backend: only check bytes_done for engines without FIO_NOIO
[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#include "hash.h"
10#include "verify.h"
11#include "trim.h"
12#include "lib/rand.h"
13
14struct io_completion_data {
15 int nr; /* input */
16
17 int error; /* output */
18 unsigned long bytes_done[DDIR_RWDIR_CNT]; /* output */
19 struct timeval time; /* output */
20};
21
22/*
23 * The ->file_map[] contains a map of blocks we have or have not done io
24 * to yet. Used to make sure we cover the entire range in a fair fashion.
25 */
26static int random_map_free(struct fio_file *f, const unsigned long long block)
27{
28 unsigned int idx = RAND_MAP_IDX(f, block);
29 unsigned int bit = RAND_MAP_BIT(f, block);
30
31 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
32
33 return (f->file_map[idx] & (1UL << bit)) == 0;
34}
35
36/*
37 * Mark a given offset as used in the map.
38 */
39static void mark_random_map(struct thread_data *td, struct io_u *io_u)
40{
41 unsigned int min_bs = td->o.rw_min_bs;
42 struct fio_file *f = io_u->file;
43 unsigned long long block;
44 unsigned int blocks, nr_blocks;
45 int busy_check;
46
47 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
48 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
49 blocks = 0;
50 busy_check = !(io_u->flags & IO_U_F_BUSY_OK);
51
52 while (nr_blocks) {
53 unsigned int idx, bit;
54 unsigned long mask, this_blocks;
55
56 /*
57 * If we have a mixed random workload, we may
58 * encounter blocks we already did IO to.
59 */
60 if (!busy_check) {
61 blocks = nr_blocks;
62 break;
63 }
64 if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block))
65 break;
66
67 idx = RAND_MAP_IDX(f, block);
68 bit = RAND_MAP_BIT(f, block);
69
70 fio_assert(td, idx < f->num_maps);
71
72 this_blocks = nr_blocks;
73 if (this_blocks + bit > BLOCKS_PER_MAP)
74 this_blocks = BLOCKS_PER_MAP - bit;
75
76 do {
77 if (this_blocks == BLOCKS_PER_MAP)
78 mask = -1UL;
79 else
80 mask = ((1UL << this_blocks) - 1) << bit;
81
82 if (!(f->file_map[idx] & mask))
83 break;
84
85 this_blocks--;
86 } while (this_blocks);
87
88 if (!this_blocks)
89 break;
90
91 f->file_map[idx] |= mask;
92 nr_blocks -= this_blocks;
93 blocks += this_blocks;
94 block += this_blocks;
95 }
96
97 if ((blocks * min_bs) < io_u->buflen)
98 io_u->buflen = blocks * min_bs;
99}
100
101static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
102 enum fio_ddir ddir)
103{
104 unsigned long long max_blocks;
105 unsigned long long max_size;
106
107 assert(ddir_rw(ddir));
108
109 /*
110 * Hmm, should we make sure that ->io_size <= ->real_file_size?
111 */
112 max_size = f->io_size;
113 if (max_size > f->real_file_size)
114 max_size = f->real_file_size;
115
116 if (td->o.zone_range)
117 max_size = td->o.zone_range;
118
119 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
120 if (!max_blocks)
121 return 0;
122
123 return max_blocks;
124}
125
126/*
127 * Return the next free block in the map.
128 */
129static int get_next_free_block(struct thread_data *td, struct fio_file *f,
130 enum fio_ddir ddir, unsigned long long *b)
131{
132 unsigned long long block, min_bs = td->o.rw_min_bs, lastb;
133 int i;
134
135 lastb = last_block(td, f, ddir);
136 if (!lastb)
137 return 1;
138
139 i = f->last_free_lookup;
140 block = i * BLOCKS_PER_MAP;
141 while (block * min_bs < f->real_file_size &&
142 block * min_bs < f->io_size) {
143 if (f->file_map[i] != -1UL) {
144 block += ffz(f->file_map[i]);
145 if (block > lastb)
146 break;
147 f->last_free_lookup = i;
148 *b = block;
149 return 0;
150 }
151
152 block += BLOCKS_PER_MAP;
153 i++;
154 }
155
156 dprint(FD_IO, "failed finding a free block\n");
157 return 1;
158}
159
160static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
161 enum fio_ddir ddir, unsigned long long *b)
162{
163 unsigned long long rmax, r, lastb;
164 int loops = 5;
165
166 lastb = last_block(td, f, ddir);
167 if (!lastb)
168 return 1;
169
170 if (f->failed_rands >= 200)
171 goto ffz;
172
173 rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
174 do {
175 if (td->o.use_os_rand)
176 r = os_random_long(&td->random_state);
177 else
178 r = __rand(&td->__random_state);
179
180 *b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0));
181
182 dprint(FD_RANDOM, "off rand %llu\n", r);
183
184
185 /*
186 * if we are not maintaining a random map, we are done.
187 */
188 if (!file_randommap(td, f))
189 goto ret_good;
190
191 /*
192 * calculate map offset and check if it's free
193 */
194 if (random_map_free(f, *b))
195 goto ret_good;
196
197 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
198 *b);
199 } while (--loops);
200
201 if (!f->failed_rands++)
202 f->last_free_lookup = 0;
203
204 /*
205 * we get here, if we didn't suceed in looking up a block. generate
206 * a random start offset into the filemap, and find the first free
207 * block from there.
208 */
209 loops = 10;
210 do {
211 f->last_free_lookup = (f->num_maps - 1) *
212 (r / ((unsigned long long) rmax + 1.0));
213 if (!get_next_free_block(td, f, ddir, b))
214 goto ret;
215
216 if (td->o.use_os_rand)
217 r = os_random_long(&td->random_state);
218 else
219 r = __rand(&td->__random_state);
220 } while (--loops);
221
222 /*
223 * that didn't work either, try exhaustive search from the start
224 */
225 f->last_free_lookup = 0;
226ffz:
227 if (!get_next_free_block(td, f, ddir, b))
228 return 0;
229 f->last_free_lookup = 0;
230 return get_next_free_block(td, f, ddir, b);
231ret_good:
232 f->failed_rands = 0;
233ret:
234 return 0;
235}
236
237static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
238 enum fio_ddir ddir, unsigned long long *b)
239{
240 if (!get_next_rand_offset(td, f, ddir, b))
241 return 0;
242
243 if (td->o.time_based) {
244 fio_file_reset(f);
245 if (!get_next_rand_offset(td, f, ddir, b))
246 return 0;
247 }
248
249 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
250 f->file_name, f->last_pos, f->real_file_size);
251 return 1;
252}
253
254static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
255 enum fio_ddir ddir, unsigned long long *offset)
256{
257 assert(ddir_rw(ddir));
258
259 if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
260 f->last_pos = f->last_pos - f->io_size;
261
262 if (f->last_pos < f->real_file_size) {
263 unsigned long long pos;
264
265 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
266 f->last_pos = f->real_file_size;
267
268 pos = f->last_pos - f->file_offset;
269 if (pos)
270 pos += td->o.ddir_seq_add;
271
272 *offset = pos;
273 return 0;
274 }
275
276 return 1;
277}
278
279static int get_next_block(struct thread_data *td, struct io_u *io_u,
280 enum fio_ddir ddir, int rw_seq)
281{
282 struct fio_file *f = io_u->file;
283 unsigned long long b, offset;
284 int ret;
285
286 assert(ddir_rw(ddir));
287
288 b = offset = -1ULL;
289
290 if (rw_seq) {
291 if (td_random(td))
292 ret = get_next_rand_block(td, f, ddir, &b);
293 else
294 ret = get_next_seq_offset(td, f, ddir, &offset);
295 } else {
296 io_u->flags |= IO_U_F_BUSY_OK;
297
298 if (td->o.rw_seq == RW_SEQ_SEQ) {
299 ret = get_next_seq_offset(td, f, ddir, &offset);
300 if (ret)
301 ret = get_next_rand_block(td, f, ddir, &b);
302 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
303 if (f->last_start != -1ULL)
304 offset = f->last_start - f->file_offset;
305 else
306 offset = 0;
307 ret = 0;
308 } else {
309 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
310 ret = 1;
311 }
312 }
313
314 if (!ret) {
315 if (offset != -1ULL)
316 io_u->offset = offset;
317 else if (b != -1ULL)
318 io_u->offset = b * td->o.ba[ddir];
319 else {
320 log_err("fio: bug in offset generation\n");
321 ret = 1;
322 }
323 }
324
325 return ret;
326}
327
328/*
329 * For random io, generate a random new block and see if it's used. Repeat
330 * until we find a free one. For sequential io, just return the end of
331 * the last io issued.
332 */
333static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
334{
335 struct fio_file *f = io_u->file;
336 enum fio_ddir ddir = io_u->ddir;
337 int rw_seq_hit = 0;
338
339 assert(ddir_rw(ddir));
340
341 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
342 rw_seq_hit = 1;
343 td->ddir_seq_nr = td->o.ddir_seq_nr;
344 }
345
346 if (get_next_block(td, io_u, ddir, rw_seq_hit))
347 return 1;
348
349 if (io_u->offset >= f->io_size) {
350 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
351 io_u->offset, f->io_size);
352 return 1;
353 }
354
355 io_u->offset += f->file_offset;
356 if (io_u->offset >= f->real_file_size) {
357 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
358 io_u->offset, f->real_file_size);
359 return 1;
360 }
361
362 return 0;
363}
364
365static int get_next_offset(struct thread_data *td, struct io_u *io_u)
366{
367 struct prof_io_ops *ops = &td->prof_io_ops;
368
369 if (ops->fill_io_u_off)
370 return ops->fill_io_u_off(td, io_u);
371
372 return __get_next_offset(td, io_u);
373}
374
375static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
376 unsigned int buflen)
377{
378 struct fio_file *f = io_u->file;
379
380 return io_u->offset + buflen <= f->io_size + get_start_offset(td);
381}
382
383static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
384{
385 const int ddir = io_u->ddir;
386 unsigned int uninitialized_var(buflen);
387 unsigned int minbs, maxbs;
388 unsigned long r, rand_max;
389
390 assert(ddir_rw(ddir));
391
392 minbs = td->o.min_bs[ddir];
393 maxbs = td->o.max_bs[ddir];
394
395 if (minbs == maxbs)
396 return minbs;
397
398 /*
399 * If we can't satisfy the min block size from here, then fail
400 */
401 if (!io_u_fits(td, io_u, minbs))
402 return 0;
403
404 if (td->o.use_os_rand)
405 rand_max = OS_RAND_MAX;
406 else
407 rand_max = FRAND_MAX;
408
409 do {
410 if (td->o.use_os_rand)
411 r = os_random_long(&td->bsrange_state);
412 else
413 r = __rand(&td->__bsrange_state);
414
415 if (!td->o.bssplit_nr[ddir]) {
416 buflen = 1 + (unsigned int) ((double) maxbs *
417 (r / (rand_max + 1.0)));
418 if (buflen < minbs)
419 buflen = minbs;
420 } else {
421 long perc = 0;
422 unsigned int i;
423
424 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
425 struct bssplit *bsp = &td->o.bssplit[ddir][i];
426
427 buflen = bsp->bs;
428 perc += bsp->perc;
429 if ((r <= ((rand_max / 100L) * perc)) &&
430 io_u_fits(td, io_u, buflen))
431 break;
432 }
433 }
434
435 if (!td->o.bs_unaligned && is_power_of_2(minbs))
436 buflen = (buflen + minbs - 1) & ~(minbs - 1);
437
438 } while (!io_u_fits(td, io_u, buflen));
439
440 return buflen;
441}
442
443static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
444{
445 struct prof_io_ops *ops = &td->prof_io_ops;
446
447 if (ops->fill_io_u_size)
448 return ops->fill_io_u_size(td, io_u);
449
450 return __get_next_buflen(td, io_u);
451}
452
453static void set_rwmix_bytes(struct thread_data *td)
454{
455 unsigned int diff;
456
457 /*
458 * we do time or byte based switch. this is needed because
459 * buffered writes may issue a lot quicker than they complete,
460 * whereas reads do not.
461 */
462 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
463 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
464}
465
466static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
467{
468 unsigned int v;
469 unsigned long r;
470
471 if (td->o.use_os_rand) {
472 r = os_random_long(&td->rwmix_state);
473 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
474 } else {
475 r = __rand(&td->__rwmix_state);
476 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
477 }
478
479 if (v <= td->o.rwmix[DDIR_READ])
480 return DDIR_READ;
481
482 return DDIR_WRITE;
483}
484
485static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
486{
487 enum fio_ddir odir = ddir ^ 1;
488 struct timeval t;
489 long usec;
490
491 assert(ddir_rw(ddir));
492
493 if (td->rate_pending_usleep[ddir] <= 0)
494 return ddir;
495
496 /*
497 * We have too much pending sleep in this direction. See if we
498 * should switch.
499 */
500 if (td_rw(td)) {
501 /*
502 * Other direction does not have too much pending, switch
503 */
504 if (td->rate_pending_usleep[odir] < 100000)
505 return odir;
506
507 /*
508 * Both directions have pending sleep. Sleep the minimum time
509 * and deduct from both.
510 */
511 if (td->rate_pending_usleep[ddir] <=
512 td->rate_pending_usleep[odir]) {
513 usec = td->rate_pending_usleep[ddir];
514 } else {
515 usec = td->rate_pending_usleep[odir];
516 ddir = odir;
517 }
518 } else
519 usec = td->rate_pending_usleep[ddir];
520
521 /*
522 * We are going to sleep, ensure that we flush anything pending as
523 * not to skew our latency numbers.
524 *
525 * Changed to only monitor 'in flight' requests here instead of the
526 * td->cur_depth, b/c td->cur_depth does not accurately represent
527 * io's that have been actually submitted to an async engine,
528 * and cur_depth is meaningless for sync engines.
529 */
530 if (td->io_u_in_flight) {
531 int fio_unused ret;
532
533 ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
534 }
535
536 fio_gettime(&t, NULL);
537 usec_sleep(td, usec);
538 usec = utime_since_now(&t);
539
540 td->rate_pending_usleep[ddir] -= usec;
541
542 odir = ddir ^ 1;
543 if (td_rw(td) && __should_check_rate(td, odir))
544 td->rate_pending_usleep[odir] -= usec;
545
546 if (ddir_trim(ddir))
547 return ddir;
548 return ddir;
549}
550
551/*
552 * Return the data direction for the next io_u. If the job is a
553 * mixed read/write workload, check the rwmix cycle and switch if
554 * necessary.
555 */
556static enum fio_ddir get_rw_ddir(struct thread_data *td)
557{
558 enum fio_ddir ddir;
559
560 /*
561 * see if it's time to fsync
562 */
563 if (td->o.fsync_blocks &&
564 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
565 td->io_issues[DDIR_WRITE] && should_fsync(td))
566 return DDIR_SYNC;
567
568 /*
569 * see if it's time to fdatasync
570 */
571 if (td->o.fdatasync_blocks &&
572 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
573 td->io_issues[DDIR_WRITE] && should_fsync(td))
574 return DDIR_DATASYNC;
575
576 /*
577 * see if it's time to sync_file_range
578 */
579 if (td->sync_file_range_nr &&
580 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
581 td->io_issues[DDIR_WRITE] && should_fsync(td))
582 return DDIR_SYNC_FILE_RANGE;
583
584 if (td_rw(td)) {
585 /*
586 * Check if it's time to seed a new data direction.
587 */
588 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
589 /*
590 * Put a top limit on how many bytes we do for
591 * one data direction, to avoid overflowing the
592 * ranges too much
593 */
594 ddir = get_rand_ddir(td);
595
596 if (ddir != td->rwmix_ddir)
597 set_rwmix_bytes(td);
598
599 td->rwmix_ddir = ddir;
600 }
601 ddir = td->rwmix_ddir;
602 } else if (td_read(td))
603 ddir = DDIR_READ;
604 else if (td_write(td))
605 ddir = DDIR_WRITE;
606 else
607 ddir = DDIR_TRIM;
608
609 td->rwmix_ddir = rate_ddir(td, ddir);
610 return td->rwmix_ddir;
611}
612
613static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
614{
615 io_u->ddir = get_rw_ddir(td);
616
617 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
618 td->o.barrier_blocks &&
619 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
620 td->io_issues[DDIR_WRITE])
621 io_u->flags |= IO_U_F_BARRIER;
622}
623
624void put_file_log(struct thread_data *td, struct fio_file *f)
625{
626 int ret = put_file(td, f);
627
628 if (ret)
629 td_verror(td, ret, "file close");
630}
631
632void put_io_u(struct thread_data *td, struct io_u *io_u)
633{
634 td_io_u_lock(td);
635
636 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
637 put_file_log(td, io_u->file);
638 io_u->file = NULL;
639 io_u->flags &= ~IO_U_F_FREE_DEF;
640 io_u->flags |= IO_U_F_FREE;
641
642 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
643 td->cur_depth--;
644 flist_del_init(&io_u->list);
645 flist_add(&io_u->list, &td->io_u_freelist);
646 td_io_u_unlock(td);
647 td_io_u_free_notify(td);
648}
649
650void clear_io_u(struct thread_data *td, struct io_u *io_u)
651{
652 io_u->flags &= ~IO_U_F_FLIGHT;
653 put_io_u(td, io_u);
654}
655
656void requeue_io_u(struct thread_data *td, struct io_u **io_u)
657{
658 struct io_u *__io_u = *io_u;
659
660 dprint(FD_IO, "requeue %p\n", __io_u);
661
662 td_io_u_lock(td);
663
664 __io_u->flags |= IO_U_F_FREE;
665 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
666 td->io_issues[__io_u->ddir]--;
667
668 __io_u->flags &= ~IO_U_F_FLIGHT;
669 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
670 td->cur_depth--;
671 flist_del(&__io_u->list);
672 flist_add_tail(&__io_u->list, &td->io_u_requeues);
673 td_io_u_unlock(td);
674 *io_u = NULL;
675}
676
677static int fill_io_u(struct thread_data *td, struct io_u *io_u)
678{
679 if (td->io_ops->flags & FIO_NOIO)
680 goto out;
681
682 set_rw_ddir(td, io_u);
683
684 /*
685 * fsync() or fdatasync() or trim etc, we are done
686 */
687 if (!ddir_rw(io_u->ddir))
688 goto out;
689
690 /*
691 * See if it's time to switch to a new zone
692 */
693 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
694 td->zone_bytes = 0;
695 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
696 io_u->file->last_pos = io_u->file->file_offset;
697 td->io_skip_bytes += td->o.zone_skip;
698 }
699
700 /*
701 * No log, let the seq/rand engine retrieve the next buflen and
702 * position.
703 */
704 if (get_next_offset(td, io_u)) {
705 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
706 return 1;
707 }
708
709 io_u->buflen = get_next_buflen(td, io_u);
710 if (!io_u->buflen) {
711 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
712 return 1;
713 }
714
715 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
716 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
717 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
718 io_u->buflen, io_u->file->real_file_size);
719 return 1;
720 }
721
722 /*
723 * mark entry before potentially trimming io_u
724 */
725 if (td_random(td) && file_randommap(td, io_u->file))
726 mark_random_map(td, io_u);
727
728 /*
729 * If using a write iolog, store this entry.
730 */
731out:
732 dprint_io_u(io_u, "fill_io_u");
733 td->zone_bytes += io_u->buflen;
734 log_io_u(td, io_u);
735 return 0;
736}
737
738static void __io_u_mark_map(unsigned int *map, unsigned int nr)
739{
740 int idx = 0;
741
742 switch (nr) {
743 default:
744 idx = 6;
745 break;
746 case 33 ... 64:
747 idx = 5;
748 break;
749 case 17 ... 32:
750 idx = 4;
751 break;
752 case 9 ... 16:
753 idx = 3;
754 break;
755 case 5 ... 8:
756 idx = 2;
757 break;
758 case 1 ... 4:
759 idx = 1;
760 case 0:
761 break;
762 }
763
764 map[idx]++;
765}
766
767void io_u_mark_submit(struct thread_data *td, unsigned int nr)
768{
769 __io_u_mark_map(td->ts.io_u_submit, nr);
770 td->ts.total_submit++;
771}
772
773void io_u_mark_complete(struct thread_data *td, unsigned int nr)
774{
775 __io_u_mark_map(td->ts.io_u_complete, nr);
776 td->ts.total_complete++;
777}
778
779void io_u_mark_depth(struct thread_data *td, unsigned int nr)
780{
781 int idx = 0;
782
783 switch (td->cur_depth) {
784 default:
785 idx = 6;
786 break;
787 case 32 ... 63:
788 idx = 5;
789 break;
790 case 16 ... 31:
791 idx = 4;
792 break;
793 case 8 ... 15:
794 idx = 3;
795 break;
796 case 4 ... 7:
797 idx = 2;
798 break;
799 case 2 ... 3:
800 idx = 1;
801 case 1:
802 break;
803 }
804
805 td->ts.io_u_map[idx] += nr;
806}
807
808static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
809{
810 int idx = 0;
811
812 assert(usec < 1000);
813
814 switch (usec) {
815 case 750 ... 999:
816 idx = 9;
817 break;
818 case 500 ... 749:
819 idx = 8;
820 break;
821 case 250 ... 499:
822 idx = 7;
823 break;
824 case 100 ... 249:
825 idx = 6;
826 break;
827 case 50 ... 99:
828 idx = 5;
829 break;
830 case 20 ... 49:
831 idx = 4;
832 break;
833 case 10 ... 19:
834 idx = 3;
835 break;
836 case 4 ... 9:
837 idx = 2;
838 break;
839 case 2 ... 3:
840 idx = 1;
841 case 0 ... 1:
842 break;
843 }
844
845 assert(idx < FIO_IO_U_LAT_U_NR);
846 td->ts.io_u_lat_u[idx]++;
847}
848
849static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
850{
851 int idx = 0;
852
853 switch (msec) {
854 default:
855 idx = 11;
856 break;
857 case 1000 ... 1999:
858 idx = 10;
859 break;
860 case 750 ... 999:
861 idx = 9;
862 break;
863 case 500 ... 749:
864 idx = 8;
865 break;
866 case 250 ... 499:
867 idx = 7;
868 break;
869 case 100 ... 249:
870 idx = 6;
871 break;
872 case 50 ... 99:
873 idx = 5;
874 break;
875 case 20 ... 49:
876 idx = 4;
877 break;
878 case 10 ... 19:
879 idx = 3;
880 break;
881 case 4 ... 9:
882 idx = 2;
883 break;
884 case 2 ... 3:
885 idx = 1;
886 case 0 ... 1:
887 break;
888 }
889
890 assert(idx < FIO_IO_U_LAT_M_NR);
891 td->ts.io_u_lat_m[idx]++;
892}
893
894static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
895{
896 if (usec < 1000)
897 io_u_mark_lat_usec(td, usec);
898 else
899 io_u_mark_lat_msec(td, usec / 1000);
900}
901
902/*
903 * Get next file to service by choosing one at random
904 */
905static struct fio_file *get_next_file_rand(struct thread_data *td,
906 enum fio_file_flags goodf,
907 enum fio_file_flags badf)
908{
909 struct fio_file *f;
910 int fno;
911
912 do {
913 int opened = 0;
914 unsigned long r;
915
916 if (td->o.use_os_rand) {
917 r = os_random_long(&td->next_file_state);
918 fno = (unsigned int) ((double) td->o.nr_files
919 * (r / (OS_RAND_MAX + 1.0)));
920 } else {
921 r = __rand(&td->__next_file_state);
922 fno = (unsigned int) ((double) td->o.nr_files
923 * (r / (FRAND_MAX + 1.0)));
924 }
925
926 f = td->files[fno];
927 if (fio_file_done(f))
928 continue;
929
930 if (!fio_file_open(f)) {
931 int err;
932
933 err = td_io_open_file(td, f);
934 if (err)
935 continue;
936 opened = 1;
937 }
938
939 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
940 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
941 return f;
942 }
943 if (opened)
944 td_io_close_file(td, f);
945 } while (1);
946}
947
948/*
949 * Get next file to service by doing round robin between all available ones
950 */
951static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
952 int badf)
953{
954 unsigned int old_next_file = td->next_file;
955 struct fio_file *f;
956
957 do {
958 int opened = 0;
959
960 f = td->files[td->next_file];
961
962 td->next_file++;
963 if (td->next_file >= td->o.nr_files)
964 td->next_file = 0;
965
966 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
967 if (fio_file_done(f)) {
968 f = NULL;
969 continue;
970 }
971
972 if (!fio_file_open(f)) {
973 int err;
974
975 err = td_io_open_file(td, f);
976 if (err) {
977 dprint(FD_FILE, "error %d on open of %s\n",
978 err, f->file_name);
979 f = NULL;
980 continue;
981 }
982 opened = 1;
983 }
984
985 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
986 f->flags);
987 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
988 break;
989
990 if (opened)
991 td_io_close_file(td, f);
992
993 f = NULL;
994 } while (td->next_file != old_next_file);
995
996 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
997 return f;
998}
999
1000static struct fio_file *__get_next_file(struct thread_data *td)
1001{
1002 struct fio_file *f;
1003
1004 assert(td->o.nr_files <= td->files_index);
1005
1006 if (td->nr_done_files >= td->o.nr_files) {
1007 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1008 " nr_files=%d\n", td->nr_open_files,
1009 td->nr_done_files,
1010 td->o.nr_files);
1011 return NULL;
1012 }
1013
1014 f = td->file_service_file;
1015 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1016 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1017 goto out;
1018 if (td->file_service_left--)
1019 goto out;
1020 }
1021
1022 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1023 td->o.file_service_type == FIO_FSERVICE_SEQ)
1024 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1025 else
1026 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1027
1028 td->file_service_file = f;
1029 td->file_service_left = td->file_service_nr - 1;
1030out:
1031 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1032 return f;
1033}
1034
1035static struct fio_file *get_next_file(struct thread_data *td)
1036{
1037 struct prof_io_ops *ops = &td->prof_io_ops;
1038
1039 if (ops->get_next_file)
1040 return ops->get_next_file(td);
1041
1042 return __get_next_file(td);
1043}
1044
1045static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1046{
1047 struct fio_file *f;
1048
1049 do {
1050 f = get_next_file(td);
1051 if (!f)
1052 return 1;
1053
1054 io_u->file = f;
1055 get_file(f);
1056
1057 if (!fill_io_u(td, io_u))
1058 break;
1059
1060 put_file_log(td, f);
1061 td_io_close_file(td, f);
1062 io_u->file = NULL;
1063 fio_file_set_done(f);
1064 td->nr_done_files++;
1065 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1066 td->nr_done_files, td->o.nr_files);
1067 } while (1);
1068
1069 return 0;
1070}
1071
1072
1073struct io_u *__get_io_u(struct thread_data *td)
1074{
1075 struct io_u *io_u = NULL;
1076
1077 td_io_u_lock(td);
1078
1079again:
1080 if (!flist_empty(&td->io_u_requeues))
1081 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1082 else if (!queue_full(td)) {
1083 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1084
1085 io_u->buflen = 0;
1086 io_u->resid = 0;
1087 io_u->file = NULL;
1088 io_u->end_io = NULL;
1089 }
1090
1091 if (io_u) {
1092 assert(io_u->flags & IO_U_F_FREE);
1093 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1094 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1095 io_u->flags &= ~IO_U_F_VER_LIST;
1096
1097 io_u->error = 0;
1098 flist_del(&io_u->list);
1099 flist_add_tail(&io_u->list, &td->io_u_busylist);
1100 td->cur_depth++;
1101 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1102 } else if (td->o.verify_async) {
1103 /*
1104 * We ran out, wait for async verify threads to finish and
1105 * return one
1106 */
1107 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1108 goto again;
1109 }
1110
1111 td_io_u_unlock(td);
1112 return io_u;
1113}
1114
1115static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1116{
1117 if (td->o.trim_backlog && td->trim_entries) {
1118 int get_trim = 0;
1119
1120 if (td->trim_batch) {
1121 td->trim_batch--;
1122 get_trim = 1;
1123 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1124 td->last_ddir != DDIR_READ) {
1125 td->trim_batch = td->o.trim_batch;
1126 if (!td->trim_batch)
1127 td->trim_batch = td->o.trim_backlog;
1128 get_trim = 1;
1129 }
1130
1131 if (get_trim && !get_next_trim(td, io_u))
1132 return 1;
1133 }
1134
1135 return 0;
1136}
1137
1138static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1139{
1140 if (td->o.verify_backlog && td->io_hist_len) {
1141 int get_verify = 0;
1142
1143 if (td->verify_batch)
1144 get_verify = 1;
1145 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1146 td->last_ddir != DDIR_READ) {
1147 td->verify_batch = td->o.verify_batch;
1148 if (!td->verify_batch)
1149 td->verify_batch = td->o.verify_backlog;
1150 get_verify = 1;
1151 }
1152
1153 if (get_verify && !get_next_verify(td, io_u)) {
1154 td->verify_batch--;
1155 return 1;
1156 }
1157 }
1158
1159 return 0;
1160}
1161
1162/*
1163 * Fill offset and start time into the buffer content, to prevent too
1164 * easy compressible data for simple de-dupe attempts. Do this for every
1165 * 512b block in the range, since that should be the smallest block size
1166 * we can expect from a device.
1167 */
1168static void small_content_scramble(struct io_u *io_u)
1169{
1170 unsigned int i, nr_blocks = io_u->buflen / 512;
1171 unsigned long long boffset;
1172 unsigned int offset;
1173 void *p, *end;
1174
1175 if (!nr_blocks)
1176 return;
1177
1178 p = io_u->xfer_buf;
1179 boffset = io_u->offset;
1180 io_u->buf_filled_len = 0;
1181
1182 for (i = 0; i < nr_blocks; i++) {
1183 /*
1184 * Fill the byte offset into a "random" start offset of
1185 * the buffer, given by the product of the usec time
1186 * and the actual offset.
1187 */
1188 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1189 offset &= ~(sizeof(unsigned long long) - 1);
1190 if (offset >= 512 - sizeof(unsigned long long))
1191 offset -= sizeof(unsigned long long);
1192 memcpy(p + offset, &boffset, sizeof(boffset));
1193
1194 end = p + 512 - sizeof(io_u->start_time);
1195 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1196 p += 512;
1197 boffset += 512;
1198 }
1199}
1200
1201/*
1202 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1203 * etc. The returned io_u is fully ready to be prepped and submitted.
1204 */
1205struct io_u *get_io_u(struct thread_data *td)
1206{
1207 struct fio_file *f;
1208 struct io_u *io_u;
1209 int do_scramble = 0;
1210
1211 io_u = __get_io_u(td);
1212 if (!io_u) {
1213 dprint(FD_IO, "__get_io_u failed\n");
1214 return NULL;
1215 }
1216
1217 if (check_get_verify(td, io_u))
1218 goto out;
1219 if (check_get_trim(td, io_u))
1220 goto out;
1221
1222 /*
1223 * from a requeue, io_u already setup
1224 */
1225 if (io_u->file)
1226 goto out;
1227
1228 /*
1229 * If using an iolog, grab next piece if any available.
1230 */
1231 if (td->o.read_iolog_file) {
1232 if (read_iolog_get(td, io_u))
1233 goto err_put;
1234 } else if (set_io_u_file(td, io_u)) {
1235 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1236 goto err_put;
1237 }
1238
1239 f = io_u->file;
1240 assert(fio_file_open(f));
1241
1242 if (ddir_rw(io_u->ddir)) {
1243 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1244 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1245 goto err_put;
1246 }
1247
1248 f->last_start = io_u->offset;
1249 f->last_pos = io_u->offset + io_u->buflen;
1250
1251 if (io_u->ddir == DDIR_WRITE) {
1252 if (td->o.refill_buffers) {
1253 io_u_fill_buffer(td, io_u,
1254 io_u->xfer_buflen, io_u->xfer_buflen);
1255 } else if (td->o.scramble_buffers)
1256 do_scramble = 1;
1257 if (td->o.verify != VERIFY_NONE) {
1258 populate_verify_io_u(td, io_u);
1259 do_scramble = 0;
1260 }
1261 } else if (io_u->ddir == DDIR_READ) {
1262 /*
1263 * Reset the buf_filled parameters so next time if the
1264 * buffer is used for writes it is refilled.
1265 */
1266 io_u->buf_filled_len = 0;
1267 }
1268 }
1269
1270 /*
1271 * Set io data pointers.
1272 */
1273 io_u->xfer_buf = io_u->buf;
1274 io_u->xfer_buflen = io_u->buflen;
1275
1276out:
1277 assert(io_u->file);
1278 if (!td_io_prep(td, io_u)) {
1279 if (!td->o.disable_slat)
1280 fio_gettime(&io_u->start_time, NULL);
1281 if (do_scramble)
1282 small_content_scramble(io_u);
1283 return io_u;
1284 }
1285err_put:
1286 dprint(FD_IO, "get_io_u failed\n");
1287 put_io_u(td, io_u);
1288 return NULL;
1289}
1290
1291void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1292{
1293 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1294 const char *msg[] = { "read", "write", "sync", "datasync",
1295 "sync_file_range", "wait", "trim" };
1296
1297 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1298 return;
1299
1300 log_err("fio: io_u error");
1301
1302 if (io_u->file)
1303 log_err(" on file %s", io_u->file->file_name);
1304
1305 log_err(": %s\n", strerror(io_u->error));
1306
1307 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1308 io_u->offset, io_u->xfer_buflen);
1309
1310 if (!td->error)
1311 td_verror(td, io_u->error, "io_u error");
1312}
1313
1314static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1315 struct io_completion_data *icd,
1316 const enum fio_ddir idx, unsigned int bytes)
1317{
1318 unsigned long uninitialized_var(lusec);
1319
1320 if (!td->o.disable_clat || !td->o.disable_bw)
1321 lusec = utime_since(&io_u->issue_time, &icd->time);
1322
1323 if (!td->o.disable_lat) {
1324 unsigned long tusec;
1325
1326 tusec = utime_since(&io_u->start_time, &icd->time);
1327 add_lat_sample(td, idx, tusec, bytes);
1328 }
1329
1330 if (!td->o.disable_clat) {
1331 add_clat_sample(td, idx, lusec, bytes);
1332 io_u_mark_latency(td, lusec);
1333 }
1334
1335 if (!td->o.disable_bw)
1336 add_bw_sample(td, idx, bytes, &icd->time);
1337
1338 add_iops_sample(td, idx, &icd->time);
1339}
1340
1341static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1342{
1343 unsigned long long secs, remainder, bps, bytes;
1344 bytes = td->this_io_bytes[ddir];
1345 bps = td->rate_bps[ddir];
1346 secs = bytes / bps;
1347 remainder = bytes % bps;
1348 return remainder * 1000000 / bps + secs * 1000000;
1349}
1350
1351static void io_completed(struct thread_data *td, struct io_u *io_u,
1352 struct io_completion_data *icd)
1353{
1354 /*
1355 * Older gcc's are too dumb to realize that usec is always used
1356 * initialized, silence that warning.
1357 */
1358 unsigned long uninitialized_var(usec);
1359 struct fio_file *f;
1360
1361 dprint_io_u(io_u, "io complete");
1362
1363 td_io_u_lock(td);
1364 assert(io_u->flags & IO_U_F_FLIGHT);
1365 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1366 td_io_u_unlock(td);
1367
1368 if (ddir_sync(io_u->ddir)) {
1369 td->last_was_sync = 1;
1370 f = io_u->file;
1371 if (f) {
1372 f->first_write = -1ULL;
1373 f->last_write = -1ULL;
1374 }
1375 return;
1376 }
1377
1378 td->last_was_sync = 0;
1379 td->last_ddir = io_u->ddir;
1380
1381 if (!io_u->error && ddir_rw(io_u->ddir)) {
1382 unsigned int bytes = io_u->buflen - io_u->resid;
1383 const enum fio_ddir idx = io_u->ddir;
1384 const enum fio_ddir odx = io_u->ddir ^ 1;
1385 int ret;
1386
1387 td->io_blocks[idx]++;
1388 td->this_io_blocks[idx]++;
1389 td->io_bytes[idx] += bytes;
1390
1391 if (!(io_u->flags & IO_U_F_VER_LIST))
1392 td->this_io_bytes[idx] += bytes;
1393
1394 if (idx == DDIR_WRITE) {
1395 f = io_u->file;
1396 if (f) {
1397 if (f->first_write == -1ULL ||
1398 io_u->offset < f->first_write)
1399 f->first_write = io_u->offset;
1400 if (f->last_write == -1ULL ||
1401 ((io_u->offset + bytes) > f->last_write))
1402 f->last_write = io_u->offset + bytes;
1403 }
1404 }
1405
1406 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1407 td->runstate == TD_VERIFYING)) {
1408 account_io_completion(td, io_u, icd, idx, bytes);
1409
1410 if (__should_check_rate(td, idx)) {
1411 td->rate_pending_usleep[idx] =
1412 (usec_for_io(td, idx) -
1413 utime_since_now(&td->start));
1414 }
1415 if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1416 td->rate_pending_usleep[odx] =
1417 (usec_for_io(td, odx) -
1418 utime_since_now(&td->start));
1419 }
1420
1421 if (td_write(td) && idx == DDIR_WRITE &&
1422 td->o.do_verify &&
1423 td->o.verify != VERIFY_NONE)
1424 log_io_piece(td, io_u);
1425
1426 icd->bytes_done[idx] += bytes;
1427
1428 if (io_u->end_io) {
1429 ret = io_u->end_io(td, io_u);
1430 if (ret && !icd->error)
1431 icd->error = ret;
1432 }
1433 } else if (io_u->error) {
1434 icd->error = io_u->error;
1435 io_u_log_error(td, io_u);
1436 }
1437 if (icd->error) {
1438 enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1439 if (!td_non_fatal_error(td, eb, icd->error))
1440 return;
1441 /*
1442 * If there is a non_fatal error, then add to the error count
1443 * and clear all the errors.
1444 */
1445 update_error_count(td, icd->error);
1446 td_clear_error(td);
1447 icd->error = 0;
1448 io_u->error = 0;
1449 }
1450}
1451
1452static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1453 int nr)
1454{
1455 int ddir;
1456 if (!td->o.disable_clat || !td->o.disable_bw)
1457 fio_gettime(&icd->time, NULL);
1458
1459 icd->nr = nr;
1460
1461 icd->error = 0;
1462 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1463 icd->bytes_done[ddir] = 0;
1464}
1465
1466static void ios_completed(struct thread_data *td,
1467 struct io_completion_data *icd)
1468{
1469 struct io_u *io_u;
1470 int i;
1471
1472 for (i = 0; i < icd->nr; i++) {
1473 io_u = td->io_ops->event(td, i);
1474
1475 io_completed(td, io_u, icd);
1476
1477 if (!(io_u->flags & IO_U_F_FREE_DEF))
1478 put_io_u(td, io_u);
1479 }
1480}
1481
1482/*
1483 * Complete a single io_u for the sync engines.
1484 */
1485int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1486 unsigned long *bytes)
1487{
1488 struct io_completion_data icd;
1489
1490 init_icd(td, &icd, 1);
1491 io_completed(td, io_u, &icd);
1492
1493 if (!(io_u->flags & IO_U_F_FREE_DEF))
1494 put_io_u(td, io_u);
1495
1496 if (icd.error) {
1497 td_verror(td, icd.error, "io_u_sync_complete");
1498 return -1;
1499 }
1500
1501 if (bytes) {
1502 int ddir;
1503
1504 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1505 bytes[ddir] += icd.bytes_done[ddir];
1506 }
1507
1508 return 0;
1509}
1510
1511/*
1512 * Called to complete min_events number of io for the async engines.
1513 */
1514int io_u_queued_complete(struct thread_data *td, int min_evts,
1515 unsigned long *bytes)
1516{
1517 struct io_completion_data icd;
1518 struct timespec *tvp = NULL;
1519 int ret;
1520 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1521
1522 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1523
1524 if (!min_evts)
1525 tvp = &ts;
1526
1527 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1528 if (ret < 0) {
1529 td_verror(td, -ret, "td_io_getevents");
1530 return ret;
1531 } else if (!ret)
1532 return ret;
1533
1534 init_icd(td, &icd, ret);
1535 ios_completed(td, &icd);
1536 if (icd.error) {
1537 td_verror(td, icd.error, "io_u_queued_complete");
1538 return -1;
1539 }
1540
1541 if (bytes) {
1542 int ddir;
1543
1544 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1545 bytes[ddir] += icd.bytes_done[ddir];
1546 }
1547
1548 return 0;
1549}
1550
1551/*
1552 * Call when io_u is really queued, to update the submission latency.
1553 */
1554void io_u_queued(struct thread_data *td, struct io_u *io_u)
1555{
1556 if (!td->o.disable_slat) {
1557 unsigned long slat_time;
1558
1559 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1560 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1561 }
1562}
1563
1564/*
1565 * "randomly" fill the buffer contents
1566 */
1567void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1568 unsigned int min_write, unsigned int max_bs)
1569{
1570 io_u->buf_filled_len = 0;
1571
1572 if (!td->o.zero_buffers) {
1573 unsigned int perc = td->o.compress_percentage;
1574
1575 if (perc) {
1576 unsigned int seg = min_write;
1577
1578 seg = min(min_write, td->o.compress_chunk);
1579 fill_random_buf_percentage(&td->buf_state, io_u->buf,
1580 perc, seg, max_bs);
1581 } else
1582 fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1583 } else
1584 memset(io_u->buf, 0, max_bs);
1585}