projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Merge branch 'atomics' of https://github.com/bvanassche/fio
[fio.git] / io_u.c
... / ...
CommitLineData
1#include <unistd.h>
2#include <string.h>
3#include <assert.h>
4
5#include "fio.h"
6#include "verify.h"
7#include "trim.h"
8#include "lib/rand.h"
9#include "lib/axmap.h"
10#include "err.h"
11#include "lib/pow2.h"
12#include "minmax.h"
13#include "zbd.h"
14
15struct io_completion_data {
16 int nr; /* input */
17
18 int error; /* output */
19 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
20 struct timespec time; /* output */
21};
22
23/*
24 * The ->io_axmap contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
26 */
27static bool random_map_free(struct fio_file *f, const uint64_t block)
28{
29 return !axmap_isset(f->io_axmap, block);
30}
31
32/*
33 * Mark a given offset as used in the map.
34 */
35static uint64_t mark_random_map(struct thread_data *td, struct io_u *io_u,
36 uint64_t offset, uint64_t buflen)
37{
38 unsigned long long min_bs = td->o.min_bs[io_u->ddir];
39 struct fio_file *f = io_u->file;
40 unsigned long long nr_blocks;
41 uint64_t block;
42
43 block = (offset - f->file_offset) / (uint64_t) min_bs;
44 nr_blocks = (buflen + min_bs - 1) / min_bs;
45 assert(nr_blocks > 0);
46
47 if (!(io_u->flags & IO_U_F_BUSY_OK)) {
48 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
49 assert(nr_blocks > 0);
50 }
51
52 if ((nr_blocks * min_bs) < buflen)
53 buflen = nr_blocks * min_bs;
54
55 return buflen;
56}
57
58static uint64_t last_block(struct thread_data *td, struct fio_file *f,
59 enum fio_ddir ddir)
60{
61 uint64_t max_blocks;
62 uint64_t max_size;
63
64 assert(ddir_rw(ddir));
65
66 /*
67 * Hmm, should we make sure that ->io_size <= ->real_file_size?
68 * -> not for now since there is code assuming it could go either.
69 */
70 max_size = f->io_size;
71 if (max_size > f->real_file_size)
72 max_size = f->real_file_size;
73
74 if (td->o.zone_mode == ZONE_MODE_STRIDED && td->o.zone_range)
75 max_size = td->o.zone_range;
76
77 if (td->o.min_bs[ddir] > td->o.ba[ddir])
78 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
79
80 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
81 if (!max_blocks)
82 return 0;
83
84 return max_blocks;
85}
86
87static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
88 enum fio_ddir ddir, uint64_t *b,
89 uint64_t lastb)
90{
91 uint64_t r;
92
93 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
94 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
95
96 r = __rand(&td->random_state);
97
98 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
99
100 *b = lastb * (r / (rand_max(&td->random_state) + 1.0));
101 } else {
102 uint64_t off = 0;
103
104 assert(fio_file_lfsr(f));
105
106 if (lfsr_next(&f->lfsr, &off))
107 return 1;
108
109 *b = off;
110 }
111
112 /*
113 * if we are not maintaining a random map, we are done.
114 */
115 if (!file_randommap(td, f))
116 goto ret;
117
118 /*
119 * calculate map offset and check if it's free
120 */
121 if (random_map_free(f, *b))
122 goto ret;
123
124 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
125 (unsigned long long) *b);
126
127 *b = axmap_next_free(f->io_axmap, *b);
128 if (*b == (uint64_t) -1ULL)
129 return 1;
130ret:
131 return 0;
132}
133
134static int __get_next_rand_offset_zipf(struct thread_data *td,
135 struct fio_file *f, enum fio_ddir ddir,
136 uint64_t *b)
137{
138 *b = zipf_next(&f->zipf);
139 return 0;
140}
141
142static int __get_next_rand_offset_pareto(struct thread_data *td,
143 struct fio_file *f, enum fio_ddir ddir,
144 uint64_t *b)
145{
146 *b = pareto_next(&f->zipf);
147 return 0;
148}
149
150static int __get_next_rand_offset_gauss(struct thread_data *td,
151 struct fio_file *f, enum fio_ddir ddir,
152 uint64_t *b)
153{
154 *b = gauss_next(&f->gauss);
155 return 0;
156}
157
158static int __get_next_rand_offset_zoned_abs(struct thread_data *td,
159 struct fio_file *f,
160 enum fio_ddir ddir, uint64_t *b)
161{
162 struct zone_split_index *zsi;
163 uint64_t lastb, send, stotal;
164 unsigned int v;
165
166 lastb = last_block(td, f, ddir);
167 if (!lastb)
168 return 1;
169
170 if (!td->o.zone_split_nr[ddir]) {
171bail:
172 return __get_next_rand_offset(td, f, ddir, b, lastb);
173 }
174
175 /*
176 * Generate a value, v, between 1 and 100, both inclusive
177 */
178 v = rand_between(&td->zone_state, 1, 100);
179
180 /*
181 * Find our generated table. 'send' is the end block of this zone,
182 * 'stotal' is our start offset.
183 */
184 zsi = &td->zone_state_index[ddir][v - 1];
185 stotal = zsi->size_prev / td->o.ba[ddir];
186 send = zsi->size / td->o.ba[ddir];
187
188 /*
189 * Should never happen
190 */
191 if (send == -1U) {
192 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
193 log_err("fio: bug in zoned generation\n");
194 goto bail;
195 } else if (send > lastb) {
196 /*
197 * This happens if the user specifies ranges that exceed
198 * the file/device size. We can't handle that gracefully,
199 * so error and exit.
200 */
201 log_err("fio: zoned_abs sizes exceed file size\n");
202 return 1;
203 }
204
205 /*
206 * Generate index from 0..send-stotal
207 */
208 if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1)
209 return 1;
210
211 *b += stotal;
212 return 0;
213}
214
215static int __get_next_rand_offset_zoned(struct thread_data *td,
216 struct fio_file *f, enum fio_ddir ddir,
217 uint64_t *b)
218{
219 unsigned int v, send, stotal;
220 uint64_t offset, lastb;
221 struct zone_split_index *zsi;
222
223 lastb = last_block(td, f, ddir);
224 if (!lastb)
225 return 1;
226
227 if (!td->o.zone_split_nr[ddir]) {
228bail:
229 return __get_next_rand_offset(td, f, ddir, b, lastb);
230 }
231
232 /*
233 * Generate a value, v, between 1 and 100, both inclusive
234 */
235 v = rand_between(&td->zone_state, 1, 100);
236
237 zsi = &td->zone_state_index[ddir][v - 1];
238 stotal = zsi->size_perc_prev;
239 send = zsi->size_perc;
240
241 /*
242 * Should never happen
243 */
244 if (send == -1U) {
245 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
246 log_err("fio: bug in zoned generation\n");
247 goto bail;
248 }
249
250 /*
251 * 'send' is some percentage below or equal to 100 that
252 * marks the end of the current IO range. 'stotal' marks
253 * the start, in percent.
254 */
255 if (stotal)
256 offset = stotal * lastb / 100ULL;
257 else
258 offset = 0;
259
260 lastb = lastb * (send - stotal) / 100ULL;
261
262 /*
263 * Generate index from 0..send-of-lastb
264 */
265 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
266 return 1;
267
268 /*
269 * Add our start offset, if any
270 */
271 if (offset)
272 *b += offset;
273
274 return 0;
275}
276
277static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
278 enum fio_ddir ddir, uint64_t *b)
279{
280 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
281 uint64_t lastb;
282
283 lastb = last_block(td, f, ddir);
284 if (!lastb)
285 return 1;
286
287 return __get_next_rand_offset(td, f, ddir, b, lastb);
288 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
289 return __get_next_rand_offset_zipf(td, f, ddir, b);
290 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
291 return __get_next_rand_offset_pareto(td, f, ddir, b);
292 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
293 return __get_next_rand_offset_gauss(td, f, ddir, b);
294 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
295 return __get_next_rand_offset_zoned(td, f, ddir, b);
296 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
297 return __get_next_rand_offset_zoned_abs(td, f, ddir, b);
298
299 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
300 return 1;
301}
302
303static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
304{
305 unsigned int v;
306
307 if (td->o.perc_rand[ddir] == 100)
308 return true;
309
310 v = rand_between(&td->seq_rand_state[ddir], 1, 100);
311
312 return v <= td->o.perc_rand[ddir];
313}
314
315static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f)
316{
317 struct thread_options *o = &td->o;
318
319 if (o->invalidate_cache && !o->odirect) {
320 int fio_unused ret;
321
322 ret = file_invalidate_cache(td, f);
323 }
324}
325
326static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
327 enum fio_ddir ddir, uint64_t *b)
328{
329 if (!get_next_rand_offset(td, f, ddir, b))
330 return 0;
331
332 if (td->o.time_based ||
333 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
334 fio_file_reset(td, f);
335 loop_cache_invalidate(td, f);
336 if (!get_next_rand_offset(td, f, ddir, b))
337 return 0;
338 }
339
340 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
341 f->file_name, (unsigned long long) f->last_pos[ddir],
342 (unsigned long long) f->real_file_size);
343 return 1;
344}
345
346static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
347 enum fio_ddir ddir, uint64_t *offset)
348{
349 struct thread_options *o = &td->o;
350
351 assert(ddir_rw(ddir));
352
353 /*
354 * If we reach the end for a time based run, reset us back to 0
355 * and invalidate the cache, if we need to.
356 */
357 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
358 o->time_based) {
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 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->o.zone_mode == ZONE_MODE_ZBD)
750 ddir = zbd_adjust_ddir(td, io_u, ddir);
751
752 if (td_trimwrite(td)) {
753 struct fio_file *f = io_u->file;
754 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
755 ddir = DDIR_TRIM;
756 else
757 ddir = DDIR_WRITE;
758 }
759
760 io_u->ddir = io_u->acct_ddir = ddir;
761
762 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
763 td->o.barrier_blocks &&
764 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
765 td->io_issues[DDIR_WRITE])
766 io_u_set(td, io_u, IO_U_F_BARRIER);
767}
768
769void put_file_log(struct thread_data *td, struct fio_file *f)
770{
771 unsigned int ret = put_file(td, f);
772
773 if (ret)
774 td_verror(td, ret, "file close");
775}
776
777void put_io_u(struct thread_data *td, struct io_u *io_u)
778{
779 const bool needs_lock = td_async_processing(td);
780
781 zbd_put_io_u(io_u);
782
783 if (td->parent)
784 td = td->parent;
785
786 if (needs_lock)
787 __td_io_u_lock(td);
788
789 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
790 put_file_log(td, io_u->file);
791
792 io_u->file = NULL;
793 io_u_set(td, io_u, IO_U_F_FREE);
794
795 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
796 td->cur_depth--;
797 assert(!(td->flags & TD_F_CHILD));
798 }
799 io_u_qpush(&td->io_u_freelist, io_u);
800 td_io_u_free_notify(td);
801
802 if (needs_lock)
803 __td_io_u_unlock(td);
804}
805
806void clear_io_u(struct thread_data *td, struct io_u *io_u)
807{
808 io_u_clear(td, io_u, IO_U_F_FLIGHT);
809 put_io_u(td, io_u);
810}
811
812void requeue_io_u(struct thread_data *td, struct io_u **io_u)
813{
814 const bool needs_lock = td_async_processing(td);
815 struct io_u *__io_u = *io_u;
816 enum fio_ddir ddir = acct_ddir(__io_u);
817
818 dprint(FD_IO, "requeue %p\n", __io_u);
819
820 if (td->parent)
821 td = td->parent;
822
823 if (needs_lock)
824 __td_io_u_lock(td);
825
826 io_u_set(td, __io_u, IO_U_F_FREE);
827 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
828 td->io_issues[ddir]--;
829
830 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
831 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
832 td->cur_depth--;
833 assert(!(td->flags & TD_F_CHILD));
834 }
835
836 io_u_rpush(&td->io_u_requeues, __io_u);
837 td_io_u_free_notify(td);
838
839 if (needs_lock)
840 __td_io_u_unlock(td);
841
842 *io_u = NULL;
843}
844
845static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u)
846{
847 struct fio_file *f = io_u->file;
848
849 assert(td->o.zone_mode == ZONE_MODE_STRIDED);
850 assert(td->o.zone_size);
851 assert(td->o.zone_range);
852
853 /*
854 * See if it's time to switch to a new zone
855 */
856 if (td->zone_bytes >= td->o.zone_size) {
857 td->zone_bytes = 0;
858 f->file_offset += td->o.zone_range + td->o.zone_skip;
859
860 /*
861 * Wrap from the beginning, if we exceed the file size
862 */
863 if (f->file_offset >= f->real_file_size)
864 f->file_offset = get_start_offset(td, f);
865
866 f->last_pos[io_u->ddir] = f->file_offset;
867 td->io_skip_bytes += td->o.zone_skip;
868 }
869
870 /*
871 * If zone_size > zone_range, then maintain the same zone until
872 * zone_bytes >= zone_size.
873 */
874 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
875 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
876 f->file_offset, f->last_pos[io_u->ddir]);
877 f->last_pos[io_u->ddir] = f->file_offset;
878 }
879
880 /*
881 * For random: if 'norandommap' is not set and zone_size > zone_range,
882 * map needs to be reset as it's done with zone_range everytime.
883 */
884 if ((td->zone_bytes % td->o.zone_range) == 0)
885 fio_file_reset(td, f);
886}
887
888static int fill_io_u(struct thread_data *td, struct io_u *io_u)
889{
890 bool is_random;
891 uint64_t offset;
892 enum io_u_action ret;
893
894 if (td_ioengine_flagged(td, FIO_NOIO))
895 goto out;
896
897 set_rw_ddir(td, io_u);
898
899 /*
900 * fsync() or fdatasync() or trim etc, we are done
901 */
902 if (!ddir_rw(io_u->ddir))
903 goto out;
904
905 if (td->o.zone_mode == ZONE_MODE_STRIDED)
906 setup_strided_zone_mode(td, io_u);
907 else if (td->o.zone_mode == ZONE_MODE_ZBD)
908 setup_zbd_zone_mode(td, io_u);
909
910 /*
911 * No log, let the seq/rand engine retrieve the next buflen and
912 * position.
913 */
914 if (get_next_offset(td, io_u, &is_random)) {
915 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
916 return 1;
917 }
918
919 io_u->buflen = get_next_buflen(td, io_u, is_random);
920 if (!io_u->buflen) {
921 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
922 return 1;
923 }
924
925 offset = io_u->offset;
926 if (td->o.zone_mode == ZONE_MODE_ZBD) {
927 ret = zbd_adjust_block(td, io_u);
928 if (ret == io_u_eof)
929 return 1;
930 }
931
932 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
933 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n",
934 io_u,
935 (unsigned long long) io_u->offset, io_u->buflen,
936 (unsigned long long) io_u->file->real_file_size);
937 return 1;
938 }
939
940 /*
941 * mark entry before potentially trimming io_u
942 */
943 if (td_random(td) && file_randommap(td, io_u->file))
944 io_u->buflen = mark_random_map(td, io_u, offset, io_u->buflen);
945
946out:
947 dprint_io_u(io_u, "fill");
948 td->zone_bytes += io_u->buflen;
949 return 0;
950}
951
952static void __io_u_mark_map(uint64_t *map, unsigned int nr)
953{
954 int idx = 0;
955
956 switch (nr) {
957 default:
958 idx = 6;
959 break;
960 case 33 ... 64:
961 idx = 5;
962 break;
963 case 17 ... 32:
964 idx = 4;
965 break;
966 case 9 ... 16:
967 idx = 3;
968 break;
969 case 5 ... 8:
970 idx = 2;
971 break;
972 case 1 ... 4:
973 idx = 1;
974 case 0:
975 break;
976 }
977
978 map[idx]++;
979}
980
981void io_u_mark_submit(struct thread_data *td, unsigned int nr)
982{
983 __io_u_mark_map(td->ts.io_u_submit, nr);
984 td->ts.total_submit++;
985}
986
987void io_u_mark_complete(struct thread_data *td, unsigned int nr)
988{
989 __io_u_mark_map(td->ts.io_u_complete, nr);
990 td->ts.total_complete++;
991}
992
993void io_u_mark_depth(struct thread_data *td, unsigned int nr)
994{
995 int idx = 0;
996
997 switch (td->cur_depth) {
998 default:
999 idx = 6;
1000 break;
1001 case 32 ... 63:
1002 idx = 5;
1003 break;
1004 case 16 ... 31:
1005 idx = 4;
1006 break;
1007 case 8 ... 15:
1008 idx = 3;
1009 break;
1010 case 4 ... 7:
1011 idx = 2;
1012 break;
1013 case 2 ... 3:
1014 idx = 1;
1015 case 1:
1016 break;
1017 }
1018
1019 td->ts.io_u_map[idx] += nr;
1020}
1021
1022static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1023{
1024 int idx = 0;
1025
1026 assert(nsec < 1000);
1027
1028 switch (nsec) {
1029 case 750 ... 999:
1030 idx = 9;
1031 break;
1032 case 500 ... 749:
1033 idx = 8;
1034 break;
1035 case 250 ... 499:
1036 idx = 7;
1037 break;
1038 case 100 ... 249:
1039 idx = 6;
1040 break;
1041 case 50 ... 99:
1042 idx = 5;
1043 break;
1044 case 20 ... 49:
1045 idx = 4;
1046 break;
1047 case 10 ... 19:
1048 idx = 3;
1049 break;
1050 case 4 ... 9:
1051 idx = 2;
1052 break;
1053 case 2 ... 3:
1054 idx = 1;
1055 case 0 ... 1:
1056 break;
1057 }
1058
1059 assert(idx < FIO_IO_U_LAT_N_NR);
1060 td->ts.io_u_lat_n[idx]++;
1061}
1062
1063static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1064{
1065 int idx = 0;
1066
1067 assert(usec < 1000 && usec >= 1);
1068
1069 switch (usec) {
1070 case 750 ... 999:
1071 idx = 9;
1072 break;
1073 case 500 ... 749:
1074 idx = 8;
1075 break;
1076 case 250 ... 499:
1077 idx = 7;
1078 break;
1079 case 100 ... 249:
1080 idx = 6;
1081 break;
1082 case 50 ... 99:
1083 idx = 5;
1084 break;
1085 case 20 ... 49:
1086 idx = 4;
1087 break;
1088 case 10 ... 19:
1089 idx = 3;
1090 break;
1091 case 4 ... 9:
1092 idx = 2;
1093 break;
1094 case 2 ... 3:
1095 idx = 1;
1096 case 0 ... 1:
1097 break;
1098 }
1099
1100 assert(idx < FIO_IO_U_LAT_U_NR);
1101 td->ts.io_u_lat_u[idx]++;
1102}
1103
1104static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1105{
1106 int idx = 0;
1107
1108 assert(msec >= 1);
1109
1110 switch (msec) {
1111 default:
1112 idx = 11;
1113 break;
1114 case 1000 ... 1999:
1115 idx = 10;
1116 break;
1117 case 750 ... 999:
1118 idx = 9;
1119 break;
1120 case 500 ... 749:
1121 idx = 8;
1122 break;
1123 case 250 ... 499:
1124 idx = 7;
1125 break;
1126 case 100 ... 249:
1127 idx = 6;
1128 break;
1129 case 50 ... 99:
1130 idx = 5;
1131 break;
1132 case 20 ... 49:
1133 idx = 4;
1134 break;
1135 case 10 ... 19:
1136 idx = 3;
1137 break;
1138 case 4 ... 9:
1139 idx = 2;
1140 break;
1141 case 2 ... 3:
1142 idx = 1;
1143 case 0 ... 1:
1144 break;
1145 }
1146
1147 assert(idx < FIO_IO_U_LAT_M_NR);
1148 td->ts.io_u_lat_m[idx]++;
1149}
1150
1151static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1152{
1153 if (nsec < 1000)
1154 io_u_mark_lat_nsec(td, nsec);
1155 else if (nsec < 1000000)
1156 io_u_mark_lat_usec(td, nsec / 1000);
1157 else
1158 io_u_mark_lat_msec(td, nsec / 1000000);
1159}
1160
1161static unsigned int __get_next_fileno_rand(struct thread_data *td)
1162{
1163 unsigned long fileno;
1164
1165 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1166 uint64_t frand_max = rand_max(&td->next_file_state);
1167 unsigned long r;
1168
1169 r = __rand(&td->next_file_state);
1170 return (unsigned int) ((double) td->o.nr_files
1171 * (r / (frand_max + 1.0)));
1172 }
1173
1174 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1175 fileno = zipf_next(&td->next_file_zipf);
1176 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1177 fileno = pareto_next(&td->next_file_zipf);
1178 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1179 fileno = gauss_next(&td->next_file_gauss);
1180 else {
1181 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1182 assert(0);
1183 return 0;
1184 }
1185
1186 return fileno >> FIO_FSERVICE_SHIFT;
1187}
1188
1189/*
1190 * Get next file to service by choosing one at random
1191 */
1192static struct fio_file *get_next_file_rand(struct thread_data *td,
1193 enum fio_file_flags goodf,
1194 enum fio_file_flags badf)
1195{
1196 struct fio_file *f;
1197 int fno;
1198
1199 do {
1200 int opened = 0;
1201
1202 fno = __get_next_fileno_rand(td);
1203
1204 f = td->files[fno];
1205 if (fio_file_done(f))
1206 continue;
1207
1208 if (!fio_file_open(f)) {
1209 int err;
1210
1211 if (td->nr_open_files >= td->o.open_files)
1212 return ERR_PTR(-EBUSY);
1213
1214 err = td_io_open_file(td, f);
1215 if (err)
1216 continue;
1217 opened = 1;
1218 }
1219
1220 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1221 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1222 return f;
1223 }
1224 if (opened)
1225 td_io_close_file(td, f);
1226 } while (1);
1227}
1228
1229/*
1230 * Get next file to service by doing round robin between all available ones
1231 */
1232static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1233 int badf)
1234{
1235 unsigned int old_next_file = td->next_file;
1236 struct fio_file *f;
1237
1238 do {
1239 int opened = 0;
1240
1241 f = td->files[td->next_file];
1242
1243 td->next_file++;
1244 if (td->next_file >= td->o.nr_files)
1245 td->next_file = 0;
1246
1247 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1248 if (fio_file_done(f)) {
1249 f = NULL;
1250 continue;
1251 }
1252
1253 if (!fio_file_open(f)) {
1254 int err;
1255
1256 if (td->nr_open_files >= td->o.open_files)
1257 return ERR_PTR(-EBUSY);
1258
1259 err = td_io_open_file(td, f);
1260 if (err) {
1261 dprint(FD_FILE, "error %d on open of %s\n",
1262 err, f->file_name);
1263 f = NULL;
1264 continue;
1265 }
1266 opened = 1;
1267 }
1268
1269 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1270 f->flags);
1271 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1272 break;
1273
1274 if (opened)
1275 td_io_close_file(td, f);
1276
1277 f = NULL;
1278 } while (td->next_file != old_next_file);
1279
1280 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1281 return f;
1282}
1283
1284static struct fio_file *__get_next_file(struct thread_data *td)
1285{
1286 struct fio_file *f;
1287
1288 assert(td->o.nr_files <= td->files_index);
1289
1290 if (td->nr_done_files >= td->o.nr_files) {
1291 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1292 " nr_files=%d\n", td->nr_open_files,
1293 td->nr_done_files,
1294 td->o.nr_files);
1295 return NULL;
1296 }
1297
1298 f = td->file_service_file;
1299 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1300 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1301 goto out;
1302 if (td->file_service_left--)
1303 goto out;
1304 }
1305
1306 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1307 td->o.file_service_type == FIO_FSERVICE_SEQ)
1308 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1309 else
1310 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1311
1312 if (IS_ERR(f))
1313 return f;
1314
1315 td->file_service_file = f;
1316 td->file_service_left = td->file_service_nr - 1;
1317out:
1318 if (f)
1319 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1320 else
1321 dprint(FD_FILE, "get_next_file: NULL\n");
1322 return f;
1323}
1324
1325static struct fio_file *get_next_file(struct thread_data *td)
1326{
1327 return __get_next_file(td);
1328}
1329
1330static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1331{
1332 struct fio_file *f;
1333
1334 do {
1335 f = get_next_file(td);
1336 if (IS_ERR_OR_NULL(f))
1337 return PTR_ERR(f);
1338
1339 io_u->file = f;
1340 get_file(f);
1341
1342 if (!fill_io_u(td, io_u))
1343 break;
1344
1345 zbd_put_io_u(io_u);
1346
1347 put_file_log(td, f);
1348 td_io_close_file(td, f);
1349 io_u->file = NULL;
1350 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1351 fio_file_reset(td, f);
1352 else {
1353 fio_file_set_done(f);
1354 td->nr_done_files++;
1355 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1356 td->nr_done_files, td->o.nr_files);
1357 }
1358 } while (1);
1359
1360 return 0;
1361}
1362
1363static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1364 unsigned long long tnsec, unsigned long long max_nsec)
1365{
1366 if (!td->error)
1367 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
1368 td_verror(td, ETIMEDOUT, "max latency exceeded");
1369 icd->error = ETIMEDOUT;
1370}
1371
1372static void lat_new_cycle(struct thread_data *td)
1373{
1374 fio_gettime(&td->latency_ts, NULL);
1375 td->latency_ios = ddir_rw_sum(td->io_blocks);
1376 td->latency_failed = 0;
1377}
1378
1379/*
1380 * We had an IO outside the latency target. Reduce the queue depth. If we
1381 * are at QD=1, then it's time to give up.
1382 */
1383static bool __lat_target_failed(struct thread_data *td)
1384{
1385 if (td->latency_qd == 1)
1386 return true;
1387
1388 td->latency_qd_high = td->latency_qd;
1389
1390 if (td->latency_qd == td->latency_qd_low)
1391 td->latency_qd_low--;
1392
1393 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1394 td->latency_stable_count = 0;
1395
1396 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1397
1398 /*
1399 * When we ramp QD down, quiesce existing IO to prevent
1400 * a storm of ramp downs due to pending higher depth.
1401 */
1402 io_u_quiesce(td);
1403 lat_new_cycle(td);
1404 return false;
1405}
1406
1407static bool lat_target_failed(struct thread_data *td)
1408{
1409 if (td->o.latency_percentile.u.f == 100.0)
1410 return __lat_target_failed(td);
1411
1412 td->latency_failed++;
1413 return false;
1414}
1415
1416void lat_target_init(struct thread_data *td)
1417{
1418 td->latency_end_run = 0;
1419
1420 if (td->o.latency_target) {
1421 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1422 fio_gettime(&td->latency_ts, NULL);
1423 td->latency_qd = 1;
1424 td->latency_qd_high = td->o.iodepth;
1425 td->latency_qd_low = 1;
1426 td->latency_ios = ddir_rw_sum(td->io_blocks);
1427 } else
1428 td->latency_qd = td->o.iodepth;
1429}
1430
1431void lat_target_reset(struct thread_data *td)
1432{
1433 if (!td->latency_end_run)
1434 lat_target_init(td);
1435}
1436
1437static void lat_target_success(struct thread_data *td)
1438{
1439 const unsigned int qd = td->latency_qd;
1440 struct thread_options *o = &td->o;
1441
1442 td->latency_qd_low = td->latency_qd;
1443
1444 if (td->latency_qd + 1 == td->latency_qd_high) {
1445 /*
1446 * latency_qd will not incease on lat_target_success(), so
1447 * called stable. If we stick with this queue depth, the
1448 * final latency is likely lower than latency_target. Fix
1449 * this by increasing latency_qd_high slowly. Use a naive
1450 * heuristic here. If we get lat_target_success() 3 times
1451 * in a row, increase latency_qd_high by 1.
1452 */
1453 if (++td->latency_stable_count >= 3) {
1454 td->latency_qd_high++;
1455 td->latency_stable_count = 0;
1456 }
1457 }
1458
1459 /*
1460 * If we haven't failed yet, we double up to a failing value instead
1461 * of bisecting from highest possible queue depth. If we have set
1462 * a limit other than td->o.iodepth, bisect between that.
1463 */
1464 if (td->latency_qd_high != o->iodepth)
1465 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1466 else
1467 td->latency_qd *= 2;
1468
1469 if (td->latency_qd > o->iodepth)
1470 td->latency_qd = o->iodepth;
1471
1472 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1473
1474 /*
1475 * Same as last one, we are done. Let it run a latency cycle, so
1476 * we get only the results from the targeted depth.
1477 */
1478 if (!o->latency_run && td->latency_qd == qd) {
1479 if (td->latency_end_run) {
1480 dprint(FD_RATE, "We are done\n");
1481 td->done = 1;
1482 } else {
1483 dprint(FD_RATE, "Quiesce and final run\n");
1484 io_u_quiesce(td);
1485 td->latency_end_run = 1;
1486 reset_all_stats(td);
1487 reset_io_stats(td);
1488 }
1489 }
1490
1491 lat_new_cycle(td);
1492}
1493
1494/*
1495 * Check if we can bump the queue depth
1496 */
1497void lat_target_check(struct thread_data *td)
1498{
1499 uint64_t usec_window;
1500 uint64_t ios;
1501 double success_ios;
1502
1503 usec_window = utime_since_now(&td->latency_ts);
1504 if (usec_window < td->o.latency_window)
1505 return;
1506
1507 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1508 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1509 success_ios *= 100.0;
1510
1511 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1512
1513 if (success_ios >= td->o.latency_percentile.u.f)
1514 lat_target_success(td);
1515 else
1516 __lat_target_failed(td);
1517}
1518
1519/*
1520 * If latency target is enabled, we might be ramping up or down and not
1521 * using the full queue depth available.
1522 */
1523bool queue_full(const struct thread_data *td)
1524{
1525 const int qempty = io_u_qempty(&td->io_u_freelist);
1526
1527 if (qempty)
1528 return true;
1529 if (!td->o.latency_target)
1530 return false;
1531
1532 return td->cur_depth >= td->latency_qd;
1533}
1534
1535struct io_u *__get_io_u(struct thread_data *td)
1536{
1537 const bool needs_lock = td_async_processing(td);
1538 struct io_u *io_u = NULL;
1539 int ret;
1540
1541 if (td->stop_io)
1542 return NULL;
1543
1544 if (needs_lock)
1545 __td_io_u_lock(td);
1546
1547again:
1548 if (!io_u_rempty(&td->io_u_requeues))
1549 io_u = io_u_rpop(&td->io_u_requeues);
1550 else if (!queue_full(td)) {
1551 io_u = io_u_qpop(&td->io_u_freelist);
1552
1553 io_u->file = NULL;
1554 io_u->buflen = 0;
1555 io_u->resid = 0;
1556 io_u->end_io = NULL;
1557 }
1558
1559 if (io_u) {
1560 assert(io_u->flags & IO_U_F_FREE);
1561 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1562 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1563 IO_U_F_VER_LIST | IO_U_F_PRIORITY);
1564
1565 io_u->error = 0;
1566 io_u->acct_ddir = -1;
1567 td->cur_depth++;
1568 assert(!(td->flags & TD_F_CHILD));
1569 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1570 io_u->ipo = NULL;
1571 } else if (td_async_processing(td)) {
1572 /*
1573 * We ran out, wait for async verify threads to finish and
1574 * return one
1575 */
1576 assert(!(td->flags & TD_F_CHILD));
1577 ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1578 assert(ret == 0);
1579 if (!td->error)
1580 goto again;
1581 }
1582
1583 if (needs_lock)
1584 __td_io_u_unlock(td);
1585
1586 return io_u;
1587}
1588
1589static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1590{
1591 if (!(td->flags & TD_F_TRIM_BACKLOG))
1592 return false;
1593 if (!td->trim_entries)
1594 return false;
1595
1596 if (td->trim_batch) {
1597 td->trim_batch--;
1598 if (get_next_trim(td, io_u))
1599 return true;
1600 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1601 td->last_ddir != DDIR_READ) {
1602 td->trim_batch = td->o.trim_batch;
1603 if (!td->trim_batch)
1604 td->trim_batch = td->o.trim_backlog;
1605 if (get_next_trim(td, io_u))
1606 return true;
1607 }
1608
1609 return false;
1610}
1611
1612static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1613{
1614 if (!(td->flags & TD_F_VER_BACKLOG))
1615 return false;
1616
1617 if (td->io_hist_len) {
1618 int get_verify = 0;
1619
1620 if (td->verify_batch)
1621 get_verify = 1;
1622 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1623 td->last_ddir != DDIR_READ) {
1624 td->verify_batch = td->o.verify_batch;
1625 if (!td->verify_batch)
1626 td->verify_batch = td->o.verify_backlog;
1627 get_verify = 1;
1628 }
1629
1630 if (get_verify && !get_next_verify(td, io_u)) {
1631 td->verify_batch--;
1632 return true;
1633 }
1634 }
1635
1636 return false;
1637}
1638
1639/*
1640 * Fill offset and start time into the buffer content, to prevent too
1641 * easy compressible data for simple de-dupe attempts. Do this for every
1642 * 512b block in the range, since that should be the smallest block size
1643 * we can expect from a device.
1644 */
1645static void small_content_scramble(struct io_u *io_u)
1646{
1647 unsigned long long i, nr_blocks = io_u->buflen >> 9;
1648 unsigned int offset;
1649 uint64_t boffset, *iptr;
1650 char *p;
1651
1652 if (!nr_blocks)
1653 return;
1654
1655 p = io_u->xfer_buf;
1656 boffset = io_u->offset;
1657
1658 if (io_u->buf_filled_len)
1659 io_u->buf_filled_len = 0;
1660
1661 /*
1662 * Generate random index between 0..7. We do chunks of 512b, if
1663 * we assume a cacheline is 64 bytes, then we have 8 of those.
1664 * Scramble content within the blocks in the same cacheline to
1665 * speed things up.
1666 */
1667 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1668
1669 for (i = 0; i < nr_blocks; i++) {
1670 /*
1671 * Fill offset into start of cacheline, time into end
1672 * of cacheline
1673 */
1674 iptr = (void *) p + (offset << 6);
1675 *iptr = boffset;
1676
1677 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1678 iptr[0] = io_u->start_time.tv_sec;
1679 iptr[1] = io_u->start_time.tv_nsec;
1680
1681 p += 512;
1682 boffset += 512;
1683 }
1684}
1685
1686/*
1687 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1688 * etc. The returned io_u is fully ready to be prepped, populated and submitted.
1689 */
1690struct io_u *get_io_u(struct thread_data *td)
1691{
1692 struct fio_file *f;
1693 struct io_u *io_u;
1694 int do_scramble = 0;
1695 long ret = 0;
1696
1697 io_u = __get_io_u(td);
1698 if (!io_u) {
1699 dprint(FD_IO, "__get_io_u failed\n");
1700 return NULL;
1701 }
1702
1703 if (check_get_verify(td, io_u))
1704 goto out;
1705 if (check_get_trim(td, io_u))
1706 goto out;
1707
1708 /*
1709 * from a requeue, io_u already setup
1710 */
1711 if (io_u->file)
1712 goto out;
1713
1714 /*
1715 * If using an iolog, grab next piece if any available.
1716 */
1717 if (td->flags & TD_F_READ_IOLOG) {
1718 if (read_iolog_get(td, io_u))
1719 goto err_put;
1720 } else if (set_io_u_file(td, io_u)) {
1721 ret = -EBUSY;
1722 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1723 goto err_put;
1724 }
1725
1726 f = io_u->file;
1727 if (!f) {
1728 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1729 goto err_put;
1730 }
1731
1732 assert(fio_file_open(f));
1733
1734 if (ddir_rw(io_u->ddir)) {
1735 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1736 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1737 goto err_put;
1738 }
1739
1740 f->last_start[io_u->ddir] = io_u->offset;
1741 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1742
1743 if (io_u->ddir == DDIR_WRITE) {
1744 if (td->flags & TD_F_REFILL_BUFFERS) {
1745 io_u_fill_buffer(td, io_u,
1746 td->o.min_bs[DDIR_WRITE],
1747 io_u->buflen);
1748 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1749 !(td->flags & TD_F_COMPRESS) &&
1750 !(td->flags & TD_F_DO_VERIFY))
1751 do_scramble = 1;
1752 } else if (io_u->ddir == DDIR_READ) {
1753 /*
1754 * Reset the buf_filled parameters so next time if the
1755 * buffer is used for writes it is refilled.
1756 */
1757 io_u->buf_filled_len = 0;
1758 }
1759 }
1760
1761 /*
1762 * Set io data pointers.
1763 */
1764 io_u->xfer_buf = io_u->buf;
1765 io_u->xfer_buflen = io_u->buflen;
1766
1767out:
1768 assert(io_u->file);
1769 if (!td_io_prep(td, io_u)) {
1770 if (!td->o.disable_lat)
1771 fio_gettime(&io_u->start_time, NULL);
1772
1773 if (do_scramble)
1774 small_content_scramble(io_u);
1775
1776 return io_u;
1777 }
1778err_put:
1779 dprint(FD_IO, "get_io_u failed\n");
1780 put_io_u(td, io_u);
1781 return ERR_PTR(ret);
1782}
1783
1784static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1785{
1786 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1787
1788 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1789 return;
1790
1791 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n",
1792 io_u->file ? " on file " : "",
1793 io_u->file ? io_u->file->file_name : "",
1794 strerror(io_u->error),
1795 io_ddir_name(io_u->ddir),
1796 io_u->offset, io_u->xfer_buflen);
1797
1798 if (td->io_ops->errdetails) {
1799 char *err = td->io_ops->errdetails(io_u);
1800
1801 log_err("fio: %s\n", err);
1802 free(err);
1803 }
1804
1805 if (!td->error)
1806 td_verror(td, io_u->error, "io_u error");
1807}
1808
1809void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1810{
1811 __io_u_log_error(td, io_u);
1812 if (td->parent)
1813 __io_u_log_error(td->parent, io_u);
1814}
1815
1816static inline bool gtod_reduce(struct thread_data *td)
1817{
1818 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1819 || td->o.gtod_reduce;
1820}
1821
1822static void trim_block_info(struct thread_data *td, struct io_u *io_u)
1823{
1824 uint32_t *info = io_u_block_info(td, io_u);
1825
1826 if (BLOCK_INFO_STATE(*info) >= BLOCK_STATE_TRIM_FAILURE)
1827 return;
1828
1829 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, BLOCK_INFO_TRIMS(*info) + 1);
1830}
1831
1832static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1833 struct io_completion_data *icd,
1834 const enum fio_ddir idx, unsigned int bytes)
1835{
1836 const int no_reduce = !gtod_reduce(td);
1837 unsigned long long llnsec = 0;
1838
1839 if (td->parent)
1840 td = td->parent;
1841
1842 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1843 return;
1844
1845 if (no_reduce)
1846 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1847
1848 if (!td->o.disable_lat) {
1849 unsigned long long tnsec;
1850
1851 tnsec = ntime_since(&io_u->start_time, &icd->time);
1852 add_lat_sample(td, idx, tnsec, bytes, io_u->offset, io_u_is_prio(io_u));
1853
1854 if (td->flags & TD_F_PROFILE_OPS) {
1855 struct prof_io_ops *ops = &td->prof_io_ops;
1856
1857 if (ops->io_u_lat)
1858 icd->error = ops->io_u_lat(td, tnsec);
1859 }
1860
1861 if (td->o.max_latency && tnsec > td->o.max_latency)
1862 lat_fatal(td, icd, tnsec, td->o.max_latency);
1863 if (td->o.latency_target && tnsec > td->o.latency_target) {
1864 if (lat_target_failed(td))
1865 lat_fatal(td, icd, tnsec, td->o.latency_target);
1866 }
1867 }
1868
1869 if (ddir_rw(idx)) {
1870 if (!td->o.disable_clat) {
1871 add_clat_sample(td, idx, llnsec, bytes, io_u->offset, io_u_is_prio(io_u));
1872 io_u_mark_latency(td, llnsec);
1873 }
1874
1875 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1876 add_bw_sample(td, io_u, bytes, llnsec);
1877
1878 if (no_reduce && per_unit_log(td->iops_log))
1879 add_iops_sample(td, io_u, bytes);
1880 } else if (ddir_sync(idx) && !td->o.disable_clat)
1881 add_sync_clat_sample(&td->ts, llnsec);
1882
1883 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM)
1884 trim_block_info(td, io_u);
1885}
1886
1887static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1888 uint64_t offset, unsigned int bytes)
1889{
1890 int idx;
1891
1892 if (!f)
1893 return;
1894
1895 if (f->first_write == -1ULL || offset < f->first_write)
1896 f->first_write = offset;
1897 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1898 f->last_write = offset + bytes;
1899
1900 if (!f->last_write_comp)
1901 return;
1902
1903 idx = f->last_write_idx++;
1904 f->last_write_comp[idx] = offset;
1905 if (f->last_write_idx == td->o.iodepth)
1906 f->last_write_idx = 0;
1907}
1908
1909static bool should_account(struct thread_data *td)
1910{
1911 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1912 td->runstate == TD_VERIFYING);
1913}
1914
1915static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1916 struct io_completion_data *icd)
1917{
1918 struct io_u *io_u = *io_u_ptr;
1919 enum fio_ddir ddir = io_u->ddir;
1920 struct fio_file *f = io_u->file;
1921
1922 dprint_io_u(io_u, "complete");
1923
1924 assert(io_u->flags & IO_U_F_FLIGHT);
1925 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1926
1927 /*
1928 * Mark IO ok to verify
1929 */
1930 if (io_u->ipo) {
1931 /*
1932 * Remove errored entry from the verification list
1933 */
1934 if (io_u->error)
1935 unlog_io_piece(td, io_u);
1936 else {
1937 atomic_store_release(&io_u->ipo->flags,
1938 io_u->ipo->flags & ~IP_F_IN_FLIGHT);
1939 }
1940 }
1941
1942 if (ddir_sync(ddir)) {
1943 td->last_was_sync = true;
1944 if (f) {
1945 f->first_write = -1ULL;
1946 f->last_write = -1ULL;
1947 }
1948 if (should_account(td))
1949 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
1950 return;
1951 }
1952
1953 td->last_was_sync = false;
1954 td->last_ddir = ddir;
1955
1956 if (!io_u->error && ddir_rw(ddir)) {
1957 unsigned long long bytes = io_u->buflen - io_u->resid;
1958 int ret;
1959
1960 td->io_blocks[ddir]++;
1961 td->io_bytes[ddir] += bytes;
1962
1963 if (!(io_u->flags & IO_U_F_VER_LIST)) {
1964 td->this_io_blocks[ddir]++;
1965 td->this_io_bytes[ddir] += bytes;
1966 }
1967
1968 if (ddir == DDIR_WRITE)
1969 file_log_write_comp(td, f, io_u->offset, bytes);
1970
1971 if (should_account(td))
1972 account_io_completion(td, io_u, icd, ddir, bytes);
1973
1974 icd->bytes_done[ddir] += bytes;
1975
1976 if (io_u->end_io) {
1977 ret = io_u->end_io(td, io_u_ptr);
1978 io_u = *io_u_ptr;
1979 if (ret && !icd->error)
1980 icd->error = ret;
1981 }
1982 } else if (io_u->error) {
1983 icd->error = io_u->error;
1984 io_u_log_error(td, io_u);
1985 }
1986 if (icd->error) {
1987 enum error_type_bit eb = td_error_type(ddir, icd->error);
1988
1989 if (!td_non_fatal_error(td, eb, icd->error))
1990 return;
1991
1992 /*
1993 * If there is a non_fatal error, then add to the error count
1994 * and clear all the errors.
1995 */
1996 update_error_count(td, icd->error);
1997 td_clear_error(td);
1998 icd->error = 0;
1999 if (io_u)
2000 io_u->error = 0;
2001 }
2002}
2003
2004static void init_icd(struct thread_data *td, struct io_completion_data *icd,
2005 int nr)
2006{
2007 int ddir;
2008
2009 if (!gtod_reduce(td))
2010 fio_gettime(&icd->time, NULL);
2011
2012 icd->nr = nr;
2013
2014 icd->error = 0;
2015 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2016 icd->bytes_done[ddir] = 0;
2017}
2018
2019static void ios_completed(struct thread_data *td,
2020 struct io_completion_data *icd)
2021{
2022 struct io_u *io_u;
2023 int i;
2024
2025 for (i = 0; i < icd->nr; i++) {
2026 io_u = td->io_ops->event(td, i);
2027
2028 io_completed(td, &io_u, icd);
2029
2030 if (io_u)
2031 put_io_u(td, io_u);
2032 }
2033}
2034
2035/*
2036 * Complete a single io_u for the sync engines.
2037 */
2038int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
2039{
2040 struct io_completion_data icd;
2041 int ddir;
2042
2043 init_icd(td, &icd, 1);
2044 io_completed(td, &io_u, &icd);
2045
2046 if (io_u)
2047 put_io_u(td, io_u);
2048
2049 if (icd.error) {
2050 td_verror(td, icd.error, "io_u_sync_complete");
2051 return -1;
2052 }
2053
2054 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2055 td->bytes_done[ddir] += icd.bytes_done[ddir];
2056
2057 return 0;
2058}
2059
2060/*
2061 * Called to complete min_events number of io for the async engines.
2062 */
2063int io_u_queued_complete(struct thread_data *td, int min_evts)
2064{
2065 struct io_completion_data icd;
2066 struct timespec *tvp = NULL;
2067 int ret, ddir;
2068 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2069
2070 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2071
2072 if (!min_evts)
2073 tvp = &ts;
2074 else if (min_evts > td->cur_depth)
2075 min_evts = td->cur_depth;
2076
2077 /* No worries, td_io_getevents fixes min and max if they are
2078 * set incorrectly */
2079 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2080 if (ret < 0) {
2081 td_verror(td, -ret, "td_io_getevents");
2082 return ret;
2083 } else if (!ret)
2084 return ret;
2085
2086 init_icd(td, &icd, ret);
2087 ios_completed(td, &icd);
2088 if (icd.error) {
2089 td_verror(td, icd.error, "io_u_queued_complete");
2090 return -1;
2091 }
2092
2093 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2094 td->bytes_done[ddir] += icd.bytes_done[ddir];
2095
2096 return ret;
2097}
2098
2099/*
2100 * Call when io_u is really queued, to update the submission latency.
2101 */
2102void io_u_queued(struct thread_data *td, struct io_u *io_u)
2103{
2104 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2105 unsigned long slat_time;
2106
2107 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2108
2109 if (td->parent)
2110 td = td->parent;
2111
2112 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2113 io_u->offset, io_u_is_prio(io_u));
2114 }
2115}
2116
2117/*
2118 * See if we should reuse the last seed, if dedupe is enabled
2119 */
2120static struct frand_state *get_buf_state(struct thread_data *td)
2121{
2122 unsigned int v;
2123
2124 if (!td->o.dedupe_percentage)
2125 return &td->buf_state;
2126 else if (td->o.dedupe_percentage == 100) {
2127 frand_copy(&td->buf_state_prev, &td->buf_state);
2128 return &td->buf_state;
2129 }
2130
2131 v = rand_between(&td->dedupe_state, 1, 100);
2132
2133 if (v <= td->o.dedupe_percentage)
2134 return &td->buf_state_prev;
2135
2136 return &td->buf_state;
2137}
2138
2139static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2140{
2141 if (td->o.dedupe_percentage == 100)
2142 frand_copy(rs, &td->buf_state_prev);
2143 else if (rs == &td->buf_state)
2144 frand_copy(&td->buf_state_prev, rs);
2145}
2146
2147void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write,
2148 unsigned long long max_bs)
2149{
2150 struct thread_options *o = &td->o;
2151
2152 if (o->mem_type == MEM_CUDA_MALLOC)
2153 return;
2154
2155 if (o->compress_percentage || o->dedupe_percentage) {
2156 unsigned int perc = td->o.compress_percentage;
2157 struct frand_state *rs;
2158 unsigned long long left = max_bs;
2159 unsigned long long this_write;
2160
2161 do {
2162 rs = get_buf_state(td);
2163
2164 min_write = min(min_write, left);
2165
2166 if (perc) {
2167 this_write = min_not_zero(min_write,
2168 (unsigned long long) td->o.compress_chunk);
2169
2170 fill_random_buf_percentage(rs, buf, perc,
2171 this_write, this_write,
2172 o->buffer_pattern,
2173 o->buffer_pattern_bytes);
2174 } else {
2175 fill_random_buf(rs, buf, min_write);
2176 this_write = min_write;
2177 }
2178
2179 buf += this_write;
2180 left -= this_write;
2181 save_buf_state(td, rs);
2182 } while (left);
2183 } else if (o->buffer_pattern_bytes)
2184 fill_buffer_pattern(td, buf, max_bs);
2185 else if (o->zero_buffers)
2186 memset(buf, 0, max_bs);
2187 else
2188 fill_random_buf(get_buf_state(td), buf, max_bs);
2189}
2190
2191/*
2192 * "randomly" fill the buffer contents
2193 */
2194void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2195 unsigned long long min_write, unsigned long long max_bs)
2196{
2197 io_u->buf_filled_len = 0;
2198 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2199}
2200
2201static int do_sync_file_range(const struct thread_data *td,
2202 struct fio_file *f)
2203{
2204 uint64_t offset, nbytes;
2205
2206 offset = f->first_write;
2207 nbytes = f->last_write - f->first_write;
2208
2209 if (!nbytes)
2210 return 0;
2211
2212 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2213}
2214
2215int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2216{
2217 int ret;
2218
2219 if (io_u->ddir == DDIR_SYNC) {
2220 ret = fsync(io_u->file->fd);
2221 } else if (io_u->ddir == DDIR_DATASYNC) {
2222#ifdef CONFIG_FDATASYNC
2223 ret = fdatasync(io_u->file->fd);
2224#else
2225 ret = io_u->xfer_buflen;
2226 io_u->error = EINVAL;
2227#endif
2228 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2229 ret = do_sync_file_range(td, io_u->file);
2230 else {
2231 ret = io_u->xfer_buflen;
2232 io_u->error = EINVAL;
2233 }
2234
2235 if (ret < 0)
2236 io_u->error = errno;
2237
2238 return ret;
2239}
2240
2241int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2242{
2243#ifndef FIO_HAVE_TRIM
2244 io_u->error = EINVAL;
2245 return 0;
2246#else
2247 struct fio_file *f = io_u->file;
2248 int ret;
2249
2250 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2251 if (!ret)
2252 return io_u->xfer_buflen;
2253
2254 io_u->error = ret;
2255 return 0;
2256#endif
2257}
fio - Flexible IO Tester