projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Random IO fixes
[fio.git] / fio.c
... / ...
CommitLineData
1/*
2 * fio - the flexible io tester
3 *
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 * Copyright (C) 2006 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 <signal.h>
28#include <time.h>
29#include <locale.h>
30#include <assert.h>
31#include <sys/stat.h>
32#include <sys/wait.h>
33#include <sys/ipc.h>
34#include <sys/shm.h>
35#include <sys/mman.h>
36
37#include "fio.h"
38#include "hash.h"
39
40unsigned long page_mask;
41unsigned long page_size;
42#define ALIGN(buf) \
43 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
44
45int groupid = 0;
46int thread_number = 0;
47int nr_process = 0;
48int nr_thread = 0;
49int shm_id = 0;
50int temp_stall_ts;
51
52static struct fio_sem *startup_sem;
53static volatile int fio_abort;
54static int exit_value;
55
56struct io_log *agg_io_log[2];
57
58#define TERMINATE_ALL (-1)
59#define JOB_START_TIMEOUT (5 * 1000)
60
61static inline void td_set_runstate(struct thread_data *td, int runstate)
62{
63 td->runstate = runstate;
64}
65
66static void terminate_threads(int group_id)
67{
68 struct thread_data *td;
69 int i;
70
71 for_each_td(td, i) {
72 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
73 /*
74 * if the thread is running, just let it exit
75 */
76 if (td->runstate < TD_RUNNING)
77 kill(td->pid, SIGQUIT);
78 td->terminate = 1;
79 td->o.start_delay = 0;
80 }
81 }
82}
83
84static void sig_handler(int sig)
85{
86 switch (sig) {
87 case SIGALRM:
88 update_io_ticks();
89 disk_util_timer_arm();
90 print_thread_status();
91 break;
92 default:
93 printf("\nfio: terminating on signal %d\n", sig);
94 fflush(stdout);
95 terminate_threads(TERMINATE_ALL);
96 break;
97 }
98}
99
100/*
101 * Check if we are above the minimum rate given.
102 */
103static int check_min_rate(struct thread_data *td, struct timeval *now)
104{
105 unsigned long long bytes = 0;
106 unsigned long iops = 0;
107 unsigned long spent;
108 unsigned long rate;
109
110 /*
111 * No minimum rate set, always ok
112 */
113 if (!td->o.ratemin && !td->o.rate_iops_min)
114 return 0;
115
116 /*
117 * allow a 2 second settle period in the beginning
118 */
119 if (mtime_since(&td->start, now) < 2000)
120 return 0;
121
122 if (td_read(td)) {
123 iops += td->io_blocks[DDIR_READ];
124 bytes += td->this_io_bytes[DDIR_READ];
125 }
126 if (td_write(td)) {
127 iops += td->io_blocks[DDIR_WRITE];
128 bytes += td->this_io_bytes[DDIR_WRITE];
129 }
130
131 /*
132 * if rate blocks is set, sample is running
133 */
134 if (td->rate_bytes || td->rate_blocks) {
135 spent = mtime_since(&td->lastrate, now);
136 if (spent < td->o.ratecycle)
137 return 0;
138
139 if (td->o.rate) {
140 /*
141 * check bandwidth specified rate
142 */
143 if (bytes < td->rate_bytes) {
144 log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
145 return 1;
146 } else {
147 rate = (bytes - td->rate_bytes) / spent;
148 if (rate < td->o.ratemin || bytes < td->rate_bytes) {
149 log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
150 return 1;
151 }
152 }
153 } else {
154 /*
155 * checks iops specified rate
156 */
157 if (iops < td->o.rate_iops) {
158 log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
159 return 1;
160 } else {
161 rate = (iops - td->rate_blocks) / spent;
162 if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
163 log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
164 }
165 }
166 }
167 }
168
169 td->rate_bytes = bytes;
170 td->rate_blocks = iops;
171 memcpy(&td->lastrate, now, sizeof(*now));
172 return 0;
173}
174
175static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
176{
177 if (!td->o.timeout)
178 return 0;
179 if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
180 return 1;
181
182 return 0;
183}
184
185/*
186 * When job exits, we can cancel the in-flight IO if we are using async
187 * io. Attempt to do so.
188 */
189static void cleanup_pending_aio(struct thread_data *td)
190{
191 struct list_head *entry, *n;
192 struct io_u *io_u;
193 int r;
194
195 /*
196 * get immediately available events, if any
197 */
198 r = io_u_queued_complete(td, 0);
199 if (r < 0)
200 return;
201
202 /*
203 * now cancel remaining active events
204 */
205 if (td->io_ops->cancel) {
206 list_for_each_safe(entry, n, &td->io_u_busylist) {
207 io_u = list_entry(entry, struct io_u, list);
208
209 /*
210 * if the io_u isn't in flight, then that generally
211 * means someone leaked an io_u. complain but fix
212 * it up, so we don't stall here.
213 */
214 if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
215 log_err("fio: non-busy IO on busy list\n");
216 put_io_u(td, io_u);
217 } else {
218 r = td->io_ops->cancel(td, io_u);
219 if (!r)
220 put_io_u(td, io_u);
221 }
222 }
223 }
224
225 if (td->cur_depth)
226 r = io_u_queued_complete(td, td->cur_depth);
227}
228
229/*
230 * Helper to handle the final sync of a file. Works just like the normal
231 * io path, just does everything sync.
232 */
233static int fio_io_sync(struct thread_data *td, struct fio_file *f)
234{
235 struct io_u *io_u = __get_io_u(td);
236 int ret;
237
238 if (!io_u)
239 return 1;
240
241 io_u->ddir = DDIR_SYNC;
242 io_u->file = f;
243
244 if (td_io_prep(td, io_u)) {
245 put_io_u(td, io_u);
246 return 1;
247 }
248
249requeue:
250 ret = td_io_queue(td, io_u);
251 if (ret < 0) {
252 td_verror(td, io_u->error, "td_io_queue");
253 put_io_u(td, io_u);
254 return 1;
255 } else if (ret == FIO_Q_QUEUED) {
256 if (io_u_queued_complete(td, 1) < 0)
257 return 1;
258 } else if (ret == FIO_Q_COMPLETED) {
259 if (io_u->error) {
260 td_verror(td, io_u->error, "td_io_queue");
261 return 1;
262 }
263
264 if (io_u_sync_complete(td, io_u) < 0)
265 return 1;
266 } else if (ret == FIO_Q_BUSY) {
267 if (td_io_commit(td))
268 return 1;
269 goto requeue;
270 }
271
272 return 0;
273}
274
275/*
276 * The main verify engine. Runs over the writes we previously submitted,
277 * reads the blocks back in, and checks the crc/md5 of the data.
278 */
279static void do_verify(struct thread_data *td)
280{
281 struct fio_file *f;
282 struct io_u *io_u;
283 int ret, min_events;
284 unsigned int i;
285
286 /*
287 * sync io first and invalidate cache, to make sure we really
288 * read from disk.
289 */
290 for_each_file(td, f, i) {
291 if (!(f->flags & FIO_FILE_OPEN))
292 continue;
293 if (fio_io_sync(td, f))
294 break;
295 if (file_invalidate_cache(td, f))
296 break;
297 }
298
299 if (td->error)
300 return;
301
302 td_set_runstate(td, TD_VERIFYING);
303
304 io_u = NULL;
305 while (!td->terminate) {
306 int ret2;
307
308 io_u = __get_io_u(td);
309 if (!io_u)
310 break;
311
312 if (runtime_exceeded(td, &io_u->start_time)) {
313 put_io_u(td, io_u);
314 td->terminate = 1;
315 break;
316 }
317
318 if (get_next_verify(td, io_u)) {
319 put_io_u(td, io_u);
320 break;
321 }
322
323 if (td_io_prep(td, io_u)) {
324 put_io_u(td, io_u);
325 break;
326 }
327
328 io_u->end_io = verify_io_u;
329
330 ret = td_io_queue(td, io_u);
331 switch (ret) {
332 case FIO_Q_COMPLETED:
333 if (io_u->error)
334 ret = -io_u->error;
335 else if (io_u->resid) {
336 int bytes = io_u->xfer_buflen - io_u->resid;
337 struct fio_file *f = io_u->file;
338
339 /*
340 * zero read, fail
341 */
342 if (!bytes) {
343 td_verror(td, ENODATA, "full resid");
344 put_io_u(td, io_u);
345 break;
346 }
347
348 io_u->xfer_buflen = io_u->resid;
349 io_u->xfer_buf += bytes;
350 io_u->offset += bytes;
351 f->last_completed_pos = io_u->offset;
352
353 td->ts.short_io_u[io_u->ddir]++;
354
355 if (io_u->offset == f->real_file_size)
356 goto sync_done;
357
358 requeue_io_u(td, &io_u);
359 } else {
360sync_done:
361 ret = io_u_sync_complete(td, io_u);
362 if (ret < 0)
363 break;
364 }
365 continue;
366 case FIO_Q_QUEUED:
367 break;
368 case FIO_Q_BUSY:
369 requeue_io_u(td, &io_u);
370 ret2 = td_io_commit(td);
371 if (ret2 < 0)
372 ret = ret2;
373 break;
374 default:
375 assert(ret < 0);
376 td_verror(td, -ret, "td_io_queue");
377 break;
378 }
379
380 if (ret < 0 || td->error)
381 break;
382
383 /*
384 * if we can queue more, do so. but check if there are
385 * completed io_u's first.
386 */
387 min_events = 0;
388 if (queue_full(td) || ret == FIO_Q_BUSY) {
389 min_events = 1;
390
391 if (td->cur_depth > td->o.iodepth_low)
392 min_events = td->cur_depth - td->o.iodepth_low;
393 }
394
395 /*
396 * Reap required number of io units, if any, and do the
397 * verification on them through the callback handler
398 */
399 if (io_u_queued_complete(td, min_events) < 0)
400 break;
401 }
402
403 if (!td->error) {
404 min_events = td->cur_depth;
405
406 if (min_events)
407 ret = io_u_queued_complete(td, min_events);
408 } else
409 cleanup_pending_aio(td);
410
411 td_set_runstate(td, TD_RUNNING);
412}
413
414/*
415 * Main IO worker function. It retrieves io_u's to process and queues
416 * and reaps them, checking for rate and errors along the way.
417 */
418static void do_io(struct thread_data *td)
419{
420 struct timeval s;
421 unsigned long usec;
422 unsigned int i;
423 int ret = 0;
424
425 td_set_runstate(td, TD_RUNNING);
426
427 while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
428 struct timeval comp_time;
429 long bytes_done = 0;
430 int min_evts = 0;
431 struct io_u *io_u;
432 int ret2;
433
434 if (td->terminate)
435 break;
436
437 io_u = get_io_u(td);
438 if (!io_u)
439 break;
440
441 memcpy(&s, &io_u->start_time, sizeof(s));
442
443 if (runtime_exceeded(td, &s)) {
444 put_io_u(td, io_u);
445 td->terminate = 1;
446 break;
447 }
448
449 /*
450 * Add verification end_io handler, if asked to verify
451 * a previously written file.
452 */
453 if (td->o.verify != VERIFY_NONE)
454 io_u->end_io = verify_io_u;
455
456 ret = td_io_queue(td, io_u);
457 switch (ret) {
458 case FIO_Q_COMPLETED:
459 if (io_u->error)
460 ret = -io_u->error;
461 else if (io_u->resid) {
462 int bytes = io_u->xfer_buflen - io_u->resid;
463 struct fio_file *f = io_u->file;
464
465 /*
466 * zero read, fail
467 */
468 if (!bytes) {
469 td_verror(td, ENODATA, "full resid");
470 put_io_u(td, io_u);
471 break;
472 }
473
474 io_u->xfer_buflen = io_u->resid;
475 io_u->xfer_buf += bytes;
476 io_u->offset += bytes;
477 f->last_completed_pos = io_u->offset;
478
479 td->ts.short_io_u[io_u->ddir]++;
480
481 if (io_u->offset == f->real_file_size)
482 goto sync_done;
483
484 requeue_io_u(td, &io_u);
485 } else {
486sync_done:
487 fio_gettime(&comp_time, NULL);
488 bytes_done = io_u_sync_complete(td, io_u);
489 if (bytes_done < 0)
490 ret = bytes_done;
491 }
492 break;
493 case FIO_Q_QUEUED:
494 /*
495 * if the engine doesn't have a commit hook,
496 * the io_u is really queued. if it does have such
497 * a hook, it has to call io_u_queued() itself.
498 */
499 if (td->io_ops->commit == NULL)
500 io_u_queued(td, io_u);
501 break;
502 case FIO_Q_BUSY:
503 requeue_io_u(td, &io_u);
504 ret2 = td_io_commit(td);
505 if (ret2 < 0)
506 ret = ret2;
507 break;
508 default:
509 assert(ret < 0);
510 put_io_u(td, io_u);
511 break;
512 }
513
514 if (ret < 0 || td->error)
515 break;
516
517 /*
518 * See if we need to complete some commands
519 */
520 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
521 min_evts = 0;
522 if (queue_full(td) || ret == FIO_Q_BUSY) {
523 min_evts = 1;
524
525 if (td->cur_depth > td->o.iodepth_low)
526 min_evts = td->cur_depth - td->o.iodepth_low;
527 }
528
529 fio_gettime(&comp_time, NULL);
530 bytes_done = io_u_queued_complete(td, min_evts);
531 if (bytes_done < 0)
532 break;
533 }
534
535 if (!bytes_done)
536 continue;
537
538 /*
539 * the rate is batched for now, it should work for batches
540 * of completions except the very first one which may look
541 * a little bursty
542 */
543 usec = utime_since(&s, &comp_time);
544
545 rate_throttle(td, usec, bytes_done);
546
547 if (check_min_rate(td, &comp_time)) {
548 if (exitall_on_terminate)
549 terminate_threads(td->groupid);
550 td_verror(td, ENODATA, "check_min_rate");
551 break;
552 }
553
554 if (td->o.thinktime) {
555 unsigned long long b;
556
557 b = td->io_blocks[0] + td->io_blocks[1];
558 if (!(b % td->o.thinktime_blocks)) {
559 int left;
560
561 if (td->o.thinktime_spin)
562 __usec_sleep(td->o.thinktime_spin);
563
564 left = td->o.thinktime - td->o.thinktime_spin;
565 if (left)
566 usec_sleep(td, left);
567 }
568 }
569 }
570
571 if (td->o.fill_device && td->error == ENOSPC) {
572 td->error = 0;
573 td->terminate = 1;
574 }
575 if (!td->error) {
576 struct fio_file *f;
577
578 i = td->cur_depth;
579 if (i)
580 ret = io_u_queued_complete(td, i);
581
582 if (should_fsync(td) && td->o.end_fsync) {
583 td_set_runstate(td, TD_FSYNCING);
584
585 for_each_file(td, f, i) {
586 if (!(f->flags & FIO_FILE_OPEN))
587 continue;
588 fio_io_sync(td, f);
589 }
590 }
591 } else
592 cleanup_pending_aio(td);
593}
594
595static void cleanup_io_u(struct thread_data *td)
596{
597 struct list_head *entry, *n;
598 struct io_u *io_u;
599
600 list_for_each_safe(entry, n, &td->io_u_freelist) {
601 io_u = list_entry(entry, struct io_u, list);
602
603 list_del(&io_u->list);
604 free(io_u);
605 }
606
607 free_io_mem(td);
608}
609
610/*
611 * "randomly" fill the buffer contents
612 */
613static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
614{
615 long *ptr = io_u->buf;
616
617 if (!td->o.zero_buffers) {
618 while ((void *) ptr - io_u->buf < max_bs) {
619 *ptr = rand() * GOLDEN_RATIO_PRIME;
620 ptr++;
621 }
622 } else
623 memset(ptr, 0, max_bs);
624}
625
626static int init_io_u(struct thread_data *td)
627{
628 struct io_u *io_u;
629 unsigned int max_bs;
630 int i, max_units;
631 char *p;
632
633 if (td->io_ops->flags & FIO_SYNCIO)
634 max_units = 1;
635 else
636 max_units = td->o.iodepth;
637
638 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
639 td->orig_buffer_size = (unsigned long long) max_bs * (unsigned long long) max_units;
640
641 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
642 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
643
644 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
645 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
646 return 1;
647 }
648
649 if (allocate_io_mem(td))
650 return 1;
651
652 if (td->o.odirect)
653 p = ALIGN(td->orig_buffer);
654 else
655 p = td->orig_buffer;
656
657 for (i = 0; i < max_units; i++) {
658 if (td->terminate)
659 return 1;
660 io_u = malloc(sizeof(*io_u));
661 memset(io_u, 0, sizeof(*io_u));
662 INIT_LIST_HEAD(&io_u->list);
663
664 if (!(td->io_ops->flags & FIO_NOIO)) {
665 io_u->buf = p + max_bs * i;
666
667 if (td_write(td))
668 fill_io_buf(td, io_u, max_bs);
669 }
670
671 io_u->index = i;
672 io_u->flags = IO_U_F_FREE;
673 list_add(&io_u->list, &td->io_u_freelist);
674 }
675
676 io_u_init_timeout();
677
678 return 0;
679}
680
681static int switch_ioscheduler(struct thread_data *td)
682{
683 char tmp[256], tmp2[128];
684 FILE *f;
685 int ret;
686
687 if (td->io_ops->flags & FIO_DISKLESSIO)
688 return 0;
689
690 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
691
692 f = fopen(tmp, "r+");
693 if (!f) {
694 if (errno == ENOENT) {
695 log_err("fio: os or kernel doesn't support IO scheduler switching\n");
696 return 0;
697 }
698 td_verror(td, errno, "fopen iosched");
699 return 1;
700 }
701
702 /*
703 * Set io scheduler.
704 */
705 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
706 if (ferror(f) || ret != 1) {
707 td_verror(td, errno, "fwrite");
708 fclose(f);
709 return 1;
710 }
711
712 rewind(f);
713
714 /*
715 * Read back and check that the selected scheduler is now the default.
716 */
717 ret = fread(tmp, 1, sizeof(tmp), f);
718 if (ferror(f) || ret < 0) {
719 td_verror(td, errno, "fread");
720 fclose(f);
721 return 1;
722 }
723
724 sprintf(tmp2, "[%s]", td->o.ioscheduler);
725 if (!strstr(tmp, tmp2)) {
726 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
727 td_verror(td, EINVAL, "iosched_switch");
728 fclose(f);
729 return 1;
730 }
731
732 fclose(f);
733 return 0;
734}
735
736static int keep_running(struct thread_data *td)
737{
738 unsigned long long io_done;
739
740 if (td->done)
741 return 0;
742 if (td->o.time_based)
743 return 1;
744 if (td->o.loops) {
745 td->o.loops--;
746 return 1;
747 }
748
749 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE] + td->io_skip_bytes;
750 if (io_done < td->o.size)
751 return 1;
752
753 return 0;
754}
755
756static int clear_io_state(struct thread_data *td)
757{
758 struct fio_file *f;
759 unsigned int i;
760 int ret;
761
762 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
763 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
764 td->zone_bytes = 0;
765 td->rate_bytes = 0;
766 td->rate_blocks = 0;
767 td->rw_end_set[0] = td->rw_end_set[1] = 0;
768
769 td->last_was_sync = 0;
770
771 if (td->o.time_based)
772 td->nr_done_files = 0;
773
774 for_each_file(td, f, i)
775 td_io_close_file(td, f);
776
777 ret = 0;
778 for_each_file(td, f, i) {
779 f->flags &= ~FIO_FILE_DONE;
780 ret = td_io_open_file(td, f);
781 if (ret)
782 break;
783 }
784
785 return ret;
786}
787
788/*
789 * Entry point for the thread based jobs. The process based jobs end up
790 * here as well, after a little setup.
791 */
792static void *thread_main(void *data)
793{
794 unsigned long long runtime[2], elapsed;
795 struct thread_data *td = data;
796 int clear_state;
797
798 if (!td->o.use_thread)
799 setsid();
800
801 td->pid = getpid();
802
803 INIT_LIST_HEAD(&td->io_u_freelist);
804 INIT_LIST_HEAD(&td->io_u_busylist);
805 INIT_LIST_HEAD(&td->io_u_requeues);
806 INIT_LIST_HEAD(&td->io_log_list);
807 INIT_LIST_HEAD(&td->io_hist_list);
808 td->io_hist_tree = RB_ROOT;
809
810 td_set_runstate(td, TD_INITIALIZED);
811 fio_sem_up(startup_sem);
812 fio_sem_down(td->mutex);
813
814 /*
815 * the ->mutex semaphore is now no longer used, close it to avoid
816 * eating a file descriptor
817 */
818 fio_sem_remove(td->mutex);
819
820 /*
821 * May alter parameters that init_io_u() will use, so we need to
822 * do this first.
823 */
824 if (init_iolog(td))
825 goto err;
826
827 if (init_io_u(td))
828 goto err;
829
830 if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
831 td_verror(td, errno, "cpu_set_affinity");
832 goto err;
833 }
834
835 if (td->ioprio_set) {
836 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
837 td_verror(td, errno, "ioprio_set");
838 goto err;
839 }
840 }
841
842 if (nice(td->o.nice) == -1) {
843 td_verror(td, errno, "nice");
844 goto err;
845 }
846
847 if (td->o.ioscheduler && switch_ioscheduler(td))
848 goto err;
849
850 if (!td->o.create_serialize && setup_files(td))
851 goto err;
852
853 if (td_io_init(td))
854 goto err;
855
856 if (open_files(td))
857 goto err;
858
859 if (init_random_map(td))
860 goto err;
861
862 if (td->o.exec_prerun) {
863 if (system(td->o.exec_prerun) < 0)
864 goto err;
865 }
866
867 fio_gettime(&td->epoch, NULL);
868 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
869 getrusage(RUSAGE_SELF, &td->ts.ru_start);
870
871 runtime[0] = runtime[1] = 0;
872 clear_state = 0;
873 while (keep_running(td)) {
874 fio_gettime(&td->start, NULL);
875 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
876
877 if (td->o.ratemin)
878 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
879
880 if (clear_state && clear_io_state(td))
881 break;
882
883 prune_io_piece_log(td);
884
885 do_io(td);
886
887 clear_state = 1;
888
889 if (td_read(td) && td->io_bytes[DDIR_READ]) {
890 if (td->rw_end_set[DDIR_READ])
891 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
892 else
893 elapsed = utime_since_now(&td->start);
894
895 runtime[DDIR_READ] += elapsed;
896 }
897 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
898 if (td->rw_end_set[DDIR_WRITE])
899 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
900 else
901 elapsed = utime_since_now(&td->start);
902
903 runtime[DDIR_WRITE] += elapsed;
904 }
905
906 if (td->error || td->terminate)
907 break;
908
909 if (!td->o.do_verify ||
910 td->o.verify == VERIFY_NONE ||
911 (td->io_ops->flags & FIO_UNIDIR))
912 continue;
913
914 if (clear_io_state(td))
915 break;
916
917 fio_gettime(&td->start, NULL);
918
919 do_verify(td);
920
921 runtime[DDIR_READ] += utime_since_now(&td->start);
922
923 if (td->error || td->terminate)
924 break;
925 }
926
927 update_rusage_stat(td);
928 td->ts.runtime[0] = (runtime[0] + 999) / 1000;
929 td->ts.runtime[1] = (runtime[1] + 999) / 1000;
930 td->ts.total_run_time = mtime_since_now(&td->epoch);
931 td->ts.io_bytes[0] = td->io_bytes[0];
932 td->ts.io_bytes[1] = td->io_bytes[1];
933
934 if (td->ts.bw_log)
935 finish_log(td, td->ts.bw_log, "bw");
936 if (td->ts.slat_log)
937 finish_log(td, td->ts.slat_log, "slat");
938 if (td->ts.clat_log)
939 finish_log(td, td->ts.clat_log, "clat");
940 if (td->o.exec_postrun) {
941 if (system(td->o.exec_postrun) < 0)
942 log_err("fio: postrun %s failed\n", td->o.exec_postrun);
943 }
944
945 if (exitall_on_terminate)
946 terminate_threads(td->groupid);
947
948err:
949 if (td->error)
950 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
951 close_files(td);
952 close_ioengine(td);
953 cleanup_io_u(td);
954
955 /*
956 * do this very late, it will log file closing as well
957 */
958 if (td->o.write_iolog_file)
959 write_iolog_close(td);
960
961 options_mem_free(td);
962 td_set_runstate(td, TD_EXITED);
963 return (void *) (unsigned long) td->error;
964}
965
966/*
967 * We cannot pass the td data into a forked process, so attach the td and
968 * pass it to the thread worker.
969 */
970static int fork_main(int shmid, int offset)
971{
972 struct thread_data *td;
973 void *data, *ret;
974
975 data = shmat(shmid, NULL, 0);
976 if (data == (void *) -1) {
977 int __err = errno;
978
979 perror("shmat");
980 return __err;
981 }
982
983 td = data + offset * sizeof(struct thread_data);
984 ret = thread_main(td);
985 shmdt(data);
986 return (int) (unsigned long) ret;
987}
988
989/*
990 * Run over the job map and reap the threads that have exited, if any.
991 */
992static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
993{
994 struct thread_data *td;
995 int i, cputhreads, realthreads, pending, status, ret;
996
997 /*
998 * reap exited threads (TD_EXITED -> TD_REAPED)
999 */
1000 realthreads = pending = cputhreads = 0;
1001 for_each_td(td, i) {
1002 int flags = 0;
1003
1004 /*
1005 * ->io_ops is NULL for a thread that has closed its
1006 * io engine
1007 */
1008 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
1009 cputhreads++;
1010 else
1011 realthreads++;
1012
1013 if (!td->pid || td->runstate == TD_REAPED)
1014 continue;
1015 if (td->o.use_thread) {
1016 if (td->runstate == TD_EXITED) {
1017 td_set_runstate(td, TD_REAPED);
1018 goto reaped;
1019 }
1020 continue;
1021 }
1022
1023 flags = WNOHANG;
1024 if (td->runstate == TD_EXITED)
1025 flags = 0;
1026
1027 /*
1028 * check if someone quit or got killed in an unusual way
1029 */
1030 ret = waitpid(td->pid, &status, flags);
1031 if (ret < 0) {
1032 if (errno == ECHILD) {
1033 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
1034 td_set_runstate(td, TD_REAPED);
1035 goto reaped;
1036 }
1037 perror("waitpid");
1038 } else if (ret == td->pid) {
1039 if (WIFSIGNALED(status)) {
1040 int sig = WTERMSIG(status);
1041
1042 if (sig != SIGQUIT)
1043 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
1044 td_set_runstate(td, TD_REAPED);
1045 goto reaped;
1046 }
1047 if (WIFEXITED(status)) {
1048 if (WEXITSTATUS(status) && !td->error)
1049 td->error = WEXITSTATUS(status);
1050
1051 td_set_runstate(td, TD_REAPED);
1052 goto reaped;
1053 }
1054 }
1055
1056 /*
1057 * thread is not dead, continue
1058 */
1059 pending++;
1060 continue;
1061reaped:
1062 if (td->o.use_thread) {
1063 long ret;
1064
1065 if (pthread_join(td->thread, (void *) &ret))
1066 perror("pthread_join");
1067 }
1068
1069 (*nr_running)--;
1070 (*m_rate) -= td->o.ratemin;
1071 (*t_rate) -= td->o.rate;
1072 pending--;
1073
1074 if (td->error)
1075 exit_value++;
1076 }
1077
1078 if (*nr_running == cputhreads && !pending && realthreads)
1079 terminate_threads(TERMINATE_ALL);
1080}
1081
1082/*
1083 * Main function for kicking off and reaping jobs, as needed.
1084 */
1085static void run_threads(void)
1086{
1087 struct thread_data *td;
1088 unsigned long spent;
1089 int i, todo, nr_running, m_rate, t_rate, nr_started;
1090
1091 if (fio_pin_memory())
1092 return;
1093
1094 if (!terse_output) {
1095 printf("Starting ");
1096 if (nr_thread)
1097 printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1098 if (nr_process) {
1099 if (nr_thread)
1100 printf(" and ");
1101 printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1102 }
1103 printf("\n");
1104 fflush(stdout);
1105 }
1106
1107 signal(SIGINT, sig_handler);
1108 signal(SIGALRM, sig_handler);
1109
1110 todo = thread_number;
1111 nr_running = 0;
1112 nr_started = 0;
1113 m_rate = t_rate = 0;
1114
1115 for_each_td(td, i) {
1116 print_status_init(td->thread_number - 1);
1117
1118 if (!td->o.create_serialize) {
1119 init_disk_util(td);
1120 continue;
1121 }
1122
1123 /*
1124 * do file setup here so it happens sequentially,
1125 * we don't want X number of threads getting their
1126 * client data interspersed on disk
1127 */
1128 if (setup_files(td)) {
1129 exit_value++;
1130 if (td->error)
1131 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1132 td_set_runstate(td, TD_REAPED);
1133 todo--;
1134 }
1135
1136 init_disk_util(td);
1137 }
1138
1139 set_genesis_time();
1140
1141 while (todo) {
1142 struct thread_data *map[MAX_JOBS];
1143 struct timeval this_start;
1144 int this_jobs = 0, left;
1145
1146 /*
1147 * create threads (TD_NOT_CREATED -> TD_CREATED)
1148 */
1149 for_each_td(td, i) {
1150 if (td->runstate != TD_NOT_CREATED)
1151 continue;
1152
1153 /*
1154 * never got a chance to start, killed by other
1155 * thread for some reason
1156 */
1157 if (td->terminate) {
1158 todo--;
1159 continue;
1160 }
1161
1162 if (td->o.start_delay) {
1163 spent = mtime_since_genesis();
1164
1165 if (td->o.start_delay * 1000 > spent)
1166 continue;
1167 }
1168
1169 if (td->o.stonewall && (nr_started || nr_running))
1170 break;
1171
1172 /*
1173 * Set state to created. Thread will transition
1174 * to TD_INITIALIZED when it's done setting up.
1175 */
1176 td_set_runstate(td, TD_CREATED);
1177 map[this_jobs++] = td;
1178 nr_started++;
1179
1180 if (td->o.use_thread) {
1181 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1182 perror("thread_create");
1183 nr_started--;
1184 break;
1185 }
1186 } else {
1187 if (!fork()) {
1188 int ret = fork_main(shm_id, i);
1189
1190 exit(ret);
1191 }
1192 }
1193 fio_sem_down(startup_sem);
1194 }
1195
1196 /*
1197 * Wait for the started threads to transition to
1198 * TD_INITIALIZED.
1199 */
1200 fio_gettime(&this_start, NULL);
1201 left = this_jobs;
1202 while (left && !fio_abort) {
1203 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1204 break;
1205
1206 usleep(100000);
1207
1208 for (i = 0; i < this_jobs; i++) {
1209 td = map[i];
1210 if (!td)
1211 continue;
1212 if (td->runstate == TD_INITIALIZED) {
1213 map[i] = NULL;
1214 left--;
1215 } else if (td->runstate >= TD_EXITED) {
1216 map[i] = NULL;
1217 left--;
1218 todo--;
1219 nr_running++; /* work-around... */
1220 }
1221 }
1222 }
1223
1224 if (left) {
1225 log_err("fio: %d jobs failed to start\n", left);
1226 for (i = 0; i < this_jobs; i++) {
1227 td = map[i];
1228 if (!td)
1229 continue;
1230 kill(td->pid, SIGTERM);
1231 }
1232 break;
1233 }
1234
1235 /*
1236 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1237 */
1238 for_each_td(td, i) {
1239 if (td->runstate != TD_INITIALIZED)
1240 continue;
1241
1242 td_set_runstate(td, TD_RUNNING);
1243 nr_running++;
1244 nr_started--;
1245 m_rate += td->o.ratemin;
1246 t_rate += td->o.rate;
1247 todo--;
1248 fio_sem_up(td->mutex);
1249 }
1250
1251 reap_threads(&nr_running, &t_rate, &m_rate);
1252
1253 if (todo)
1254 usleep(100000);
1255 }
1256
1257 while (nr_running) {
1258 reap_threads(&nr_running, &t_rate, &m_rate);
1259 usleep(10000);
1260 }
1261
1262 update_io_ticks();
1263 fio_unpin_memory();
1264}
1265
1266int main(int argc, char *argv[])
1267{
1268 long ps;
1269
1270 /*
1271 * We need locale for number printing, if it isn't set then just
1272 * go with the US format.
1273 */
1274 if (!getenv("LC_NUMERIC"))
1275 setlocale(LC_NUMERIC, "en_US");
1276
1277 if (parse_options(argc, argv))
1278 return 1;
1279
1280 if (!thread_number)
1281 return 0;
1282
1283 ps = sysconf(_SC_PAGESIZE);
1284 if (ps < 0) {
1285 log_err("Failed to get page size\n");
1286 return 1;
1287 }
1288
1289 page_size = ps;
1290 page_mask = ps - 1;
1291
1292 if (write_bw_log) {
1293 setup_log(&agg_io_log[DDIR_READ]);
1294 setup_log(&agg_io_log[DDIR_WRITE]);
1295 }
1296
1297 startup_sem = fio_sem_init(0);
1298
1299 set_genesis_time();
1300
1301 disk_util_timer_arm();
1302
1303 run_threads();
1304
1305 if (!fio_abort) {
1306 show_run_stats();
1307 if (write_bw_log) {
1308 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1309 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1310 }
1311 }
1312
1313 fio_sem_remove(startup_sem);
1314 return exit_value;
1315}
fio - Flexible IO Tester