projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Fix bad interaction with file open/close and queuing
[fio.git] / io_u.c
... / ...
CommitLineData
1#include <unistd.h>
2#include <fcntl.h>
3#include <string.h>
4#include <signal.h>
5#include <time.h>
6#include <assert.h>
7
8#include "fio.h"
9
10/*
11 * Change this define to play with the timeout handling
12 */
13#undef FIO_USE_TIMEOUT
14
15struct io_completion_data {
16 int nr; /* input */
17
18 int error; /* output */
19 unsigned long bytes_done[2]; /* output */
20 struct timeval time; /* output */
21};
22
23/*
24 * The ->file_map[] contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
26 */
27static int random_map_free(struct fio_file *f, const unsigned long long block)
28{
29 unsigned int idx = RAND_MAP_IDX(f, block);
30 unsigned int bit = RAND_MAP_BIT(f, block);
31
32 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
33
34 return (f->file_map[idx] & (1UL << bit)) == 0;
35}
36
37/*
38 * Mark a given offset as used in the map.
39 */
40static void mark_random_map(struct thread_data *td, struct io_u *io_u)
41{
42 unsigned int min_bs = td->o.rw_min_bs;
43 struct fio_file *f = io_u->file;
44 unsigned long long block;
45 unsigned int blocks;
46 unsigned int nr_blocks;
47
48 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
49 blocks = 0;
50 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
51
52 while (blocks < nr_blocks) {
53 unsigned int idx, bit;
54
55 /*
56 * If we have a mixed random workload, we may
57 * encounter blocks we already did IO to.
58 */
59 if ((td->o.ddir_nr == 1) && !random_map_free(f, block))
60 break;
61
62 idx = RAND_MAP_IDX(f, block);
63 bit = RAND_MAP_BIT(f, block);
64
65 fio_assert(td, idx < f->num_maps);
66
67 f->file_map[idx] |= (1UL << bit);
68 block++;
69 blocks++;
70 }
71
72 if ((blocks * min_bs) < io_u->buflen)
73 io_u->buflen = blocks * min_bs;
74}
75
76static inline unsigned long long last_block(struct thread_data *td,
77 struct fio_file *f,
78 enum fio_ddir ddir)
79{
80 unsigned long long max_blocks;
81
82 max_blocks = f->io_size / (unsigned long long) td->o.min_bs[ddir];
83 if (!max_blocks)
84 return 0;
85
86 return max_blocks;
87}
88
89/*
90 * Return the next free block in the map.
91 */
92static int get_next_free_block(struct thread_data *td, struct fio_file *f,
93 enum fio_ddir ddir, unsigned long long *b)
94{
95 unsigned long long min_bs = td->o.rw_min_bs;
96 int i;
97
98 i = f->last_free_lookup;
99 *b = (i * BLOCKS_PER_MAP);
100 while ((*b) * min_bs < f->real_file_size) {
101 if (f->file_map[i] != -1UL) {
102 *b += fio_ffz(f->file_map[i]);
103 if (*b > last_block(td, f, ddir))
104 break;
105 f->last_free_lookup = i;
106 return 0;
107 }
108
109 *b += BLOCKS_PER_MAP;
110 i++;
111 }
112
113 dprint(FD_IO, "failed finding a free block\n");
114 return 1;
115}
116
117static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
118 enum fio_ddir ddir, unsigned long long *b)
119{
120 unsigned long long r;
121 int loops = 5;
122
123 do {
124 r = os_random_long(&td->random_state);
125 dprint(FD_RANDOM, "off rand %llu\n", r);
126 *b = (last_block(td, f, ddir) - 1)
127 * (r / ((unsigned long long) RAND_MAX + 1.0));
128
129 /*
130 * if we are not maintaining a random map, we are done.
131 */
132 if (!file_randommap(td, f))
133 return 0;
134
135 /*
136 * calculate map offset and check if it's free
137 */
138 if (random_map_free(f, *b))
139 return 0;
140
141 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
142 *b);
143 } while (--loops);
144
145 /*
146 * we get here, if we didn't suceed in looking up a block. generate
147 * a random start offset into the filemap, and find the first free
148 * block from there.
149 */
150 loops = 10;
151 do {
152 f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
153 if (!get_next_free_block(td, f, ddir, b))
154 return 0;
155
156 r = os_random_long(&td->random_state);
157 } while (--loops);
158
159 /*
160 * that didn't work either, try exhaustive search from the start
161 */
162 f->last_free_lookup = 0;
163 return get_next_free_block(td, f, ddir, b);
164}
165
166/*
167 * For random io, generate a random new block and see if it's used. Repeat
168 * until we find a free one. For sequential io, just return the end of
169 * the last io issued.
170 */
171static int get_next_offset(struct thread_data *td, struct io_u *io_u)
172{
173 struct fio_file *f = io_u->file;
174 unsigned long long b;
175 enum fio_ddir ddir = io_u->ddir;
176
177 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
178 td->ddir_nr = td->o.ddir_nr;
179
180 if (get_next_rand_offset(td, f, ddir, &b))
181 return 1;
182 } else {
183 if (f->last_pos >= f->real_file_size) {
184 if (!td_random(td) ||
185 get_next_rand_offset(td, f, ddir, &b))
186 return 1;
187 } else
188 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
189 }
190
191 io_u->offset = b * td->o.min_bs[ddir];
192 if (io_u->offset >= f->io_size) {
193 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
194 io_u->offset, f->io_size);
195 return 1;
196 }
197
198 io_u->offset += f->file_offset;
199 if (io_u->offset >= f->real_file_size) {
200 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
201 io_u->offset, f->real_file_size);
202 return 1;
203 }
204
205 return 0;
206}
207
208static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
209{
210 const int ddir = io_u->ddir;
211 unsigned int buflen;
212 long r;
213
214 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
215 buflen = td->o.min_bs[ddir];
216 else {
217 r = os_random_long(&td->bsrange_state);
218 if (!td->o.bssplit_nr) {
219 buflen = (unsigned int)
220 (1 + (double) (td->o.max_bs[ddir] - 1)
221 * r / (RAND_MAX + 1.0));
222 } else {
223 long perc = 0;
224 unsigned int i;
225
226 for (i = 0; i < td->o.bssplit_nr; i++) {
227 struct bssplit *bsp = &td->o.bssplit[i];
228
229 buflen = bsp->bs;
230 perc += bsp->perc;
231 if (r <= ((LONG_MAX / 100L) * perc))
232 break;
233 }
234 }
235 if (!td->o.bs_unaligned) {
236 buflen = (buflen + td->o.min_bs[ddir] - 1)
237 & ~(td->o.min_bs[ddir] - 1);
238 }
239 }
240
241 if (io_u->offset + buflen > io_u->file->real_file_size) {
242 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
243 td->o.min_bs[ddir], ddir);
244 buflen = td->o.min_bs[ddir];
245 }
246
247 return buflen;
248}
249
250static void set_rwmix_bytes(struct thread_data *td)
251{
252 unsigned long issues;
253 unsigned int diff;
254
255 /*
256 * we do time or byte based switch. this is needed because
257 * buffered writes may issue a lot quicker than they complete,
258 * whereas reads do not.
259 */
260 issues = td->io_issues[td->rwmix_ddir] - td->rwmix_issues;
261 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
262
263 td->rwmix_issues = td->io_issues[td->rwmix_ddir]
264 + (issues * ((100 - diff)) / diff);
265}
266
267static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
268{
269 unsigned int v;
270 long r;
271
272 r = os_random_long(&td->rwmix_state);
273 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
274 if (v < td->o.rwmix[DDIR_READ])
275 return DDIR_READ;
276
277 return DDIR_WRITE;
278}
279
280/*
281 * Return the data direction for the next io_u. If the job is a
282 * mixed read/write workload, check the rwmix cycle and switch if
283 * necessary.
284 */
285static enum fio_ddir get_rw_ddir(struct thread_data *td)
286{
287 if (td_rw(td)) {
288 /*
289 * Check if it's time to seed a new data direction.
290 */
291 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
292 unsigned long long max_bytes;
293 enum fio_ddir ddir;
294
295 /*
296 * Put a top limit on how many bytes we do for
297 * one data direction, to avoid overflowing the
298 * ranges too much
299 */
300 ddir = get_rand_ddir(td);
301 max_bytes = td->this_io_bytes[ddir];
302 if (max_bytes >=
303 (td->o.size * td->o.rwmix[ddir] / 100)) {
304 if (!td->rw_end_set[ddir])
305 td->rw_end_set[ddir] = 1;
306
307 ddir ^= 1;
308 }
309
310 if (ddir != td->rwmix_ddir)
311 set_rwmix_bytes(td);
312
313 td->rwmix_ddir = ddir;
314 }
315 return td->rwmix_ddir;
316 } else if (td_read(td))
317 return DDIR_READ;
318 else
319 return DDIR_WRITE;
320}
321
322void put_io_u(struct thread_data *td, struct io_u *io_u)
323{
324 assert((io_u->flags & IO_U_F_FREE) == 0);
325 io_u->flags |= IO_U_F_FREE;
326
327 if (io_u->file) {
328 int ret = put_file(td, io_u->file);
329
330 if (ret)
331 td_verror(td, ret, "file close");
332 }
333
334 io_u->file = NULL;
335 list_del(&io_u->list);
336 list_add(&io_u->list, &td->io_u_freelist);
337 td->cur_depth--;
338}
339
340void requeue_io_u(struct thread_data *td, struct io_u **io_u)
341{
342 struct io_u *__io_u = *io_u;
343
344 __io_u->flags |= IO_U_F_FREE;
345 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
346 td->io_issues[__io_u->ddir]--;
347
348 __io_u->flags &= ~IO_U_F_FLIGHT;
349
350 list_del(&__io_u->list);
351 list_add_tail(&__io_u->list, &td->io_u_requeues);
352 td->cur_depth--;
353 *io_u = NULL;
354}
355
356static int fill_io_u(struct thread_data *td, struct io_u *io_u)
357{
358 if (td->io_ops->flags & FIO_NOIO)
359 goto out;
360
361 /*
362 * see if it's time to sync
363 */
364 if (td->o.fsync_blocks &&
365 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
366 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
367 io_u->ddir = DDIR_SYNC;
368 goto out;
369 }
370
371 io_u->ddir = get_rw_ddir(td);
372
373 /*
374 * See if it's time to switch to a new zone
375 */
376 if (td->zone_bytes >= td->o.zone_size) {
377 td->zone_bytes = 0;
378 io_u->file->last_pos += td->o.zone_skip;
379 td->io_skip_bytes += td->o.zone_skip;
380 }
381
382 /*
383 * No log, let the seq/rand engine retrieve the next buflen and
384 * position.
385 */
386 if (get_next_offset(td, io_u)) {
387 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
388 return 1;
389 }
390
391 io_u->buflen = get_next_buflen(td, io_u);
392 if (!io_u->buflen) {
393 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
394 return 1;
395 }
396
397 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
398 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
399 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
400 io_u->buflen, io_u->file->real_file_size);
401 return 1;
402 }
403
404 /*
405 * mark entry before potentially trimming io_u
406 */
407 if (td_random(td) && file_randommap(td, io_u->file))
408 mark_random_map(td, io_u);
409
410 /*
411 * If using a write iolog, store this entry.
412 */
413out:
414 dprint_io_u(io_u, "fill_io_u");
415 td->zone_bytes += io_u->buflen;
416 log_io_u(td, io_u);
417 return 0;
418}
419
420void io_u_mark_depth(struct thread_data *td, unsigned int nr)
421{
422 int index = 0;
423
424 switch (td->cur_depth) {
425 default:
426 index = 6;
427 break;
428 case 32 ... 63:
429 index = 5;
430 break;
431 case 16 ... 31:
432 index = 4;
433 break;
434 case 8 ... 15:
435 index = 3;
436 break;
437 case 4 ... 7:
438 index = 2;
439 break;
440 case 2 ... 3:
441 index = 1;
442 case 1:
443 break;
444 }
445
446 td->ts.io_u_map[index] += nr;
447}
448
449static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
450{
451 int index = 0;
452
453 assert(usec < 1000);
454
455 switch (usec) {
456 case 750 ... 999:
457 index = 9;
458 break;
459 case 500 ... 749:
460 index = 8;
461 break;
462 case 250 ... 499:
463 index = 7;
464 break;
465 case 100 ... 249:
466 index = 6;
467 break;
468 case 50 ... 99:
469 index = 5;
470 break;
471 case 20 ... 49:
472 index = 4;
473 break;
474 case 10 ... 19:
475 index = 3;
476 break;
477 case 4 ... 9:
478 index = 2;
479 break;
480 case 2 ... 3:
481 index = 1;
482 case 0 ... 1:
483 break;
484 }
485
486 assert(index < FIO_IO_U_LAT_U_NR);
487 td->ts.io_u_lat_u[index]++;
488}
489
490static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
491{
492 int index = 0;
493
494 switch (msec) {
495 default:
496 index = 11;
497 break;
498 case 1000 ... 1999:
499 index = 10;
500 break;
501 case 750 ... 999:
502 index = 9;
503 break;
504 case 500 ... 749:
505 index = 8;
506 break;
507 case 250 ... 499:
508 index = 7;
509 break;
510 case 100 ... 249:
511 index = 6;
512 break;
513 case 50 ... 99:
514 index = 5;
515 break;
516 case 20 ... 49:
517 index = 4;
518 break;
519 case 10 ... 19:
520 index = 3;
521 break;
522 case 4 ... 9:
523 index = 2;
524 break;
525 case 2 ... 3:
526 index = 1;
527 case 0 ... 1:
528 break;
529 }
530
531 assert(index < FIO_IO_U_LAT_M_NR);
532 td->ts.io_u_lat_m[index]++;
533}
534
535static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
536{
537 if (usec < 1000)
538 io_u_mark_lat_usec(td, usec);
539 else
540 io_u_mark_lat_msec(td, usec / 1000);
541}
542
543/*
544 * Get next file to service by choosing one at random
545 */
546static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
547 int badf)
548{
549 struct fio_file *f;
550 int fno;
551
552 do {
553 long r = os_random_long(&td->next_file_state);
554
555 fno = (unsigned int) ((double) td->o.nr_files
556 * (r / (RAND_MAX + 1.0)));
557 f = td->files[fno];
558 if (f->flags & FIO_FILE_DONE)
559 continue;
560
561 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
562 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
563 return f;
564 }
565 } while (1);
566}
567
568/*
569 * Get next file to service by doing round robin between all available ones
570 */
571static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
572 int badf)
573{
574 unsigned int old_next_file = td->next_file;
575 struct fio_file *f;
576
577 do {
578 f = td->files[td->next_file];
579
580 td->next_file++;
581 if (td->next_file >= td->o.nr_files)
582 td->next_file = 0;
583
584 if (f->flags & FIO_FILE_DONE) {
585 f = NULL;
586 continue;
587 }
588
589 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
590 break;
591
592 f = NULL;
593 } while (td->next_file != old_next_file);
594
595 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
596 return f;
597}
598
599static struct fio_file *get_next_file(struct thread_data *td)
600{
601 struct fio_file *f;
602
603 assert(td->o.nr_files <= td->files_index);
604
605 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
606 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
607 " nr_files=%d\n", td->nr_open_files,
608 td->nr_done_files,
609 td->o.nr_files);
610 return NULL;
611 }
612
613 f = td->file_service_file;
614 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
615 goto out;
616
617 if (td->o.file_service_type == FIO_FSERVICE_RR)
618 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
619 else
620 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
621
622 td->file_service_file = f;
623 td->file_service_left = td->file_service_nr - 1;
624out:
625 dprint(FD_FILE, "get_next_file: %p\n", f);
626 return f;
627}
628
629static struct fio_file *find_next_new_file(struct thread_data *td)
630{
631 struct fio_file *f;
632
633 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
634 return NULL;
635
636 if (td->o.file_service_type == FIO_FSERVICE_RR)
637 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
638 else
639 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
640
641 return f;
642}
643
644static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
645{
646 struct fio_file *f;
647
648 do {
649 f = get_next_file(td);
650 if (!f)
651 return 1;
652
653set_file:
654 io_u->file = f;
655 get_file(f);
656
657 if (!fill_io_u(td, io_u))
658 break;
659
660 /*
661 * optimization to prevent close/open of the same file. This
662 * way we preserve queueing etc.
663 */
664 if (td->o.nr_files == 1 && td->o.time_based) {
665 put_file(td, f);
666 fio_file_reset(f);
667 goto set_file;
668 }
669
670 /*
671 * td_io_close() does a put_file() as well, so no need to
672 * do that here.
673 */
674 io_u->file = NULL;
675 td_io_close_file(td, f);
676 f->flags |= FIO_FILE_DONE;
677 td->nr_done_files++;
678
679 /*
680 * probably not the right place to do this, but see
681 * if we need to open a new file
682 */
683 if (td->nr_open_files < td->o.open_files &&
684 td->o.open_files != td->o.nr_files) {
685 f = find_next_new_file(td);
686
687 if (!f || td_io_open_file(td, f))
688 return 1;
689
690 goto set_file;
691 }
692 } while (1);
693
694 return 0;
695}
696
697
698struct io_u *__get_io_u(struct thread_data *td)
699{
700 struct io_u *io_u = NULL;
701
702 if (!list_empty(&td->io_u_requeues))
703 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
704 else if (!queue_full(td)) {
705 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
706
707 io_u->buflen = 0;
708 io_u->resid = 0;
709 io_u->file = NULL;
710 io_u->end_io = NULL;
711 }
712
713 if (io_u) {
714 assert(io_u->flags & IO_U_F_FREE);
715 io_u->flags &= ~IO_U_F_FREE;
716
717 io_u->error = 0;
718 list_del(&io_u->list);
719 list_add(&io_u->list, &td->io_u_busylist);
720 td->cur_depth++;
721 }
722
723 return io_u;
724}
725
726/*
727 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
728 * etc. The returned io_u is fully ready to be prepped and submitted.
729 */
730struct io_u *get_io_u(struct thread_data *td)
731{
732 struct fio_file *f;
733 struct io_u *io_u;
734
735 io_u = __get_io_u(td);
736 if (!io_u) {
737 dprint(FD_IO, "__get_io_u failed\n");
738 return NULL;
739 }
740
741 /*
742 * from a requeue, io_u already setup
743 */
744 if (io_u->file)
745 goto out;
746
747 /*
748 * If using an iolog, grab next piece if any available.
749 */
750 if (td->o.read_iolog_file) {
751 if (read_iolog_get(td, io_u))
752 goto err_put;
753 } else if (set_io_u_file(td, io_u)) {
754 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
755 goto err_put;
756 }
757
758 f = io_u->file;
759 assert(f->flags & FIO_FILE_OPEN);
760
761 if (io_u->ddir != DDIR_SYNC) {
762 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
763 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
764 goto err_put;
765 }
766
767 f->last_pos = io_u->offset + io_u->buflen;
768
769 if (td->o.verify != VERIFY_NONE)
770 populate_verify_io_u(td, io_u);
771 }
772
773 /*
774 * Set io data pointers.
775 */
776 io_u->endpos = io_u->offset + io_u->buflen;
777 io_u->xfer_buf = io_u->buf;
778 io_u->xfer_buflen = io_u->buflen;
779out:
780 if (!td_io_prep(td, io_u)) {
781 fio_gettime(&io_u->start_time, NULL);
782 return io_u;
783 }
784err_put:
785 dprint(FD_IO, "get_io_u failed\n");
786 put_io_u(td, io_u);
787 return NULL;
788}
789
790void io_u_log_error(struct thread_data *td, struct io_u *io_u)
791{
792 const char *msg[] = { "read", "write", "sync" };
793
794 log_err("fio: io_u error");
795
796 if (io_u->file)
797 log_err(" on file %s", io_u->file->file_name);
798
799 log_err(": %s\n", strerror(io_u->error));
800
801 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
802 io_u->offset, io_u->xfer_buflen);
803
804 if (!td->error)
805 td_verror(td, io_u->error, "io_u error");
806}
807
808static void io_completed(struct thread_data *td, struct io_u *io_u,
809 struct io_completion_data *icd)
810{
811 unsigned long usec;
812
813 dprint_io_u(io_u, "io complete");
814
815 assert(io_u->flags & IO_U_F_FLIGHT);
816 io_u->flags &= ~IO_U_F_FLIGHT;
817
818 if (io_u->ddir == DDIR_SYNC) {
819 td->last_was_sync = 1;
820 return;
821 }
822
823 td->last_was_sync = 0;
824
825 if (!io_u->error) {
826 unsigned int bytes = io_u->buflen - io_u->resid;
827 const enum fio_ddir idx = io_u->ddir;
828 int ret;
829
830 td->io_blocks[idx]++;
831 td->io_bytes[idx] += bytes;
832 td->this_io_bytes[idx] += bytes;
833
834 usec = utime_since(&io_u->issue_time, &icd->time);
835
836 add_clat_sample(td, idx, usec);
837 add_bw_sample(td, idx, &icd->time);
838 io_u_mark_latency(td, usec);
839
840 if (td_write(td) && idx == DDIR_WRITE &&
841 td->o.do_verify &&
842 td->o.verify != VERIFY_NONE)
843 log_io_piece(td, io_u);
844
845 icd->bytes_done[idx] += bytes;
846
847 if (io_u->end_io) {
848 ret = io_u->end_io(td, io_u);
849 if (ret && !icd->error)
850 icd->error = ret;
851 }
852 } else {
853 icd->error = io_u->error;
854 io_u_log_error(td, io_u);
855 }
856}
857
858static void init_icd(struct io_completion_data *icd, int nr)
859{
860 fio_gettime(&icd->time, NULL);
861
862 icd->nr = nr;
863
864 icd->error = 0;
865 icd->bytes_done[0] = icd->bytes_done[1] = 0;
866}
867
868static void ios_completed(struct thread_data *td,
869 struct io_completion_data *icd)
870{
871 struct io_u *io_u;
872 int i;
873
874 for (i = 0; i < icd->nr; i++) {
875 io_u = td->io_ops->event(td, i);
876
877 io_completed(td, io_u, icd);
878 put_io_u(td, io_u);
879 }
880}
881
882/*
883 * Complete a single io_u for the sync engines.
884 */
885long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
886{
887 struct io_completion_data icd;
888
889 init_icd(&icd, 1);
890 io_completed(td, io_u, &icd);
891 put_io_u(td, io_u);
892
893 if (!icd.error)
894 return icd.bytes_done[0] + icd.bytes_done[1];
895
896 td_verror(td, icd.error, "io_u_sync_complete");
897 return -1;
898}
899
900/*
901 * Called to complete min_events number of io for the async engines.
902 */
903long io_u_queued_complete(struct thread_data *td, int min_events)
904{
905 struct io_completion_data icd;
906 struct timespec *tvp = NULL;
907 int ret;
908 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
909
910 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
911
912 if (!min_events)
913 tvp = &ts;
914
915 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
916 if (ret < 0) {
917 td_verror(td, -ret, "td_io_getevents");
918 return ret;
919 } else if (!ret)
920 return ret;
921
922 init_icd(&icd, ret);
923 ios_completed(td, &icd);
924 if (!icd.error)
925 return icd.bytes_done[0] + icd.bytes_done[1];
926
927 td_verror(td, icd.error, "io_u_queued_complete");
928 return -1;
929}
930
931/*
932 * Call when io_u is really queued, to update the submission latency.
933 */
934void io_u_queued(struct thread_data *td, struct io_u *io_u)
935{
936 unsigned long slat_time;
937
938 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
939 add_slat_sample(td, io_u->ddir, slat_time);
940}
941
942#ifdef FIO_USE_TIMEOUT
943void io_u_set_timeout(struct thread_data *td)
944{
945 assert(td->cur_depth);
946
947 td->timer.it_interval.tv_sec = 0;
948 td->timer.it_interval.tv_usec = 0;
949 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
950 td->timer.it_value.tv_usec = 0;
951 setitimer(ITIMER_REAL, &td->timer, NULL);
952 fio_gettime(&td->timeout_end, NULL);
953}
954
955static void io_u_dump(struct io_u *io_u)
956{
957 unsigned long t_start = mtime_since_now(&io_u->start_time);
958 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
959
960 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
961 log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf,
962 io_u->xfer_buf, io_u->buflen,
963 io_u->xfer_buflen,
964 io_u->offset);
965 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
966}
967#else
968void io_u_set_timeout(struct thread_data fio_unused *td)
969{
970}
971#endif
972
973#ifdef FIO_USE_TIMEOUT
974static void io_u_timeout_handler(int fio_unused sig)
975{
976 struct thread_data *td, *__td;
977 pid_t pid = getpid();
978 struct list_head *entry;
979 struct io_u *io_u;
980 int i;
981
982 log_err("fio: io_u timeout\n");
983
984 /*
985 * TLS would be nice...
986 */
987 td = NULL;
988 for_each_td(__td, i) {
989 if (__td->pid == pid) {
990 td = __td;
991 break;
992 }
993 }
994
995 if (!td) {
996 log_err("fio: io_u timeout, can't find job\n");
997 exit(1);
998 }
999
1000 if (!td->cur_depth) {
1001 log_err("fio: timeout without pending work?\n");
1002 return;
1003 }
1004
1005 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
1006
1007 list_for_each(entry, &td->io_u_busylist) {
1008 io_u = list_entry(entry, struct io_u, list);
1009
1010 io_u_dump(io_u);
1011 }
1012
1013 td_verror(td, ETIMEDOUT, "io_u timeout");
1014 exit(1);
1015}
1016#endif
1017
1018void io_u_init_timeout(void)
1019{
1020#ifdef FIO_USE_TIMEOUT
1021 signal(SIGALRM, io_u_timeout_handler);
1022#endif
1023}
fio - Flexible IO Tester