projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Merge branch 'master' of https://github.com/celestinechen/fio
[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 && o->nr_files == 1) {
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(td, 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 fio_fallthrough;
997 case 0:
998 break;
999 }
1000
1001 map[idx]++;
1002}
1003
1004void io_u_mark_submit(struct thread_data *td, unsigned int nr)
1005{
1006 __io_u_mark_map(td->ts.io_u_submit, nr);
1007 td->ts.total_submit++;
1008}
1009
1010void io_u_mark_complete(struct thread_data *td, unsigned int nr)
1011{
1012 __io_u_mark_map(td->ts.io_u_complete, nr);
1013 td->ts.total_complete++;
1014}
1015
1016void io_u_mark_depth(struct thread_data *td, unsigned int nr)
1017{
1018 int idx = 0;
1019
1020 switch (td->cur_depth) {
1021 default:
1022 idx = 6;
1023 break;
1024 case 32 ... 63:
1025 idx = 5;
1026 break;
1027 case 16 ... 31:
1028 idx = 4;
1029 break;
1030 case 8 ... 15:
1031 idx = 3;
1032 break;
1033 case 4 ... 7:
1034 idx = 2;
1035 break;
1036 case 2 ... 3:
1037 idx = 1;
1038 fio_fallthrough;
1039 case 1:
1040 break;
1041 }
1042
1043 td->ts.io_u_map[idx] += nr;
1044}
1045
1046static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1047{
1048 int idx = 0;
1049
1050 assert(nsec < 1000);
1051
1052 switch (nsec) {
1053 case 750 ... 999:
1054 idx = 9;
1055 break;
1056 case 500 ... 749:
1057 idx = 8;
1058 break;
1059 case 250 ... 499:
1060 idx = 7;
1061 break;
1062 case 100 ... 249:
1063 idx = 6;
1064 break;
1065 case 50 ... 99:
1066 idx = 5;
1067 break;
1068 case 20 ... 49:
1069 idx = 4;
1070 break;
1071 case 10 ... 19:
1072 idx = 3;
1073 break;
1074 case 4 ... 9:
1075 idx = 2;
1076 break;
1077 case 2 ... 3:
1078 idx = 1;
1079 fio_fallthrough;
1080 case 0 ... 1:
1081 break;
1082 }
1083
1084 assert(idx < FIO_IO_U_LAT_N_NR);
1085 td->ts.io_u_lat_n[idx]++;
1086}
1087
1088static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1089{
1090 int idx = 0;
1091
1092 assert(usec < 1000 && usec >= 1);
1093
1094 switch (usec) {
1095 case 750 ... 999:
1096 idx = 9;
1097 break;
1098 case 500 ... 749:
1099 idx = 8;
1100 break;
1101 case 250 ... 499:
1102 idx = 7;
1103 break;
1104 case 100 ... 249:
1105 idx = 6;
1106 break;
1107 case 50 ... 99:
1108 idx = 5;
1109 break;
1110 case 20 ... 49:
1111 idx = 4;
1112 break;
1113 case 10 ... 19:
1114 idx = 3;
1115 break;
1116 case 4 ... 9:
1117 idx = 2;
1118 break;
1119 case 2 ... 3:
1120 idx = 1;
1121 fio_fallthrough;
1122 case 0 ... 1:
1123 break;
1124 }
1125
1126 assert(idx < FIO_IO_U_LAT_U_NR);
1127 td->ts.io_u_lat_u[idx]++;
1128}
1129
1130static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1131{
1132 int idx = 0;
1133
1134 assert(msec >= 1);
1135
1136 switch (msec) {
1137 default:
1138 idx = 11;
1139 break;
1140 case 1000 ... 1999:
1141 idx = 10;
1142 break;
1143 case 750 ... 999:
1144 idx = 9;
1145 break;
1146 case 500 ... 749:
1147 idx = 8;
1148 break;
1149 case 250 ... 499:
1150 idx = 7;
1151 break;
1152 case 100 ... 249:
1153 idx = 6;
1154 break;
1155 case 50 ... 99:
1156 idx = 5;
1157 break;
1158 case 20 ... 49:
1159 idx = 4;
1160 break;
1161 case 10 ... 19:
1162 idx = 3;
1163 break;
1164 case 4 ... 9:
1165 idx = 2;
1166 break;
1167 case 2 ... 3:
1168 idx = 1;
1169 fio_fallthrough;
1170 case 0 ... 1:
1171 break;
1172 }
1173
1174 assert(idx < FIO_IO_U_LAT_M_NR);
1175 td->ts.io_u_lat_m[idx]++;
1176}
1177
1178static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1179{
1180 if (nsec < 1000)
1181 io_u_mark_lat_nsec(td, nsec);
1182 else if (nsec < 1000000)
1183 io_u_mark_lat_usec(td, nsec / 1000);
1184 else
1185 io_u_mark_lat_msec(td, nsec / 1000000);
1186}
1187
1188static unsigned int __get_next_fileno_rand(struct thread_data *td)
1189{
1190 unsigned long fileno;
1191
1192 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1193 uint64_t frand_max = rand_max(&td->next_file_state);
1194 unsigned long r;
1195
1196 r = __rand(&td->next_file_state);
1197 return (unsigned int) ((double) td->o.nr_files
1198 * (r / (frand_max + 1.0)));
1199 }
1200
1201 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1202 fileno = zipf_next(&td->next_file_zipf);
1203 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1204 fileno = pareto_next(&td->next_file_zipf);
1205 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1206 fileno = gauss_next(&td->next_file_gauss);
1207 else {
1208 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1209 assert(0);
1210 return 0;
1211 }
1212
1213 return fileno >> FIO_FSERVICE_SHIFT;
1214}
1215
1216/*
1217 * Get next file to service by choosing one at random
1218 */
1219static struct fio_file *get_next_file_rand(struct thread_data *td,
1220 enum fio_file_flags goodf,
1221 enum fio_file_flags badf)
1222{
1223 struct fio_file *f;
1224 int fno;
1225
1226 do {
1227 int opened = 0;
1228
1229 fno = __get_next_fileno_rand(td);
1230
1231 f = td->files[fno];
1232 if (fio_file_done(f))
1233 continue;
1234
1235 if (!fio_file_open(f)) {
1236 int err;
1237
1238 if (td->nr_open_files >= td->o.open_files)
1239 return ERR_PTR(-EBUSY);
1240
1241 err = td_io_open_file(td, f);
1242 if (err)
1243 continue;
1244 opened = 1;
1245 }
1246
1247 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1248 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1249 return f;
1250 }
1251 if (opened)
1252 td_io_close_file(td, f);
1253 } while (1);
1254}
1255
1256/*
1257 * Get next file to service by doing round robin between all available ones
1258 */
1259static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1260 int badf)
1261{
1262 unsigned int old_next_file = td->next_file;
1263 struct fio_file *f;
1264
1265 do {
1266 int opened = 0;
1267
1268 f = td->files[td->next_file];
1269
1270 td->next_file++;
1271 if (td->next_file >= td->o.nr_files)
1272 td->next_file = 0;
1273
1274 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1275 if (fio_file_done(f)) {
1276 f = NULL;
1277 continue;
1278 }
1279
1280 if (!fio_file_open(f)) {
1281 int err;
1282
1283 if (td->nr_open_files >= td->o.open_files)
1284 return ERR_PTR(-EBUSY);
1285
1286 err = td_io_open_file(td, f);
1287 if (err) {
1288 dprint(FD_FILE, "error %d on open of %s\n",
1289 err, f->file_name);
1290 f = NULL;
1291 continue;
1292 }
1293 opened = 1;
1294 }
1295
1296 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1297 f->flags);
1298 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1299 break;
1300
1301 if (opened)
1302 td_io_close_file(td, f);
1303
1304 f = NULL;
1305 } while (td->next_file != old_next_file);
1306
1307 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1308 return f;
1309}
1310
1311static struct fio_file *__get_next_file(struct thread_data *td)
1312{
1313 struct fio_file *f;
1314
1315 assert(td->o.nr_files <= td->files_index);
1316
1317 if (td->nr_done_files >= td->o.nr_files) {
1318 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1319 " nr_files=%d\n", td->nr_open_files,
1320 td->nr_done_files,
1321 td->o.nr_files);
1322 return NULL;
1323 }
1324
1325 f = td->file_service_file;
1326 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1327 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1328 goto out;
1329 if (td->file_service_left) {
1330 td->file_service_left--;
1331 goto out;
1332 }
1333 }
1334
1335 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1336 td->o.file_service_type == FIO_FSERVICE_SEQ)
1337 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1338 else
1339 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1340
1341 if (IS_ERR(f))
1342 return f;
1343
1344 td->file_service_file = f;
1345 td->file_service_left = td->file_service_nr - 1;
1346out:
1347 if (f)
1348 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1349 else
1350 dprint(FD_FILE, "get_next_file: NULL\n");
1351 return f;
1352}
1353
1354static struct fio_file *get_next_file(struct thread_data *td)
1355{
1356 return __get_next_file(td);
1357}
1358
1359static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1360{
1361 struct fio_file *f;
1362
1363 do {
1364 f = get_next_file(td);
1365 if (IS_ERR_OR_NULL(f))
1366 return PTR_ERR(f);
1367
1368 io_u->file = f;
1369 get_file(f);
1370
1371 if (!fill_io_u(td, io_u))
1372 break;
1373
1374 zbd_put_io_u(td, io_u);
1375
1376 put_file_log(td, f);
1377 td_io_close_file(td, f);
1378 io_u->file = NULL;
1379 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1380 fio_file_reset(td, f);
1381 else {
1382 fio_file_set_done(f);
1383 td->nr_done_files++;
1384 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1385 td->nr_done_files, td->o.nr_files);
1386 }
1387 } while (1);
1388
1389 return 0;
1390}
1391
1392static void lat_fatal(struct thread_data *td, struct io_u *io_u, struct io_completion_data *icd,
1393 unsigned long long tnsec, unsigned long long max_nsec)
1394{
1395 if (!td->error) {
1396 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec): %s %s %llu %llu\n",
1397 tnsec, max_nsec,
1398 io_u->file->file_name,
1399 io_ddir_name(io_u->ddir),
1400 io_u->offset, io_u->buflen);
1401 }
1402 td_verror(td, ETIMEDOUT, "max latency exceeded");
1403 icd->error = ETIMEDOUT;
1404}
1405
1406static void lat_new_cycle(struct thread_data *td)
1407{
1408 fio_gettime(&td->latency_ts, NULL);
1409 td->latency_ios = ddir_rw_sum(td->io_blocks);
1410 td->latency_failed = 0;
1411}
1412
1413/*
1414 * We had an IO outside the latency target. Reduce the queue depth. If we
1415 * are at QD=1, then it's time to give up.
1416 */
1417static bool __lat_target_failed(struct thread_data *td)
1418{
1419 if (td->latency_qd == 1)
1420 return true;
1421
1422 td->latency_qd_high = td->latency_qd;
1423
1424 if (td->latency_qd == td->latency_qd_low)
1425 td->latency_qd_low--;
1426
1427 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1428 td->latency_stable_count = 0;
1429
1430 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1431
1432 /*
1433 * When we ramp QD down, quiesce existing IO to prevent
1434 * a storm of ramp downs due to pending higher depth.
1435 */
1436 io_u_quiesce(td);
1437 lat_new_cycle(td);
1438 return false;
1439}
1440
1441static bool lat_target_failed(struct thread_data *td)
1442{
1443 if (td->o.latency_percentile.u.f == 100.0)
1444 return __lat_target_failed(td);
1445
1446 td->latency_failed++;
1447 return false;
1448}
1449
1450void lat_target_init(struct thread_data *td)
1451{
1452 td->latency_end_run = 0;
1453
1454 if (td->o.latency_target) {
1455 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1456 fio_gettime(&td->latency_ts, NULL);
1457 td->latency_qd = 1;
1458 td->latency_qd_high = td->o.iodepth;
1459 td->latency_qd_low = 1;
1460 td->latency_ios = ddir_rw_sum(td->io_blocks);
1461 } else
1462 td->latency_qd = td->o.iodepth;
1463}
1464
1465void lat_target_reset(struct thread_data *td)
1466{
1467 if (!td->latency_end_run)
1468 lat_target_init(td);
1469}
1470
1471static void lat_target_success(struct thread_data *td)
1472{
1473 const unsigned int qd = td->latency_qd;
1474 struct thread_options *o = &td->o;
1475
1476 td->latency_qd_low = td->latency_qd;
1477
1478 if (td->latency_qd + 1 == td->latency_qd_high) {
1479 /*
1480 * latency_qd will not incease on lat_target_success(), so
1481 * called stable. If we stick with this queue depth, the
1482 * final latency is likely lower than latency_target. Fix
1483 * this by increasing latency_qd_high slowly. Use a naive
1484 * heuristic here. If we get lat_target_success() 3 times
1485 * in a row, increase latency_qd_high by 1.
1486 */
1487 if (++td->latency_stable_count >= 3) {
1488 td->latency_qd_high++;
1489 td->latency_stable_count = 0;
1490 }
1491 }
1492
1493 /*
1494 * If we haven't failed yet, we double up to a failing value instead
1495 * of bisecting from highest possible queue depth. If we have set
1496 * a limit other than td->o.iodepth, bisect between that.
1497 */
1498 if (td->latency_qd_high != o->iodepth)
1499 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1500 else
1501 td->latency_qd *= 2;
1502
1503 if (td->latency_qd > o->iodepth)
1504 td->latency_qd = o->iodepth;
1505
1506 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1507
1508 /*
1509 * Same as last one, we are done. Let it run a latency cycle, so
1510 * we get only the results from the targeted depth.
1511 */
1512 if (!o->latency_run && td->latency_qd == qd) {
1513 if (td->latency_end_run) {
1514 dprint(FD_RATE, "We are done\n");
1515 td->done = 1;
1516 } else {
1517 dprint(FD_RATE, "Quiesce and final run\n");
1518 io_u_quiesce(td);
1519 td->latency_end_run = 1;
1520 reset_all_stats(td);
1521 reset_io_stats(td);
1522 }
1523 }
1524
1525 lat_new_cycle(td);
1526}
1527
1528/*
1529 * Check if we can bump the queue depth
1530 */
1531void lat_target_check(struct thread_data *td)
1532{
1533 uint64_t usec_window;
1534 uint64_t ios;
1535 double success_ios;
1536
1537 usec_window = utime_since_now(&td->latency_ts);
1538 if (usec_window < td->o.latency_window)
1539 return;
1540
1541 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1542 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1543 success_ios *= 100.0;
1544
1545 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1546
1547 if (success_ios >= td->o.latency_percentile.u.f)
1548 lat_target_success(td);
1549 else
1550 __lat_target_failed(td);
1551}
1552
1553/*
1554 * If latency target is enabled, we might be ramping up or down and not
1555 * using the full queue depth available.
1556 */
1557bool queue_full(const struct thread_data *td)
1558{
1559 const int qempty = io_u_qempty(&td->io_u_freelist);
1560
1561 if (qempty)
1562 return true;
1563 if (!td->o.latency_target)
1564 return false;
1565
1566 return td->cur_depth >= td->latency_qd;
1567}
1568
1569struct io_u *__get_io_u(struct thread_data *td)
1570{
1571 const bool needs_lock = td_async_processing(td);
1572 struct io_u *io_u = NULL;
1573 int ret;
1574
1575 if (td->stop_io)
1576 return NULL;
1577
1578 if (needs_lock)
1579 __td_io_u_lock(td);
1580
1581again:
1582 if (!io_u_rempty(&td->io_u_requeues)) {
1583 io_u = io_u_rpop(&td->io_u_requeues);
1584 io_u->resid = 0;
1585 } else if (!queue_full(td)) {
1586 io_u = io_u_qpop(&td->io_u_freelist);
1587
1588 io_u->file = NULL;
1589 io_u->buflen = 0;
1590 io_u->resid = 0;
1591 io_u->end_io = NULL;
1592 }
1593
1594 if (io_u) {
1595 assert(io_u->flags & IO_U_F_FREE);
1596 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1597 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1598 IO_U_F_VER_LIST);
1599
1600 io_u->error = 0;
1601 io_u->acct_ddir = -1;
1602 td->cur_depth++;
1603 assert(!(td->flags & TD_F_CHILD));
1604 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1605 io_u->ipo = NULL;
1606 } else if (td_async_processing(td)) {
1607 /*
1608 * We ran out, wait for async verify threads to finish and
1609 * return one
1610 */
1611 assert(!(td->flags & TD_F_CHILD));
1612 ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1613 assert(ret == 0);
1614 if (!td->error)
1615 goto again;
1616 }
1617
1618 if (needs_lock)
1619 __td_io_u_unlock(td);
1620
1621 return io_u;
1622}
1623
1624static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1625{
1626 if (!(td->flags & TD_F_TRIM_BACKLOG))
1627 return false;
1628 if (!td->trim_entries)
1629 return false;
1630
1631 if (td->trim_batch) {
1632 td->trim_batch--;
1633 if (get_next_trim(td, io_u))
1634 return true;
1635 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1636 td->last_ddir != DDIR_READ) {
1637 td->trim_batch = td->o.trim_batch;
1638 if (!td->trim_batch)
1639 td->trim_batch = td->o.trim_backlog;
1640 if (get_next_trim(td, io_u))
1641 return true;
1642 }
1643
1644 return false;
1645}
1646
1647static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1648{
1649 if (!(td->flags & TD_F_VER_BACKLOG))
1650 return false;
1651
1652 if (td->io_hist_len) {
1653 int get_verify = 0;
1654
1655 if (td->verify_batch)
1656 get_verify = 1;
1657 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1658 td->last_ddir != DDIR_READ) {
1659 td->verify_batch = td->o.verify_batch;
1660 if (!td->verify_batch)
1661 td->verify_batch = td->o.verify_backlog;
1662 get_verify = 1;
1663 }
1664
1665 if (get_verify && !get_next_verify(td, io_u)) {
1666 td->verify_batch--;
1667 return true;
1668 }
1669 }
1670
1671 return false;
1672}
1673
1674/*
1675 * Fill offset and start time into the buffer content, to prevent too
1676 * easy compressible data for simple de-dupe attempts. Do this for every
1677 * 512b block in the range, since that should be the smallest block size
1678 * we can expect from a device.
1679 */
1680static void small_content_scramble(struct io_u *io_u)
1681{
1682 unsigned long long i, nr_blocks = io_u->buflen >> 9;
1683 unsigned int offset;
1684 uint64_t boffset, *iptr;
1685 char *p;
1686
1687 if (!nr_blocks)
1688 return;
1689
1690 p = io_u->xfer_buf;
1691 boffset = io_u->offset;
1692
1693 if (io_u->buf_filled_len)
1694 io_u->buf_filled_len = 0;
1695
1696 /*
1697 * Generate random index between 0..7. We do chunks of 512b, if
1698 * we assume a cacheline is 64 bytes, then we have 8 of those.
1699 * Scramble content within the blocks in the same cacheline to
1700 * speed things up.
1701 */
1702 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1703
1704 for (i = 0; i < nr_blocks; i++) {
1705 /*
1706 * Fill offset into start of cacheline, time into end
1707 * of cacheline
1708 */
1709 iptr = (void *) p + (offset << 6);
1710 *iptr = boffset;
1711
1712 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1713 iptr[0] = io_u->start_time.tv_sec;
1714 iptr[1] = io_u->start_time.tv_nsec;
1715
1716 p += 512;
1717 boffset += 512;
1718 }
1719}
1720
1721/*
1722 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1723 * etc. The returned io_u is fully ready to be prepped, populated and submitted.
1724 */
1725struct io_u *get_io_u(struct thread_data *td)
1726{
1727 struct fio_file *f;
1728 struct io_u *io_u;
1729 int do_scramble = 0;
1730 long ret = 0;
1731
1732 io_u = __get_io_u(td);
1733 if (!io_u) {
1734 dprint(FD_IO, "__get_io_u failed\n");
1735 return NULL;
1736 }
1737
1738 if (check_get_verify(td, io_u))
1739 goto out;
1740 if (check_get_trim(td, io_u))
1741 goto out;
1742
1743 /*
1744 * from a requeue, io_u already setup
1745 */
1746 if (io_u->file)
1747 goto out;
1748
1749 /*
1750 * If using an iolog, grab next piece if any available.
1751 */
1752 if (td->flags & TD_F_READ_IOLOG) {
1753 if (read_iolog_get(td, io_u))
1754 goto err_put;
1755 } else if (set_io_u_file(td, io_u)) {
1756 ret = -EBUSY;
1757 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1758 goto err_put;
1759 }
1760
1761 f = io_u->file;
1762 if (!f) {
1763 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1764 goto err_put;
1765 }
1766
1767 assert(fio_file_open(f));
1768
1769 if (ddir_rw(io_u->ddir)) {
1770 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1771 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1772 goto err_put;
1773 }
1774
1775 f->last_start[io_u->ddir] = io_u->offset;
1776 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1777
1778 if (io_u->ddir == DDIR_WRITE) {
1779 if (td->flags & TD_F_REFILL_BUFFERS) {
1780 io_u_fill_buffer(td, io_u,
1781 td->o.min_bs[DDIR_WRITE],
1782 io_u->buflen);
1783 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1784 !(td->flags & TD_F_COMPRESS) &&
1785 !(td->flags & TD_F_DO_VERIFY))
1786 do_scramble = 1;
1787 } else if (io_u->ddir == DDIR_READ) {
1788 /*
1789 * Reset the buf_filled parameters so next time if the
1790 * buffer is used for writes it is refilled.
1791 */
1792 io_u->buf_filled_len = 0;
1793 }
1794 }
1795
1796 /*
1797 * Set io data pointers.
1798 */
1799 io_u->xfer_buf = io_u->buf;
1800 io_u->xfer_buflen = io_u->buflen;
1801
1802 /*
1803 * Remember the issuing context priority. The IO engine may change this.
1804 */
1805 io_u->ioprio = td->ioprio;
1806 io_u->clat_prio_index = 0;
1807out:
1808 assert(io_u->file);
1809 if (!td_io_prep(td, io_u)) {
1810 if (!td->o.disable_lat)
1811 fio_gettime(&io_u->start_time, NULL);
1812
1813 if (do_scramble)
1814 small_content_scramble(io_u);
1815
1816 return io_u;
1817 }
1818err_put:
1819 dprint(FD_IO, "get_io_u failed\n");
1820 put_io_u(td, io_u);
1821 return ERR_PTR(ret);
1822}
1823
1824static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1825{
1826 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1827
1828 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1829 return;
1830
1831 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n",
1832 io_u->file ? " on file " : "",
1833 io_u->file ? io_u->file->file_name : "",
1834 strerror(io_u->error),
1835 io_ddir_name(io_u->ddir),
1836 io_u->offset, io_u->xfer_buflen);
1837
1838 if (td->io_ops->errdetails) {
1839 char *err = td->io_ops->errdetails(io_u);
1840
1841 log_err("fio: %s\n", err);
1842 free(err);
1843 }
1844
1845 if (!td->error)
1846 td_verror(td, io_u->error, "io_u error");
1847}
1848
1849void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1850{
1851 __io_u_log_error(td, io_u);
1852 if (td->parent)
1853 __io_u_log_error(td->parent, io_u);
1854}
1855
1856static inline bool gtod_reduce(struct thread_data *td)
1857{
1858 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1859 || td->o.gtod_reduce;
1860}
1861
1862static void trim_block_info(struct thread_data *td, struct io_u *io_u)
1863{
1864 uint32_t *info = io_u_block_info(td, io_u);
1865
1866 if (BLOCK_INFO_STATE(*info) >= BLOCK_STATE_TRIM_FAILURE)
1867 return;
1868
1869 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, BLOCK_INFO_TRIMS(*info) + 1);
1870}
1871
1872static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1873 struct io_completion_data *icd,
1874 const enum fio_ddir idx, unsigned int bytes)
1875{
1876 const int no_reduce = !gtod_reduce(td);
1877 unsigned long long llnsec = 0;
1878
1879 if (td->parent)
1880 td = td->parent;
1881
1882 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1883 return;
1884
1885 if (no_reduce)
1886 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1887
1888 if (!td->o.disable_lat) {
1889 unsigned long long tnsec;
1890
1891 tnsec = ntime_since(&io_u->start_time, &icd->time);
1892 add_lat_sample(td, idx, tnsec, bytes, io_u->offset,
1893 io_u->ioprio, io_u->clat_prio_index);
1894
1895 if (td->flags & TD_F_PROFILE_OPS) {
1896 struct prof_io_ops *ops = &td->prof_io_ops;
1897
1898 if (ops->io_u_lat)
1899 icd->error = ops->io_u_lat(td, tnsec);
1900 }
1901
1902 if (ddir_rw(idx)) {
1903 if (td->o.max_latency[idx] && tnsec > td->o.max_latency[idx])
1904 lat_fatal(td, io_u, icd, tnsec, td->o.max_latency[idx]);
1905 if (td->o.latency_target && tnsec > td->o.latency_target) {
1906 if (lat_target_failed(td))
1907 lat_fatal(td, io_u, icd, tnsec, td->o.latency_target);
1908 }
1909 }
1910 }
1911
1912 if (ddir_rw(idx)) {
1913 if (!td->o.disable_clat) {
1914 add_clat_sample(td, idx, llnsec, bytes, io_u->offset,
1915 io_u->ioprio, io_u->clat_prio_index);
1916 io_u_mark_latency(td, llnsec);
1917 }
1918
1919 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1920 add_bw_sample(td, io_u, bytes, llnsec);
1921
1922 if (no_reduce && per_unit_log(td->iops_log))
1923 add_iops_sample(td, io_u, bytes);
1924 } else if (ddir_sync(idx) && !td->o.disable_clat)
1925 add_sync_clat_sample(&td->ts, llnsec);
1926
1927 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM)
1928 trim_block_info(td, io_u);
1929}
1930
1931static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1932 uint64_t offset, unsigned int bytes)
1933{
1934 int idx;
1935
1936 if (!f)
1937 return;
1938
1939 if (f->first_write == -1ULL || offset < f->first_write)
1940 f->first_write = offset;
1941 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1942 f->last_write = offset + bytes;
1943
1944 if (!f->last_write_comp)
1945 return;
1946
1947 idx = f->last_write_idx++;
1948 f->last_write_comp[idx] = offset;
1949 if (f->last_write_idx == td->o.iodepth)
1950 f->last_write_idx = 0;
1951}
1952
1953static bool should_account(struct thread_data *td)
1954{
1955 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1956 td->runstate == TD_VERIFYING);
1957}
1958
1959static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1960 struct io_completion_data *icd)
1961{
1962 struct io_u *io_u = *io_u_ptr;
1963 enum fio_ddir ddir = io_u->ddir;
1964 struct fio_file *f = io_u->file;
1965
1966 dprint_io_u(io_u, "complete");
1967
1968 assert(io_u->flags & IO_U_F_FLIGHT);
1969 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1970
1971 /*
1972 * Mark IO ok to verify
1973 */
1974 if (io_u->ipo) {
1975 /*
1976 * Remove errored entry from the verification list
1977 */
1978 if (io_u->error)
1979 unlog_io_piece(td, io_u);
1980 else {
1981 atomic_store_release(&io_u->ipo->flags,
1982 io_u->ipo->flags & ~IP_F_IN_FLIGHT);
1983 }
1984 }
1985
1986 if (ddir_sync(ddir)) {
1987 td->last_was_sync = true;
1988 if (f) {
1989 f->first_write = -1ULL;
1990 f->last_write = -1ULL;
1991 }
1992 if (should_account(td))
1993 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
1994 return;
1995 }
1996
1997 td->last_was_sync = false;
1998 td->last_ddir = ddir;
1999
2000 if (!io_u->error && ddir_rw(ddir)) {
2001 unsigned long long bytes = io_u->xfer_buflen - io_u->resid;
2002 int ret;
2003
2004 /*
2005 * Make sure we notice short IO from here, and requeue them
2006 * appropriately!
2007 */
2008 if (bytes && io_u->resid) {
2009 io_u->xfer_buflen = io_u->resid;
2010 io_u->xfer_buf += bytes;
2011 io_u->offset += bytes;
2012 td->ts.short_io_u[io_u->ddir]++;
2013 if (io_u->offset < io_u->file->real_file_size) {
2014 requeue_io_u(td, io_u_ptr);
2015 return;
2016 }
2017 }
2018
2019 td->io_blocks[ddir]++;
2020 td->io_bytes[ddir] += bytes;
2021
2022 if (!(io_u->flags & IO_U_F_VER_LIST)) {
2023 td->this_io_blocks[ddir]++;
2024 td->this_io_bytes[ddir] += bytes;
2025 }
2026
2027 if (ddir == DDIR_WRITE)
2028 file_log_write_comp(td, f, io_u->offset, bytes);
2029
2030 if (should_account(td))
2031 account_io_completion(td, io_u, icd, ddir, bytes);
2032
2033 icd->bytes_done[ddir] += bytes;
2034
2035 if (io_u->end_io) {
2036 ret = io_u->end_io(td, io_u_ptr);
2037 io_u = *io_u_ptr;
2038 if (ret && !icd->error)
2039 icd->error = ret;
2040 }
2041 } else if (io_u->error) {
2042 icd->error = io_u->error;
2043 io_u_log_error(td, io_u);
2044 }
2045 if (icd->error) {
2046 enum error_type_bit eb = td_error_type(ddir, icd->error);
2047
2048 if (!td_non_fatal_error(td, eb, icd->error))
2049 return;
2050
2051 /*
2052 * If there is a non_fatal error, then add to the error count
2053 * and clear all the errors.
2054 */
2055 update_error_count(td, icd->error);
2056 td_clear_error(td);
2057 icd->error = 0;
2058 if (io_u)
2059 io_u->error = 0;
2060 }
2061}
2062
2063static void init_icd(struct thread_data *td, struct io_completion_data *icd,
2064 int nr)
2065{
2066 int ddir;
2067
2068 if (!gtod_reduce(td))
2069 fio_gettime(&icd->time, NULL);
2070
2071 icd->nr = nr;
2072
2073 icd->error = 0;
2074 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2075 icd->bytes_done[ddir] = 0;
2076}
2077
2078static void ios_completed(struct thread_data *td,
2079 struct io_completion_data *icd)
2080{
2081 struct io_u *io_u;
2082 int i;
2083
2084 for (i = 0; i < icd->nr; i++) {
2085 io_u = td->io_ops->event(td, i);
2086
2087 io_completed(td, &io_u, icd);
2088
2089 if (io_u)
2090 put_io_u(td, io_u);
2091 }
2092}
2093
2094/*
2095 * Complete a single io_u for the sync engines.
2096 */
2097int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
2098{
2099 struct io_completion_data icd;
2100 int ddir;
2101
2102 init_icd(td, &icd, 1);
2103 io_completed(td, &io_u, &icd);
2104
2105 if (io_u)
2106 put_io_u(td, io_u);
2107
2108 if (icd.error) {
2109 td_verror(td, icd.error, "io_u_sync_complete");
2110 return -1;
2111 }
2112
2113 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2114 td->bytes_done[ddir] += icd.bytes_done[ddir];
2115
2116 return 0;
2117}
2118
2119/*
2120 * Called to complete min_events number of io for the async engines.
2121 */
2122int io_u_queued_complete(struct thread_data *td, int min_evts)
2123{
2124 struct io_completion_data icd;
2125 struct timespec *tvp = NULL;
2126 int ret, ddir;
2127 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2128
2129 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2130
2131 if (!min_evts)
2132 tvp = &ts;
2133 else if (min_evts > td->cur_depth)
2134 min_evts = td->cur_depth;
2135
2136 /* No worries, td_io_getevents fixes min and max if they are
2137 * set incorrectly */
2138 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2139 if (ret < 0) {
2140 td_verror(td, -ret, "td_io_getevents");
2141 return ret;
2142 } else if (!ret)
2143 return ret;
2144
2145 init_icd(td, &icd, ret);
2146 ios_completed(td, &icd);
2147 if (icd.error) {
2148 td_verror(td, icd.error, "io_u_queued_complete");
2149 return -1;
2150 }
2151
2152 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2153 td->bytes_done[ddir] += icd.bytes_done[ddir];
2154
2155 return ret;
2156}
2157
2158/*
2159 * Call when io_u is really queued, to update the submission latency.
2160 */
2161void io_u_queued(struct thread_data *td, struct io_u *io_u)
2162{
2163 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2164 unsigned long slat_time;
2165
2166 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2167
2168 if (td->parent)
2169 td = td->parent;
2170
2171 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2172 io_u->offset, io_u->ioprio);
2173 }
2174}
2175
2176/*
2177 * See if we should reuse the last seed, if dedupe is enabled
2178 */
2179static struct frand_state *get_buf_state(struct thread_data *td)
2180{
2181 unsigned int v;
2182 unsigned long long i;
2183
2184 if (!td->o.dedupe_percentage)
2185 return &td->buf_state;
2186 else if (td->o.dedupe_percentage == 100) {
2187 frand_copy(&td->buf_state_prev, &td->buf_state);
2188 return &td->buf_state;
2189 }
2190
2191 v = rand_between(&td->dedupe_state, 1, 100);
2192
2193 if (v <= td->o.dedupe_percentage)
2194 switch (td->o.dedupe_mode) {
2195 case DEDUPE_MODE_REPEAT:
2196 /*
2197 * The caller advances the returned frand_state.
2198 * A copy of prev should be returned instead since
2199 * a subsequent intention to generate a deduped buffer
2200 * might result in generating a unique one
2201 */
2202 frand_copy(&td->buf_state_ret, &td->buf_state_prev);
2203 return &td->buf_state_ret;
2204 case DEDUPE_MODE_WORKING_SET:
2205 i = rand_between(&td->dedupe_working_set_index_state, 0, td->num_unique_pages - 1);
2206 frand_copy(&td->buf_state_ret, &td->dedupe_working_set_states[i]);
2207 return &td->buf_state_ret;
2208 default:
2209 log_err("unexpected dedupe mode %u\n", td->o.dedupe_mode);
2210 assert(0);
2211 }
2212
2213 return &td->buf_state;
2214}
2215
2216static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2217{
2218 if (td->o.dedupe_percentage == 100)
2219 frand_copy(rs, &td->buf_state_prev);
2220 else if (rs == &td->buf_state)
2221 frand_copy(&td->buf_state_prev, rs);
2222}
2223
2224void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write,
2225 unsigned long long max_bs)
2226{
2227 struct thread_options *o = &td->o;
2228
2229 if (o->mem_type == MEM_CUDA_MALLOC)
2230 return;
2231
2232 if (o->compress_percentage || o->dedupe_percentage) {
2233 unsigned int perc = td->o.compress_percentage;
2234 struct frand_state *rs = NULL;
2235 unsigned long long left = max_bs;
2236 unsigned long long this_write;
2237
2238 do {
2239 /*
2240 * Buffers are either entirely dedupe-able or not.
2241 * If we choose to dedup, the buffer should undergo
2242 * the same manipulation as the original write. Which
2243 * means we should retrack the steps we took for compression
2244 * as well.
2245 */
2246 if (!rs)
2247 rs = get_buf_state(td);
2248
2249 min_write = min(min_write, left);
2250
2251 this_write = min_not_zero(min_write,
2252 (unsigned long long) td->o.compress_chunk);
2253
2254 fill_random_buf_percentage(rs, buf, perc,
2255 this_write, this_write,
2256 o->buffer_pattern,
2257 o->buffer_pattern_bytes);
2258
2259 buf += this_write;
2260 left -= this_write;
2261 save_buf_state(td, rs);
2262 } while (left);
2263 } else if (o->buffer_pattern_bytes)
2264 fill_buffer_pattern(td, buf, max_bs);
2265 else if (o->zero_buffers)
2266 memset(buf, 0, max_bs);
2267 else
2268 fill_random_buf(get_buf_state(td), buf, max_bs);
2269}
2270
2271/*
2272 * "randomly" fill the buffer contents
2273 */
2274void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2275 unsigned long long min_write, unsigned long long max_bs)
2276{
2277 io_u->buf_filled_len = 0;
2278 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2279}
2280
2281static int do_sync_file_range(const struct thread_data *td,
2282 struct fio_file *f)
2283{
2284 uint64_t offset, nbytes;
2285
2286 offset = f->first_write;
2287 nbytes = f->last_write - f->first_write;
2288
2289 if (!nbytes)
2290 return 0;
2291
2292 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2293}
2294
2295int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2296{
2297 int ret;
2298
2299 if (io_u->ddir == DDIR_SYNC) {
2300#ifdef CONFIG_FCNTL_SYNC
2301 ret = fcntl(io_u->file->fd, F_FULLFSYNC);
2302#else
2303 ret = fsync(io_u->file->fd);
2304#endif
2305 } else if (io_u->ddir == DDIR_DATASYNC) {
2306#ifdef CONFIG_FDATASYNC
2307 ret = fdatasync(io_u->file->fd);
2308#else
2309 ret = io_u->xfer_buflen;
2310 io_u->error = EINVAL;
2311#endif
2312 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2313 ret = do_sync_file_range(td, io_u->file);
2314 else {
2315 ret = io_u->xfer_buflen;
2316 io_u->error = EINVAL;
2317 }
2318
2319 if (ret < 0)
2320 io_u->error = errno;
2321
2322 return ret;
2323}
2324
2325int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2326{
2327#ifndef FIO_HAVE_TRIM
2328 io_u->error = EINVAL;
2329 return 0;
2330#else
2331 struct fio_file *f = io_u->file;
2332 int ret;
2333
2334 if (td->o.zone_mode == ZONE_MODE_ZBD) {
2335 ret = zbd_do_io_u_trim(td, io_u);
2336 if (ret == io_u_completed)
2337 return io_u->xfer_buflen;
2338 if (ret)
2339 goto err;
2340 }
2341
2342 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2343 if (!ret)
2344 return io_u->xfer_buflen;
2345
2346err:
2347 io_u->error = ret;
2348 return 0;
2349#endif
2350}
fio - Flexible IO Tester