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