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