projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Fix problem with io_u being beyond EOF
[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 ret = td_io_queue(td, io_u);
450 switch (ret) {
451 case FIO_Q_COMPLETED:
452 if (io_u->error)
453 ret = -io_u->error;
454 else if (io_u->resid) {
455 int bytes = io_u->xfer_buflen - io_u->resid;
456 struct fio_file *f = io_u->file;
457
458 /*
459 * zero read, fail
460 */
461 if (!bytes) {
462 td_verror(td, ENODATA, "full resid");
463 put_io_u(td, io_u);
464 break;
465 }
466
467 io_u->xfer_buflen = io_u->resid;
468 io_u->xfer_buf += bytes;
469 io_u->offset += bytes;
470 f->last_completed_pos = io_u->offset;
471
472 td->ts.short_io_u[io_u->ddir]++;
473
474 if (io_u->offset == f->real_file_size)
475 goto sync_done;
476
477 requeue_io_u(td, &io_u);
478 } else {
479sync_done:
480 fio_gettime(&comp_time, NULL);
481 bytes_done = io_u_sync_complete(td, io_u);
482 if (bytes_done < 0)
483 ret = bytes_done;
484 }
485 break;
486 case FIO_Q_QUEUED:
487 /*
488 * if the engine doesn't have a commit hook,
489 * the io_u is really queued. if it does have such
490 * a hook, it has to call io_u_queued() itself.
491 */
492 if (td->io_ops->commit == NULL)
493 io_u_queued(td, io_u);
494 break;
495 case FIO_Q_BUSY:
496 requeue_io_u(td, &io_u);
497 ret2 = td_io_commit(td);
498 if (ret2 < 0)
499 ret = ret2;
500 break;
501 default:
502 assert(ret < 0);
503 put_io_u(td, io_u);
504 break;
505 }
506
507 if (ret < 0 || td->error)
508 break;
509
510 /*
511 * See if we need to complete some commands
512 */
513 if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
514 min_evts = 0;
515 if (queue_full(td) || ret == FIO_Q_BUSY) {
516 min_evts = 1;
517
518 if (td->cur_depth > td->o.iodepth_low)
519 min_evts = td->cur_depth - td->o.iodepth_low;
520 }
521
522 fio_gettime(&comp_time, NULL);
523 bytes_done = io_u_queued_complete(td, min_evts);
524 if (bytes_done < 0)
525 break;
526 }
527
528 if (!bytes_done)
529 continue;
530
531 /*
532 * the rate is batched for now, it should work for batches
533 * of completions except the very first one which may look
534 * a little bursty
535 */
536 usec = utime_since(&s, &comp_time);
537
538 rate_throttle(td, usec, bytes_done);
539
540 if (check_min_rate(td, &comp_time)) {
541 if (exitall_on_terminate)
542 terminate_threads(td->groupid);
543 td_verror(td, ENODATA, "check_min_rate");
544 break;
545 }
546
547 if (td->o.thinktime) {
548 unsigned long long b;
549
550 b = td->io_blocks[0] + td->io_blocks[1];
551 if (!(b % td->o.thinktime_blocks)) {
552 int left;
553
554 if (td->o.thinktime_spin)
555 __usec_sleep(td->o.thinktime_spin);
556
557 left = td->o.thinktime - td->o.thinktime_spin;
558 if (left)
559 usec_sleep(td, left);
560 }
561 }
562 }
563
564 if (!td->error) {
565 struct fio_file *f;
566
567 i = td->cur_depth;
568 if (i)
569 ret = io_u_queued_complete(td, i);
570
571 if (should_fsync(td) && td->o.end_fsync) {
572 td_set_runstate(td, TD_FSYNCING);
573
574 for_each_file(td, f, i) {
575 if (!(f->flags & FIO_FILE_OPEN))
576 continue;
577 fio_io_sync(td, f);
578 }
579 }
580 } else
581 cleanup_pending_aio(td);
582}
583
584static void cleanup_io_u(struct thread_data *td)
585{
586 struct list_head *entry, *n;
587 struct io_u *io_u;
588
589 list_for_each_safe(entry, n, &td->io_u_freelist) {
590 io_u = list_entry(entry, struct io_u, list);
591
592 list_del(&io_u->list);
593 free(io_u);
594 }
595
596 free_io_mem(td);
597}
598
599/*
600 * "randomly" fill the buffer contents
601 */
602static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
603{
604 long *ptr = io_u->buf;
605
606 if (!td->o.zero_buffers) {
607 while ((void *) ptr - io_u->buf < max_bs) {
608 *ptr = rand() * GOLDEN_RATIO_PRIME;
609 ptr++;
610 }
611 } else
612 memset(ptr, 0, max_bs);
613}
614
615static int init_io_u(struct thread_data *td)
616{
617 struct io_u *io_u;
618 unsigned int max_bs;
619 int i, max_units;
620 char *p;
621
622 if (td->io_ops->flags & FIO_SYNCIO)
623 max_units = 1;
624 else
625 max_units = td->o.iodepth;
626
627 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
628 td->orig_buffer_size = (unsigned long long) max_bs * (unsigned long long) max_units;
629
630 if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
631 td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
632
633 if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
634 log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
635 return 1;
636 }
637
638 if (allocate_io_mem(td))
639 return 1;
640
641 if (td->o.odirect)
642 p = ALIGN(td->orig_buffer);
643 else
644 p = td->orig_buffer;
645
646 for (i = 0; i < max_units; i++) {
647 if (td->terminate)
648 return 1;
649 io_u = malloc(sizeof(*io_u));
650 memset(io_u, 0, sizeof(*io_u));
651 INIT_LIST_HEAD(&io_u->list);
652
653 io_u->buf = p + max_bs * i;
654
655 if (td_write(td))
656 fill_io_buf(td, io_u, max_bs);
657
658 io_u->index = i;
659 io_u->flags = IO_U_F_FREE;
660 list_add(&io_u->list, &td->io_u_freelist);
661 }
662
663 io_u_init_timeout();
664
665 return 0;
666}
667
668static int switch_ioscheduler(struct thread_data *td)
669{
670 char tmp[256], tmp2[128];
671 FILE *f;
672 int ret;
673
674 if (td->io_ops->flags & FIO_DISKLESSIO)
675 return 0;
676
677 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
678
679 f = fopen(tmp, "r+");
680 if (!f) {
681 if (errno == ENOENT) {
682 log_err("fio: os or kernel doesn't support IO scheduler switching\n");
683 return 0;
684 }
685 td_verror(td, errno, "fopen iosched");
686 return 1;
687 }
688
689 /*
690 * Set io scheduler.
691 */
692 ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
693 if (ferror(f) || ret != 1) {
694 td_verror(td, errno, "fwrite");
695 fclose(f);
696 return 1;
697 }
698
699 rewind(f);
700
701 /*
702 * Read back and check that the selected scheduler is now the default.
703 */
704 ret = fread(tmp, 1, sizeof(tmp), f);
705 if (ferror(f) || ret < 0) {
706 td_verror(td, errno, "fread");
707 fclose(f);
708 return 1;
709 }
710
711 sprintf(tmp2, "[%s]", td->o.ioscheduler);
712 if (!strstr(tmp, tmp2)) {
713 log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
714 td_verror(td, EINVAL, "iosched_switch");
715 fclose(f);
716 return 1;
717 }
718
719 fclose(f);
720 return 0;
721}
722
723static int keep_running(struct thread_data *td)
724{
725 unsigned long long io_done;
726
727 if (td->done)
728 return 0;
729 if (td->o.time_based)
730 return 1;
731 if (td->o.loops) {
732 td->o.loops--;
733 return 1;
734 }
735
736 io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE] + td->io_skip_bytes;
737 if (io_done < td->o.size)
738 return 1;
739
740 return 0;
741}
742
743static int clear_io_state(struct thread_data *td)
744{
745 struct fio_file *f;
746 unsigned int i;
747 int ret;
748
749 td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
750 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
751 td->zone_bytes = 0;
752 td->rate_bytes = 0;
753 td->rate_blocks = 0;
754 td->rw_end_set[0] = td->rw_end_set[1] = 0;
755
756 td->last_was_sync = 0;
757
758 if (td->o.time_based)
759 td->nr_done_files = 0;
760
761 for_each_file(td, f, i)
762 td_io_close_file(td, f);
763
764 ret = 0;
765 for_each_file(td, f, i) {
766 f->flags &= ~FIO_FILE_DONE;
767 ret = td_io_open_file(td, f);
768 if (ret)
769 break;
770 }
771
772 return ret;
773}
774
775/*
776 * Entry point for the thread based jobs. The process based jobs end up
777 * here as well, after a little setup.
778 */
779static void *thread_main(void *data)
780{
781 unsigned long long runtime[2];
782 struct thread_data *td = data;
783 unsigned long elapsed;
784 int clear_state;
785
786 if (!td->o.use_thread)
787 setsid();
788
789 td->pid = getpid();
790
791 INIT_LIST_HEAD(&td->io_u_freelist);
792 INIT_LIST_HEAD(&td->io_u_busylist);
793 INIT_LIST_HEAD(&td->io_u_requeues);
794 INIT_LIST_HEAD(&td->io_log_list);
795 INIT_LIST_HEAD(&td->io_hist_list);
796 td->io_hist_tree = RB_ROOT;
797
798 td_set_runstate(td, TD_INITIALIZED);
799 fio_sem_up(startup_sem);
800 fio_sem_down(td->mutex);
801
802 /*
803 * the ->mutex semaphore is now no longer used, close it to avoid
804 * eating a file descriptor
805 */
806 fio_sem_remove(td->mutex);
807
808 /*
809 * May alter parameters that init_io_u() will use, so we need to
810 * do this first.
811 */
812 if (init_iolog(td))
813 goto err;
814
815 if (init_io_u(td))
816 goto err;
817
818 if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
819 td_verror(td, errno, "cpu_set_affinity");
820 goto err;
821 }
822
823 if (td->ioprio) {
824 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
825 td_verror(td, errno, "ioprio_set");
826 goto err;
827 }
828 }
829
830 if (nice(td->o.nice) == -1) {
831 td_verror(td, errno, "nice");
832 goto err;
833 }
834
835 if (td->o.ioscheduler && switch_ioscheduler(td))
836 goto err;
837
838 if (!td->o.create_serialize && setup_files(td))
839 goto err;
840
841 if (td_io_init(td))
842 goto err;
843
844 if (open_files(td))
845 goto err;
846
847 if (init_random_map(td))
848 goto err;
849
850 if (td->o.exec_prerun) {
851 if (system(td->o.exec_prerun) < 0)
852 goto err;
853 }
854
855 fio_gettime(&td->epoch, NULL);
856 memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
857 getrusage(RUSAGE_SELF, &td->ts.ru_start);
858
859 runtime[0] = runtime[1] = 0;
860 clear_state = 0;
861 while (keep_running(td)) {
862 fio_gettime(&td->start, NULL);
863 memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
864
865 if (td->o.ratemin)
866 memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
867
868 if (clear_state && clear_io_state(td))
869 break;
870
871 prune_io_piece_log(td);
872
873 do_io(td);
874
875 clear_state = 1;
876
877 if (td_read(td) && td->io_bytes[DDIR_READ]) {
878 if (td->rw_end_set[DDIR_READ])
879 elapsed = utime_since(&td->start, &td->rw_end[DDIR_READ]);
880 else
881 elapsed = utime_since_now(&td->start);
882
883 runtime[DDIR_READ] += elapsed;
884 }
885 if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
886 if (td->rw_end_set[DDIR_WRITE])
887 elapsed = utime_since(&td->start, &td->rw_end[DDIR_WRITE]);
888 else
889 elapsed = utime_since_now(&td->start);
890
891 runtime[DDIR_WRITE] += elapsed;
892 }
893
894 if (td->error || td->terminate)
895 break;
896
897 if (td->o.verify == VERIFY_NONE)
898 continue;
899
900 if (clear_io_state(td))
901 break;
902
903 fio_gettime(&td->start, NULL);
904
905 do_verify(td);
906
907 runtime[DDIR_READ] += utime_since_now(&td->start);
908
909 if (td->error || td->terminate)
910 break;
911 }
912
913 update_rusage_stat(td);
914 td->ts.runtime[0] = runtime[0] / 1000;
915 td->ts.runtime[1] = runtime[1] / 1000;
916 td->ts.total_run_time = mtime_since_now(&td->epoch);
917 td->ts.io_bytes[0] = td->io_bytes[0];
918 td->ts.io_bytes[1] = td->io_bytes[1];
919
920 if (td->ts.bw_log)
921 finish_log(td, td->ts.bw_log, "bw");
922 if (td->ts.slat_log)
923 finish_log(td, td->ts.slat_log, "slat");
924 if (td->ts.clat_log)
925 finish_log(td, td->ts.clat_log, "clat");
926 if (td->o.exec_postrun) {
927 if (system(td->o.exec_postrun) < 0)
928 log_err("fio: postrun %s failed\n", td->o.exec_postrun);
929 }
930
931 if (exitall_on_terminate)
932 terminate_threads(td->groupid);
933
934err:
935 if (td->error)
936 printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
937 close_files(td);
938 close_ioengine(td);
939 cleanup_io_u(td);
940
941 /*
942 * do this very late, it will log file closing as well
943 */
944 if (td->o.write_iolog_file)
945 write_iolog_close(td);
946
947 options_mem_free(td);
948 td_set_runstate(td, TD_EXITED);
949 return (void *) (unsigned long) td->error;
950}
951
952/*
953 * We cannot pass the td data into a forked process, so attach the td and
954 * pass it to the thread worker.
955 */
956static int fork_main(int shmid, int offset)
957{
958 struct thread_data *td;
959 void *data, *ret;
960
961 data = shmat(shmid, NULL, 0);
962 if (data == (void *) -1) {
963 int __err = errno;
964
965 perror("shmat");
966 return __err;
967 }
968
969 td = data + offset * sizeof(struct thread_data);
970 ret = thread_main(td);
971 shmdt(data);
972 return (int) (unsigned long) ret;
973}
974
975/*
976 * Run over the job map and reap the threads that have exited, if any.
977 */
978static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
979{
980 struct thread_data *td;
981 int i, cputhreads, pending, status, ret;
982
983 /*
984 * reap exited threads (TD_EXITED -> TD_REAPED)
985 */
986 pending = cputhreads = 0;
987 for_each_td(td, i) {
988 int flags = 0;
989
990 /*
991 * ->io_ops is NULL for a thread that has closed its
992 * io engine
993 */
994 if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
995 cputhreads++;
996
997 if (!td->pid || td->runstate == TD_REAPED)
998 continue;
999 if (td->o.use_thread) {
1000 if (td->runstate == TD_EXITED) {
1001 td_set_runstate(td, TD_REAPED);
1002 goto reaped;
1003 }
1004 continue;
1005 }
1006
1007 flags = WNOHANG;
1008 if (td->runstate == TD_EXITED)
1009 flags = 0;
1010
1011 /*
1012 * check if someone quit or got killed in an unusual way
1013 */
1014 ret = waitpid(td->pid, &status, flags);
1015 if (ret < 0) {
1016 if (errno == ECHILD) {
1017 log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
1018 td_set_runstate(td, TD_REAPED);
1019 goto reaped;
1020 }
1021 perror("waitpid");
1022 } else if (ret == td->pid) {
1023 if (WIFSIGNALED(status)) {
1024 int sig = WTERMSIG(status);
1025
1026 if (sig != SIGQUIT)
1027 log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
1028 td_set_runstate(td, TD_REAPED);
1029 goto reaped;
1030 }
1031 if (WIFEXITED(status)) {
1032 if (WEXITSTATUS(status) && !td->error)
1033 td->error = WEXITSTATUS(status);
1034
1035 td_set_runstate(td, TD_REAPED);
1036 goto reaped;
1037 }
1038 }
1039
1040 /*
1041 * thread is not dead, continue
1042 */
1043 pending++;
1044 continue;
1045reaped:
1046 if (td->o.use_thread) {
1047 long ret;
1048
1049 if (pthread_join(td->thread, (void *) &ret))
1050 perror("pthread_join");
1051 }
1052
1053 (*nr_running)--;
1054 (*m_rate) -= td->o.ratemin;
1055 (*t_rate) -= td->o.rate;
1056 pending--;
1057
1058 if (td->error)
1059 exit_value++;
1060 }
1061
1062 if (*nr_running == cputhreads && !pending)
1063 terminate_threads(TERMINATE_ALL);
1064}
1065
1066/*
1067 * Main function for kicking off and reaping jobs, as needed.
1068 */
1069static void run_threads(void)
1070{
1071 struct thread_data *td;
1072 unsigned long spent;
1073 int i, todo, nr_running, m_rate, t_rate, nr_started;
1074
1075 if (fio_pin_memory())
1076 return;
1077
1078 if (!terse_output) {
1079 printf("Starting ");
1080 if (nr_thread)
1081 printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
1082 if (nr_process) {
1083 if (nr_thread)
1084 printf(" and ");
1085 printf("%d process%s", nr_process, nr_process > 1 ? "es" : "");
1086 }
1087 printf("\n");
1088 fflush(stdout);
1089 }
1090
1091 signal(SIGINT, sig_handler);
1092 signal(SIGALRM, sig_handler);
1093
1094 todo = thread_number;
1095 nr_running = 0;
1096 nr_started = 0;
1097 m_rate = t_rate = 0;
1098
1099 for_each_td(td, i) {
1100 print_status_init(td->thread_number - 1);
1101
1102 if (!td->o.create_serialize) {
1103 init_disk_util(td);
1104 continue;
1105 }
1106
1107 /*
1108 * do file setup here so it happens sequentially,
1109 * we don't want X number of threads getting their
1110 * client data interspersed on disk
1111 */
1112 if (setup_files(td)) {
1113 exit_value++;
1114 if (td->error)
1115 log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
1116 td_set_runstate(td, TD_REAPED);
1117 todo--;
1118 }
1119
1120 init_disk_util(td);
1121 }
1122
1123 set_genesis_time();
1124
1125 while (todo) {
1126 struct thread_data *map[MAX_JOBS];
1127 struct timeval this_start;
1128 int this_jobs = 0, left;
1129
1130 /*
1131 * create threads (TD_NOT_CREATED -> TD_CREATED)
1132 */
1133 for_each_td(td, i) {
1134 if (td->runstate != TD_NOT_CREATED)
1135 continue;
1136
1137 /*
1138 * never got a chance to start, killed by other
1139 * thread for some reason
1140 */
1141 if (td->terminate) {
1142 todo--;
1143 continue;
1144 }
1145
1146 if (td->o.start_delay) {
1147 spent = mtime_since_genesis();
1148
1149 if (td->o.start_delay * 1000 > spent)
1150 continue;
1151 }
1152
1153 if (td->o.stonewall && (nr_started || nr_running))
1154 break;
1155
1156 /*
1157 * Set state to created. Thread will transition
1158 * to TD_INITIALIZED when it's done setting up.
1159 */
1160 td_set_runstate(td, TD_CREATED);
1161 map[this_jobs++] = td;
1162 nr_started++;
1163
1164 if (td->o.use_thread) {
1165 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1166 perror("thread_create");
1167 nr_started--;
1168 break;
1169 }
1170 } else {
1171 if (!fork()) {
1172 int ret = fork_main(shm_id, i);
1173
1174 exit(ret);
1175 }
1176 }
1177 fio_sem_down(startup_sem);
1178 }
1179
1180 /*
1181 * Wait for the started threads to transition to
1182 * TD_INITIALIZED.
1183 */
1184 fio_gettime(&this_start, NULL);
1185 left = this_jobs;
1186 while (left && !fio_abort) {
1187 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1188 break;
1189
1190 usleep(100000);
1191
1192 for (i = 0; i < this_jobs; i++) {
1193 td = map[i];
1194 if (!td)
1195 continue;
1196 if (td->runstate == TD_INITIALIZED) {
1197 map[i] = NULL;
1198 left--;
1199 } else if (td->runstate >= TD_EXITED) {
1200 map[i] = NULL;
1201 left--;
1202 todo--;
1203 nr_running++; /* work-around... */
1204 }
1205 }
1206 }
1207
1208 if (left) {
1209 log_err("fio: %d jobs failed to start\n", left);
1210 for (i = 0; i < this_jobs; i++) {
1211 td = map[i];
1212 if (!td)
1213 continue;
1214 kill(td->pid, SIGTERM);
1215 }
1216 break;
1217 }
1218
1219 /*
1220 * start created threads (TD_INITIALIZED -> TD_RUNNING).
1221 */
1222 for_each_td(td, i) {
1223 if (td->runstate != TD_INITIALIZED)
1224 continue;
1225
1226 td_set_runstate(td, TD_RUNNING);
1227 nr_running++;
1228 nr_started--;
1229 m_rate += td->o.ratemin;
1230 t_rate += td->o.rate;
1231 todo--;
1232 fio_sem_up(td->mutex);
1233 }
1234
1235 reap_threads(&nr_running, &t_rate, &m_rate);
1236
1237 if (todo)
1238 usleep(100000);
1239 }
1240
1241 while (nr_running) {
1242 reap_threads(&nr_running, &t_rate, &m_rate);
1243 usleep(10000);
1244 }
1245
1246 update_io_ticks();
1247 fio_unpin_memory();
1248}
1249
1250int main(int argc, char *argv[])
1251{
1252 long ps;
1253
1254 /*
1255 * We need locale for number printing, if it isn't set then just
1256 * go with the US format.
1257 */
1258 if (!getenv("LC_NUMERIC"))
1259 setlocale(LC_NUMERIC, "en_US");
1260
1261 if (parse_options(argc, argv))
1262 return 1;
1263
1264 if (!thread_number)
1265 return 0;
1266
1267 ps = sysconf(_SC_PAGESIZE);
1268 if (ps < 0) {
1269 log_err("Failed to get page size\n");
1270 return 1;
1271 }
1272
1273 page_size = ps;
1274 page_mask = ps - 1;
1275
1276 if (write_bw_log) {
1277 setup_log(&agg_io_log[DDIR_READ]);
1278 setup_log(&agg_io_log[DDIR_WRITE]);
1279 }
1280
1281 startup_sem = fio_sem_init(0);
1282
1283 set_genesis_time();
1284
1285 disk_util_timer_arm();
1286
1287 run_threads();
1288
1289 if (!fio_abort) {
1290 show_run_stats();
1291 if (write_bw_log) {
1292 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1293 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1294 }
1295 }
1296
1297 fio_sem_remove(startup_sem);
1298 return exit_value;
1299}
fio - Flexible IO Tester