projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
io_u: move to next zone even if zoneskip is unset
[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;
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 break;
629 }
630
631 if (td->flags & TD_F_REGROW_LOGS)
632 regrow_logs(td);
633
634 if (completed)
635 return completed;
636
637 return ret;
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->start);
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 usec_sleep(td, usec);
684 return ddir;
685}
686
687/*
688 * Return the data direction for the next io_u. If the job is a
689 * mixed read/write workload, check the rwmix cycle and switch if
690 * necessary.
691 */
692static enum fio_ddir get_rw_ddir(struct thread_data *td)
693{
694 enum fio_ddir ddir;
695
696 /*
697 * See if it's time to fsync/fdatasync/sync_file_range first,
698 * and if not then move on to check regular I/Os.
699 */
700 if (should_fsync(td)) {
701 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
702 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
703 return DDIR_SYNC;
704
705 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
706 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
707 return DDIR_DATASYNC;
708
709 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
710 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
711 return DDIR_SYNC_FILE_RANGE;
712 }
713
714 if (td_rw(td)) {
715 /*
716 * Check if it's time to seed a new data direction.
717 */
718 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
719 /*
720 * Put a top limit on how many bytes we do for
721 * one data direction, to avoid overflowing the
722 * ranges too much
723 */
724 ddir = get_rand_ddir(td);
725
726 if (ddir != td->rwmix_ddir)
727 set_rwmix_bytes(td);
728
729 td->rwmix_ddir = ddir;
730 }
731 ddir = td->rwmix_ddir;
732 } else if (td_read(td))
733 ddir = DDIR_READ;
734 else if (td_write(td))
735 ddir = DDIR_WRITE;
736 else if (td_trim(td))
737 ddir = DDIR_TRIM;
738 else
739 ddir = DDIR_INVAL;
740
741 td->rwmix_ddir = rate_ddir(td, ddir);
742 return td->rwmix_ddir;
743}
744
745static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
746{
747 enum fio_ddir ddir = get_rw_ddir(td);
748
749 if (td_trimwrite(td)) {
750 struct fio_file *f = io_u->file;
751 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
752 ddir = DDIR_TRIM;
753 else
754 ddir = DDIR_WRITE;
755 }
756
757 io_u->ddir = io_u->acct_ddir = ddir;
758
759 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
760 td->o.barrier_blocks &&
761 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
762 td->io_issues[DDIR_WRITE])
763 io_u_set(td, io_u, IO_U_F_BARRIER);
764}
765
766void put_file_log(struct thread_data *td, struct fio_file *f)
767{
768 unsigned int ret = put_file(td, f);
769
770 if (ret)
771 td_verror(td, ret, "file close");
772}
773
774void put_io_u(struct thread_data *td, struct io_u *io_u)
775{
776 const bool needs_lock = td_async_processing(td);
777
778 zbd_put_io_u(io_u);
779
780 if (td->parent)
781 td = td->parent;
782
783 if (needs_lock)
784 __td_io_u_lock(td);
785
786 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
787 put_file_log(td, io_u->file);
788
789 io_u->file = NULL;
790 io_u_set(td, io_u, IO_U_F_FREE);
791
792 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
793 td->cur_depth--;
794 assert(!(td->flags & TD_F_CHILD));
795 }
796 io_u_qpush(&td->io_u_freelist, io_u);
797 td_io_u_free_notify(td);
798
799 if (needs_lock)
800 __td_io_u_unlock(td);
801}
802
803void clear_io_u(struct thread_data *td, struct io_u *io_u)
804{
805 io_u_clear(td, io_u, IO_U_F_FLIGHT);
806 put_io_u(td, io_u);
807}
808
809void requeue_io_u(struct thread_data *td, struct io_u **io_u)
810{
811 const bool needs_lock = td_async_processing(td);
812 struct io_u *__io_u = *io_u;
813 enum fio_ddir ddir = acct_ddir(__io_u);
814
815 dprint(FD_IO, "requeue %p\n", __io_u);
816
817 if (td->parent)
818 td = td->parent;
819
820 if (needs_lock)
821 __td_io_u_lock(td);
822
823 io_u_set(td, __io_u, IO_U_F_FREE);
824 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
825 td->io_issues[ddir]--;
826
827 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
828 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
829 td->cur_depth--;
830 assert(!(td->flags & TD_F_CHILD));
831 }
832
833 io_u_rpush(&td->io_u_requeues, __io_u);
834 td_io_u_free_notify(td);
835
836 if (needs_lock)
837 __td_io_u_unlock(td);
838
839 *io_u = NULL;
840}
841
842static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u)
843{
844 struct fio_file *f = io_u->file;
845
846 assert(td->o.zone_mode == ZONE_MODE_STRIDED);
847 assert(td->o.zone_size);
848 assert(td->o.zone_range);
849
850 /*
851 * See if it's time to switch to a new zone
852 */
853 if (td->zone_bytes >= td->o.zone_size) {
854 td->zone_bytes = 0;
855 f->file_offset += td->o.zone_range + td->o.zone_skip;
856
857 /*
858 * Wrap from the beginning, if we exceed the file size
859 */
860 if (f->file_offset >= f->real_file_size)
861 f->file_offset = get_start_offset(td, f);
862
863 f->last_pos[io_u->ddir] = f->file_offset;
864 td->io_skip_bytes += td->o.zone_skip;
865 }
866
867 /*
868 * If zone_size > zone_range, then maintain the same zone until
869 * zone_bytes >= zone_size.
870 */
871 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
872 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
873 f->file_offset, f->last_pos[io_u->ddir]);
874 f->last_pos[io_u->ddir] = f->file_offset;
875 }
876
877 /*
878 * For random: if 'norandommap' is not set and zone_size > zone_range,
879 * map needs to be reset as it's done with zone_range everytime.
880 */
881 if ((td->zone_bytes % td->o.zone_range) == 0)
882 fio_file_reset(td, f);
883}
884
885static int fill_io_u(struct thread_data *td, struct io_u *io_u)
886{
887 bool is_random;
888 uint64_t offset;
889 enum io_u_action ret;
890
891 if (td_ioengine_flagged(td, FIO_NOIO))
892 goto out;
893
894 set_rw_ddir(td, io_u);
895
896 /*
897 * fsync() or fdatasync() or trim etc, we are done
898 */
899 if (!ddir_rw(io_u->ddir))
900 goto out;
901
902 if (td->o.zone_mode == ZONE_MODE_STRIDED)
903 setup_strided_zone_mode(td, io_u);
904 else if (td->o.zone_mode == ZONE_MODE_ZBD)
905 setup_zbd_zone_mode(td, io_u);
906
907 /*
908 * No log, let the seq/rand engine retrieve the next buflen and
909 * position.
910 */
911 if (get_next_offset(td, io_u, &is_random)) {
912 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
913 return 1;
914 }
915
916 io_u->buflen = get_next_buflen(td, io_u, is_random);
917 if (!io_u->buflen) {
918 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
919 return 1;
920 }
921
922 offset = io_u->offset;
923 if (td->o.zone_mode == ZONE_MODE_ZBD) {
924 ret = zbd_adjust_block(td, io_u);
925 if (ret == io_u_eof)
926 return 1;
927 }
928
929 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
930 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n",
931 io_u,
932 (unsigned long long) io_u->offset, io_u->buflen,
933 (unsigned long long) io_u->file->real_file_size);
934 return 1;
935 }
936
937 /*
938 * mark entry before potentially trimming io_u
939 */
940 if (td_random(td) && file_randommap(td, io_u->file))
941 io_u->buflen = mark_random_map(td, io_u, offset, io_u->buflen);
942
943out:
944 dprint_io_u(io_u, "fill");
945 td->zone_bytes += io_u->buflen;
946 return 0;
947}
948
949static void __io_u_mark_map(uint64_t *map, unsigned int nr)
950{
951 int idx = 0;
952
953 switch (nr) {
954 default:
955 idx = 6;
956 break;
957 case 33 ... 64:
958 idx = 5;
959 break;
960 case 17 ... 32:
961 idx = 4;
962 break;
963 case 9 ... 16:
964 idx = 3;
965 break;
966 case 5 ... 8:
967 idx = 2;
968 break;
969 case 1 ... 4:
970 idx = 1;
971 case 0:
972 break;
973 }
974
975 map[idx]++;
976}
977
978void io_u_mark_submit(struct thread_data *td, unsigned int nr)
979{
980 __io_u_mark_map(td->ts.io_u_submit, nr);
981 td->ts.total_submit++;
982}
983
984void io_u_mark_complete(struct thread_data *td, unsigned int nr)
985{
986 __io_u_mark_map(td->ts.io_u_complete, nr);
987 td->ts.total_complete++;
988}
989
990void io_u_mark_depth(struct thread_data *td, unsigned int nr)
991{
992 int idx = 0;
993
994 switch (td->cur_depth) {
995 default:
996 idx = 6;
997 break;
998 case 32 ... 63:
999 idx = 5;
1000 break;
1001 case 16 ... 31:
1002 idx = 4;
1003 break;
1004 case 8 ... 15:
1005 idx = 3;
1006 break;
1007 case 4 ... 7:
1008 idx = 2;
1009 break;
1010 case 2 ... 3:
1011 idx = 1;
1012 case 1:
1013 break;
1014 }
1015
1016 td->ts.io_u_map[idx] += nr;
1017}
1018
1019static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1020{
1021 int idx = 0;
1022
1023 assert(nsec < 1000);
1024
1025 switch (nsec) {
1026 case 750 ... 999:
1027 idx = 9;
1028 break;
1029 case 500 ... 749:
1030 idx = 8;
1031 break;
1032 case 250 ... 499:
1033 idx = 7;
1034 break;
1035 case 100 ... 249:
1036 idx = 6;
1037 break;
1038 case 50 ... 99:
1039 idx = 5;
1040 break;
1041 case 20 ... 49:
1042 idx = 4;
1043 break;
1044 case 10 ... 19:
1045 idx = 3;
1046 break;
1047 case 4 ... 9:
1048 idx = 2;
1049 break;
1050 case 2 ... 3:
1051 idx = 1;
1052 case 0 ... 1:
1053 break;
1054 }
1055
1056 assert(idx < FIO_IO_U_LAT_N_NR);
1057 td->ts.io_u_lat_n[idx]++;
1058}
1059
1060static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1061{
1062 int idx = 0;
1063
1064 assert(usec < 1000 && usec >= 1);
1065
1066 switch (usec) {
1067 case 750 ... 999:
1068 idx = 9;
1069 break;
1070 case 500 ... 749:
1071 idx = 8;
1072 break;
1073 case 250 ... 499:
1074 idx = 7;
1075 break;
1076 case 100 ... 249:
1077 idx = 6;
1078 break;
1079 case 50 ... 99:
1080 idx = 5;
1081 break;
1082 case 20 ... 49:
1083 idx = 4;
1084 break;
1085 case 10 ... 19:
1086 idx = 3;
1087 break;
1088 case 4 ... 9:
1089 idx = 2;
1090 break;
1091 case 2 ... 3:
1092 idx = 1;
1093 case 0 ... 1:
1094 break;
1095 }
1096
1097 assert(idx < FIO_IO_U_LAT_U_NR);
1098 td->ts.io_u_lat_u[idx]++;
1099}
1100
1101static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1102{
1103 int idx = 0;
1104
1105 assert(msec >= 1);
1106
1107 switch (msec) {
1108 default:
1109 idx = 11;
1110 break;
1111 case 1000 ... 1999:
1112 idx = 10;
1113 break;
1114 case 750 ... 999:
1115 idx = 9;
1116 break;
1117 case 500 ... 749:
1118 idx = 8;
1119 break;
1120 case 250 ... 499:
1121 idx = 7;
1122 break;
1123 case 100 ... 249:
1124 idx = 6;
1125 break;
1126 case 50 ... 99:
1127 idx = 5;
1128 break;
1129 case 20 ... 49:
1130 idx = 4;
1131 break;
1132 case 10 ... 19:
1133 idx = 3;
1134 break;
1135 case 4 ... 9:
1136 idx = 2;
1137 break;
1138 case 2 ... 3:
1139 idx = 1;
1140 case 0 ... 1:
1141 break;
1142 }
1143
1144 assert(idx < FIO_IO_U_LAT_M_NR);
1145 td->ts.io_u_lat_m[idx]++;
1146}
1147
1148static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1149{
1150 if (nsec < 1000)
1151 io_u_mark_lat_nsec(td, nsec);
1152 else if (nsec < 1000000)
1153 io_u_mark_lat_usec(td, nsec / 1000);
1154 else
1155 io_u_mark_lat_msec(td, nsec / 1000000);
1156}
1157
1158static unsigned int __get_next_fileno_rand(struct thread_data *td)
1159{
1160 unsigned long fileno;
1161
1162 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1163 uint64_t frand_max = rand_max(&td->next_file_state);
1164 unsigned long r;
1165
1166 r = __rand(&td->next_file_state);
1167 return (unsigned int) ((double) td->o.nr_files
1168 * (r / (frand_max + 1.0)));
1169 }
1170
1171 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1172 fileno = zipf_next(&td->next_file_zipf);
1173 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1174 fileno = pareto_next(&td->next_file_zipf);
1175 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1176 fileno = gauss_next(&td->next_file_gauss);
1177 else {
1178 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1179 assert(0);
1180 return 0;
1181 }
1182
1183 return fileno >> FIO_FSERVICE_SHIFT;
1184}
1185
1186/*
1187 * Get next file to service by choosing one at random
1188 */
1189static struct fio_file *get_next_file_rand(struct thread_data *td,
1190 enum fio_file_flags goodf,
1191 enum fio_file_flags badf)
1192{
1193 struct fio_file *f;
1194 int fno;
1195
1196 do {
1197 int opened = 0;
1198
1199 fno = __get_next_fileno_rand(td);
1200
1201 f = td->files[fno];
1202 if (fio_file_done(f))
1203 continue;
1204
1205 if (!fio_file_open(f)) {
1206 int err;
1207
1208 if (td->nr_open_files >= td->o.open_files)
1209 return ERR_PTR(-EBUSY);
1210
1211 err = td_io_open_file(td, f);
1212 if (err)
1213 continue;
1214 opened = 1;
1215 }
1216
1217 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1218 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1219 return f;
1220 }
1221 if (opened)
1222 td_io_close_file(td, f);
1223 } while (1);
1224}
1225
1226/*
1227 * Get next file to service by doing round robin between all available ones
1228 */
1229static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1230 int badf)
1231{
1232 unsigned int old_next_file = td->next_file;
1233 struct fio_file *f;
1234
1235 do {
1236 int opened = 0;
1237
1238 f = td->files[td->next_file];
1239
1240 td->next_file++;
1241 if (td->next_file >= td->o.nr_files)
1242 td->next_file = 0;
1243
1244 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1245 if (fio_file_done(f)) {
1246 f = NULL;
1247 continue;
1248 }
1249
1250 if (!fio_file_open(f)) {
1251 int err;
1252
1253 if (td->nr_open_files >= td->o.open_files)
1254 return ERR_PTR(-EBUSY);
1255
1256 err = td_io_open_file(td, f);
1257 if (err) {
1258 dprint(FD_FILE, "error %d on open of %s\n",
1259 err, f->file_name);
1260 f = NULL;
1261 continue;
1262 }
1263 opened = 1;
1264 }
1265
1266 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1267 f->flags);
1268 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1269 break;
1270
1271 if (opened)
1272 td_io_close_file(td, f);
1273
1274 f = NULL;
1275 } while (td->next_file != old_next_file);
1276
1277 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1278 return f;
1279}
1280
1281static struct fio_file *__get_next_file(struct thread_data *td)
1282{
1283 struct fio_file *f;
1284
1285 assert(td->o.nr_files <= td->files_index);
1286
1287 if (td->nr_done_files >= td->o.nr_files) {
1288 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1289 " nr_files=%d\n", td->nr_open_files,
1290 td->nr_done_files,
1291 td->o.nr_files);
1292 return NULL;
1293 }
1294
1295 f = td->file_service_file;
1296 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1297 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1298 goto out;
1299 if (td->file_service_left--)
1300 goto out;
1301 }
1302
1303 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1304 td->o.file_service_type == FIO_FSERVICE_SEQ)
1305 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1306 else
1307 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1308
1309 if (IS_ERR(f))
1310 return f;
1311
1312 td->file_service_file = f;
1313 td->file_service_left = td->file_service_nr - 1;
1314out:
1315 if (f)
1316 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1317 else
1318 dprint(FD_FILE, "get_next_file: NULL\n");
1319 return f;
1320}
1321
1322static struct fio_file *get_next_file(struct thread_data *td)
1323{
1324 return __get_next_file(td);
1325}
1326
1327static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1328{
1329 struct fio_file *f;
1330
1331 do {
1332 f = get_next_file(td);
1333 if (IS_ERR_OR_NULL(f))
1334 return PTR_ERR(f);
1335
1336 io_u->file = f;
1337 get_file(f);
1338
1339 if (!fill_io_u(td, io_u))
1340 break;
1341
1342 zbd_put_io_u(io_u);
1343
1344 put_file_log(td, f);
1345 td_io_close_file(td, f);
1346 io_u->file = NULL;
1347 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1348 fio_file_reset(td, f);
1349 else {
1350 fio_file_set_done(f);
1351 td->nr_done_files++;
1352 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1353 td->nr_done_files, td->o.nr_files);
1354 }
1355 } while (1);
1356
1357 return 0;
1358}
1359
1360static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1361 unsigned long long tnsec, unsigned long long max_nsec)
1362{
1363 if (!td->error)
1364 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
1365 td_verror(td, ETIMEDOUT, "max latency exceeded");
1366 icd->error = ETIMEDOUT;
1367}
1368
1369static void lat_new_cycle(struct thread_data *td)
1370{
1371 fio_gettime(&td->latency_ts, NULL);
1372 td->latency_ios = ddir_rw_sum(td->io_blocks);
1373 td->latency_failed = 0;
1374}
1375
1376/*
1377 * We had an IO outside the latency target. Reduce the queue depth. If we
1378 * are at QD=1, then it's time to give up.
1379 */
1380static bool __lat_target_failed(struct thread_data *td)
1381{
1382 if (td->latency_qd == 1)
1383 return true;
1384
1385 td->latency_qd_high = td->latency_qd;
1386
1387 if (td->latency_qd == td->latency_qd_low)
1388 td->latency_qd_low--;
1389
1390 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1391
1392 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1393
1394 /*
1395 * When we ramp QD down, quiesce existing IO to prevent
1396 * a storm of ramp downs due to pending higher depth.
1397 */
1398 io_u_quiesce(td);
1399 lat_new_cycle(td);
1400 return false;
1401}
1402
1403static bool lat_target_failed(struct thread_data *td)
1404{
1405 if (td->o.latency_percentile.u.f == 100.0)
1406 return __lat_target_failed(td);
1407
1408 td->latency_failed++;
1409 return false;
1410}
1411
1412void lat_target_init(struct thread_data *td)
1413{
1414 td->latency_end_run = 0;
1415
1416 if (td->o.latency_target) {
1417 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1418 fio_gettime(&td->latency_ts, NULL);
1419 td->latency_qd = 1;
1420 td->latency_qd_high = td->o.iodepth;
1421 td->latency_qd_low = 1;
1422 td->latency_ios = ddir_rw_sum(td->io_blocks);
1423 } else
1424 td->latency_qd = td->o.iodepth;
1425}
1426
1427void lat_target_reset(struct thread_data *td)
1428{
1429 if (!td->latency_end_run)
1430 lat_target_init(td);
1431}
1432
1433static void lat_target_success(struct thread_data *td)
1434{
1435 const unsigned int qd = td->latency_qd;
1436 struct thread_options *o = &td->o;
1437
1438 td->latency_qd_low = td->latency_qd;
1439
1440 /*
1441 * If we haven't failed yet, we double up to a failing value instead
1442 * of bisecting from highest possible queue depth. If we have set
1443 * a limit other than td->o.iodepth, bisect between that.
1444 */
1445 if (td->latency_qd_high != o->iodepth)
1446 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1447 else
1448 td->latency_qd *= 2;
1449
1450 if (td->latency_qd > o->iodepth)
1451 td->latency_qd = o->iodepth;
1452
1453 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1454
1455 /*
1456 * Same as last one, we are done. Let it run a latency cycle, so
1457 * we get only the results from the targeted depth.
1458 */
1459 if (td->latency_qd == qd) {
1460 if (td->latency_end_run) {
1461 dprint(FD_RATE, "We are done\n");
1462 td->done = 1;
1463 } else {
1464 dprint(FD_RATE, "Quiesce and final run\n");
1465 io_u_quiesce(td);
1466 td->latency_end_run = 1;
1467 reset_all_stats(td);
1468 reset_io_stats(td);
1469 }
1470 }
1471
1472 lat_new_cycle(td);
1473}
1474
1475/*
1476 * Check if we can bump the queue depth
1477 */
1478void lat_target_check(struct thread_data *td)
1479{
1480 uint64_t usec_window;
1481 uint64_t ios;
1482 double success_ios;
1483
1484 usec_window = utime_since_now(&td->latency_ts);
1485 if (usec_window < td->o.latency_window)
1486 return;
1487
1488 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1489 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1490 success_ios *= 100.0;
1491
1492 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1493
1494 if (success_ios >= td->o.latency_percentile.u.f)
1495 lat_target_success(td);
1496 else
1497 __lat_target_failed(td);
1498}
1499
1500/*
1501 * If latency target is enabled, we might be ramping up or down and not
1502 * using the full queue depth available.
1503 */
1504bool queue_full(const struct thread_data *td)
1505{
1506 const int qempty = io_u_qempty(&td->io_u_freelist);
1507
1508 if (qempty)
1509 return true;
1510 if (!td->o.latency_target)
1511 return false;
1512
1513 return td->cur_depth >= td->latency_qd;
1514}
1515
1516struct io_u *__get_io_u(struct thread_data *td)
1517{
1518 const bool needs_lock = td_async_processing(td);
1519 struct io_u *io_u = NULL;
1520 int ret;
1521
1522 if (td->stop_io)
1523 return NULL;
1524
1525 if (needs_lock)
1526 __td_io_u_lock(td);
1527
1528again:
1529 if (!io_u_rempty(&td->io_u_requeues))
1530 io_u = io_u_rpop(&td->io_u_requeues);
1531 else if (!queue_full(td)) {
1532 io_u = io_u_qpop(&td->io_u_freelist);
1533
1534 io_u->file = NULL;
1535 io_u->buflen = 0;
1536 io_u->resid = 0;
1537 io_u->end_io = NULL;
1538 }
1539
1540 if (io_u) {
1541 assert(io_u->flags & IO_U_F_FREE);
1542 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1543 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1544 IO_U_F_VER_LIST);
1545
1546 io_u->error = 0;
1547 io_u->acct_ddir = -1;
1548 td->cur_depth++;
1549 assert(!(td->flags & TD_F_CHILD));
1550 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1551 io_u->ipo = NULL;
1552 } else if (td_async_processing(td)) {
1553 /*
1554 * We ran out, wait for async verify threads to finish and
1555 * return one
1556 */
1557 assert(!(td->flags & TD_F_CHILD));
1558 ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1559 assert(ret == 0);
1560 if (!td->error)
1561 goto again;
1562 }
1563
1564 if (needs_lock)
1565 __td_io_u_unlock(td);
1566
1567 return io_u;
1568}
1569
1570static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1571{
1572 if (!(td->flags & TD_F_TRIM_BACKLOG))
1573 return false;
1574 if (!td->trim_entries)
1575 return false;
1576
1577 if (td->trim_batch) {
1578 td->trim_batch--;
1579 if (get_next_trim(td, io_u))
1580 return true;
1581 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1582 td->last_ddir != DDIR_READ) {
1583 td->trim_batch = td->o.trim_batch;
1584 if (!td->trim_batch)
1585 td->trim_batch = td->o.trim_backlog;
1586 if (get_next_trim(td, io_u))
1587 return true;
1588 }
1589
1590 return false;
1591}
1592
1593static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1594{
1595 if (!(td->flags & TD_F_VER_BACKLOG))
1596 return false;
1597
1598 if (td->io_hist_len) {
1599 int get_verify = 0;
1600
1601 if (td->verify_batch)
1602 get_verify = 1;
1603 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1604 td->last_ddir != DDIR_READ) {
1605 td->verify_batch = td->o.verify_batch;
1606 if (!td->verify_batch)
1607 td->verify_batch = td->o.verify_backlog;
1608 get_verify = 1;
1609 }
1610
1611 if (get_verify && !get_next_verify(td, io_u)) {
1612 td->verify_batch--;
1613 return true;
1614 }
1615 }
1616
1617 return false;
1618}
1619
1620/*
1621 * Fill offset and start time into the buffer content, to prevent too
1622 * easy compressible data for simple de-dupe attempts. Do this for every
1623 * 512b block in the range, since that should be the smallest block size
1624 * we can expect from a device.
1625 */
1626static void small_content_scramble(struct io_u *io_u)
1627{
1628 unsigned long long i, nr_blocks = io_u->buflen >> 9;
1629 unsigned int offset;
1630 uint64_t boffset, *iptr;
1631 char *p;
1632
1633 if (!nr_blocks)
1634 return;
1635
1636 p = io_u->xfer_buf;
1637 boffset = io_u->offset;
1638
1639 if (io_u->buf_filled_len)
1640 io_u->buf_filled_len = 0;
1641
1642 /*
1643 * Generate random index between 0..7. We do chunks of 512b, if
1644 * we assume a cacheline is 64 bytes, then we have 8 of those.
1645 * Scramble content within the blocks in the same cacheline to
1646 * speed things up.
1647 */
1648 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1649
1650 for (i = 0; i < nr_blocks; i++) {
1651 /*
1652 * Fill offset into start of cacheline, time into end
1653 * of cacheline
1654 */
1655 iptr = (void *) p + (offset << 6);
1656 *iptr = boffset;
1657
1658 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1659 iptr[0] = io_u->start_time.tv_sec;
1660 iptr[1] = io_u->start_time.tv_nsec;
1661
1662 p += 512;
1663 boffset += 512;
1664 }
1665}
1666
1667/*
1668 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1669 * etc. The returned io_u is fully ready to be prepped, populated and submitted.
1670 */
1671struct io_u *get_io_u(struct thread_data *td)
1672{
1673 struct fio_file *f;
1674 struct io_u *io_u;
1675 int do_scramble = 0;
1676 long ret = 0;
1677
1678 io_u = __get_io_u(td);
1679 if (!io_u) {
1680 dprint(FD_IO, "__get_io_u failed\n");
1681 return NULL;
1682 }
1683
1684 if (check_get_verify(td, io_u))
1685 goto out;
1686 if (check_get_trim(td, io_u))
1687 goto out;
1688
1689 /*
1690 * from a requeue, io_u already setup
1691 */
1692 if (io_u->file)
1693 goto out;
1694
1695 /*
1696 * If using an iolog, grab next piece if any available.
1697 */
1698 if (td->flags & TD_F_READ_IOLOG) {
1699 if (read_iolog_get(td, io_u))
1700 goto err_put;
1701 } else if (set_io_u_file(td, io_u)) {
1702 ret = -EBUSY;
1703 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1704 goto err_put;
1705 }
1706
1707 f = io_u->file;
1708 if (!f) {
1709 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1710 goto err_put;
1711 }
1712
1713 assert(fio_file_open(f));
1714
1715 if (ddir_rw(io_u->ddir)) {
1716 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1717 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1718 goto err_put;
1719 }
1720
1721 f->last_start[io_u->ddir] = io_u->offset;
1722 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1723
1724 if (io_u->ddir == DDIR_WRITE) {
1725 if (td->flags & TD_F_REFILL_BUFFERS) {
1726 io_u_fill_buffer(td, io_u,
1727 td->o.min_bs[DDIR_WRITE],
1728 io_u->buflen);
1729 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1730 !(td->flags & TD_F_COMPRESS) &&
1731 !(td->flags & TD_F_DO_VERIFY))
1732 do_scramble = 1;
1733 } else if (io_u->ddir == DDIR_READ) {
1734 /*
1735 * Reset the buf_filled parameters so next time if the
1736 * buffer is used for writes it is refilled.
1737 */
1738 io_u->buf_filled_len = 0;
1739 }
1740 }
1741
1742 /*
1743 * Set io data pointers.
1744 */
1745 io_u->xfer_buf = io_u->buf;
1746 io_u->xfer_buflen = io_u->buflen;
1747
1748out:
1749 assert(io_u->file);
1750 if (!td_io_prep(td, io_u)) {
1751 if (!td->o.disable_lat)
1752 fio_gettime(&io_u->start_time, NULL);
1753
1754 if (do_scramble)
1755 small_content_scramble(io_u);
1756
1757 return io_u;
1758 }
1759err_put:
1760 dprint(FD_IO, "get_io_u failed\n");
1761 put_io_u(td, io_u);
1762 return ERR_PTR(ret);
1763}
1764
1765static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1766{
1767 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1768
1769 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1770 return;
1771
1772 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n",
1773 io_u->file ? " on file " : "",
1774 io_u->file ? io_u->file->file_name : "",
1775 strerror(io_u->error),
1776 io_ddir_name(io_u->ddir),
1777 io_u->offset, io_u->xfer_buflen);
1778
1779 if (td->io_ops->errdetails) {
1780 char *err = td->io_ops->errdetails(io_u);
1781
1782 log_err("fio: %s\n", err);
1783 free(err);
1784 }
1785
1786 if (!td->error)
1787 td_verror(td, io_u->error, "io_u error");
1788}
1789
1790void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1791{
1792 __io_u_log_error(td, io_u);
1793 if (td->parent)
1794 __io_u_log_error(td->parent, io_u);
1795}
1796
1797static inline bool gtod_reduce(struct thread_data *td)
1798{
1799 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1800 || td->o.gtod_reduce;
1801}
1802
1803static void trim_block_info(struct thread_data *td, struct io_u *io_u)
1804{
1805 uint32_t *info = io_u_block_info(td, io_u);
1806
1807 if (BLOCK_INFO_STATE(*info) >= BLOCK_STATE_TRIM_FAILURE)
1808 return;
1809
1810 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, BLOCK_INFO_TRIMS(*info) + 1);
1811}
1812
1813static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1814 struct io_completion_data *icd,
1815 const enum fio_ddir idx, unsigned int bytes)
1816{
1817 const int no_reduce = !gtod_reduce(td);
1818 unsigned long long llnsec = 0;
1819
1820 if (td->parent)
1821 td = td->parent;
1822
1823 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1824 return;
1825
1826 if (no_reduce)
1827 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1828
1829 if (!td->o.disable_lat) {
1830 unsigned long long tnsec;
1831
1832 tnsec = ntime_since(&io_u->start_time, &icd->time);
1833 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
1834
1835 if (td->flags & TD_F_PROFILE_OPS) {
1836 struct prof_io_ops *ops = &td->prof_io_ops;
1837
1838 if (ops->io_u_lat)
1839 icd->error = ops->io_u_lat(td, tnsec);
1840 }
1841
1842 if (td->o.max_latency && tnsec > td->o.max_latency)
1843 lat_fatal(td, icd, tnsec, td->o.max_latency);
1844 if (td->o.latency_target && tnsec > td->o.latency_target) {
1845 if (lat_target_failed(td))
1846 lat_fatal(td, icd, tnsec, td->o.latency_target);
1847 }
1848 }
1849
1850 if (ddir_rw(idx)) {
1851 if (!td->o.disable_clat) {
1852 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1853 io_u_mark_latency(td, llnsec);
1854 }
1855
1856 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1857 add_bw_sample(td, io_u, bytes, llnsec);
1858
1859 if (no_reduce && per_unit_log(td->iops_log))
1860 add_iops_sample(td, io_u, bytes);
1861 } else if (ddir_sync(idx) && !td->o.disable_clat)
1862 add_sync_clat_sample(&td->ts, llnsec);
1863
1864 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM)
1865 trim_block_info(td, io_u);
1866}
1867
1868static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1869 uint64_t offset, unsigned int bytes)
1870{
1871 int idx;
1872
1873 if (!f)
1874 return;
1875
1876 if (f->first_write == -1ULL || offset < f->first_write)
1877 f->first_write = offset;
1878 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1879 f->last_write = offset + bytes;
1880
1881 if (!f->last_write_comp)
1882 return;
1883
1884 idx = f->last_write_idx++;
1885 f->last_write_comp[idx] = offset;
1886 if (f->last_write_idx == td->o.iodepth)
1887 f->last_write_idx = 0;
1888}
1889
1890static bool should_account(struct thread_data *td)
1891{
1892 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1893 td->runstate == TD_VERIFYING);
1894}
1895
1896static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1897 struct io_completion_data *icd)
1898{
1899 struct io_u *io_u = *io_u_ptr;
1900 enum fio_ddir ddir = io_u->ddir;
1901 struct fio_file *f = io_u->file;
1902
1903 dprint_io_u(io_u, "complete");
1904
1905 assert(io_u->flags & IO_U_F_FLIGHT);
1906 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1907
1908 /*
1909 * Mark IO ok to verify
1910 */
1911 if (io_u->ipo) {
1912 /*
1913 * Remove errored entry from the verification list
1914 */
1915 if (io_u->error)
1916 unlog_io_piece(td, io_u);
1917 else {
1918 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1919 write_barrier();
1920 }
1921 }
1922
1923 if (ddir_sync(ddir)) {
1924 td->last_was_sync = true;
1925 if (f) {
1926 f->first_write = -1ULL;
1927 f->last_write = -1ULL;
1928 }
1929 if (should_account(td))
1930 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
1931 return;
1932 }
1933
1934 td->last_was_sync = false;
1935 td->last_ddir = ddir;
1936
1937 if (!io_u->error && ddir_rw(ddir)) {
1938 unsigned long long bytes = io_u->buflen - io_u->resid;
1939 int ret;
1940
1941 td->io_blocks[ddir]++;
1942 td->io_bytes[ddir] += bytes;
1943
1944 if (!(io_u->flags & IO_U_F_VER_LIST)) {
1945 td->this_io_blocks[ddir]++;
1946 td->this_io_bytes[ddir] += bytes;
1947 }
1948
1949 if (ddir == DDIR_WRITE)
1950 file_log_write_comp(td, f, io_u->offset, bytes);
1951
1952 if (should_account(td))
1953 account_io_completion(td, io_u, icd, ddir, bytes);
1954
1955 icd->bytes_done[ddir] += bytes;
1956
1957 if (io_u->end_io) {
1958 ret = io_u->end_io(td, io_u_ptr);
1959 io_u = *io_u_ptr;
1960 if (ret && !icd->error)
1961 icd->error = ret;
1962 }
1963 } else if (io_u->error) {
1964 icd->error = io_u->error;
1965 io_u_log_error(td, io_u);
1966 }
1967 if (icd->error) {
1968 enum error_type_bit eb = td_error_type(ddir, icd->error);
1969
1970 if (!td_non_fatal_error(td, eb, icd->error))
1971 return;
1972
1973 /*
1974 * If there is a non_fatal error, then add to the error count
1975 * and clear all the errors.
1976 */
1977 update_error_count(td, icd->error);
1978 td_clear_error(td);
1979 icd->error = 0;
1980 if (io_u)
1981 io_u->error = 0;
1982 }
1983}
1984
1985static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1986 int nr)
1987{
1988 int ddir;
1989
1990 if (!gtod_reduce(td))
1991 fio_gettime(&icd->time, NULL);
1992
1993 icd->nr = nr;
1994
1995 icd->error = 0;
1996 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1997 icd->bytes_done[ddir] = 0;
1998}
1999
2000static void ios_completed(struct thread_data *td,
2001 struct io_completion_data *icd)
2002{
2003 struct io_u *io_u;
2004 int i;
2005
2006 for (i = 0; i < icd->nr; i++) {
2007 io_u = td->io_ops->event(td, i);
2008
2009 io_completed(td, &io_u, icd);
2010
2011 if (io_u)
2012 put_io_u(td, io_u);
2013 }
2014}
2015
2016/*
2017 * Complete a single io_u for the sync engines.
2018 */
2019int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
2020{
2021 struct io_completion_data icd;
2022 int ddir;
2023
2024 init_icd(td, &icd, 1);
2025 io_completed(td, &io_u, &icd);
2026
2027 if (io_u)
2028 put_io_u(td, io_u);
2029
2030 if (icd.error) {
2031 td_verror(td, icd.error, "io_u_sync_complete");
2032 return -1;
2033 }
2034
2035 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2036 td->bytes_done[ddir] += icd.bytes_done[ddir];
2037
2038 return 0;
2039}
2040
2041/*
2042 * Called to complete min_events number of io for the async engines.
2043 */
2044int io_u_queued_complete(struct thread_data *td, int min_evts)
2045{
2046 struct io_completion_data icd;
2047 struct timespec *tvp = NULL;
2048 int ret, ddir;
2049 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2050
2051 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2052
2053 if (!min_evts)
2054 tvp = &ts;
2055 else if (min_evts > td->cur_depth)
2056 min_evts = td->cur_depth;
2057
2058 /* No worries, td_io_getevents fixes min and max if they are
2059 * set incorrectly */
2060 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2061 if (ret < 0) {
2062 td_verror(td, -ret, "td_io_getevents");
2063 return ret;
2064 } else if (!ret)
2065 return ret;
2066
2067 init_icd(td, &icd, ret);
2068 ios_completed(td, &icd);
2069 if (icd.error) {
2070 td_verror(td, icd.error, "io_u_queued_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 ret;
2078}
2079
2080/*
2081 * Call when io_u is really queued, to update the submission latency.
2082 */
2083void io_u_queued(struct thread_data *td, struct io_u *io_u)
2084{
2085 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2086 unsigned long slat_time;
2087
2088 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2089
2090 if (td->parent)
2091 td = td->parent;
2092
2093 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2094 io_u->offset);
2095 }
2096}
2097
2098/*
2099 * See if we should reuse the last seed, if dedupe is enabled
2100 */
2101static struct frand_state *get_buf_state(struct thread_data *td)
2102{
2103 unsigned int v;
2104
2105 if (!td->o.dedupe_percentage)
2106 return &td->buf_state;
2107 else if (td->o.dedupe_percentage == 100) {
2108 frand_copy(&td->buf_state_prev, &td->buf_state);
2109 return &td->buf_state;
2110 }
2111
2112 v = rand_between(&td->dedupe_state, 1, 100);
2113
2114 if (v <= td->o.dedupe_percentage)
2115 return &td->buf_state_prev;
2116
2117 return &td->buf_state;
2118}
2119
2120static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2121{
2122 if (td->o.dedupe_percentage == 100)
2123 frand_copy(rs, &td->buf_state_prev);
2124 else if (rs == &td->buf_state)
2125 frand_copy(&td->buf_state_prev, rs);
2126}
2127
2128void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write,
2129 unsigned long long max_bs)
2130{
2131 struct thread_options *o = &td->o;
2132
2133 if (o->mem_type == MEM_CUDA_MALLOC)
2134 return;
2135
2136 if (o->compress_percentage || o->dedupe_percentage) {
2137 unsigned int perc = td->o.compress_percentage;
2138 struct frand_state *rs;
2139 unsigned long long left = max_bs;
2140 unsigned long long this_write;
2141
2142 do {
2143 rs = get_buf_state(td);
2144
2145 min_write = min(min_write, left);
2146
2147 if (perc) {
2148 this_write = min_not_zero(min_write,
2149 (unsigned long long) td->o.compress_chunk);
2150
2151 fill_random_buf_percentage(rs, buf, perc,
2152 this_write, this_write,
2153 o->buffer_pattern,
2154 o->buffer_pattern_bytes);
2155 } else {
2156 fill_random_buf(rs, buf, min_write);
2157 this_write = min_write;
2158 }
2159
2160 buf += this_write;
2161 left -= this_write;
2162 save_buf_state(td, rs);
2163 } while (left);
2164 } else if (o->buffer_pattern_bytes)
2165 fill_buffer_pattern(td, buf, max_bs);
2166 else if (o->zero_buffers)
2167 memset(buf, 0, max_bs);
2168 else
2169 fill_random_buf(get_buf_state(td), buf, max_bs);
2170}
2171
2172/*
2173 * "randomly" fill the buffer contents
2174 */
2175void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2176 unsigned long long min_write, unsigned long long max_bs)
2177{
2178 io_u->buf_filled_len = 0;
2179 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2180}
2181
2182static int do_sync_file_range(const struct thread_data *td,
2183 struct fio_file *f)
2184{
2185 off64_t offset, nbytes;
2186
2187 offset = f->first_write;
2188 nbytes = f->last_write - f->first_write;
2189
2190 if (!nbytes)
2191 return 0;
2192
2193 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2194}
2195
2196int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2197{
2198 int ret;
2199
2200 if (io_u->ddir == DDIR_SYNC) {
2201 ret = fsync(io_u->file->fd);
2202 } else if (io_u->ddir == DDIR_DATASYNC) {
2203#ifdef CONFIG_FDATASYNC
2204 ret = fdatasync(io_u->file->fd);
2205#else
2206 ret = io_u->xfer_buflen;
2207 io_u->error = EINVAL;
2208#endif
2209 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2210 ret = do_sync_file_range(td, io_u->file);
2211 else {
2212 ret = io_u->xfer_buflen;
2213 io_u->error = EINVAL;
2214 }
2215
2216 if (ret < 0)
2217 io_u->error = errno;
2218
2219 return ret;
2220}
2221
2222int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2223{
2224#ifndef FIO_HAVE_TRIM
2225 io_u->error = EINVAL;
2226 return 0;
2227#else
2228 struct fio_file *f = io_u->file;
2229 int ret;
2230
2231 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2232 if (!ret)
2233 return io_u->xfer_buflen;
2234
2235 io_u->error = ret;
2236 return 0;
2237#endif
2238}
fio - Flexible IO Tester