_runstate(td, TD_INITIALIZED);
1431 dprint(FD_MUTEX, "up startup_mutex\n");
1432 fio_mutex_up(startup_mutex);
1433 dprint(FD_MUTEX, "wait on td->mutex\n");
1434 fio_mutex_down(td->mutex);
1435 dprint(FD_MUTEX, "done waiting on td->mutex\n");
1436
2e1df07d
JA
1437 /*
1438 * A new gid requires privilege, so we need to do this before setting
1439 * the uid.
1440 */
4896473e 1441 if (o->gid != -1U && setgid(o->gid)) {
2e1df07d
JA
1442 td_verror(td, errno, "setgid");
1443 goto err;
1444 }
4896473e 1445 if (o->uid != -1U && setuid(o->uid)) {
2e1df07d
JA
1446 td_verror(td, errno, "setuid");
1447 goto err;
1448 }
1449
1450 /*
1451 * If we have a gettimeofday() thread, make sure we exclude that
1452 * thread from this job
1453 */
4896473e
JA
1454 if (o->gtod_cpu)
1455 fio_cpu_clear(&o->cpumask, o->gtod_cpu);
2e1df07d
JA
1456
1457 /*
1458 * Set affinity first, in case it has an impact on the memory
1459 * allocations.
1460 */
b2a9e649 1461 if (fio_option_is_set(o, cpumask)) {
c2acfbac 1462 if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
30cb4c65 1463 ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
c2acfbac
JA
1464 if (!ret) {
1465 log_err("fio: no CPUs set\n");
1466 log_err("fio: Try increasing number of available CPUs\n");
1467 td_verror(td, EINVAL, "cpus_split");
1468 goto err;
1469 }
1470 }
28727df7
JA
1471 ret = fio_setaffinity(td->pid, o->cpumask);
1472 if (ret == -1) {
4896473e
JA
1473 td_verror(td, errno, "cpu_set_affinity");
1474 goto err;
1475 }
2e1df07d
JA
1476 }
1477
67bf9823 1478#ifdef CONFIG_LIBNUMA
d0b937ed 1479 /* numa node setup */
b2a9e649
JA
1480 if (fio_option_is_set(o, numa_cpunodes) ||
1481 fio_option_is_set(o, numa_memnodes)) {
43522848 1482 struct bitmask *mask;
d0b937ed
YR
1483
1484 if (numa_available() < 0) {
1485 td_verror(td, errno, "Does not support NUMA API\n");
1486 goto err;
1487 }
1488
b2a9e649 1489 if (fio_option_is_set(o, numa_cpunodes)) {
43522848
DG
1490 mask = numa_parse_nodestring(o->numa_cpunodes);
1491 ret = numa_run_on_node_mask(mask);
1492 numa_free_nodemask(mask);
d0b937ed
YR
1493 if (ret == -1) {
1494 td_verror(td, errno, \
1495 "numa_run_on_node_mask failed\n");
1496 goto err;
1497 }
1498 }
1499
b2a9e649 1500 if (fio_option_is_set(o, numa_memnodes)) {
43522848
DG
1501 mask = NULL;
1502 if (o->numa_memnodes)
1503 mask = numa_parse_nodestring(o->numa_memnodes);
1504
4896473e 1505 switch (o->numa_mem_mode) {
d0b937ed 1506 case MPOL_INTERLEAVE:
43522848 1507 numa_set_interleave_mask(mask);
d0b937ed
YR
1508 break;
1509 case MPOL_BIND:
43522848 1510 numa_set_membind(mask);
d0b937ed
YR
1511 break;
1512 case MPOL_LOCAL:
1513 numa_set_localalloc();
1514 break;
1515 case MPOL_PREFERRED:
4896473e 1516 numa_set_preferred(o->numa_mem_prefer_node);
d0b937ed
YR
1517 break;
1518 case MPOL_DEFAULT:
1519 default:
1520 break;
1521 }
1522
43522848
DG
1523 if (mask)
1524 numa_free_nodemask(mask);
1525
d0b937ed
YR
1526 }
1527 }
1528#endif
1529
9a3f1100
JA
1530 if (fio_pin_memory(td))
1531 goto err;
1532
2e1df07d
JA
1533 /*
1534 * May alter parameters that init_io_u() will use, so we need to
1535 * do this first.
1536 */
1537 if (init_iolog(td))
1538 goto err;
1539
1540 if (init_io_u(td))
1541 goto err;
1542
4896473e 1543 if (o->verify_async && verify_async_init(td))
2e1df07d
JA
1544 goto err;
1545
27d74836
JA
1546 if (fio_option_is_set(o, ioprio) ||
1547 fio_option_is_set(o, ioprio_class)) {
28727df7
JA
1548 ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio);
1549 if (ret == -1) {
2e1df07d
JA
1550 td_verror(td, errno, "ioprio_set");
1551 goto err;
1552 }
1553 }
1554
4896473e 1555 if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
2e1df07d
JA
1556 goto err;
1557
649c10c5 1558 errno = 0;
4896473e 1559 if (nice(o->nice) == -1 && errno != 0) {
2e1df07d
JA
1560 td_verror(td, errno, "nice");
1561 goto err;
1562 }
1563
4896473e 1564 if (o->ioscheduler && switch_ioscheduler(td))
2e1df07d
JA
1565 goto err;
1566
4896473e 1567 if (!o->create_serialize && setup_files(td))
2e1df07d
JA
1568 goto err;
1569
1570 if (td_io_init(td))
1571 goto err;
1572
1573 if (init_random_map(td))
1574 goto err;
1575
ce486495 1576 if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
4896473e 1577 goto err;
2e1df07d 1578
4896473e 1579 if (o->pre_read) {
2e1df07d
JA
1580 if (pre_read_files(td) < 0)
1581 goto err;
1582 }
1583
24660963 1584 if (iolog_compress_init(td, sk_out))
155f2f02 1585 goto err;
aee2ab67 1586
dc5bfbb2
JA
1587 fio_verify_init(td);
1588
24660963 1589 if (rate_submit_init(td, sk_out))
a9da8ab2
JA
1590 goto err;
1591
2e1df07d 1592 fio_gettime(&td->epoch, NULL);
44404c5a 1593 fio_getrusage(&td->ru_start);
ac28d905
JA
1594 memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
1595 memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
1596
1597 if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
1598 o->ratemin[DDIR_TRIM]) {
1599 memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
1600 sizeof(td->bw_sample_time));
1601 memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
1602 sizeof(td->bw_sample_time));
1603 memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
1604 sizeof(td->bw_sample_time));
1605 }
1606
2e1df07d
JA
1607 clear_state = 0;
1608 while (keep_running(td)) {
100f49f1
JA
1609 uint64_t verify_bytes;
1610
2e1df07d 1611 fio_gettime(&td->start, NULL);
2e1df07d
JA
1612 memcpy(&td->tv_cache, &td->start, sizeof(td->start));
1613
2e1df07d 1614 if (clear_state)
ac28d905 1615 clear_io_state(td, 0);
2e1df07d
JA
1616
1617 prune_io_piece_log(td);
1618
62167762
JC
1619 if (td->o.verify_only && (td_write(td) || td_rw(td)))
1620 verify_bytes = do_dry_run(td);
095196b1 1621 else {
95b03be0
JA
1622 uint64_t bytes_done[DDIR_RWDIR_CNT];
1623
1624 do_io(td, bytes_done);
1625
1626 if (!ddir_rw_sum(bytes_done)) {
095196b1 1627 fio_mark_td_terminate(td);
95b03be0
JA
1628 verify_bytes = 0;
1629 } else {
1630 verify_bytes = bytes_done[DDIR_WRITE] +
1631 bytes_done[DDIR_TRIM];
1632 }
095196b1 1633 }
2e1df07d
JA
1634
1635 clear_state = 1;
1636
40c666c8
JA
1637 /*
1638 * Make sure we've successfully updated the rusage stats
1639 * before waiting on the stat mutex. Otherwise we could have
1640 * the stat thread holding stat mutex and waiting for
1641 * the rusage_sem, which would never get upped because
1642 * this thread is waiting for the stat mutex.
1643 */
1644 check_update_rusage(td);
1645
e5437a07 1646 fio_mutex_down(stat_mutex);
95603b74
BF
1647 if (td_read(td) && td->io_bytes[DDIR_READ])
1648 update_runtime(td, elapsed_us, DDIR_READ);
1649 if (td_write(td) && td->io_bytes[DDIR_WRITE])
1650 update_runtime(td, elapsed_us, DDIR_WRITE);
1651 if (td_trim(td) && td->io_bytes[DDIR_TRIM])
1652 update_runtime(td, elapsed_us, DDIR_TRIM);
e5437a07
VT
1653 fio_gettime(&td->start, NULL);
1654 fio_mutex_up(stat_mutex);
2e1df07d
JA
1655
1656 if (td->error || td->terminate)
1657 break;
1658
4896473e
JA
1659 if (!o->do_verify ||
1660 o->verify == VERIFY_NONE ||
2e1df07d
JA
1661 (td->io_ops->flags & FIO_UNIDIR))
1662 continue;
1663
ac28d905 1664 clear_io_state(td, 0);
2e1df07d
JA
1665
1666 fio_gettime(&td->start, NULL);
1667
100f49f1 1668 do_verify(td, verify_bytes);
2e1df07d 1669
40c666c8
JA
1670 /*
1671 * See comment further up for why this is done here.
1672 */
1673 check_update_rusage(td);
1674
e5437a07 1675 fio_mutex_down(stat_mutex);
95603b74 1676 update_runtime(td, elapsed_us, DDIR_READ);
e5437a07
VT
1677 fio_gettime(&td->start, NULL);
1678 fio_mutex_up(stat_mutex);
2e1df07d
JA
1679
1680 if (td->error || td->terminate)
1681 break;
1682 }
1683
1684 update_rusage_stat(td);
2e1df07d 1685 td->ts.total_run_time = mtime_since_now(&td->epoch);
6eaf09d6
SL
1686 td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
1687 td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
1688 td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
2e1df07d 1689
ca09be4b 1690 if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) &&
d264264a
JA
1691 (td->o.verify != VERIFY_NONE && td_write(td)))
1692 verify_save_state(td->thread_number);
ca09be4b 1693
9a3f1100
JA
1694 fio_unpin_memory(td);
1695
905e3d4f 1696 fio_writeout_logs(td);
2e1df07d 1697
155f2f02 1698 iolog_compress_exit(td);
103b174e 1699 rate_submit_exit(td);
aee2ab67 1700
4896473e 1701 if (o->exec_postrun)
ce486495 1702 exec_string(o, o->exec_postrun, (const char *)"postrun");
2e1df07d 1703
f9cafb12 1704 if (exitall_on_terminate || (o->exitall_error && td->error))
2e1df07d
JA
1705 fio_terminate_threads(td->groupid);
1706
1707err:
1708 if (td->error)
1709 log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
1710 td->verror);
1711
4896473e 1712 if (o->verify_async)
2e1df07d
JA
1713 verify_async_exit(td);
1714
1715 close_and_free_files(td);
2e1df07d 1716 cleanup_io_u(td);
32dbca2c 1717 close_ioengine(td);
2e1df07d 1718 cgroup_shutdown(td, &cgroup_mnt);
ca09be4b 1719 verify_free_state(td);
2e1df07d 1720
7c4d0fb7
JA
1721 if (td->zone_state_index) {
1722 int i;
1723
1724 for (i = 0; i < DDIR_RWDIR_CNT; i++)
1725 free(td->zone_state_index[i]);
1726 free(td->zone_state_index);
1727 td->zone_state_index = NULL;
1728 }
1729
b2a9e649 1730 if (fio_option_is_set(o, cpumask)) {
8a1db9a1
JA
1731 ret = fio_cpuset_exit(&o->cpumask);
1732 if (ret)
1733 td_verror(td, ret, "fio_cpuset_exit");
2e1df07d
JA
1734 }
1735
1736 /*
1737 * do this very late, it will log file closing as well
1738 */
4896473e 1739 if (o->write_iolog_file)
2e1df07d
JA
1740 write_iolog_close(td);
1741
ea66e04f
JA
1742 fio_mutex_remove(td->mutex);
1743 td->mutex = NULL;
1744
2e1df07d 1745 td_set_runstate(td, TD_EXITED);
fda2cfac
JA
1746
1747 /*
1748 * Do this last after setting our runstate to exited, so we
1749 * know that the stat thread is signaled.
1750 */
1751 check_update_rusage(td);
1752
24660963 1753 sk_out_drop();
e43606c2 1754 return (void *) (uintptr_t) td->error;
2e1df07d
JA
1755}
1756
1757
1758/*
1759 * We cannot pass the td data into a forked process, so attach the td and
1760 * pass it to the thread worker.
1761 */
24660963 1762static int fork_main(struct sk_out *sk_out, int shmid, int offset)
2e1df07d 1763{
24660963 1764 struct fork_data *fd;
2e1df07d
JA
1765 void *data, *ret;
1766
c8931876 1767#if !defined(__hpux) && !defined(CONFIG_NO_SHM)
2e1df07d
JA
1768 data = shmat(shmid, NULL, 0);
1769 if (data == (void *) -1) {
1770 int __err = errno;
1771
1772 perror("shmat");
1773 return __err;
1774 }
1775#else
1776 /*
1777 * HP-UX inherits shm mappings?
1778 */
1779 data = threads;
1780#endif
1781
24660963
JA
1782 fd = calloc(1, sizeof(*fd));
1783 fd->td = data + offset * sizeof(struct thread_data);
1784 fd->sk_out = sk_out;
1785 ret = thread_main(fd);
2e1df07d 1786 shmdt(data);
e43606c2 1787 return (int) (uintptr_t) ret;
2e1df07d
JA
1788}
1789
cba5460c
JA
1790static void dump_td_info(struct thread_data *td)
1791{
1792 log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
1793 "be stuck. Doing forceful exit of this job.\n", td->o.name,
1794 (unsigned long) time_since_now(&td->terminate_time));
1795}
1796
2e1df07d
JA
1797/*
1798 * Run over the job map and reap the threads that have exited, if any.
1799 */
1800static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
1801 unsigned int *m_rate)
1802{
1803 struct thread_data *td;
1804 unsigned int cputhreads, realthreads, pending;
1805 int i, status, ret;
1806
1807 /*
1808 * reap exited threads (TD_EXITED -> TD_REAPED)
1809 */
1810 realthreads = pending = cputhreads = 0;
1811 for_each_td(td, i) {
1812 int flags = 0;
1813
1814 /*
1815 * ->io_ops is NULL for a thread that has closed its
1816 * io engine
1817 */
1818 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1819 cputhreads++;
1820 else
1821 realthreads++;
1822
1823 if (!td->pid) {
1824 pending++;
1825 continue;
1826 }
1827 if (td->runstate == TD_REAPED)
1828 continue;
1829 if (td->o.use_thread) {
1830 if (td->runstate == TD_EXITED) {
1831 td_set_runstate(td, TD_REAPED);
1832 goto reaped;
1833 }
1834 continue;
1835 }
1836
1837 flags = WNOHANG;
1838 if (td->runstate == TD_EXITED)
1839 flags = 0;
1840
1841 /*
1842 * check if someone quit or got killed in an unusual way
1843 */
1844 ret = waitpid(td->pid, &status, flags);
1845 if (ret < 0) {
1846 if (errno == ECHILD) {
1847 log_err("fio: pid=%d disappeared %d\n",
1848 (int) td->pid, td->runstate);
a5e371a6 1849 td->sig = ECHILD;
2e1df07d
JA
1850 td_set_runstate(td, TD_REAPED);
1851 goto reaped;
1852 }
1853 perror("waitpid");
1854 } else if (ret == td->pid) {
1855 if (WIFSIGNALED(status)) {
1856 int sig = WTERMSIG(status);
1857
36d80bc7 1858 if (sig != SIGTERM && sig != SIGUSR2)
2e1df07d
JA
1859 log_err("fio: pid=%d, got signal=%d\n",
1860 (int) td->pid, sig);
a5e371a6 1861 td->sig = sig;
2e1df07d
JA
1862 td_set_runstate(td, TD_REAPED);
1863 goto reaped;
1864 }
1865 if (WIFEXITED(status)) {
1866 if (WEXITSTATUS(status) && !td->error)
1867 td->error = WEXITSTATUS(status);
1868
1869 td_set_runstate(td, TD_REAPED);
1870 goto reaped;
1871 }
1872 }
1873
cba5460c
JA
1874 /*
1875 * If the job is stuck, do a forceful timeout of it and
1876 * move on.
1877 */
1878 if (td->terminate &&
1879 time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
1880 dump_td_info(td);
1881 td_set_runstate(td, TD_REAPED);
1882 goto reaped;
1883 }
1884
2e1df07d
JA
1885 /*
1886 * thread is not dead, continue
1887 */
1888 pending++;
1889 continue;
1890reaped:
1891 (*nr_running)--;
342f4be4
JA
1892 (*m_rate) -= ddir_rw_sum(td->o.ratemin);
1893 (*t_rate) -= ddir_rw_sum(td->o.rate);
2e1df07d
JA
1894 if (!td->pid)
1895 pending--;
1896
1897 if (td->error)
1898 exit_value++;
1899
1900 done_secs += mtime_since_now(&td->epoch) / 1000;
4a88752a 1901 profile_td_exit(td);
2e1df07d
JA
1902 }
1903
1904 if (*nr_running == cputhreads && !pending && realthreads)
1905 fio_terminate_threads(TERMINATE_ALL);
1906}
1907
e39c0676 1908static bool __check_trigger_file(void)
ca09be4b
JA
1909{
1910 struct stat sb;
1911
1912 if (!trigger_file)
e39c0676 1913 return false;
ca09be4b
JA
1914
1915 if (stat(trigger_file, &sb))
e39c0676 1916 return false;
ca09be4b
JA
1917
1918 if (unlink(trigger_file) < 0)
1919 log_err("fio: failed to unlink %s: %s\n", trigger_file,
1920 strerror(errno));
1921
e39c0676 1922 return true;
ca09be4b
JA
1923}
1924
e39c0676 1925static bool trigger_timedout(void)
ca09be4b
JA
1926{
1927 if (trigger_timeout)
1928 return time_since_genesis() >= trigger_timeout;
1929
e39c0676 1930 return false;
ca09be4b
JA
1931}
1932
1933void exec_trigger(const char *cmd)
1934{
1935 int ret;
1936
1937 if (!cmd)
1938 return;
1939
1940 ret = system(cmd);
1941 if (ret == -1)
1942 log_err("fio: failed executing %s trigger\n", cmd);
1943}
1944
1945void check_trigger_file(void)
1946{
1947 if (__check_trigger_file() || trigger_timedout()) {
796fb3ce
JA
1948 if (nr_clients)
1949 fio_clients_send_trigger(trigger_remote_cmd);
1950 else {
d264264a 1951 verify_save_state(IO_LIST_ALL);
ca09be4b
JA
1952 fio_terminate_threads(TERMINATE_ALL);
1953 exec_trigger(trigger_cmd);
1954 }
1955 }
1956}
1957
1958static int fio_verify_load_state(struct thread_data *td)
1959{
1960 int ret;
1961
1962 if (!td->o.verify_state)
1963 return 0;
1964
1965 if (is_backend) {
1966 void *data;
c3546b53 1967 int ver;
ca09be4b
JA
1968
1969 ret = fio_server_get_verify_state(td->o.name,
c3546b53 1970 td->thread_number - 1, &data, &ver);
ca09be4b 1971 if (!ret)
c3546b53 1972 verify_convert_assign_state(td, data, ver);
ca09be4b
JA
1973 } else
1974 ret = verify_load_state(td, "local");
1975
1976 return ret;
1977}
1978
06464907
JA
1979static void do_usleep(unsigned int usecs)
1980{
1981 check_for_running_stats();
ca09be4b 1982 check_trigger_file();
06464907
JA
1983 usleep(usecs);
1984}
1985
e39c0676 1986static bool check_mount_writes(struct thread_data *td)
e81ecca3
JA
1987{
1988 struct fio_file *f;
1989 unsigned int i;
1990
1991 if (!td_write(td) || td->o.allow_mounted_write)
e39c0676 1992 return false;
e81ecca3
JA
1993
1994 for_each_file(td, f, i) {
1995 if (f->filetype != FIO_TYPE_BD)
1996 continue;
1997 if (device_is_mounted(f->file_name))
1998 goto mounted;
1999 }
2000
e39c0676 2001 return false;
e81ecca3
JA
2002mounted:
2003 log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
e39c0676 2004 return true;
e81ecca3
JA
2005}
2006
9cc8cb91
AK
2007static bool waitee_running(struct thread_data *me)
2008{
2009 const char *waitee = me->o.wait_for;
2010 const char *self = me->o.name;
2011 struct thread_data *td;
2012 int i;
2013
2014 if (!waitee)
2015 return false;
2016
2017 for_each_td(td, i) {
2018 if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee))
2019 continue;
2020
2021 if (td->runstate < TD_EXITED) {
2022 dprint(FD_PROCESS, "%s fenced by %s(%s)\n",
2023 self, td->o.name,
2024 runstate_to_name(td->runstate));
2025 return true;
2026 }
2027 }
2028
2029 dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee);
2030 return false;
2031}
2032
2e1df07d
JA
2033/*
2034 * Main function for kicking off and reaping jobs, as needed.
2035 */
24660963 2036static void run_threads(struct sk_out *sk_out)
2e1df07d
JA
2037{
2038 struct thread_data *td;
2e1df07d 2039 unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
0de5b26f 2040 uint64_t spent;
2e1df07d 2041
2e1df07d
JA
2042 if (fio_gtod_offload && fio_start_gtod_thread())
2043 return;
334185e9 2044
f2a2ce0e 2045 fio_idle_prof_init();
2e1df07d
JA
2046
2047 set_sig_handlers();
2048
3a5f6bde
JA
2049 nr_thread = nr_process = 0;
2050 for_each_td(td, i) {
e81ecca3
JA
2051 if (check_mount_writes(td))
2052 return;
3a5f6bde
JA
2053 if (td->o.use_thread)
2054 nr_thread++;
2055 else
2056 nr_process++;
2057 }
2058
01cfefcc 2059 if (output_format & FIO_OUTPUT_NORMAL) {
2e1df07d
JA
2060 log_info("Starting ");
2061 if (nr_thread)
2062 log_info("%d thread%s", nr_thread,
2063 nr_thread > 1 ? "s" : "");
2064 if (nr_process) {
2065 if (nr_thread)
2066 log_info(" and ");
2067 log_info("%d process%s", nr_process,
2068 nr_process > 1 ? "es" : "");
2069 }
2070 log_info("\n");
e411c301 2071 log_info_flush();
2e1df07d
JA
2072 }
2073
2074 todo = thread_number;
2075 nr_running = 0;
2076 nr_started = 0;
2077 m_rate = t_rate = 0;
2078
2079 for_each_td(td, i) {
2080 print_status_init(td->thread_number - 1);
2081
2082 if (!td->o.create_serialize)
2083 continue;
2084
ca09be4b
JA
2085 if (fio_verify_load_state(td))
2086 goto reap;
2087
2e1df07d
JA
2088 /*
2089 * do file setup here so it happens sequentially,
2090 * we don't want X number of threads getting their
2091 * client data interspersed on disk
2092 */
2093 if (setup_files(td)) {
ca09be4b 2094reap:
2e1df07d
JA
2095 exit_value++;
2096 if (td->error)
2097 log_err("fio: pid=%d, err=%d/%s\n",
2098 (int) td->pid, td->error, td->verror);
2099 td_set_runstate(td, TD_REAPED);
2100 todo--;
2101 } else {
2102 struct fio_file *f;
2103 unsigned int j;
2104
2105 /*
2106 * for sharing to work, each job must always open
2107 * its own files. so close them, if we opened them
2108 * for creation
2109 */
2110 for_each_file(td, f, j) {
2111 if (fio_file_open(f))
2112 td_io_close_file(td, f);
2113 }
2114 }
2115 }
2116
f2a2ce0e
HL
2117 /* start idle threads before io threads start to run */
2118 fio_idle_prof_start();
2119
2e1df07d
JA
2120 set_genesis_time();
2121
2122 while (todo) {
2123 struct thread_data *map[REAL_MAX_JOBS];
2124 struct timeval this_start;
2125 int this_jobs = 0, left;
2126
2127 /*
2128 * create threads (TD_NOT_CREATED -> TD_CREATED)
2129 */
2130 for_each_td(td, i) {
2131 if (td->runstate != TD_NOT_CREATED)
2132 continue;
2133
2134 /*
2135 * never got a chance to start, killed by other
2136 * thread for some reason
2137 */
2138 if (td->terminate) {
2139 todo--;
2140 continue;
2141 }
2142
2143 if (td->o.start_delay) {
0de5b26f 2144 spent = utime_since_genesis();
2e1df07d 2145
74454ce4 2146 if (td->o.start_delay > spent)
2e1df07d
JA
2147 continue;
2148 }
2149
2150 if (td->o.stonewall && (nr_started || nr_running)) {
2151 dprint(FD_PROCESS, "%s: stonewall wait\n",
2152 td->o.name);
2153 break;
2154 }
2155
9cc8cb91
AK
2156 if (waitee_running(td)) {
2157 dprint(FD_PROCESS, "%s: waiting for %s\n",
2158 td->o.name, td->o.wait_for);
5c74fc76 2159 continue;
9cc8cb91
AK
2160 }
2161
2e1df07d
JA
2162 init_disk_util(td);
2163
c97f1ad6
JA
2164 td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
2165 td->update_rusage = 0;
2166
2e1df07d
JA
2167 /*
2168 * Set state to created. Thread will transition
2169 * to TD_INITIALIZED when it's done setting up.
2170 */
2171 td_set_runstate(td, TD_CREATED);
2172 map[this_jobs++] = td;
2173 nr_started++;
2174
2175 if (td->o.use_thread) {
24660963 2176 struct fork_data *fd;
2e1df07d
JA
2177 int ret;
2178
24660963
JA
2179 fd = calloc(1, sizeof(*fd));
2180 fd->td = td;
2181 fd->sk_out = sk_out;
2182
2e1df07d
JA
2183 dprint(FD_PROCESS, "will pthread_create\n");
2184 ret = pthread_create(&td->thread, NULL,
24660963 2185 thread_main, fd);
2e1df07d
JA
2186 if (ret) {
2187 log_err("pthread_create: %s\n",
2188 strerror(ret));
24660963 2189 free(fd);
2e1df07d
JA
2190 nr_started--;
2191 break;
2192 }
2193 ret = pthread_detach(td->thread);
2194 if (ret)
2195 log_err("pthread_detach: %s",
2196 strerror(ret));
2197 } else {
2198 pid_t pid;
2199 dprint(FD_PROCESS, "will fork\n");
2200 pid = fork();
2201 if (!pid) {
24660963 2202 int ret = fork_main(sk_out, shm_id, i);
2e1df07d
JA
2203
2204 _exit(ret);
2205 } else if (i == fio_debug_jobno)
2206 *fio_debug_jobp = pid;
2207 }
2208 dprint(FD_MUTEX, "wait on startup_mutex\n");
09400a60 2209 if (fio_mutex_down_timeout(startup_mutex, 10000)) {
2e1df07d
JA
2210 log_err("fio: job startup hung? exiting.\n");
2211 fio_terminate_threads(TERMINATE_ALL);
2212 fio_abort = 1;
2213 nr_started--;
2214 break;
2215 }
2216 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2217 }
2218
2219 /*
2220 * Wait for the started threads to transition to
2221 * TD_INITIALIZED.
2222 */
2223 fio_gettime(&this_start, NULL);
2224 left = this_jobs;
2225 while (left && !fio_abort) {
2226 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
2227 break;
2228
06464907 2229 do_usleep(100000);
2e1df07d
JA
2230
2231 for (i = 0; i < this_jobs; i++) {
2232 td = map[i];
2233 if (!td)
2234 continue;
2235 if (td->runstate == TD_INITIALIZED) {
2236 map[i] = NULL;
2237 left--;
2238 } else if (td->runstate >= TD_EXITED) {
2239 map[i] = NULL;
2240 left--;
2241 todo--;
2242 nr_running++; /* work-around... */
2243 }
2244 }
2245 }
2246
2247 if (left) {
4e87c37a
JA
2248 log_err("fio: %d job%s failed to start\n", left,
2249 left > 1 ? "s" : "");
2e1df07d
JA
2250 for (i = 0; i < this_jobs; i++) {
2251 td = map[i];
2252 if (!td)
2253 continue;
2254 kill(td->pid, SIGTERM);
2255 }
2256 break;
2257 }
2258
2259 /*
2260 * start created threads (TD_INITIALIZED -> TD_RUNNING).
2261 */
2262 for_each_td(td, i) {
2263 if (td->runstate != TD_INITIALIZED)
2264 continue;
2265
2266 if (in_ramp_time(td))
2267 td_set_runstate(td, TD_RAMP);
2268 else
2269 td_set_runstate(td, TD_RUNNING);
2270 nr_running++;
2271 nr_started--;
342f4be4
JA
2272 m_rate += ddir_rw_sum(td->o.ratemin);
2273 t_rate += ddir_rw_sum(td->o.rate);
2e1df07d
JA
2274 todo--;
2275 fio_mutex_up(td->mutex);
2276 }
2277
2278 reap_threads(&nr_running, &t_rate, &m_rate);
2279
122c7725 2280 if (todo)
06464907 2281 do_usleep(100000);
2e1df07d
JA
2282 }
2283
2284 while (nr_running) {
2285 reap_threads(&nr_running, &t_rate, &m_rate);
06464907 2286 do_usleep(10000);
2e1df07d
JA
2287 }
2288
f2a2ce0e
HL
2289 fio_idle_prof_stop();
2290
2e1df07d 2291 update_io_ticks();
2e1df07d
JA
2292}
2293
5ddc6707 2294static void wait_for_helper_thread_exit(void)
9ec7779f 2295{
8aab824f
JA
2296 void *ret;
2297
98b4b0a2 2298 helper_exit = 1;
5ddc6707
JA
2299 pthread_cond_signal(&helper_cond);
2300 pthread_join(helper_thread, &ret);
9ec7779f
JA
2301}
2302
27357187
JA
2303static void free_disk_util(void)
2304{
27357187 2305 disk_util_prune_entries();
8aab824f 2306
5ddc6707 2307 pthread_cond_destroy(&helper_cond);
27357187
JA
2308}
2309
5ddc6707 2310static void *helper_thread_main(void *data)
2e1df07d 2311{
24660963 2312 struct sk_out *sk_out = data;
9ec7779f
JA
2313 int ret = 0;
2314
24660963 2315 sk_out_assign(sk_out);
2a2804c9 2316
2e1df07d
JA
2317 fio_mutex_up(startup_mutex);
2318
8aab824f
JA
2319 while (!ret) {
2320 uint64_t sec = DISK_UTIL_MSEC / 1000;
2321 uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
2322 struct timespec ts;
2323 struct timeval tv;
2324
2325 gettimeofday(&tv, NULL);
2326 ts.tv_sec = tv.tv_sec + sec;
2327 ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
f893b76d 2328
823952af 2329 if (ts.tv_nsec >= 1000000000ULL) {
8aab824f
JA
2330 ts.tv_nsec -= 1000000000ULL;
2331 ts.tv_sec++;
2332 }
2333
5ddc6707 2334 pthread_cond_timedwait(&helper_cond, &helper_lock, &ts);
8aab824f 2335
9ec7779f 2336 ret = update_io_ticks();
2e1df07d 2337
5ddc6707
JA
2338 if (helper_do_stat) {
2339 helper_do_stat = 0;
2340 __show_running_run_stats();
2341 }
2342
2e1df07d
JA
2343 if (!is_backend)
2344 print_thread_status();
2345 }
2346
24660963 2347 sk_out_drop();
2e1df07d
JA
2348 return NULL;
2349}
2350
24660963 2351static int create_helper_thread(struct sk_out *sk_out)
2e1df07d
JA
2352{
2353 int ret;
2354
9ec7779f
JA
2355 setup_disk_util();
2356
5ddc6707
JA
2357 pthread_cond_init(&helper_cond, NULL);
2358 pthread_mutex_init(&helper_lock, NULL);
8aab824f 2359
24660963 2360 ret = pthread_create(&helper_thread, NULL, helper_thread_main, sk_out);
2e1df07d 2361 if (ret) {
5ddc6707 2362 log_err("Can't create helper thread: %s\n", strerror(ret));
2e1df07d
JA
2363 return 1;
2364 }
2365
2e1df07d
JA
2366 dprint(FD_MUTEX, "wait on startup_mutex\n");
2367 fio_mutex_down(startup_mutex);
2368 dprint(FD_MUTEX, "done waiting on startup_mutex\n");
2369 return 0;
2370}
2371
24660963 2372int fio_backend(struct sk_out *sk_out)
2e1df07d
JA
2373{
2374 struct thread_data *td;
2375 int i;
2376
2377 if (exec_profile) {
2378 if (load_profile(exec_profile))
2379 return 1;
2380 free(exec_profile);
2381 exec_profile = NULL;
2382 }
2383 if (!thread_number)
2384 return 0;
2385
2386 if (write_bw_log) {
aee2ab67
JA
2387 struct log_params p = {
2388 .log_type = IO_LOG_TYPE_BW,
2389 };
2390
2391 setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
2392 setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
2393 setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
2e1df07d
JA
2394 }
2395
521da527 2396 startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
2e1df07d
JA
2397 if (startup_mutex == NULL)
2398 return 1;
2e1df07d
JA
2399
2400 set_genesis_time();
cef9175e 2401 stat_init();
24660963 2402 create_helper_thread(sk_out);
2e1df07d
JA
2403
2404 cgroup_list = smalloc(sizeof(*cgroup_list));
2405 INIT_FLIST_HEAD(cgroup_list);
2406
24660963 2407 run_threads(sk_out);
2e1df07d 2408
5ddc6707 2409 wait_for_helper_thread_exit();
8aab824f 2410
2e1df07d 2411 if (!fio_abort) {
83f7b64e 2412 __show_run_stats();
2e1df07d 2413 if (write_bw_log) {
cb7e0ace
JA
2414 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2415 struct io_log *log = agg_io_log[i];
2416
3a5db920 2417 flush_log(log, 0);
518dac09 2418 free_log(log);
cb7e0ace 2419 }
2e1df07d
JA
2420 }
2421 }
2422
fda2cfac 2423 for_each_td(td, i) {
2e1df07d 2424 fio_options_free(td);
8049adc1
JA
2425 if (td->rusage_sem) {
2426 fio_mutex_remove(td->rusage_sem);
2427 td->rusage_sem = NULL;
2428 }
fda2cfac 2429 }
2e1df07d 2430
a462baef 2431 free_disk_util();
2e1df07d
JA
2432 cgroup_kill(cgroup_list);
2433 sfree(cgroup_list);
2434 sfree(cgroup_mnt);
2435
2436 fio_mutex_remove(startup_mutex);
cef9175e 2437 stat_exit();
2e1df07d
JA
2438 return exit_value;
2439}