projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
engines/libaio: update for newer io_setup2() system call
[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 buflen = bsp->bs;
561 perc += bsp->perc;
562 if (!perc)
563 break;
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 } while (!io_u_fits(td, io_u, buflen));
576
577 return buflen;
578}
579
580static void set_rwmix_bytes(struct thread_data *td)
581{
582 unsigned int diff;
583
584 /*
585 * we do time or byte based switch. this is needed because
586 * buffered writes may issue a lot quicker than they complete,
587 * whereas reads do not.
588 */
589 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
590 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
591}
592
593static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
594{
595 unsigned int v;
596
597 v = rand_between(&td->rwmix_state, 1, 100);
598
599 if (v <= td->o.rwmix[DDIR_READ])
600 return DDIR_READ;
601
602 return DDIR_WRITE;
603}
604
605int io_u_quiesce(struct thread_data *td)
606{
607 int ret = 0, completed = 0;
608
609 /*
610 * We are going to sleep, ensure that we flush anything pending as
611 * not to skew our latency numbers.
612 *
613 * Changed to only monitor 'in flight' requests here instead of the
614 * td->cur_depth, b/c td->cur_depth does not accurately represent
615 * io's that have been actually submitted to an async engine,
616 * and cur_depth is meaningless for sync engines.
617 */
618 if (td->io_u_queued || td->cur_depth)
619 td_io_commit(td);
620
621 while (td->io_u_in_flight) {
622 ret = io_u_queued_complete(td, 1);
623 if (ret > 0)
624 completed += ret;
625 else if (ret < 0)
626 break;
627 }
628
629 if (td->flags & TD_F_REGROW_LOGS)
630 regrow_logs(td);
631
632 if (completed)
633 return completed;
634
635 return ret;
636}
637
638static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
639{
640 enum fio_ddir odir = ddir ^ 1;
641 uint64_t usec;
642 uint64_t now;
643
644 assert(ddir_rw(ddir));
645 now = utime_since_now(&td->start);
646
647 /*
648 * if rate_next_io_time is in the past, need to catch up to rate
649 */
650 if (td->rate_next_io_time[ddir] <= now)
651 return ddir;
652
653 /*
654 * We are ahead of rate in this direction. See if we
655 * should switch.
656 */
657 if (td_rw(td) && td->o.rwmix[odir]) {
658 /*
659 * Other direction is behind rate, switch
660 */
661 if (td->rate_next_io_time[odir] <= now)
662 return odir;
663
664 /*
665 * Both directions are ahead of rate. sleep the min,
666 * switch if necessary
667 */
668 if (td->rate_next_io_time[ddir] <=
669 td->rate_next_io_time[odir]) {
670 usec = td->rate_next_io_time[ddir] - now;
671 } else {
672 usec = td->rate_next_io_time[odir] - now;
673 ddir = odir;
674 }
675 } else
676 usec = td->rate_next_io_time[ddir] - now;
677
678 if (td->o.io_submit_mode == IO_MODE_INLINE)
679 io_u_quiesce(td);
680
681 usec_sleep(td, usec);
682 return ddir;
683}
684
685/*
686 * Return the data direction for the next io_u. If the job is a
687 * mixed read/write workload, check the rwmix cycle and switch if
688 * necessary.
689 */
690static enum fio_ddir get_rw_ddir(struct thread_data *td)
691{
692 enum fio_ddir ddir;
693
694 /*
695 * See if it's time to fsync/fdatasync/sync_file_range first,
696 * and if not then move on to check regular I/Os.
697 */
698 if (should_fsync(td)) {
699 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
700 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
701 return DDIR_SYNC;
702
703 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
704 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
705 return DDIR_DATASYNC;
706
707 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
708 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
709 return DDIR_SYNC_FILE_RANGE;
710 }
711
712 if (td_rw(td)) {
713 /*
714 * Check if it's time to seed a new data direction.
715 */
716 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
717 /*
718 * Put a top limit on how many bytes we do for
719 * one data direction, to avoid overflowing the
720 * ranges too much
721 */
722 ddir = get_rand_ddir(td);
723
724 if (ddir != td->rwmix_ddir)
725 set_rwmix_bytes(td);
726
727 td->rwmix_ddir = ddir;
728 }
729 ddir = td->rwmix_ddir;
730 } else if (td_read(td))
731 ddir = DDIR_READ;
732 else if (td_write(td))
733 ddir = DDIR_WRITE;
734 else if (td_trim(td))
735 ddir = DDIR_TRIM;
736 else
737 ddir = DDIR_INVAL;
738
739 td->rwmix_ddir = rate_ddir(td, ddir);
740 return td->rwmix_ddir;
741}
742
743static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
744{
745 enum fio_ddir ddir = get_rw_ddir(td);
746
747 if (td_trimwrite(td)) {
748 struct fio_file *f = io_u->file;
749 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
750 ddir = DDIR_TRIM;
751 else
752 ddir = DDIR_WRITE;
753 }
754
755 io_u->ddir = io_u->acct_ddir = ddir;
756
757 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
758 td->o.barrier_blocks &&
759 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
760 td->io_issues[DDIR_WRITE])
761 io_u_set(td, io_u, IO_U_F_BARRIER);
762}
763
764void put_file_log(struct thread_data *td, struct fio_file *f)
765{
766 unsigned int ret = put_file(td, f);
767
768 if (ret)
769 td_verror(td, ret, "file close");
770}
771
772void put_io_u(struct thread_data *td, struct io_u *io_u)
773{
774 const bool needs_lock = td_async_processing(td);
775
776 if (io_u->post_submit) {
777 io_u->post_submit(io_u, io_u->error == 0);
778 io_u->post_submit = NULL;
779 }
780
781 if (td->parent)
782 td = td->parent;
783
784 if (needs_lock)
785 __td_io_u_lock(td);
786
787 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
788 put_file_log(td, io_u->file);
789
790 io_u->file = NULL;
791 io_u_set(td, io_u, IO_U_F_FREE);
792
793 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
794 td->cur_depth--;
795 assert(!(td->flags & TD_F_CHILD));
796 }
797 io_u_qpush(&td->io_u_freelist, io_u);
798 td_io_u_free_notify(td);
799
800 if (needs_lock)
801 __td_io_u_unlock(td);
802}
803
804void clear_io_u(struct thread_data *td, struct io_u *io_u)
805{
806 io_u_clear(td, io_u, IO_U_F_FLIGHT);
807 put_io_u(td, io_u);
808}
809
810void requeue_io_u(struct thread_data *td, struct io_u **io_u)
811{
812 const bool needs_lock = td_async_processing(td);
813 struct io_u *__io_u = *io_u;
814 enum fio_ddir ddir = acct_ddir(__io_u);
815
816 dprint(FD_IO, "requeue %p\n", __io_u);
817
818 if (td->parent)
819 td = td->parent;
820
821 if (needs_lock)
822 __td_io_u_lock(td);
823
824 io_u_set(td, __io_u, IO_U_F_FREE);
825 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
826 td->io_issues[ddir]--;
827
828 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
829 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
830 td->cur_depth--;
831 assert(!(td->flags & TD_F_CHILD));
832 }
833
834 io_u_rpush(&td->io_u_requeues, __io_u);
835 td_io_u_free_notify(td);
836
837 if (needs_lock)
838 __td_io_u_unlock(td);
839
840 *io_u = NULL;
841}
842
843static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u)
844{
845 struct fio_file *f = io_u->file;
846
847 assert(td->o.zone_mode == ZONE_MODE_STRIDED);
848 assert(td->o.zone_size);
849 assert(td->o.zone_range);
850
851 /*
852 * See if it's time to switch to a new zone
853 */
854 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
855 td->zone_bytes = 0;
856 f->file_offset += td->o.zone_range + td->o.zone_skip;
857
858 /*
859 * Wrap from the beginning, if we exceed the file size
860 */
861 if (f->file_offset >= f->real_file_size)
862 f->file_offset = get_start_offset(td, f);
863
864 f->last_pos[io_u->ddir] = f->file_offset;
865 td->io_skip_bytes += td->o.zone_skip;
866 }
867
868 /*
869 * If zone_size > zone_range, then maintain the same zone until
870 * zone_bytes >= zone_size.
871 */
872 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
873 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
874 f->file_offset, f->last_pos[io_u->ddir]);
875 f->last_pos[io_u->ddir] = f->file_offset;
876 }
877
878 /*
879 * For random: if 'norandommap' is not set and zone_size > zone_range,
880 * map needs to be reset as it's done with zone_range everytime.
881 */
882 if ((td->zone_bytes % td->o.zone_range) == 0)
883 fio_file_reset(td, f);
884}
885
886static int fill_io_u(struct thread_data *td, struct io_u *io_u)
887{
888 bool is_random;
889 uint64_t offset;
890 enum io_u_action ret;
891
892 if (td_ioengine_flagged(td, FIO_NOIO))
893 goto out;
894
895 set_rw_ddir(td, io_u);
896
897 /*
898 * fsync() or fdatasync() or trim etc, we are done
899 */
900 if (!ddir_rw(io_u->ddir))
901 goto out;
902
903 if (td->o.zone_mode == ZONE_MODE_STRIDED)
904 setup_strided_zone_mode(td, io_u);
905
906 /*
907 * No log, let the seq/rand engine retrieve the next buflen and
908 * position.
909 */
910 if (get_next_offset(td, io_u, &is_random)) {
911 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
912 return 1;
913 }
914
915 io_u->buflen = get_next_buflen(td, io_u, is_random);
916 if (!io_u->buflen) {
917 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
918 return 1;
919 }
920
921 offset = io_u->offset;
922 if (td->o.zone_mode == ZONE_MODE_ZBD) {
923 ret = zbd_adjust_block(td, io_u);
924 if (ret == io_u_eof)
925 return 1;
926 }
927
928 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
929 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n",
930 io_u,
931 (unsigned long long) io_u->offset, io_u->buflen,
932 (unsigned long long) io_u->file->real_file_size);
933 return 1;
934 }
935
936 /*
937 * mark entry before potentially trimming io_u
938 */
939 if (td_random(td) && file_randommap(td, io_u->file))
940 io_u->buflen = mark_random_map(td, io_u, offset, io_u->buflen);
941
942out:
943 dprint_io_u(io_u, "fill");
944 td->zone_bytes += io_u->buflen;
945 return 0;
946}
947
948static void __io_u_mark_map(uint64_t *map, unsigned int nr)
949{
950 int idx = 0;
951
952 switch (nr) {
953 default:
954 idx = 6;
955 break;
956 case 33 ... 64:
957 idx = 5;
958 break;
959 case 17 ... 32:
960 idx = 4;
961 break;
962 case 9 ... 16:
963 idx = 3;
964 break;
965 case 5 ... 8:
966 idx = 2;
967 break;
968 case 1 ... 4:
969 idx = 1;
970 case 0:
971 break;
972 }
973
974 map[idx]++;
975}
976
977void io_u_mark_submit(struct thread_data *td, unsigned int nr)
978{
979 __io_u_mark_map(td->ts.io_u_submit, nr);
980 td->ts.total_submit++;
981}
982
983void io_u_mark_complete(struct thread_data *td, unsigned int nr)
984{
985 __io_u_mark_map(td->ts.io_u_complete, nr);
986 td->ts.total_complete++;
987}
988
989void io_u_mark_depth(struct thread_data *td, unsigned int nr)
990{
991 int idx = 0;
992
993 switch (td->cur_depth) {
994 default:
995 idx = 6;
996 break;
997 case 32 ... 63:
998 idx = 5;
999 break;
1000 case 16 ... 31:
1001 idx = 4;
1002 break;
1003 case 8 ... 15:
1004 idx = 3;
1005 break;
1006 case 4 ... 7:
1007 idx = 2;
1008 break;
1009 case 2 ... 3:
1010 idx = 1;
1011 case 1:
1012 break;
1013 }
1014
1015 td->ts.io_u_map[idx] += nr;
1016}
1017
1018static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1019{
1020 int idx = 0;
1021
1022 assert(nsec < 1000);
1023
1024 switch (nsec) {
1025 case 750 ... 999:
1026 idx = 9;
1027 break;
1028 case 500 ... 749:
1029 idx = 8;
1030 break;
1031 case 250 ... 499:
1032 idx = 7;
1033 break;
1034 case 100 ... 249:
1035 idx = 6;
1036 break;
1037 case 50 ... 99:
1038 idx = 5;
1039 break;
1040 case 20 ... 49:
1041 idx = 4;
1042 break;
1043 case 10 ... 19:
1044 idx = 3;
1045 break;
1046 case 4 ... 9:
1047 idx = 2;
1048 break;
1049 case 2 ... 3:
1050 idx = 1;
1051 case 0 ... 1:
1052 break;
1053 }
1054
1055 assert(idx < FIO_IO_U_LAT_N_NR);
1056 td->ts.io_u_lat_n[idx]++;
1057}
1058
1059static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1060{
1061 int idx = 0;
1062
1063 assert(usec < 1000 && usec >= 1);
1064
1065 switch (usec) {
1066 case 750 ... 999:
1067 idx = 9;
1068 break;
1069 case 500 ... 749:
1070 idx = 8;
1071 break;
1072 case 250 ... 499:
1073 idx = 7;
1074 break;
1075 case 100 ... 249:
1076 idx = 6;
1077 break;
1078 case 50 ... 99:
1079 idx = 5;
1080 break;
1081 case 20 ... 49:
1082 idx = 4;
1083 break;
1084 case 10 ... 19:
1085 idx = 3;
1086 break;
1087 case 4 ... 9:
1088 idx = 2;
1089 break;
1090 case 2 ... 3:
1091 idx = 1;
1092 case 0 ... 1:
1093 break;
1094 }
1095
1096 assert(idx < FIO_IO_U_LAT_U_NR);
1097 td->ts.io_u_lat_u[idx]++;
1098}
1099
1100static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1101{
1102 int idx = 0;
1103
1104 assert(msec >= 1);
1105
1106 switch (msec) {
1107 default:
1108 idx = 11;
1109 break;
1110 case 1000 ... 1999:
1111 idx = 10;
1112 break;
1113 case 750 ... 999:
1114 idx = 9;
1115 break;
1116 case 500 ... 749:
1117 idx = 8;
1118 break;
1119 case 250 ... 499:
1120 idx = 7;
1121 break;
1122 case 100 ... 249:
1123 idx = 6;
1124 break;
1125 case 50 ... 99:
1126 idx = 5;
1127 break;
1128 case 20 ... 49:
1129 idx = 4;
1130 break;
1131 case 10 ... 19:
1132 idx = 3;
1133 break;
1134 case 4 ... 9:
1135 idx = 2;
1136 break;
1137 case 2 ... 3:
1138 idx = 1;
1139 case 0 ... 1:
1140 break;
1141 }
1142
1143 assert(idx < FIO_IO_U_LAT_M_NR);
1144 td->ts.io_u_lat_m[idx]++;
1145}
1146
1147static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1148{
1149 if (nsec < 1000)
1150 io_u_mark_lat_nsec(td, nsec);
1151 else if (nsec < 1000000)
1152 io_u_mark_lat_usec(td, nsec / 1000);
1153 else
1154 io_u_mark_lat_msec(td, nsec / 1000000);
1155}
1156
1157static unsigned int __get_next_fileno_rand(struct thread_data *td)
1158{
1159 unsigned long fileno;
1160
1161 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1162 uint64_t frand_max = rand_max(&td->next_file_state);
1163 unsigned long r;
1164
1165 r = __rand(&td->next_file_state);
1166 return (unsigned int) ((double) td->o.nr_files
1167 * (r / (frand_max + 1.0)));
1168 }
1169
1170 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1171 fileno = zipf_next(&td->next_file_zipf);
1172 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1173 fileno = pareto_next(&td->next_file_zipf);
1174 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1175 fileno = gauss_next(&td->next_file_gauss);
1176 else {
1177 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1178 assert(0);
1179 return 0;
1180 }
1181
1182 return fileno >> FIO_FSERVICE_SHIFT;
1183}
1184
1185/*
1186 * Get next file to service by choosing one at random
1187 */
1188static struct fio_file *get_next_file_rand(struct thread_data *td,
1189 enum fio_file_flags goodf,
1190 enum fio_file_flags badf)
1191{
1192 struct fio_file *f;
1193 int fno;
1194
1195 do {
1196 int opened = 0;
1197
1198 fno = __get_next_fileno_rand(td);
1199
1200 f = td->files[fno];
1201 if (fio_file_done(f))
1202 continue;
1203
1204 if (!fio_file_open(f)) {
1205 int err;
1206
1207 if (td->nr_open_files >= td->o.open_files)
1208 return ERR_PTR(-EBUSY);
1209
1210 err = td_io_open_file(td, f);
1211 if (err)
1212 continue;
1213 opened = 1;
1214 }
1215
1216 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1217 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1218 return f;
1219 }
1220 if (opened)
1221 td_io_close_file(td, f);
1222 } while (1);
1223}
1224
1225/*
1226 * Get next file to service by doing round robin between all available ones
1227 */
1228static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1229 int badf)
1230{
1231 unsigned int old_next_file = td->next_file;
1232 struct fio_file *f;
1233
1234 do {
1235 int opened = 0;
1236
1237 f = td->files[td->next_file];
1238
1239 td->next_file++;
1240 if (td->next_file >= td->o.nr_files)
1241 td->next_file = 0;
1242
1243 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1244 if (fio_file_done(f)) {
1245 f = NULL;
1246 continue;
1247 }
1248
1249 if (!fio_file_open(f)) {
1250 int err;
1251
1252 if (td->nr_open_files >= td->o.open_files)
1253 return ERR_PTR(-EBUSY);
1254
1255 err = td_io_open_file(td, f);
1256 if (err) {
1257 dprint(FD_FILE, "error %d on open of %s\n",
1258 err, f->file_name);
1259 f = NULL;
1260 continue;
1261 }
1262 opened = 1;
1263 }
1264
1265 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1266 f->flags);
1267 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1268 break;
1269
1270 if (opened)
1271 td_io_close_file(td, f);
1272
1273 f = NULL;
1274 } while (td->next_file != old_next_file);
1275
1276 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1277 return f;
1278}
1279
1280static struct fio_file *__get_next_file(struct thread_data *td)
1281{
1282 struct fio_file *f;
1283
1284 assert(td->o.nr_files <= td->files_index);
1285
1286 if (td->nr_done_files >= td->o.nr_files) {
1287 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1288 " nr_files=%d\n", td->nr_open_files,
1289 td->nr_done_files,
1290 td->o.nr_files);
1291 return NULL;
1292 }
1293
1294 f = td->file_service_file;
1295 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1296 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1297 goto out;
1298 if (td->file_service_left--)
1299 goto out;
1300 }
1301
1302 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1303 td->o.file_service_type == FIO_FSERVICE_SEQ)
1304 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1305 else
1306 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1307
1308 if (IS_ERR(f))
1309 return f;
1310
1311 td->file_service_file = f;
1312 td->file_service_left = td->file_service_nr - 1;
1313out:
1314 if (f)
1315 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1316 else
1317 dprint(FD_FILE, "get_next_file: NULL\n");
1318 return f;
1319}
1320
1321static struct fio_file *get_next_file(struct thread_data *td)
1322{
1323 return __get_next_file(td);
1324}
1325
1326static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1327{
1328 struct fio_file *f;
1329
1330 do {
1331 f = get_next_file(td);
1332 if (IS_ERR_OR_NULL(f))
1333 return PTR_ERR(f);
1334
1335 io_u->file = f;
1336 get_file(f);
1337
1338 if (!fill_io_u(td, io_u))
1339 break;
1340
1341 if (io_u->post_submit) {
1342 io_u->post_submit(io_u, false);
1343 io_u->post_submit = NULL;
1344 }
1345
1346 put_file_log(td, f);
1347 td_io_close_file(td, f);
1348 io_u->file = NULL;
1349 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1350 fio_file_reset(td, f);
1351 else {
1352 fio_file_set_done(f);
1353 td->nr_done_files++;
1354 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1355 td->nr_done_files, td->o.nr_files);
1356 }
1357 } while (1);
1358
1359 return 0;
1360}
1361
1362static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1363 unsigned long long tnsec, unsigned long long max_nsec)
1364{
1365 if (!td->error)
1366 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
1367 td_verror(td, ETIMEDOUT, "max latency exceeded");
1368 icd->error = ETIMEDOUT;
1369}
1370
1371static void lat_new_cycle(struct thread_data *td)
1372{
1373 fio_gettime(&td->latency_ts, NULL);
1374 td->latency_ios = ddir_rw_sum(td->io_blocks);
1375 td->latency_failed = 0;
1376}
1377
1378/*
1379 * We had an IO outside the latency target. Reduce the queue depth. If we
1380 * are at QD=1, then it's time to give up.
1381 */
1382static bool __lat_target_failed(struct thread_data *td)
1383{
1384 if (td->latency_qd == 1)
1385 return true;
1386
1387 td->latency_qd_high = td->latency_qd;
1388
1389 if (td->latency_qd == td->latency_qd_low)
1390 td->latency_qd_low--;
1391
1392 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1393
1394 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1395
1396 /*
1397 * When we ramp QD down, quiesce existing IO to prevent
1398 * a storm of ramp downs due to pending higher depth.
1399 */
1400 io_u_quiesce(td);
1401 lat_new_cycle(td);
1402 return false;
1403}
1404
1405static bool lat_target_failed(struct thread_data *td)
1406{
1407 if (td->o.latency_percentile.u.f == 100.0)
1408 return __lat_target_failed(td);
1409
1410 td->latency_failed++;
1411 return false;
1412}
1413
1414void lat_target_init(struct thread_data *td)
1415{
1416 td->latency_end_run = 0;
1417
1418 if (td->o.latency_target) {
1419 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1420 fio_gettime(&td->latency_ts, NULL);
1421 td->latency_qd = 1;
1422 td->latency_qd_high = td->o.iodepth;
1423 td->latency_qd_low = 1;
1424 td->latency_ios = ddir_rw_sum(td->io_blocks);
1425 } else
1426 td->latency_qd = td->o.iodepth;
1427}
1428
1429void lat_target_reset(struct thread_data *td)
1430{
1431 if (!td->latency_end_run)
1432 lat_target_init(td);
1433}
1434
1435static void lat_target_success(struct thread_data *td)
1436{
1437 const unsigned int qd = td->latency_qd;
1438 struct thread_options *o = &td->o;
1439
1440 td->latency_qd_low = td->latency_qd;
1441
1442 /*
1443 * If we haven't failed yet, we double up to a failing value instead
1444 * of bisecting from highest possible queue depth. If we have set
1445 * a limit other than td->o.iodepth, bisect between that.
1446 */
1447 if (td->latency_qd_high != o->iodepth)
1448 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1449 else
1450 td->latency_qd *= 2;
1451
1452 if (td->latency_qd > o->iodepth)
1453 td->latency_qd = o->iodepth;
1454
1455 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1456
1457 /*
1458 * Same as last one, we are done. Let it run a latency cycle, so
1459 * we get only the results from the targeted depth.
1460 */
1461 if (td->latency_qd == qd) {
1462 if (td->latency_end_run) {
1463 dprint(FD_RATE, "We are done\n");
1464 td->done = 1;
1465 } else {
1466 dprint(FD_RATE, "Quiesce and final run\n");
1467 io_u_quiesce(td);
1468 td->latency_end_run = 1;
1469 reset_all_stats(td);
1470 reset_io_stats(td);
1471 }
1472 }
1473
1474 lat_new_cycle(td);
1475}
1476
1477/*
1478 * Check if we can bump the queue depth
1479 */
1480void lat_target_check(struct thread_data *td)
1481{
1482 uint64_t usec_window;
1483 uint64_t ios;
1484 double success_ios;
1485
1486 usec_window = utime_since_now(&td->latency_ts);
1487 if (usec_window < td->o.latency_window)
1488 return;
1489
1490 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1491 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1492 success_ios *= 100.0;
1493
1494 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1495
1496 if (success_ios >= td->o.latency_percentile.u.f)
1497 lat_target_success(td);
1498 else
1499 __lat_target_failed(td);
1500}
1501
1502/*
1503 * If latency target is enabled, we might be ramping up or down and not
1504 * using the full queue depth available.
1505 */
1506bool queue_full(const struct thread_data *td)
1507{
1508 const int qempty = io_u_qempty(&td->io_u_freelist);
1509
1510 if (qempty)
1511 return true;
1512 if (!td->o.latency_target)
1513 return false;
1514
1515 return td->cur_depth >= td->latency_qd;
1516}
1517
1518struct io_u *__get_io_u(struct thread_data *td)
1519{
1520 const bool needs_lock = td_async_processing(td);
1521 struct io_u *io_u = NULL;
1522 int ret;
1523
1524 if (td->stop_io)
1525 return NULL;
1526
1527 if (needs_lock)
1528 __td_io_u_lock(td);
1529
1530again:
1531 if (!io_u_rempty(&td->io_u_requeues))
1532 io_u = io_u_rpop(&td->io_u_requeues);
1533 else if (!queue_full(td)) {
1534 io_u = io_u_qpop(&td->io_u_freelist);
1535
1536 io_u->file = NULL;
1537 io_u->buflen = 0;
1538 io_u->resid = 0;
1539 io_u->end_io = NULL;
1540 }
1541
1542 if (io_u) {
1543 assert(io_u->flags & IO_U_F_FREE);
1544 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1545 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1546 IO_U_F_VER_LIST);
1547
1548 io_u->error = 0;
1549 io_u->acct_ddir = -1;
1550 td->cur_depth++;
1551 assert(!(td->flags & TD_F_CHILD));
1552 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1553 io_u->ipo = NULL;
1554 } else if (td_async_processing(td)) {
1555 /*
1556 * We ran out, wait for async verify threads to finish and
1557 * return one
1558 */
1559 assert(!(td->flags & TD_F_CHILD));
1560 ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1561 assert(ret == 0);
1562 if (!td->error)
1563 goto again;
1564 }
1565
1566 if (needs_lock)
1567 __td_io_u_unlock(td);
1568
1569 return io_u;
1570}
1571
1572static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1573{
1574 if (!(td->flags & TD_F_TRIM_BACKLOG))
1575 return false;
1576 if (!td->trim_entries)
1577 return false;
1578
1579 if (td->trim_batch) {
1580 td->trim_batch--;
1581 if (get_next_trim(td, io_u))
1582 return true;
1583 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1584 td->last_ddir != DDIR_READ) {
1585 td->trim_batch = td->o.trim_batch;
1586 if (!td->trim_batch)
1587 td->trim_batch = td->o.trim_backlog;
1588 if (get_next_trim(td, io_u))
1589 return true;
1590 }
1591
1592 return false;
1593}
1594
1595static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1596{
1597 if (!(td->flags & TD_F_VER_BACKLOG))
1598 return false;
1599
1600 if (td->io_hist_len) {
1601 int get_verify = 0;
1602
1603 if (td->verify_batch)
1604 get_verify = 1;
1605 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1606 td->last_ddir != DDIR_READ) {
1607 td->verify_batch = td->o.verify_batch;
1608 if (!td->verify_batch)
1609 td->verify_batch = td->o.verify_backlog;
1610 get_verify = 1;
1611 }
1612
1613 if (get_verify && !get_next_verify(td, io_u)) {
1614 td->verify_batch--;
1615 return true;
1616 }
1617 }
1618
1619 return false;
1620}
1621
1622/*
1623 * Fill offset and start time into the buffer content, to prevent too
1624 * easy compressible data for simple de-dupe attempts. Do this for every
1625 * 512b block in the range, since that should be the smallest block size
1626 * we can expect from a device.
1627 */
1628static void small_content_scramble(struct io_u *io_u)
1629{
1630 unsigned long long i, nr_blocks = io_u->buflen >> 9;
1631 unsigned int offset;
1632 uint64_t boffset, *iptr;
1633 char *p;
1634
1635 if (!nr_blocks)
1636 return;
1637
1638 p = io_u->xfer_buf;
1639 boffset = io_u->offset;
1640
1641 if (io_u->buf_filled_len)
1642 io_u->buf_filled_len = 0;
1643
1644 /*
1645 * Generate random index between 0..7. We do chunks of 512b, if
1646 * we assume a cacheline is 64 bytes, then we have 8 of those.
1647 * Scramble content within the blocks in the same cacheline to
1648 * speed things up.
1649 */
1650 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1651
1652 for (i = 0; i < nr_blocks; i++) {
1653 /*
1654 * Fill offset into start of cacheline, time into end
1655 * of cacheline
1656 */
1657 iptr = (void *) p + (offset << 6);
1658 *iptr = boffset;
1659
1660 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1661 iptr[0] = io_u->start_time.tv_sec;
1662 iptr[1] = io_u->start_time.tv_nsec;
1663
1664 p += 512;
1665 boffset += 512;
1666 }
1667}
1668
1669/*
1670 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1671 * etc. The returned io_u is fully ready to be prepped, populated and submitted.
1672 */
1673struct io_u *get_io_u(struct thread_data *td)
1674{
1675 struct fio_file *f;
1676 struct io_u *io_u;
1677 int do_scramble = 0;
1678 long ret = 0;
1679
1680 io_u = __get_io_u(td);
1681 if (!io_u) {
1682 dprint(FD_IO, "__get_io_u failed\n");
1683 return NULL;
1684 }
1685
1686 if (check_get_verify(td, io_u))
1687 goto out;
1688 if (check_get_trim(td, io_u))
1689 goto out;
1690
1691 /*
1692 * from a requeue, io_u already setup
1693 */
1694 if (io_u->file)
1695 goto out;
1696
1697 /*
1698 * If using an iolog, grab next piece if any available.
1699 */
1700 if (td->flags & TD_F_READ_IOLOG) {
1701 if (read_iolog_get(td, io_u))
1702 goto err_put;
1703 } else if (set_io_u_file(td, io_u)) {
1704 ret = -EBUSY;
1705 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1706 goto err_put;
1707 }
1708
1709 f = io_u->file;
1710 if (!f) {
1711 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1712 goto err_put;
1713 }
1714
1715 assert(fio_file_open(f));
1716
1717 if (ddir_rw(io_u->ddir)) {
1718 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1719 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1720 goto err_put;
1721 }
1722
1723 f->last_start[io_u->ddir] = io_u->offset;
1724 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1725
1726 if (io_u->ddir == DDIR_WRITE) {
1727 if (td->flags & TD_F_REFILL_BUFFERS) {
1728 io_u_fill_buffer(td, io_u,
1729 td->o.min_bs[DDIR_WRITE],
1730 io_u->buflen);
1731 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1732 !(td->flags & TD_F_COMPRESS) &&
1733 !(td->flags & TD_F_DO_VERIFY))
1734 do_scramble = 1;
1735 } else if (io_u->ddir == DDIR_READ) {
1736 /*
1737 * Reset the buf_filled parameters so next time if the
1738 * buffer is used for writes it is refilled.
1739 */
1740 io_u->buf_filled_len = 0;
1741 }
1742 }
1743
1744 /*
1745 * Set io data pointers.
1746 */
1747 io_u->xfer_buf = io_u->buf;
1748 io_u->xfer_buflen = io_u->buflen;
1749
1750out:
1751 assert(io_u->file);
1752 if (!td_io_prep(td, io_u)) {
1753 if (!td->o.disable_lat)
1754 fio_gettime(&io_u->start_time, NULL);
1755
1756 if (do_scramble)
1757 small_content_scramble(io_u);
1758
1759 return io_u;
1760 }
1761err_put:
1762 dprint(FD_IO, "get_io_u failed\n");
1763 put_io_u(td, io_u);
1764 return ERR_PTR(ret);
1765}
1766
1767static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1768{
1769 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1770
1771 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1772 return;
1773
1774 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n",
1775 io_u->file ? " on file " : "",
1776 io_u->file ? io_u->file->file_name : "",
1777 strerror(io_u->error),
1778 io_ddir_name(io_u->ddir),
1779 io_u->offset, io_u->xfer_buflen);
1780
1781 if (td->io_ops->errdetails) {
1782 char *err = td->io_ops->errdetails(io_u);
1783
1784 log_err("fio: %s\n", err);
1785 free(err);
1786 }
1787
1788 if (!td->error)
1789 td_verror(td, io_u->error, "io_u error");
1790}
1791
1792void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1793{
1794 __io_u_log_error(td, io_u);
1795 if (td->parent)
1796 __io_u_log_error(td->parent, io_u);
1797}
1798
1799static inline bool gtod_reduce(struct thread_data *td)
1800{
1801 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1802 || td->o.gtod_reduce;
1803}
1804
1805static void trim_block_info(struct thread_data *td, struct io_u *io_u)
1806{
1807 uint32_t *info = io_u_block_info(td, io_u);
1808
1809 if (BLOCK_INFO_STATE(*info) >= BLOCK_STATE_TRIM_FAILURE)
1810 return;
1811
1812 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, BLOCK_INFO_TRIMS(*info) + 1);
1813}
1814
1815static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1816 struct io_completion_data *icd,
1817 const enum fio_ddir idx, unsigned int bytes)
1818{
1819 const int no_reduce = !gtod_reduce(td);
1820 unsigned long long llnsec = 0;
1821
1822 if (td->parent)
1823 td = td->parent;
1824
1825 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1826 return;
1827
1828 if (no_reduce)
1829 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1830
1831 if (!td->o.disable_lat) {
1832 unsigned long long tnsec;
1833
1834 tnsec = ntime_since(&io_u->start_time, &icd->time);
1835 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
1836
1837 if (td->flags & TD_F_PROFILE_OPS) {
1838 struct prof_io_ops *ops = &td->prof_io_ops;
1839
1840 if (ops->io_u_lat)
1841 icd->error = ops->io_u_lat(td, tnsec);
1842 }
1843
1844 if (td->o.max_latency && tnsec > td->o.max_latency)
1845 lat_fatal(td, icd, tnsec, td->o.max_latency);
1846 if (td->o.latency_target && tnsec > td->o.latency_target) {
1847 if (lat_target_failed(td))
1848 lat_fatal(td, icd, tnsec, td->o.latency_target);
1849 }
1850 }
1851
1852 if (ddir_rw(idx)) {
1853 if (!td->o.disable_clat) {
1854 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1855 io_u_mark_latency(td, llnsec);
1856 }
1857
1858 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1859 add_bw_sample(td, io_u, bytes, llnsec);
1860
1861 if (no_reduce && per_unit_log(td->iops_log))
1862 add_iops_sample(td, io_u, bytes);
1863 } else if (ddir_sync(idx) && !td->o.disable_clat)
1864 add_sync_clat_sample(&td->ts, llnsec);
1865
1866 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM)
1867 trim_block_info(td, io_u);
1868}
1869
1870static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1871 uint64_t offset, unsigned int bytes)
1872{
1873 int idx;
1874
1875 if (!f)
1876 return;
1877
1878 if (f->first_write == -1ULL || offset < f->first_write)
1879 f->first_write = offset;
1880 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1881 f->last_write = offset + bytes;
1882
1883 if (!f->last_write_comp)
1884 return;
1885
1886 idx = f->last_write_idx++;
1887 f->last_write_comp[idx] = offset;
1888 if (f->last_write_idx == td->o.iodepth)
1889 f->last_write_idx = 0;
1890}
1891
1892static bool should_account(struct thread_data *td)
1893{
1894 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1895 td->runstate == TD_VERIFYING);
1896}
1897
1898static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1899 struct io_completion_data *icd)
1900{
1901 struct io_u *io_u = *io_u_ptr;
1902 enum fio_ddir ddir = io_u->ddir;
1903 struct fio_file *f = io_u->file;
1904
1905 dprint_io_u(io_u, "complete");
1906
1907 assert(io_u->flags & IO_U_F_FLIGHT);
1908 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1909
1910 /*
1911 * Mark IO ok to verify
1912 */
1913 if (io_u->ipo) {
1914 /*
1915 * Remove errored entry from the verification list
1916 */
1917 if (io_u->error)
1918 unlog_io_piece(td, io_u);
1919 else {
1920 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1921 write_barrier();
1922 }
1923 }
1924
1925 if (ddir_sync(ddir)) {
1926 td->last_was_sync = true;
1927 if (f) {
1928 f->first_write = -1ULL;
1929 f->last_write = -1ULL;
1930 }
1931 if (should_account(td))
1932 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
1933 return;
1934 }
1935
1936 td->last_was_sync = false;
1937 td->last_ddir = ddir;
1938
1939 if (!io_u->error && ddir_rw(ddir)) {
1940 unsigned long long bytes = io_u->buflen - io_u->resid;
1941 int ret;
1942
1943 td->io_blocks[ddir]++;
1944 td->io_bytes[ddir] += bytes;
1945
1946 if (!(io_u->flags & IO_U_F_VER_LIST)) {
1947 td->this_io_blocks[ddir]++;
1948 td->this_io_bytes[ddir] += bytes;
1949 }
1950
1951 if (ddir == DDIR_WRITE)
1952 file_log_write_comp(td, f, io_u->offset, bytes);
1953
1954 if (should_account(td))
1955 account_io_completion(td, io_u, icd, ddir, bytes);
1956
1957 icd->bytes_done[ddir] += bytes;
1958
1959 if (io_u->end_io) {
1960 ret = io_u->end_io(td, io_u_ptr);
1961 io_u = *io_u_ptr;
1962 if (ret && !icd->error)
1963 icd->error = ret;
1964 }
1965 } else if (io_u->error) {
1966 icd->error = io_u->error;
1967 io_u_log_error(td, io_u);
1968 }
1969 if (icd->error) {
1970 enum error_type_bit eb = td_error_type(ddir, icd->error);
1971
1972 if (!td_non_fatal_error(td, eb, icd->error))
1973 return;
1974
1975 /*
1976 * If there is a non_fatal error, then add to the error count
1977 * and clear all the errors.
1978 */
1979 update_error_count(td, icd->error);
1980 td_clear_error(td);
1981 icd->error = 0;
1982 if (io_u)
1983 io_u->error = 0;
1984 }
1985}
1986
1987static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1988 int nr)
1989{
1990 int ddir;
1991
1992 if (!gtod_reduce(td))
1993 fio_gettime(&icd->time, NULL);
1994
1995 icd->nr = nr;
1996
1997 icd->error = 0;
1998 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1999 icd->bytes_done[ddir] = 0;
2000}
2001
2002static void ios_completed(struct thread_data *td,
2003 struct io_completion_data *icd)
2004{
2005 struct io_u *io_u;
2006 int i;
2007
2008 for (i = 0; i < icd->nr; i++) {
2009 io_u = td->io_ops->event(td, i);
2010
2011 io_completed(td, &io_u, icd);
2012
2013 if (io_u)
2014 put_io_u(td, io_u);
2015 }
2016}
2017
2018/*
2019 * Complete a single io_u for the sync engines.
2020 */
2021int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
2022{
2023 struct io_completion_data icd;
2024 int ddir;
2025
2026 init_icd(td, &icd, 1);
2027 io_completed(td, &io_u, &icd);
2028
2029 if (io_u)
2030 put_io_u(td, io_u);
2031
2032 if (icd.error) {
2033 td_verror(td, icd.error, "io_u_sync_complete");
2034 return -1;
2035 }
2036
2037 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2038 td->bytes_done[ddir] += icd.bytes_done[ddir];
2039
2040 return 0;
2041}
2042
2043/*
2044 * Called to complete min_events number of io for the async engines.
2045 */
2046int io_u_queued_complete(struct thread_data *td, int min_evts)
2047{
2048 struct io_completion_data icd;
2049 struct timespec *tvp = NULL;
2050 int ret, ddir;
2051 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2052
2053 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2054
2055 if (!min_evts)
2056 tvp = &ts;
2057 else if (min_evts > td->cur_depth)
2058 min_evts = td->cur_depth;
2059
2060 /* No worries, td_io_getevents fixes min and max if they are
2061 * set incorrectly */
2062 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2063 if (ret < 0) {
2064 td_verror(td, -ret, "td_io_getevents");
2065 return ret;
2066 } else if (!ret)
2067 return ret;
2068
2069 init_icd(td, &icd, ret);
2070 ios_completed(td, &icd);
2071 if (icd.error) {
2072 td_verror(td, icd.error, "io_u_queued_complete");
2073 return -1;
2074 }
2075
2076 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2077 td->bytes_done[ddir] += icd.bytes_done[ddir];
2078
2079 return ret;
2080}
2081
2082/*
2083 * Call when io_u is really queued, to update the submission latency.
2084 */
2085void io_u_queued(struct thread_data *td, struct io_u *io_u)
2086{
2087 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2088 unsigned long slat_time;
2089
2090 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2091
2092 if (td->parent)
2093 td = td->parent;
2094
2095 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2096 io_u->offset);
2097 }
2098}
2099
2100/*
2101 * See if we should reuse the last seed, if dedupe is enabled
2102 */
2103static struct frand_state *get_buf_state(struct thread_data *td)
2104{
2105 unsigned int v;
2106
2107 if (!td->o.dedupe_percentage)
2108 return &td->buf_state;
2109 else if (td->o.dedupe_percentage == 100) {
2110 frand_copy(&td->buf_state_prev, &td->buf_state);
2111 return &td->buf_state;
2112 }
2113
2114 v = rand_between(&td->dedupe_state, 1, 100);
2115
2116 if (v <= td->o.dedupe_percentage)
2117 return &td->buf_state_prev;
2118
2119 return &td->buf_state;
2120}
2121
2122static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2123{
2124 if (td->o.dedupe_percentage == 100)
2125 frand_copy(rs, &td->buf_state_prev);
2126 else if (rs == &td->buf_state)
2127 frand_copy(&td->buf_state_prev, rs);
2128}
2129
2130void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write,
2131 unsigned long long max_bs)
2132{
2133 struct thread_options *o = &td->o;
2134
2135 if (o->mem_type == MEM_CUDA_MALLOC)
2136 return;
2137
2138 if (o->compress_percentage || o->dedupe_percentage) {
2139 unsigned int perc = td->o.compress_percentage;
2140 struct frand_state *rs;
2141 unsigned long long left = max_bs;
2142 unsigned long long this_write;
2143
2144 do {
2145 rs = get_buf_state(td);
2146
2147 min_write = min(min_write, left);
2148
2149 if (perc) {
2150 this_write = min_not_zero(min_write,
2151 (unsigned long long) td->o.compress_chunk);
2152
2153 fill_random_buf_percentage(rs, buf, perc,
2154 this_write, this_write,
2155 o->buffer_pattern,
2156 o->buffer_pattern_bytes);
2157 } else {
2158 fill_random_buf(rs, buf, min_write);
2159 this_write = min_write;
2160 }
2161
2162 buf += this_write;
2163 left -= this_write;
2164 save_buf_state(td, rs);
2165 } while (left);
2166 } else if (o->buffer_pattern_bytes)
2167 fill_buffer_pattern(td, buf, max_bs);
2168 else if (o->zero_buffers)
2169 memset(buf, 0, max_bs);
2170 else
2171 fill_random_buf(get_buf_state(td), buf, max_bs);
2172}
2173
2174/*
2175 * "randomly" fill the buffer contents
2176 */
2177void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2178 unsigned long long min_write, unsigned long long max_bs)
2179{
2180 io_u->buf_filled_len = 0;
2181 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2182}
2183
2184static int do_sync_file_range(const struct thread_data *td,
2185 struct fio_file *f)
2186{
2187 off64_t offset, nbytes;
2188
2189 offset = f->first_write;
2190 nbytes = f->last_write - f->first_write;
2191
2192 if (!nbytes)
2193 return 0;
2194
2195 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2196}
2197
2198int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2199{
2200 int ret;
2201
2202 if (io_u->ddir == DDIR_SYNC) {
2203 ret = fsync(io_u->file->fd);
2204 } else if (io_u->ddir == DDIR_DATASYNC) {
2205#ifdef CONFIG_FDATASYNC
2206 ret = fdatasync(io_u->file->fd);
2207#else
2208 ret = io_u->xfer_buflen;
2209 io_u->error = EINVAL;
2210#endif
2211 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2212 ret = do_sync_file_range(td, io_u->file);
2213 else {
2214 ret = io_u->xfer_buflen;
2215 io_u->error = EINVAL;
2216 }
2217
2218 if (ret < 0)
2219 io_u->error = errno;
2220
2221 return ret;
2222}
2223
2224int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2225{
2226#ifndef FIO_HAVE_TRIM
2227 io_u->error = EINVAL;
2228 return 0;
2229#else
2230 struct fio_file *f = io_u->file;
2231 int ret;
2232
2233 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2234 if (!ret)
2235 return io_u->xfer_buflen;
2236
2237 io_u->error = ret;
2238 return 0;
2239#endif
2240}
fio - Flexible IO Tester