projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Merge branch 'asprintf1' of https://github.com/kusumi/fio into master
[fio.git] / io_u.c
... / ...
CommitLineData
1#include <unistd.h>
2#include <string.h>
3#include <assert.h>
4
5#include "fio.h"
6#include "verify.h"
7#include "trim.h"
8#include "lib/rand.h"
9#include "lib/axmap.h"
10#include "err.h"
11#include "lib/pow2.h"
12#include "minmax.h"
13#include "zbd.h"
14
15struct io_completion_data {
16 int nr; /* input */
17
18 int error; /* output */
19 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
20 struct timespec time; /* output */
21};
22
23/*
24 * The ->io_axmap 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 bool random_map_free(struct fio_file *f, const uint64_t block)
28{
29 return !axmap_isset(f->io_axmap, block);
30}
31
32/*
33 * Mark a given offset as used in the map.
34 */
35static uint64_t mark_random_map(struct thread_data *td, struct io_u *io_u,
36 uint64_t offset, uint64_t buflen)
37{
38 unsigned long long min_bs = td->o.min_bs[io_u->ddir];
39 struct fio_file *f = io_u->file;
40 unsigned long long nr_blocks;
41 uint64_t block;
42
43 block = (offset - f->file_offset) / (uint64_t) min_bs;
44 nr_blocks = (buflen + min_bs - 1) / min_bs;
45 assert(nr_blocks > 0);
46
47 if (!(io_u->flags & IO_U_F_BUSY_OK)) {
48 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
49 assert(nr_blocks > 0);
50 }
51
52 if ((nr_blocks * min_bs) < buflen)
53 buflen = nr_blocks * min_bs;
54
55 return buflen;
56}
57
58static uint64_t last_block(struct thread_data *td, struct fio_file *f,
59 enum fio_ddir ddir)
60{
61 uint64_t max_blocks;
62 uint64_t max_size;
63
64 assert(ddir_rw(ddir));
65
66 /*
67 * Hmm, should we make sure that ->io_size <= ->real_file_size?
68 * -> not for now since there is code assuming it could go either.
69 */
70 max_size = f->io_size;
71 if (max_size > f->real_file_size)
72 max_size = f->real_file_size;
73
74 if (td->o.zone_mode == ZONE_MODE_STRIDED && td->o.zone_range)
75 max_size = td->o.zone_range;
76
77 if (td->o.min_bs[ddir] > td->o.ba[ddir])
78 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
79
80 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
81 if (!max_blocks)
82 return 0;
83
84 return max_blocks;
85}
86
87static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
88 enum fio_ddir ddir, uint64_t *b,
89 uint64_t lastb)
90{
91 uint64_t r;
92
93 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
94 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
95
96 r = __rand(&td->random_state);
97
98 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
99
100 *b = lastb * (r / (rand_max(&td->random_state) + 1.0));
101 } else {
102 uint64_t off = 0;
103
104 assert(fio_file_lfsr(f));
105
106 if (lfsr_next(&f->lfsr, &off))
107 return 1;
108
109 *b = off;
110 }
111
112 /*
113 * if we are not maintaining a random map, we are done.
114 */
115 if (!file_randommap(td, f))
116 goto ret;
117
118 /*
119 * calculate map offset and check if it's free
120 */
121 if (random_map_free(f, *b))
122 goto ret;
123
124 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
125 (unsigned long long) *b);
126
127 *b = axmap_next_free(f->io_axmap, *b);
128 if (*b == (uint64_t) -1ULL)
129 return 1;
130ret:
131 return 0;
132}
133
134static int __get_next_rand_offset_zipf(struct thread_data *td,
135 struct fio_file *f, enum fio_ddir ddir,
136 uint64_t *b)
137{
138 *b = zipf_next(&f->zipf);
139 return 0;
140}
141
142static int __get_next_rand_offset_pareto(struct thread_data *td,
143 struct fio_file *f, enum fio_ddir ddir,
144 uint64_t *b)
145{
146 *b = pareto_next(&f->zipf);
147 return 0;
148}
149
150static int __get_next_rand_offset_gauss(struct thread_data *td,
151 struct fio_file *f, enum fio_ddir ddir,
152 uint64_t *b)
153{
154 *b = gauss_next(&f->gauss);
155 return 0;
156}
157
158static int __get_next_rand_offset_zoned_abs(struct thread_data *td,
159 struct fio_file *f,
160 enum fio_ddir ddir, uint64_t *b)
161{
162 struct zone_split_index *zsi;
163 uint64_t lastb, send, stotal;
164 unsigned int v;
165
166 lastb = last_block(td, f, ddir);
167 if (!lastb)
168 return 1;
169
170 if (!td->o.zone_split_nr[ddir]) {
171bail:
172 return __get_next_rand_offset(td, f, ddir, b, lastb);
173 }
174
175 /*
176 * Generate a value, v, between 1 and 100, both inclusive
177 */
178 v = rand_between(&td->zone_state, 1, 100);
179
180 /*
181 * Find our generated table. 'send' is the end block of this zone,
182 * 'stotal' is our start offset.
183 */
184 zsi = &td->zone_state_index[ddir][v - 1];
185 stotal = zsi->size_prev / td->o.ba[ddir];
186 send = zsi->size / td->o.ba[ddir];
187
188 /*
189 * Should never happen
190 */
191 if (send == -1U) {
192 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
193 log_err("fio: bug in zoned generation\n");
194 goto bail;
195 } else if (send > lastb) {
196 /*
197 * This happens if the user specifies ranges that exceed
198 * the file/device size. We can't handle that gracefully,
199 * so error and exit.
200 */
201 log_err("fio: zoned_abs sizes exceed file size\n");
202 return 1;
203 }
204
205 /*
206 * Generate index from 0..send-stotal
207 */
208 if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1)
209 return 1;
210
211 *b += stotal;
212 return 0;
213}
214
215static int __get_next_rand_offset_zoned(struct thread_data *td,
216 struct fio_file *f, enum fio_ddir ddir,
217 uint64_t *b)
218{
219 unsigned int v, send, stotal;
220 uint64_t offset, lastb;
221 struct zone_split_index *zsi;
222
223 lastb = last_block(td, f, ddir);
224 if (!lastb)
225 return 1;
226
227 if (!td->o.zone_split_nr[ddir]) {
228bail:
229 return __get_next_rand_offset(td, f, ddir, b, lastb);
230 }
231
232 /*
233 * Generate a value, v, between 1 and 100, both inclusive
234 */
235 v = rand_between(&td->zone_state, 1, 100);
236
237 zsi = &td->zone_state_index[ddir][v - 1];
238 stotal = zsi->size_perc_prev;
239 send = zsi->size_perc;
240
241 /*
242 * Should never happen
243 */
244 if (send == -1U) {
245 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
246 log_err("fio: bug in zoned generation\n");
247 goto bail;
248 }
249
250 /*
251 * 'send' is some percentage below or equal to 100 that
252 * marks the end of the current IO range. 'stotal' marks
253 * the start, in percent.
254 */
255 if (stotal)
256 offset = stotal * lastb / 100ULL;
257 else
258 offset = 0;
259
260 lastb = lastb * (send - stotal) / 100ULL;
261
262 /*
263 * Generate index from 0..send-of-lastb
264 */
265 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
266 return 1;
267
268 /*
269 * Add our start offset, if any
270 */
271 if (offset)
272 *b += offset;
273
274 return 0;
275}
276
277static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
278 enum fio_ddir ddir, uint64_t *b)
279{
280 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
281 uint64_t lastb;
282
283 lastb = last_block(td, f, ddir);
284 if (!lastb)
285 return 1;
286
287 return __get_next_rand_offset(td, f, ddir, b, lastb);
288 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
289 return __get_next_rand_offset_zipf(td, f, ddir, b);
290 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
291 return __get_next_rand_offset_pareto(td, f, ddir, b);
292 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
293 return __get_next_rand_offset_gauss(td, f, ddir, b);
294 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
295 return __get_next_rand_offset_zoned(td, f, ddir, b);
296 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
297 return __get_next_rand_offset_zoned_abs(td, f, ddir, b);
298
299 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
300 return 1;
301}
302
303static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
304{
305 unsigned int v;
306
307 if (td->o.perc_rand[ddir] == 100)
308 return true;
309
310 v = rand_between(&td->seq_rand_state[ddir], 1, 100);
311
312 return v <= td->o.perc_rand[ddir];
313}
314
315static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f)
316{
317 struct thread_options *o = &td->o;
318
319 if (o->invalidate_cache && !o->odirect) {
320 int fio_unused ret;
321
322 ret = file_invalidate_cache(td, f);
323 }
324}
325
326static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
327 enum fio_ddir ddir, uint64_t *b)
328{
329 if (!get_next_rand_offset(td, f, ddir, b))
330 return 0;
331
332 if (td->o.time_based ||
333 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
334 fio_file_reset(td, f);
335 loop_cache_invalidate(td, f);
336 if (!get_next_rand_offset(td, f, ddir, b))
337 return 0;
338 }
339
340 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
341 f->file_name, (unsigned long long) f->last_pos[ddir],
342 (unsigned long long) f->real_file_size);
343 return 1;
344}
345
346static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
347 enum fio_ddir ddir, uint64_t *offset)
348{
349 struct thread_options *o = &td->o;
350
351 assert(ddir_rw(ddir));
352
353 /*
354 * If we reach the end for a time based run, reset us back to 0
355 * and invalidate the cache, if we need to.
356 */
357 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
358 o->time_based) {
359 f->last_pos[ddir] = f->file_offset;
360 loop_cache_invalidate(td, f);
361 }
362
363 if (f->last_pos[ddir] < f->real_file_size) {
364 uint64_t pos;
365
366 /*
367 * Only rewind if we already hit the end
368 */
369 if (f->last_pos[ddir] == f->file_offset &&
370 f->file_offset && o->ddir_seq_add < 0) {
371 if (f->real_file_size > f->io_size)
372 f->last_pos[ddir] = f->io_size;
373 else
374 f->last_pos[ddir] = f->real_file_size;
375 }
376
377 pos = f->last_pos[ddir] - f->file_offset;
378 if (pos && o->ddir_seq_add) {
379 pos += o->ddir_seq_add;
380
381 /*
382 * If we reach beyond the end of the file
383 * with holed IO, wrap around to the
384 * beginning again. If we're doing backwards IO,
385 * wrap to the end.
386 */
387 if (pos >= f->real_file_size) {
388 if (o->ddir_seq_add > 0)
389 pos = f->file_offset;
390 else {
391 if (f->real_file_size > f->io_size)
392 pos = f->io_size;
393 else
394 pos = f->real_file_size;
395
396 pos += o->ddir_seq_add;
397 }
398 }
399 }
400
401 *offset = pos;
402 return 0;
403 }
404
405 return 1;
406}
407
408static int get_next_block(struct thread_data *td, struct io_u *io_u,
409 enum fio_ddir ddir, int rw_seq,
410 bool *is_random)
411{
412 struct fio_file *f = io_u->file;
413 uint64_t b, offset;
414 int ret;
415
416 assert(ddir_rw(ddir));
417
418 b = offset = -1ULL;
419
420 if (rw_seq) {
421 if (td_random(td)) {
422 if (should_do_random(td, ddir)) {
423 ret = get_next_rand_block(td, f, ddir, &b);
424 *is_random = true;
425 } else {
426 *is_random = false;
427 io_u_set(td, io_u, IO_U_F_BUSY_OK);
428 ret = get_next_seq_offset(td, f, ddir, &offset);
429 if (ret)
430 ret = get_next_rand_block(td, f, ddir, &b);
431 }
432 } else {
433 *is_random = false;
434 ret = get_next_seq_offset(td, f, ddir, &offset);
435 }
436 } else {
437 io_u_set(td, io_u, IO_U_F_BUSY_OK);
438 *is_random = false;
439
440 if (td->o.rw_seq == RW_SEQ_SEQ) {
441 ret = get_next_seq_offset(td, f, ddir, &offset);
442 if (ret) {
443 ret = get_next_rand_block(td, f, ddir, &b);
444 *is_random = false;
445 }
446 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
447 if (f->last_start[ddir] != -1ULL)
448 offset = f->last_start[ddir] - f->file_offset;
449 else
450 offset = 0;
451 ret = 0;
452 } else {
453 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
454 ret = 1;
455 }
456 }
457
458 if (!ret) {
459 if (offset != -1ULL)
460 io_u->offset = offset;
461 else if (b != -1ULL)
462 io_u->offset = b * td->o.ba[ddir];
463 else {
464 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
465 ret = 1;
466 }
467 io_u->verify_offset = io_u->offset;
468 }
469
470 return ret;
471}
472
473/*
474 * For random io, generate a random new block and see if it's used. Repeat
475 * until we find a free one. For sequential io, just return the end of
476 * the last io issued.
477 */
478static int get_next_offset(struct thread_data *td, struct io_u *io_u,
479 bool *is_random)
480{
481 struct fio_file *f = io_u->file;
482 enum fio_ddir ddir = io_u->ddir;
483 int rw_seq_hit = 0;
484
485 assert(ddir_rw(ddir));
486
487 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
488 rw_seq_hit = 1;
489 td->ddir_seq_nr = td->o.ddir_seq_nr;
490 }
491
492 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
493 return 1;
494
495 if (io_u->offset >= f->io_size) {
496 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
497 (unsigned long long) io_u->offset,
498 (unsigned long long) f->io_size);
499 return 1;
500 }
501
502 io_u->offset += f->file_offset;
503 if (io_u->offset >= f->real_file_size) {
504 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
505 (unsigned long long) io_u->offset,
506 (unsigned long long) f->real_file_size);
507 return 1;
508 }
509
510 io_u->verify_offset = io_u->offset;
511 return 0;
512}
513
514static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
515 unsigned long long buflen)
516{
517 struct fio_file *f = io_u->file;
518
519 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
520}
521
522static unsigned long long get_next_buflen(struct thread_data *td, struct io_u *io_u,
523 bool is_random)
524{
525 int ddir = io_u->ddir;
526 unsigned long long buflen = 0;
527 unsigned long long minbs, maxbs;
528 uint64_t frand_max, r;
529 bool power_2;
530
531 assert(ddir_rw(ddir));
532
533 if (td->o.bs_is_seq_rand)
534 ddir = is_random ? DDIR_WRITE : DDIR_READ;
535
536 minbs = td->o.min_bs[ddir];
537 maxbs = td->o.max_bs[ddir];
538
539 if (minbs == maxbs)
540 return minbs;
541
542 /*
543 * If we can't satisfy the min block size from here, then fail
544 */
545 if (!io_u_fits(td, io_u, minbs))
546 return 0;
547
548 frand_max = rand_max(&td->bsrange_state[ddir]);
549 do {
550 r = __rand(&td->bsrange_state[ddir]);
551
552 if (!td->o.bssplit_nr[ddir]) {
553 buflen = minbs + (unsigned long long) ((double) maxbs *
554 (r / (frand_max + 1.0)));
555 } else {
556 long long perc = 0;
557 unsigned int i;
558
559 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
560 struct bssplit *bsp = &td->o.bssplit[ddir][i];
561
562 if (!bsp->perc)
563 continue;
564 buflen = bsp->bs;
565 perc += bsp->perc;
566 if ((r / perc <= frand_max / 100ULL) &&
567 io_u_fits(td, io_u, buflen))
568 break;
569 }
570 }
571
572 power_2 = is_power_of_2(minbs);
573 if (!td->o.bs_unaligned && power_2)
574 buflen &= ~(minbs - 1);
575 else if (!td->o.bs_unaligned && !power_2)
576 buflen -= buflen % minbs;
577 if (buflen > maxbs)
578 buflen = maxbs;
579 } while (!io_u_fits(td, io_u, buflen));
580
581 return buflen;
582}
583
584static void set_rwmix_bytes(struct thread_data *td)
585{
586 unsigned int diff;
587
588 /*
589 * we do time or byte based switch. this is needed because
590 * buffered writes may issue a lot quicker than they complete,
591 * whereas reads do not.
592 */
593 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
594 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
595}
596
597static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
598{
599 unsigned int v;
600
601 v = rand_between(&td->rwmix_state, 1, 100);
602
603 if (v <= td->o.rwmix[DDIR_READ])
604 return DDIR_READ;
605
606 return DDIR_WRITE;
607}
608
609int io_u_quiesce(struct thread_data *td)
610{
611 int ret = 0, completed = 0, err = 0;
612
613 /*
614 * We are going to sleep, ensure that we flush anything pending as
615 * not to skew our latency numbers.
616 *
617 * Changed to only monitor 'in flight' requests here instead of the
618 * td->cur_depth, b/c td->cur_depth does not accurately represent
619 * io's that have been actually submitted to an async engine,
620 * and cur_depth is meaningless for sync engines.
621 */
622 if (td->io_u_queued || td->cur_depth)
623 td_io_commit(td);
624
625 while (td->io_u_in_flight) {
626 ret = io_u_queued_complete(td, 1);
627 if (ret > 0)
628 completed += ret;
629 else if (ret < 0)
630 err = ret;
631 }
632
633 if (td->flags & TD_F_REGROW_LOGS)
634 regrow_logs(td);
635
636 if (completed)
637 return completed;
638
639 return err;
640}
641
642static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
643{
644 enum fio_ddir odir = ddir ^ 1;
645 uint64_t usec;
646 uint64_t now;
647
648 assert(ddir_rw(ddir));
649 now = utime_since_now(&td->epoch);
650
651 /*
652 * if rate_next_io_time is in the past, need to catch up to rate
653 */
654 if (td->rate_next_io_time[ddir] <= now)
655 return ddir;
656
657 /*
658 * We are ahead of rate in this direction. See if we
659 * should switch.
660 */
661 if (td_rw(td) && td->o.rwmix[odir]) {
662 /*
663 * Other direction is behind rate, switch
664 */
665 if (td->rate_next_io_time[odir] <= now)
666 return odir;
667
668 /*
669 * Both directions are ahead of rate. sleep the min,
670 * switch if necessary
671 */
672 if (td->rate_next_io_time[ddir] <=
673 td->rate_next_io_time[odir]) {
674 usec = td->rate_next_io_time[ddir] - now;
675 } else {
676 usec = td->rate_next_io_time[odir] - now;
677 ddir = odir;
678 }
679 } else
680 usec = td->rate_next_io_time[ddir] - now;
681
682 if (td->o.io_submit_mode == IO_MODE_INLINE)
683 io_u_quiesce(td);
684
685 if (td->o.timeout && ((usec + now) > td->o.timeout)) {
686 /*
687 * check if the usec is capable of taking negative values
688 */
689 if (now > td->o.timeout) {
690 ddir = DDIR_INVAL;
691 return ddir;
692 }
693 usec = td->o.timeout - now;
694 }
695 usec_sleep(td, usec);
696
697 now = utime_since_now(&td->epoch);
698 if ((td->o.timeout && (now > td->o.timeout)) || td->terminate)
699 ddir = DDIR_INVAL;
700
701 return ddir;
702}
703
704/*
705 * Return the data direction for the next io_u. If the job is a
706 * mixed read/write workload, check the rwmix cycle and switch if
707 * necessary.
708 */
709static enum fio_ddir get_rw_ddir(struct thread_data *td)
710{
711 enum fio_ddir ddir;
712
713 /*
714 * See if it's time to fsync/fdatasync/sync_file_range first,
715 * and if not then move on to check regular I/Os.
716 */
717 if (should_fsync(td)) {
718 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
719 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
720 return DDIR_SYNC;
721
722 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
723 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
724 return DDIR_DATASYNC;
725
726 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
727 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
728 return DDIR_SYNC_FILE_RANGE;
729 }
730
731 if (td_rw(td)) {
732 /*
733 * Check if it's time to seed a new data direction.
734 */
735 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
736 /*
737 * Put a top limit on how many bytes we do for
738 * one data direction, to avoid overflowing the
739 * ranges too much
740 */
741 ddir = get_rand_ddir(td);
742
743 if (ddir != td->rwmix_ddir)
744 set_rwmix_bytes(td);
745
746 td->rwmix_ddir = ddir;
747 }
748 ddir = td->rwmix_ddir;
749 } else if (td_read(td))
750 ddir = DDIR_READ;
751 else if (td_write(td))
752 ddir = DDIR_WRITE;
753 else if (td_trim(td))
754 ddir = DDIR_TRIM;
755 else
756 ddir = DDIR_INVAL;
757
758 td->rwmix_ddir = rate_ddir(td, ddir);
759 return td->rwmix_ddir;
760}
761
762static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
763{
764 enum fio_ddir ddir = get_rw_ddir(td);
765
766 if (td->o.zone_mode == ZONE_MODE_ZBD)
767 ddir = zbd_adjust_ddir(td, io_u, ddir);
768
769 if (td_trimwrite(td)) {
770 struct fio_file *f = io_u->file;
771 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
772 ddir = DDIR_TRIM;
773 else
774 ddir = DDIR_WRITE;
775 }
776
777 io_u->ddir = io_u->acct_ddir = ddir;
778
779 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
780 td->o.barrier_blocks &&
781 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
782 td->io_issues[DDIR_WRITE])
783 io_u_set(td, io_u, IO_U_F_BARRIER);
784}
785
786void put_file_log(struct thread_data *td, struct fio_file *f)
787{
788 unsigned int ret = put_file(td, f);
789
790 if (ret)
791 td_verror(td, ret, "file close");
792}
793
794void put_io_u(struct thread_data *td, struct io_u *io_u)
795{
796 const bool needs_lock = td_async_processing(td);
797
798 zbd_put_io_u(io_u);
799
800 if (td->parent)
801 td = td->parent;
802
803 if (needs_lock)
804 __td_io_u_lock(td);
805
806 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
807 put_file_log(td, io_u->file);
808
809 io_u->file = NULL;
810 io_u_set(td, io_u, IO_U_F_FREE);
811
812 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
813 td->cur_depth--;
814 assert(!(td->flags & TD_F_CHILD));
815 }
816 io_u_qpush(&td->io_u_freelist, io_u);
817 td_io_u_free_notify(td);
818
819 if (needs_lock)
820 __td_io_u_unlock(td);
821}
822
823void clear_io_u(struct thread_data *td, struct io_u *io_u)
824{
825 io_u_clear(td, io_u, IO_U_F_FLIGHT);
826 put_io_u(td, io_u);
827}
828
829void requeue_io_u(struct thread_data *td, struct io_u **io_u)
830{
831 const bool needs_lock = td_async_processing(td);
832 struct io_u *__io_u = *io_u;
833 enum fio_ddir ddir = acct_ddir(__io_u);
834
835 dprint(FD_IO, "requeue %p\n", __io_u);
836
837 if (td->parent)
838 td = td->parent;
839
840 if (needs_lock)
841 __td_io_u_lock(td);
842
843 io_u_set(td, __io_u, IO_U_F_FREE);
844 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
845 td->io_issues[ddir]--;
846
847 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
848 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
849 td->cur_depth--;
850 assert(!(td->flags & TD_F_CHILD));
851 }
852
853 io_u_rpush(&td->io_u_requeues, __io_u);
854 td_io_u_free_notify(td);
855
856 if (needs_lock)
857 __td_io_u_unlock(td);
858
859 *io_u = NULL;
860}
861
862static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u)
863{
864 struct fio_file *f = io_u->file;
865
866 assert(td->o.zone_mode == ZONE_MODE_STRIDED);
867 assert(td->o.zone_size);
868 assert(td->o.zone_range);
869
870 /*
871 * See if it's time to switch to a new zone
872 */
873 if (td->zone_bytes >= td->o.zone_size) {
874 td->zone_bytes = 0;
875 f->file_offset += td->o.zone_range + td->o.zone_skip;
876
877 /*
878 * Wrap from the beginning, if we exceed the file size
879 */
880 if (f->file_offset >= f->real_file_size)
881 f->file_offset = get_start_offset(td, f);
882
883 f->last_pos[io_u->ddir] = f->file_offset;
884 td->io_skip_bytes += td->o.zone_skip;
885 }
886
887 /*
888 * If zone_size > zone_range, then maintain the same zone until
889 * zone_bytes >= zone_size.
890 */
891 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
892 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
893 f->file_offset, f->last_pos[io_u->ddir]);
894 f->last_pos[io_u->ddir] = f->file_offset;
895 }
896
897 /*
898 * For random: if 'norandommap' is not set and zone_size > zone_range,
899 * map needs to be reset as it's done with zone_range everytime.
900 */
901 if ((td->zone_bytes % td->o.zone_range) == 0)
902 fio_file_reset(td, f);
903}
904
905static int fill_io_u(struct thread_data *td, struct io_u *io_u)
906{
907 bool is_random;
908 uint64_t offset;
909 enum io_u_action ret;
910
911 if (td_ioengine_flagged(td, FIO_NOIO))
912 goto out;
913
914 set_rw_ddir(td, io_u);
915
916 if (io_u->ddir == DDIR_INVAL) {
917 dprint(FD_IO, "invalid direction received ddir = %d", io_u->ddir);
918 return 1;
919 }
920 /*
921 * fsync() or fdatasync() or trim etc, we are done
922 */
923 if (!ddir_rw(io_u->ddir))
924 goto out;
925
926 if (td->o.zone_mode == ZONE_MODE_STRIDED)
927 setup_strided_zone_mode(td, io_u);
928 else if (td->o.zone_mode == ZONE_MODE_ZBD)
929 setup_zbd_zone_mode(td, io_u);
930
931 /*
932 * No log, let the seq/rand engine retrieve the next buflen and
933 * position.
934 */
935 if (get_next_offset(td, io_u, &is_random)) {
936 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
937 return 1;
938 }
939
940 io_u->buflen = get_next_buflen(td, io_u, is_random);
941 if (!io_u->buflen) {
942 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
943 return 1;
944 }
945
946 offset = io_u->offset;
947 if (td->o.zone_mode == ZONE_MODE_ZBD) {
948 ret = zbd_adjust_block(td, io_u);
949 if (ret == io_u_eof)
950 return 1;
951 }
952
953 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
954 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n",
955 io_u,
956 (unsigned long long) io_u->offset, io_u->buflen,
957 (unsigned long long) io_u->file->real_file_size);
958 return 1;
959 }
960
961 /*
962 * mark entry before potentially trimming io_u
963 */
964 if (td_random(td) && file_randommap(td, io_u->file))
965 io_u->buflen = mark_random_map(td, io_u, offset, io_u->buflen);
966
967out:
968 dprint_io_u(io_u, "fill");
969 io_u->verify_offset = io_u->offset;
970 td->zone_bytes += io_u->buflen;
971 return 0;
972}
973
974static void __io_u_mark_map(uint64_t *map, unsigned int nr)
975{
976 int idx = 0;
977
978 switch (nr) {
979 default:
980 idx = 6;
981 break;
982 case 33 ... 64:
983 idx = 5;
984 break;
985 case 17 ... 32:
986 idx = 4;
987 break;
988 case 9 ... 16:
989 idx = 3;
990 break;
991 case 5 ... 8:
992 idx = 2;
993 break;
994 case 1 ... 4:
995 idx = 1;
996 case 0:
997 break;
998 }
999
1000 map[idx]++;
1001}
1002
1003void io_u_mark_submit(struct thread_data *td, unsigned int nr)
1004{
1005 __io_u_mark_map(td->ts.io_u_submit, nr);
1006 td->ts.total_submit++;
1007}
1008
1009void io_u_mark_complete(struct thread_data *td, unsigned int nr)
1010{
1011 __io_u_mark_map(td->ts.io_u_complete, nr);
1012 td->ts.total_complete++;
1013}
1014
1015void io_u_mark_depth(struct thread_data *td, unsigned int nr)
1016{
1017 int idx = 0;
1018
1019 switch (td->cur_depth) {
1020 default:
1021 idx = 6;
1022 break;
1023 case 32 ... 63:
1024 idx = 5;
1025 break;
1026 case 16 ... 31:
1027 idx = 4;
1028 break;
1029 case 8 ... 15:
1030 idx = 3;
1031 break;
1032 case 4 ... 7:
1033 idx = 2;
1034 break;
1035 case 2 ... 3:
1036 idx = 1;
1037 case 1:
1038 break;
1039 }
1040
1041 td->ts.io_u_map[idx] += nr;
1042}
1043
1044static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1045{
1046 int idx = 0;
1047
1048 assert(nsec < 1000);
1049
1050 switch (nsec) {
1051 case 750 ... 999:
1052 idx = 9;
1053 break;
1054 case 500 ... 749:
1055 idx = 8;
1056 break;
1057 case 250 ... 499:
1058 idx = 7;
1059 break;
1060 case 100 ... 249:
1061 idx = 6;
1062 break;
1063 case 50 ... 99:
1064 idx = 5;
1065 break;
1066 case 20 ... 49:
1067 idx = 4;
1068 break;
1069 case 10 ... 19:
1070 idx = 3;
1071 break;
1072 case 4 ... 9:
1073 idx = 2;
1074 break;
1075 case 2 ... 3:
1076 idx = 1;
1077 case 0 ... 1:
1078 break;
1079 }
1080
1081 assert(idx < FIO_IO_U_LAT_N_NR);
1082 td->ts.io_u_lat_n[idx]++;
1083}
1084
1085static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1086{
1087 int idx = 0;
1088
1089 assert(usec < 1000 && usec >= 1);
1090
1091 switch (usec) {
1092 case 750 ... 999:
1093 idx = 9;
1094 break;
1095 case 500 ... 749:
1096 idx = 8;
1097 break;
1098 case 250 ... 499:
1099 idx = 7;
1100 break;
1101 case 100 ... 249:
1102 idx = 6;
1103 break;
1104 case 50 ... 99:
1105 idx = 5;
1106 break;
1107 case 20 ... 49:
1108 idx = 4;
1109 break;
1110 case 10 ... 19:
1111 idx = 3;
1112 break;
1113 case 4 ... 9:
1114 idx = 2;
1115 break;
1116 case 2 ... 3:
1117 idx = 1;
1118 case 0 ... 1:
1119 break;
1120 }
1121
1122 assert(idx < FIO_IO_U_LAT_U_NR);
1123 td->ts.io_u_lat_u[idx]++;
1124}
1125
1126static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1127{
1128 int idx = 0;
1129
1130 assert(msec >= 1);
1131
1132 switch (msec) {
1133 default:
1134 idx = 11;
1135 break;
1136 case 1000 ... 1999:
1137 idx = 10;
1138 break;
1139 case 750 ... 999:
1140 idx = 9;
1141 break;
1142 case 500 ... 749:
1143 idx = 8;
1144 break;
1145 case 250 ... 499:
1146 idx = 7;
1147 break;
1148 case 100 ... 249:
1149 idx = 6;
1150 break;
1151 case 50 ... 99:
1152 idx = 5;
1153 break;
1154 case 20 ... 49:
1155 idx = 4;
1156 break;
1157 case 10 ... 19:
1158 idx = 3;
1159 break;
1160 case 4 ... 9:
1161 idx = 2;
1162 break;
1163 case 2 ... 3:
1164 idx = 1;
1165 case 0 ... 1:
1166 break;
1167 }
1168
1169 assert(idx < FIO_IO_U_LAT_M_NR);
1170 td->ts.io_u_lat_m[idx]++;
1171}
1172
1173static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1174{
1175 if (nsec < 1000)
1176 io_u_mark_lat_nsec(td, nsec);
1177 else if (nsec < 1000000)
1178 io_u_mark_lat_usec(td, nsec / 1000);
1179 else
1180 io_u_mark_lat_msec(td, nsec / 1000000);
1181}
1182
1183static unsigned int __get_next_fileno_rand(struct thread_data *td)
1184{
1185 unsigned long fileno;
1186
1187 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1188 uint64_t frand_max = rand_max(&td->next_file_state);
1189 unsigned long r;
1190
1191 r = __rand(&td->next_file_state);
1192 return (unsigned int) ((double) td->o.nr_files
1193 * (r / (frand_max + 1.0)));
1194 }
1195
1196 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1197 fileno = zipf_next(&td->next_file_zipf);
1198 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1199 fileno = pareto_next(&td->next_file_zipf);
1200 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1201 fileno = gauss_next(&td->next_file_gauss);
1202 else {
1203 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1204 assert(0);
1205 return 0;
1206 }
1207
1208 return fileno >> FIO_FSERVICE_SHIFT;
1209}
1210
1211/*
1212 * Get next file to service by choosing one at random
1213 */
1214static struct fio_file *get_next_file_rand(struct thread_data *td,
1215 enum fio_file_flags goodf,
1216 enum fio_file_flags badf)
1217{
1218 struct fio_file *f;
1219 int fno;
1220
1221 do {
1222 int opened = 0;
1223
1224 fno = __get_next_fileno_rand(td);
1225
1226 f = td->files[fno];
1227 if (fio_file_done(f))
1228 continue;
1229
1230 if (!fio_file_open(f)) {
1231 int err;
1232
1233 if (td->nr_open_files >= td->o.open_files)
1234 return ERR_PTR(-EBUSY);
1235
1236 err = td_io_open_file(td, f);
1237 if (err)
1238 continue;
1239 opened = 1;
1240 }
1241
1242 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1243 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1244 return f;
1245 }
1246 if (opened)
1247 td_io_close_file(td, f);
1248 } while (1);
1249}
1250
1251/*
1252 * Get next file to service by doing round robin between all available ones
1253 */
1254static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1255 int badf)
1256{
1257 unsigned int old_next_file = td->next_file;
1258 struct fio_file *f;
1259
1260 do {
1261 int opened = 0;
1262
1263 f = td->files[td->next_file];
1264
1265 td->next_file++;
1266 if (td->next_file >= td->o.nr_files)
1267 td->next_file = 0;
1268
1269 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1270 if (fio_file_done(f)) {
1271 f = NULL;
1272 continue;
1273 }
1274
1275 if (!fio_file_open(f)) {
1276 int err;
1277
1278 if (td->nr_open_files >= td->o.open_files)
1279 return ERR_PTR(-EBUSY);
1280
1281 err = td_io_open_file(td, f);
1282 if (err) {
1283 dprint(FD_FILE, "error %d on open of %s\n",
1284 err, f->file_name);
1285 f = NULL;
1286 continue;
1287 }
1288 opened = 1;
1289 }
1290
1291 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1292 f->flags);
1293 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1294 break;
1295
1296 if (opened)
1297 td_io_close_file(td, f);
1298
1299 f = NULL;
1300 } while (td->next_file != old_next_file);
1301
1302 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1303 return f;
1304}
1305
1306static struct fio_file *__get_next_file(struct thread_data *td)
1307{
1308 struct fio_file *f;
1309
1310 assert(td->o.nr_files <= td->files_index);
1311
1312 if (td->nr_done_files >= td->o.nr_files) {
1313 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1314 " nr_files=%d\n", td->nr_open_files,
1315 td->nr_done_files,
1316 td->o.nr_files);
1317 return NULL;
1318 }
1319
1320 f = td->file_service_file;
1321 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1322 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1323 goto out;
1324 if (td->file_service_left--)
1325 goto out;
1326 }
1327
1328 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1329 td->o.file_service_type == FIO_FSERVICE_SEQ)
1330 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1331 else
1332 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1333
1334 if (IS_ERR(f))
1335 return f;
1336
1337 td->file_service_file = f;
1338 td->file_service_left = td->file_service_nr - 1;
1339out:
1340 if (f)
1341 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1342 else
1343 dprint(FD_FILE, "get_next_file: NULL\n");
1344 return f;
1345}
1346
1347static struct fio_file *get_next_file(struct thread_data *td)
1348{
1349 return __get_next_file(td);
1350}
1351
1352static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1353{
1354 struct fio_file *f;
1355
1356 do {
1357 f = get_next_file(td);
1358 if (IS_ERR_OR_NULL(f))
1359 return PTR_ERR(f);
1360
1361 io_u->file = f;
1362 get_file(f);
1363
1364 if (!fill_io_u(td, io_u))
1365 break;
1366
1367 zbd_put_io_u(io_u);
1368
1369 put_file_log(td, f);
1370 td_io_close_file(td, f);
1371 io_u->file = NULL;
1372 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1373 fio_file_reset(td, f);
1374 else {
1375 fio_file_set_done(f);
1376 td->nr_done_files++;
1377 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1378 td->nr_done_files, td->o.nr_files);
1379 }
1380 } while (1);
1381
1382 return 0;
1383}
1384
1385static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1386 unsigned long long tnsec, unsigned long long max_nsec)
1387{
1388 if (!td->error)
1389 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
1390 td_verror(td, ETIMEDOUT, "max latency exceeded");
1391 icd->error = ETIMEDOUT;
1392}
1393
1394static void lat_new_cycle(struct thread_data *td)
1395{
1396 fio_gettime(&td->latency_ts, NULL);
1397 td->latency_ios = ddir_rw_sum(td->io_blocks);
1398 td->latency_failed = 0;
1399}
1400
1401/*
1402 * We had an IO outside the latency target. Reduce the queue depth. If we
1403 * are at QD=1, then it's time to give up.
1404 */
1405static bool __lat_target_failed(struct thread_data *td)
1406{
1407 if (td->latency_qd == 1)
1408 return true;
1409
1410 td->latency_qd_high = td->latency_qd;
1411
1412 if (td->latency_qd == td->latency_qd_low)
1413 td->latency_qd_low--;
1414
1415 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1416 td->latency_stable_count = 0;
1417
1418 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1419
1420 /*
1421 * When we ramp QD down, quiesce existing IO to prevent
1422 * a storm of ramp downs due to pending higher depth.
1423 */
1424 io_u_quiesce(td);
1425 lat_new_cycle(td);
1426 return false;
1427}
1428
1429static bool lat_target_failed(struct thread_data *td)
1430{
1431 if (td->o.latency_percentile.u.f == 100.0)
1432 return __lat_target_failed(td);
1433
1434 td->latency_failed++;
1435 return false;
1436}
1437
1438void lat_target_init(struct thread_data *td)
1439{
1440 td->latency_end_run = 0;
1441
1442 if (td->o.latency_target) {
1443 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1444 fio_gettime(&td->latency_ts, NULL);
1445 td->latency_qd = 1;
1446 td->latency_qd_high = td->o.iodepth;
1447 td->latency_qd_low = 1;
1448 td->latency_ios = ddir_rw_sum(td->io_blocks);
1449 } else
1450 td->latency_qd = td->o.iodepth;
1451}
1452
1453void lat_target_reset(struct thread_data *td)
1454{
1455 if (!td->latency_end_run)
1456 lat_target_init(td);
1457}
1458
1459static void lat_target_success(struct thread_data *td)
1460{
1461 const unsigned int qd = td->latency_qd;
1462 struct thread_options *o = &td->o;
1463
1464 td->latency_qd_low = td->latency_qd;
1465
1466 if (td->latency_qd + 1 == td->latency_qd_high) {
1467 /*
1468 * latency_qd will not incease on lat_target_success(), so
1469 * called stable. If we stick with this queue depth, the
1470 * final latency is likely lower than latency_target. Fix
1471 * this by increasing latency_qd_high slowly. Use a naive
1472 * heuristic here. If we get lat_target_success() 3 times
1473 * in a row, increase latency_qd_high by 1.
1474 */
1475 if (++td->latency_stable_count >= 3) {
1476 td->latency_qd_high++;
1477 td->latency_stable_count = 0;
1478 }
1479 }
1480
1481 /*
1482 * If we haven't failed yet, we double up to a failing value instead
1483 * of bisecting from highest possible queue depth. If we have set
1484 * a limit other than td->o.iodepth, bisect between that.
1485 */
1486 if (td->latency_qd_high != o->iodepth)
1487 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1488 else
1489 td->latency_qd *= 2;
1490
1491 if (td->latency_qd > o->iodepth)
1492 td->latency_qd = o->iodepth;
1493
1494 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1495
1496 /*
1497 * Same as last one, we are done. Let it run a latency cycle, so
1498 * we get only the results from the targeted depth.
1499 */
1500 if (!o->latency_run && td->latency_qd == qd) {
1501 if (td->latency_end_run) {
1502 dprint(FD_RATE, "We are done\n");
1503 td->done = 1;
1504 } else {
1505 dprint(FD_RATE, "Quiesce and final run\n");
1506 io_u_quiesce(td);
1507 td->latency_end_run = 1;
1508 reset_all_stats(td);
1509 reset_io_stats(td);
1510 }
1511 }
1512
1513 lat_new_cycle(td);
1514}
1515
1516/*
1517 * Check if we can bump the queue depth
1518 */
1519void lat_target_check(struct thread_data *td)
1520{
1521 uint64_t usec_window;
1522 uint64_t ios;
1523 double success_ios;
1524
1525 usec_window = utime_since_now(&td->latency_ts);
1526 if (usec_window < td->o.latency_window)
1527 return;
1528
1529 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1530 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1531 success_ios *= 100.0;
1532
1533 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1534
1535 if (success_ios >= td->o.latency_percentile.u.f)
1536 lat_target_success(td);
1537 else
1538 __lat_target_failed(td);
1539}
1540
1541/*
1542 * If latency target is enabled, we might be ramping up or down and not
1543 * using the full queue depth available.
1544 */
1545bool queue_full(const struct thread_data *td)
1546{
1547 const int qempty = io_u_qempty(&td->io_u_freelist);
1548
1549 if (qempty)
1550 return true;
1551 if (!td->o.latency_target)
1552 return false;
1553
1554 return td->cur_depth >= td->latency_qd;
1555}
1556
1557struct io_u *__get_io_u(struct thread_data *td)
1558{
1559 const bool needs_lock = td_async_processing(td);
1560 struct io_u *io_u = NULL;
1561 int ret;
1562
1563 if (td->stop_io)
1564 return NULL;
1565
1566 if (needs_lock)
1567 __td_io_u_lock(td);
1568
1569again:
1570 if (!io_u_rempty(&td->io_u_requeues)) {
1571 io_u = io_u_rpop(&td->io_u_requeues);
1572 io_u->resid = 0;
1573 } else if (!queue_full(td)) {
1574 io_u = io_u_qpop(&td->io_u_freelist);
1575
1576 io_u->file = NULL;
1577 io_u->buflen = 0;
1578 io_u->resid = 0;
1579 io_u->end_io = NULL;
1580 }
1581
1582 if (io_u) {
1583 assert(io_u->flags & IO_U_F_FREE);
1584 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1585 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1586 IO_U_F_VER_LIST | IO_U_F_PRIORITY);
1587
1588 io_u->error = 0;
1589 io_u->acct_ddir = -1;
1590 td->cur_depth++;
1591 assert(!(td->flags & TD_F_CHILD));
1592 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1593 io_u->ipo = NULL;
1594 } else if (td_async_processing(td)) {
1595 /*
1596 * We ran out, wait for async verify threads to finish and
1597 * return one
1598 */
1599 assert(!(td->flags & TD_F_CHILD));
1600 ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1601 assert(ret == 0);
1602 if (!td->error)
1603 goto again;
1604 }
1605
1606 if (needs_lock)
1607 __td_io_u_unlock(td);
1608
1609 return io_u;
1610}
1611
1612static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1613{
1614 if (!(td->flags & TD_F_TRIM_BACKLOG))
1615 return false;
1616 if (!td->trim_entries)
1617 return false;
1618
1619 if (td->trim_batch) {
1620 td->trim_batch--;
1621 if (get_next_trim(td, io_u))
1622 return true;
1623 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1624 td->last_ddir != DDIR_READ) {
1625 td->trim_batch = td->o.trim_batch;
1626 if (!td->trim_batch)
1627 td->trim_batch = td->o.trim_backlog;
1628 if (get_next_trim(td, io_u))
1629 return true;
1630 }
1631
1632 return false;
1633}
1634
1635static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1636{
1637 if (!(td->flags & TD_F_VER_BACKLOG))
1638 return false;
1639
1640 if (td->io_hist_len) {
1641 int get_verify = 0;
1642
1643 if (td->verify_batch)
1644 get_verify = 1;
1645 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1646 td->last_ddir != DDIR_READ) {
1647 td->verify_batch = td->o.verify_batch;
1648 if (!td->verify_batch)
1649 td->verify_batch = td->o.verify_backlog;
1650 get_verify = 1;
1651 }
1652
1653 if (get_verify && !get_next_verify(td, io_u)) {
1654 td->verify_batch--;
1655 return true;
1656 }
1657 }
1658
1659 return false;
1660}
1661
1662/*
1663 * Fill offset and start time into the buffer content, to prevent too
1664 * easy compressible data for simple de-dupe attempts. Do this for every
1665 * 512b block in the range, since that should be the smallest block size
1666 * we can expect from a device.
1667 */
1668static void small_content_scramble(struct io_u *io_u)
1669{
1670 unsigned long long i, nr_blocks = io_u->buflen >> 9;
1671 unsigned int offset;
1672 uint64_t boffset, *iptr;
1673 char *p;
1674
1675 if (!nr_blocks)
1676 return;
1677
1678 p = io_u->xfer_buf;
1679 boffset = io_u->offset;
1680
1681 if (io_u->buf_filled_len)
1682 io_u->buf_filled_len = 0;
1683
1684 /*
1685 * Generate random index between 0..7. We do chunks of 512b, if
1686 * we assume a cacheline is 64 bytes, then we have 8 of those.
1687 * Scramble content within the blocks in the same cacheline to
1688 * speed things up.
1689 */
1690 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1691
1692 for (i = 0; i < nr_blocks; i++) {
1693 /*
1694 * Fill offset into start of cacheline, time into end
1695 * of cacheline
1696 */
1697 iptr = (void *) p + (offset << 6);
1698 *iptr = boffset;
1699
1700 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1701 iptr[0] = io_u->start_time.tv_sec;
1702 iptr[1] = io_u->start_time.tv_nsec;
1703
1704 p += 512;
1705 boffset += 512;
1706 }
1707}
1708
1709/*
1710 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1711 * etc. The returned io_u is fully ready to be prepped, populated and submitted.
1712 */
1713struct io_u *get_io_u(struct thread_data *td)
1714{
1715 struct fio_file *f;
1716 struct io_u *io_u;
1717 int do_scramble = 0;
1718 long ret = 0;
1719
1720 io_u = __get_io_u(td);
1721 if (!io_u) {
1722 dprint(FD_IO, "__get_io_u failed\n");
1723 return NULL;
1724 }
1725
1726 if (check_get_verify(td, io_u))
1727 goto out;
1728 if (check_get_trim(td, io_u))
1729 goto out;
1730
1731 /*
1732 * from a requeue, io_u already setup
1733 */
1734 if (io_u->file)
1735 goto out;
1736
1737 /*
1738 * If using an iolog, grab next piece if any available.
1739 */
1740 if (td->flags & TD_F_READ_IOLOG) {
1741 if (read_iolog_get(td, io_u))
1742 goto err_put;
1743 } else if (set_io_u_file(td, io_u)) {
1744 ret = -EBUSY;
1745 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1746 goto err_put;
1747 }
1748
1749 f = io_u->file;
1750 if (!f) {
1751 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1752 goto err_put;
1753 }
1754
1755 assert(fio_file_open(f));
1756
1757 if (ddir_rw(io_u->ddir)) {
1758 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1759 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1760 goto err_put;
1761 }
1762
1763 f->last_start[io_u->ddir] = io_u->offset;
1764 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1765
1766 if (io_u->ddir == DDIR_WRITE) {
1767 if (td->flags & TD_F_REFILL_BUFFERS) {
1768 io_u_fill_buffer(td, io_u,
1769 td->o.min_bs[DDIR_WRITE],
1770 io_u->buflen);
1771 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1772 !(td->flags & TD_F_COMPRESS) &&
1773 !(td->flags & TD_F_DO_VERIFY))
1774 do_scramble = 1;
1775 } else if (io_u->ddir == DDIR_READ) {
1776 /*
1777 * Reset the buf_filled parameters so next time if the
1778 * buffer is used for writes it is refilled.
1779 */
1780 io_u->buf_filled_len = 0;
1781 }
1782 }
1783
1784 /*
1785 * Set io data pointers.
1786 */
1787 io_u->xfer_buf = io_u->buf;
1788 io_u->xfer_buflen = io_u->buflen;
1789
1790out:
1791 assert(io_u->file);
1792 if (!td_io_prep(td, io_u)) {
1793 if (!td->o.disable_lat)
1794 fio_gettime(&io_u->start_time, NULL);
1795
1796 if (do_scramble)
1797 small_content_scramble(io_u);
1798
1799 return io_u;
1800 }
1801err_put:
1802 dprint(FD_IO, "get_io_u failed\n");
1803 put_io_u(td, io_u);
1804 return ERR_PTR(ret);
1805}
1806
1807static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1808{
1809 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1810
1811 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1812 return;
1813
1814 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n",
1815 io_u->file ? " on file " : "",
1816 io_u->file ? io_u->file->file_name : "",
1817 strerror(io_u->error),
1818 io_ddir_name(io_u->ddir),
1819 io_u->offset, io_u->xfer_buflen);
1820
1821 if (td->io_ops->errdetails) {
1822 char *err = td->io_ops->errdetails(io_u);
1823
1824 log_err("fio: %s\n", err);
1825 free(err);
1826 }
1827
1828 if (!td->error)
1829 td_verror(td, io_u->error, "io_u error");
1830}
1831
1832void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1833{
1834 __io_u_log_error(td, io_u);
1835 if (td->parent)
1836 __io_u_log_error(td->parent, io_u);
1837}
1838
1839static inline bool gtod_reduce(struct thread_data *td)
1840{
1841 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1842 || td->o.gtod_reduce;
1843}
1844
1845static void trim_block_info(struct thread_data *td, struct io_u *io_u)
1846{
1847 uint32_t *info = io_u_block_info(td, io_u);
1848
1849 if (BLOCK_INFO_STATE(*info) >= BLOCK_STATE_TRIM_FAILURE)
1850 return;
1851
1852 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, BLOCK_INFO_TRIMS(*info) + 1);
1853}
1854
1855static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1856 struct io_completion_data *icd,
1857 const enum fio_ddir idx, unsigned int bytes)
1858{
1859 const int no_reduce = !gtod_reduce(td);
1860 unsigned long long llnsec = 0;
1861
1862 if (td->parent)
1863 td = td->parent;
1864
1865 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1866 return;
1867
1868 if (no_reduce)
1869 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1870
1871 if (!td->o.disable_lat) {
1872 unsigned long long tnsec;
1873
1874 tnsec = ntime_since(&io_u->start_time, &icd->time);
1875 add_lat_sample(td, idx, tnsec, bytes, io_u->offset, io_u_is_prio(io_u));
1876
1877 if (td->flags & TD_F_PROFILE_OPS) {
1878 struct prof_io_ops *ops = &td->prof_io_ops;
1879
1880 if (ops->io_u_lat)
1881 icd->error = ops->io_u_lat(td, tnsec);
1882 }
1883
1884 if (td->o.max_latency && tnsec > td->o.max_latency)
1885 lat_fatal(td, icd, tnsec, td->o.max_latency);
1886 if (td->o.latency_target && tnsec > td->o.latency_target) {
1887 if (lat_target_failed(td))
1888 lat_fatal(td, icd, tnsec, td->o.latency_target);
1889 }
1890 }
1891
1892 if (ddir_rw(idx)) {
1893 if (!td->o.disable_clat) {
1894 add_clat_sample(td, idx, llnsec, bytes, io_u->offset, io_u_is_prio(io_u));
1895 io_u_mark_latency(td, llnsec);
1896 }
1897
1898 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1899 add_bw_sample(td, io_u, bytes, llnsec);
1900
1901 if (no_reduce && per_unit_log(td->iops_log))
1902 add_iops_sample(td, io_u, bytes);
1903 } else if (ddir_sync(idx) && !td->o.disable_clat)
1904 add_sync_clat_sample(&td->ts, llnsec);
1905
1906 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM)
1907 trim_block_info(td, io_u);
1908}
1909
1910static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1911 uint64_t offset, unsigned int bytes)
1912{
1913 int idx;
1914
1915 if (!f)
1916 return;
1917
1918 if (f->first_write == -1ULL || offset < f->first_write)
1919 f->first_write = offset;
1920 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1921 f->last_write = offset + bytes;
1922
1923 if (!f->last_write_comp)
1924 return;
1925
1926 idx = f->last_write_idx++;
1927 f->last_write_comp[idx] = offset;
1928 if (f->last_write_idx == td->o.iodepth)
1929 f->last_write_idx = 0;
1930}
1931
1932static bool should_account(struct thread_data *td)
1933{
1934 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1935 td->runstate == TD_VERIFYING);
1936}
1937
1938static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1939 struct io_completion_data *icd)
1940{
1941 struct io_u *io_u = *io_u_ptr;
1942 enum fio_ddir ddir = io_u->ddir;
1943 struct fio_file *f = io_u->file;
1944
1945 dprint_io_u(io_u, "complete");
1946
1947 assert(io_u->flags & IO_U_F_FLIGHT);
1948 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1949
1950 /*
1951 * Mark IO ok to verify
1952 */
1953 if (io_u->ipo) {
1954 /*
1955 * Remove errored entry from the verification list
1956 */
1957 if (io_u->error)
1958 unlog_io_piece(td, io_u);
1959 else {
1960 atomic_store_release(&io_u->ipo->flags,
1961 io_u->ipo->flags & ~IP_F_IN_FLIGHT);
1962 }
1963 }
1964
1965 if (ddir_sync(ddir)) {
1966 td->last_was_sync = true;
1967 if (f) {
1968 f->first_write = -1ULL;
1969 f->last_write = -1ULL;
1970 }
1971 if (should_account(td))
1972 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
1973 return;
1974 }
1975
1976 td->last_was_sync = false;
1977 td->last_ddir = ddir;
1978
1979 if (!io_u->error && ddir_rw(ddir)) {
1980 unsigned long long bytes = io_u->xfer_buflen - io_u->resid;
1981 int ret;
1982
1983 /*
1984 * Make sure we notice short IO from here, and requeue them
1985 * appropriately!
1986 */
1987 if (io_u->resid) {
1988 io_u->xfer_buflen = io_u->resid;
1989 io_u->xfer_buf += bytes;
1990 io_u->offset += bytes;
1991 td->ts.short_io_u[io_u->ddir]++;
1992 if (io_u->offset < io_u->file->real_file_size) {
1993 requeue_io_u(td, io_u_ptr);
1994 return;
1995 }
1996 }
1997
1998 td->io_blocks[ddir]++;
1999 td->io_bytes[ddir] += bytes;
2000
2001 if (!(io_u->flags & IO_U_F_VER_LIST)) {
2002 td->this_io_blocks[ddir]++;
2003 td->this_io_bytes[ddir] += bytes;
2004 }
2005
2006 if (ddir == DDIR_WRITE)
2007 file_log_write_comp(td, f, io_u->offset, bytes);
2008
2009 if (should_account(td))
2010 account_io_completion(td, io_u, icd, ddir, bytes);
2011
2012 icd->bytes_done[ddir] += bytes;
2013
2014 if (io_u->end_io) {
2015 ret = io_u->end_io(td, io_u_ptr);
2016 io_u = *io_u_ptr;
2017 if (ret && !icd->error)
2018 icd->error = ret;
2019 }
2020 } else if (io_u->error) {
2021 icd->error = io_u->error;
2022 io_u_log_error(td, io_u);
2023 }
2024 if (icd->error) {
2025 enum error_type_bit eb = td_error_type(ddir, icd->error);
2026
2027 if (!td_non_fatal_error(td, eb, icd->error))
2028 return;
2029
2030 /*
2031 * If there is a non_fatal error, then add to the error count
2032 * and clear all the errors.
2033 */
2034 update_error_count(td, icd->error);
2035 td_clear_error(td);
2036 icd->error = 0;
2037 if (io_u)
2038 io_u->error = 0;
2039 }
2040}
2041
2042static void init_icd(struct thread_data *td, struct io_completion_data *icd,
2043 int nr)
2044{
2045 int ddir;
2046
2047 if (!gtod_reduce(td))
2048 fio_gettime(&icd->time, NULL);
2049
2050 icd->nr = nr;
2051
2052 icd->error = 0;
2053 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2054 icd->bytes_done[ddir] = 0;
2055}
2056
2057static void ios_completed(struct thread_data *td,
2058 struct io_completion_data *icd)
2059{
2060 struct io_u *io_u;
2061 int i;
2062
2063 for (i = 0; i < icd->nr; i++) {
2064 io_u = td->io_ops->event(td, i);
2065
2066 io_completed(td, &io_u, icd);
2067
2068 if (io_u)
2069 put_io_u(td, io_u);
2070 }
2071}
2072
2073/*
2074 * Complete a single io_u for the sync engines.
2075 */
2076int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
2077{
2078 struct io_completion_data icd;
2079 int ddir;
2080
2081 init_icd(td, &icd, 1);
2082 io_completed(td, &io_u, &icd);
2083
2084 if (io_u)
2085 put_io_u(td, io_u);
2086
2087 if (icd.error) {
2088 td_verror(td, icd.error, "io_u_sync_complete");
2089 return -1;
2090 }
2091
2092 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2093 td->bytes_done[ddir] += icd.bytes_done[ddir];
2094
2095 return 0;
2096}
2097
2098/*
2099 * Called to complete min_events number of io for the async engines.
2100 */
2101int io_u_queued_complete(struct thread_data *td, int min_evts)
2102{
2103 struct io_completion_data icd;
2104 struct timespec *tvp = NULL;
2105 int ret, ddir;
2106 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2107
2108 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2109
2110 if (!min_evts)
2111 tvp = &ts;
2112 else if (min_evts > td->cur_depth)
2113 min_evts = td->cur_depth;
2114
2115 /* No worries, td_io_getevents fixes min and max if they are
2116 * set incorrectly */
2117 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2118 if (ret < 0) {
2119 td_verror(td, -ret, "td_io_getevents");
2120 return ret;
2121 } else if (!ret)
2122 return ret;
2123
2124 init_icd(td, &icd, ret);
2125 ios_completed(td, &icd);
2126 if (icd.error) {
2127 td_verror(td, icd.error, "io_u_queued_complete");
2128 return -1;
2129 }
2130
2131 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2132 td->bytes_done[ddir] += icd.bytes_done[ddir];
2133
2134 return ret;
2135}
2136
2137/*
2138 * Call when io_u is really queued, to update the submission latency.
2139 */
2140void io_u_queued(struct thread_data *td, struct io_u *io_u)
2141{
2142 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2143 unsigned long slat_time;
2144
2145 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2146
2147 if (td->parent)
2148 td = td->parent;
2149
2150 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2151 io_u->offset, io_u_is_prio(io_u));
2152 }
2153}
2154
2155/*
2156 * See if we should reuse the last seed, if dedupe is enabled
2157 */
2158static struct frand_state *get_buf_state(struct thread_data *td)
2159{
2160 unsigned int v;
2161
2162 if (!td->o.dedupe_percentage)
2163 return &td->buf_state;
2164 else if (td->o.dedupe_percentage == 100) {
2165 frand_copy(&td->buf_state_prev, &td->buf_state);
2166 return &td->buf_state;
2167 }
2168
2169 v = rand_between(&td->dedupe_state, 1, 100);
2170
2171 if (v <= td->o.dedupe_percentage)
2172 return &td->buf_state_prev;
2173
2174 return &td->buf_state;
2175}
2176
2177static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2178{
2179 if (td->o.dedupe_percentage == 100)
2180 frand_copy(rs, &td->buf_state_prev);
2181 else if (rs == &td->buf_state)
2182 frand_copy(&td->buf_state_prev, rs);
2183}
2184
2185void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write,
2186 unsigned long long max_bs)
2187{
2188 struct thread_options *o = &td->o;
2189
2190 if (o->mem_type == MEM_CUDA_MALLOC)
2191 return;
2192
2193 if (o->compress_percentage || o->dedupe_percentage) {
2194 unsigned int perc = td->o.compress_percentage;
2195 struct frand_state *rs;
2196 unsigned long long left = max_bs;
2197 unsigned long long this_write;
2198
2199 do {
2200 rs = get_buf_state(td);
2201
2202 min_write = min(min_write, left);
2203
2204 if (perc) {
2205 this_write = min_not_zero(min_write,
2206 (unsigned long long) td->o.compress_chunk);
2207
2208 fill_random_buf_percentage(rs, buf, perc,
2209 this_write, this_write,
2210 o->buffer_pattern,
2211 o->buffer_pattern_bytes);
2212 } else {
2213 fill_random_buf(rs, buf, min_write);
2214 this_write = min_write;
2215 }
2216
2217 buf += this_write;
2218 left -= this_write;
2219 save_buf_state(td, rs);
2220 } while (left);
2221 } else if (o->buffer_pattern_bytes)
2222 fill_buffer_pattern(td, buf, max_bs);
2223 else if (o->zero_buffers)
2224 memset(buf, 0, max_bs);
2225 else
2226 fill_random_buf(get_buf_state(td), buf, max_bs);
2227}
2228
2229/*
2230 * "randomly" fill the buffer contents
2231 */
2232void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2233 unsigned long long min_write, unsigned long long max_bs)
2234{
2235 io_u->buf_filled_len = 0;
2236 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2237}
2238
2239static int do_sync_file_range(const struct thread_data *td,
2240 struct fio_file *f)
2241{
2242 uint64_t offset, nbytes;
2243
2244 offset = f->first_write;
2245 nbytes = f->last_write - f->first_write;
2246
2247 if (!nbytes)
2248 return 0;
2249
2250 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2251}
2252
2253int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2254{
2255 int ret;
2256
2257 if (io_u->ddir == DDIR_SYNC) {
2258 ret = fsync(io_u->file->fd);
2259 } else if (io_u->ddir == DDIR_DATASYNC) {
2260#ifdef CONFIG_FDATASYNC
2261 ret = fdatasync(io_u->file->fd);
2262#else
2263 ret = io_u->xfer_buflen;
2264 io_u->error = EINVAL;
2265#endif
2266 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2267 ret = do_sync_file_range(td, io_u->file);
2268 else {
2269 ret = io_u->xfer_buflen;
2270 io_u->error = EINVAL;
2271 }
2272
2273 if (ret < 0)
2274 io_u->error = errno;
2275
2276 return ret;
2277}
2278
2279int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2280{
2281#ifndef FIO_HAVE_TRIM
2282 io_u->error = EINVAL;
2283 return 0;
2284#else
2285 struct fio_file *f = io_u->file;
2286 int ret;
2287
2288 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2289 if (!ret)
2290 return io_u->xfer_buflen;
2291
2292 io_u->error = ret;
2293 return 0;
2294#endif
2295}
fio - Flexible IO Tester