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