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