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