projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
null: drop unneeded casts from void* to non-void*
[fio.git] / io_u.c
... / ...
CommitLineData
1#include <unistd.h>
2#include <fcntl.h>
3#include <string.h>
4#include <signal.h>
5#include <time.h>
6#include <assert.h>
7
8#include "fio.h"
9#include "hash.h"
10#include "verify.h"
11#include "trim.h"
12#include "lib/rand.h"
13#include "lib/axmap.h"
14#include "err.h"
15#include "lib/pow2.h"
16#include "minmax.h"
17
18struct io_completion_data {
19 int nr; /* input */
20
21 int error; /* output */
22 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
23 struct timespec time; /* output */
24};
25
26/*
27 * The ->io_axmap contains a map of blocks we have or have not done io
28 * to yet. Used to make sure we cover the entire range in a fair fashion.
29 */
30static bool random_map_free(struct fio_file *f, const uint64_t block)
31{
32 return !axmap_isset(f->io_axmap, block);
33}
34
35/*
36 * Mark a given offset as used in the map.
37 */
38static void mark_random_map(struct thread_data *td, struct io_u *io_u)
39{
40 unsigned int min_bs = td->o.min_bs[io_u->ddir];
41 struct fio_file *f = io_u->file;
42 unsigned int nr_blocks;
43 uint64_t block;
44
45 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
46 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
47
48 if (!(io_u->flags & IO_U_F_BUSY_OK))
49 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
50
51 if ((nr_blocks * min_bs) < io_u->buflen)
52 io_u->buflen = nr_blocks * min_bs;
53}
54
55static uint64_t last_block(struct thread_data *td, struct fio_file *f,
56 enum fio_ddir ddir)
57{
58 uint64_t max_blocks;
59 uint64_t max_size;
60
61 assert(ddir_rw(ddir));
62
63 /*
64 * Hmm, should we make sure that ->io_size <= ->real_file_size?
65 * -> not for now since there is code assuming it could go either.
66 */
67 max_size = f->io_size;
68 if (max_size > f->real_file_size)
69 max_size = f->real_file_size;
70
71 if (td->o.zone_range)
72 max_size = td->o.zone_range;
73
74 if (td->o.min_bs[ddir] > td->o.ba[ddir])
75 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
76
77 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
78 if (!max_blocks)
79 return 0;
80
81 return max_blocks;
82}
83
84struct rand_off {
85 struct flist_head list;
86 uint64_t off;
87};
88
89static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
90 enum fio_ddir ddir, uint64_t *b,
91 uint64_t lastb)
92{
93 uint64_t r;
94
95 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
96 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
97
98 r = __rand(&td->random_state);
99
100 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
101
102 *b = lastb * (r / (rand_max(&td->random_state) + 1.0));
103 } else {
104 uint64_t off = 0;
105
106 assert(fio_file_lfsr(f));
107
108 if (lfsr_next(&f->lfsr, &off))
109 return 1;
110
111 *b = off;
112 }
113
114 /*
115 * if we are not maintaining a random map, we are done.
116 */
117 if (!file_randommap(td, f))
118 goto ret;
119
120 /*
121 * calculate map offset and check if it's free
122 */
123 if (random_map_free(f, *b))
124 goto ret;
125
126 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
127 (unsigned long long) *b);
128
129 *b = axmap_next_free(f->io_axmap, *b);
130 if (*b == (uint64_t) -1ULL)
131 return 1;
132ret:
133 return 0;
134}
135
136static int __get_next_rand_offset_zipf(struct thread_data *td,
137 struct fio_file *f, enum fio_ddir ddir,
138 uint64_t *b)
139{
140 *b = zipf_next(&f->zipf);
141 return 0;
142}
143
144static int __get_next_rand_offset_pareto(struct thread_data *td,
145 struct fio_file *f, enum fio_ddir ddir,
146 uint64_t *b)
147{
148 *b = pareto_next(&f->zipf);
149 return 0;
150}
151
152static int __get_next_rand_offset_gauss(struct thread_data *td,
153 struct fio_file *f, enum fio_ddir ddir,
154 uint64_t *b)
155{
156 *b = gauss_next(&f->gauss);
157 return 0;
158}
159
160static int __get_next_rand_offset_zoned_abs(struct thread_data *td,
161 struct fio_file *f,
162 enum fio_ddir ddir, uint64_t *b)
163{
164 struct zone_split_index *zsi;
165 uint64_t lastb, send, stotal;
166 static int warned;
167 unsigned int v;
168
169 lastb = last_block(td, f, ddir);
170 if (!lastb)
171 return 1;
172
173 if (!td->o.zone_split_nr[ddir]) {
174bail:
175 return __get_next_rand_offset(td, f, ddir, b, lastb);
176 }
177
178 /*
179 * Generate a value, v, between 1 and 100, both inclusive
180 */
181 v = rand32_between(&td->zone_state, 1, 100);
182
183 /*
184 * Find our generated table. 'send' is the end block of this zone,
185 * 'stotal' is our start offset.
186 */
187 zsi = &td->zone_state_index[ddir][v - 1];
188 stotal = zsi->size_prev / td->o.ba[ddir];
189 send = zsi->size / td->o.ba[ddir];
190
191 /*
192 * Should never happen
193 */
194 if (send == -1U) {
195 if (!warned) {
196 log_err("fio: bug in zoned generation\n");
197 warned = 1;
198 }
199 goto bail;
200 } else if (send > lastb) {
201 /*
202 * This happens if the user specifies ranges that exceed
203 * the file/device size. We can't handle that gracefully,
204 * so error and exit.
205 */
206 log_err("fio: zoned_abs sizes exceed file size\n");
207 return 1;
208 }
209
210 /*
211 * Generate index from 0..send-stotal
212 */
213 if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1)
214 return 1;
215
216 *b += stotal;
217 return 0;
218}
219
220static int __get_next_rand_offset_zoned(struct thread_data *td,
221 struct fio_file *f, enum fio_ddir ddir,
222 uint64_t *b)
223{
224 unsigned int v, send, stotal;
225 uint64_t offset, lastb;
226 static int warned;
227 struct zone_split_index *zsi;
228
229 lastb = last_block(td, f, ddir);
230 if (!lastb)
231 return 1;
232
233 if (!td->o.zone_split_nr[ddir]) {
234bail:
235 return __get_next_rand_offset(td, f, ddir, b, lastb);
236 }
237
238 /*
239 * Generate a value, v, between 1 and 100, both inclusive
240 */
241 v = rand32_between(&td->zone_state, 1, 100);
242
243 zsi = &td->zone_state_index[ddir][v - 1];
244 stotal = zsi->size_perc_prev;
245 send = zsi->size_perc;
246
247 /*
248 * Should never happen
249 */
250 if (send == -1U) {
251 if (!warned) {
252 log_err("fio: bug in zoned generation\n");
253 warned = 1;
254 }
255 goto bail;
256 }
257
258 /*
259 * 'send' is some percentage below or equal to 100 that
260 * marks the end of the current IO range. 'stotal' marks
261 * the start, in percent.
262 */
263 if (stotal)
264 offset = stotal * lastb / 100ULL;
265 else
266 offset = 0;
267
268 lastb = lastb * (send - stotal) / 100ULL;
269
270 /*
271 * Generate index from 0..send-of-lastb
272 */
273 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
274 return 1;
275
276 /*
277 * Add our start offset, if any
278 */
279 if (offset)
280 *b += offset;
281
282 return 0;
283}
284
285static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
286{
287 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
288 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
289
290 return r1->off - r2->off;
291}
292
293static int get_off_from_method(struct thread_data *td, struct fio_file *f,
294 enum fio_ddir ddir, uint64_t *b)
295{
296 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
297 uint64_t lastb;
298
299 lastb = last_block(td, f, ddir);
300 if (!lastb)
301 return 1;
302
303 return __get_next_rand_offset(td, f, ddir, b, lastb);
304 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
305 return __get_next_rand_offset_zipf(td, f, ddir, b);
306 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
307 return __get_next_rand_offset_pareto(td, f, ddir, b);
308 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
309 return __get_next_rand_offset_gauss(td, f, ddir, b);
310 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
311 return __get_next_rand_offset_zoned(td, f, ddir, b);
312 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
313 return __get_next_rand_offset_zoned_abs(td, f, ddir, b);
314
315 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
316 return 1;
317}
318
319/*
320 * Sort the reads for a verify phase in batches of verifysort_nr, if
321 * specified.
322 */
323static inline bool should_sort_io(struct thread_data *td)
324{
325 if (!td->o.verifysort_nr || !td->o.do_verify)
326 return false;
327 if (!td_random(td))
328 return false;
329 if (td->runstate != TD_VERIFYING)
330 return false;
331 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
332 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
333 return false;
334
335 return true;
336}
337
338static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
339{
340 unsigned int v;
341
342 if (td->o.perc_rand[ddir] == 100)
343 return true;
344
345 v = rand32_between(&td->seq_rand_state[ddir], 1, 100);
346
347 return v <= td->o.perc_rand[ddir];
348}
349
350static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
351 enum fio_ddir ddir, uint64_t *b)
352{
353 struct rand_off *r;
354 int i, ret = 1;
355
356 if (!should_sort_io(td))
357 return get_off_from_method(td, f, ddir, b);
358
359 if (!flist_empty(&td->next_rand_list)) {
360fetch:
361 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
362 flist_del(&r->list);
363 *b = r->off;
364 free(r);
365 return 0;
366 }
367
368 for (i = 0; i < td->o.verifysort_nr; i++) {
369 r = malloc(sizeof(*r));
370
371 ret = get_off_from_method(td, f, ddir, &r->off);
372 if (ret) {
373 free(r);
374 break;
375 }
376
377 flist_add(&r->list, &td->next_rand_list);
378 }
379
380 if (ret && !i)
381 return ret;
382
383 assert(!flist_empty(&td->next_rand_list));
384 flist_sort(NULL, &td->next_rand_list, flist_cmp);
385 goto fetch;
386}
387
388static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f)
389{
390 struct thread_options *o = &td->o;
391
392 if (o->invalidate_cache && !o->odirect) {
393 int fio_unused ret;
394
395 ret = file_invalidate_cache(td, f);
396 }
397}
398
399static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
400 enum fio_ddir ddir, uint64_t *b)
401{
402 if (!get_next_rand_offset(td, f, ddir, b))
403 return 0;
404
405 if (td->o.time_based ||
406 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
407 fio_file_reset(td, f);
408 if (!get_next_rand_offset(td, f, ddir, b))
409 return 0;
410 loop_cache_invalidate(td, f);
411 }
412
413 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
414 f->file_name, (unsigned long long) f->last_pos[ddir],
415 (unsigned long long) f->real_file_size);
416 return 1;
417}
418
419static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
420 enum fio_ddir ddir, uint64_t *offset)
421{
422 struct thread_options *o = &td->o;
423
424 assert(ddir_rw(ddir));
425
426 /*
427 * If we reach the end for a time based run, reset us back to 0
428 * and invalidate the cache, if we need to.
429 */
430 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
431 o->time_based) {
432 f->last_pos[ddir] = f->file_offset;
433 loop_cache_invalidate(td, f);
434 }
435
436 if (f->last_pos[ddir] < f->real_file_size) {
437 uint64_t pos;
438
439 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) {
440 if (f->real_file_size > f->io_size)
441 f->last_pos[ddir] = f->io_size;
442 else
443 f->last_pos[ddir] = f->real_file_size;
444 }
445
446 pos = f->last_pos[ddir] - f->file_offset;
447 if (pos && o->ddir_seq_add) {
448 pos += o->ddir_seq_add;
449
450 /*
451 * If we reach beyond the end of the file
452 * with holed IO, wrap around to the
453 * beginning again. If we're doing backwards IO,
454 * wrap to the end.
455 */
456 if (pos >= f->real_file_size) {
457 if (o->ddir_seq_add > 0)
458 pos = f->file_offset;
459 else {
460 if (f->real_file_size > f->io_size)
461 pos = f->io_size;
462 else
463 pos = f->real_file_size;
464
465 pos += o->ddir_seq_add;
466 }
467 }
468 }
469
470 *offset = pos;
471 return 0;
472 }
473
474 return 1;
475}
476
477static int get_next_block(struct thread_data *td, struct io_u *io_u,
478 enum fio_ddir ddir, int rw_seq,
479 unsigned int *is_random)
480{
481 struct fio_file *f = io_u->file;
482 uint64_t b, offset;
483 int ret;
484
485 assert(ddir_rw(ddir));
486
487 b = offset = -1ULL;
488
489 if (rw_seq) {
490 if (td_random(td)) {
491 if (should_do_random(td, ddir)) {
492 ret = get_next_rand_block(td, f, ddir, &b);
493 *is_random = 1;
494 } else {
495 *is_random = 0;
496 io_u_set(td, io_u, IO_U_F_BUSY_OK);
497 ret = get_next_seq_offset(td, f, ddir, &offset);
498 if (ret)
499 ret = get_next_rand_block(td, f, ddir, &b);
500 }
501 } else {
502 *is_random = 0;
503 ret = get_next_seq_offset(td, f, ddir, &offset);
504 }
505 } else {
506 io_u_set(td, io_u, IO_U_F_BUSY_OK);
507 *is_random = 0;
508
509 if (td->o.rw_seq == RW_SEQ_SEQ) {
510 ret = get_next_seq_offset(td, f, ddir, &offset);
511 if (ret) {
512 ret = get_next_rand_block(td, f, ddir, &b);
513 *is_random = 0;
514 }
515 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
516 if (f->last_start[ddir] != -1ULL)
517 offset = f->last_start[ddir] - f->file_offset;
518 else
519 offset = 0;
520 ret = 0;
521 } else {
522 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
523 ret = 1;
524 }
525 }
526
527 if (!ret) {
528 if (offset != -1ULL)
529 io_u->offset = offset;
530 else if (b != -1ULL)
531 io_u->offset = b * td->o.ba[ddir];
532 else {
533 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
534 ret = 1;
535 }
536 }
537
538 return ret;
539}
540
541/*
542 * For random io, generate a random new block and see if it's used. Repeat
543 * until we find a free one. For sequential io, just return the end of
544 * the last io issued.
545 */
546static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
547 unsigned int *is_random)
548{
549 struct fio_file *f = io_u->file;
550 enum fio_ddir ddir = io_u->ddir;
551 int rw_seq_hit = 0;
552
553 assert(ddir_rw(ddir));
554
555 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
556 rw_seq_hit = 1;
557 td->ddir_seq_nr = td->o.ddir_seq_nr;
558 }
559
560 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
561 return 1;
562
563 if (io_u->offset >= f->io_size) {
564 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
565 (unsigned long long) io_u->offset,
566 (unsigned long long) f->io_size);
567 return 1;
568 }
569
570 io_u->offset += f->file_offset;
571 if (io_u->offset >= f->real_file_size) {
572 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
573 (unsigned long long) io_u->offset,
574 (unsigned long long) f->real_file_size);
575 return 1;
576 }
577
578 return 0;
579}
580
581static int get_next_offset(struct thread_data *td, struct io_u *io_u,
582 unsigned int *is_random)
583{
584 if (td->flags & TD_F_PROFILE_OPS) {
585 struct prof_io_ops *ops = &td->prof_io_ops;
586
587 if (ops->fill_io_u_off)
588 return ops->fill_io_u_off(td, io_u, is_random);
589 }
590
591 return __get_next_offset(td, io_u, is_random);
592}
593
594static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
595 unsigned int buflen)
596{
597 struct fio_file *f = io_u->file;
598
599 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
600}
601
602static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
603 unsigned int is_random)
604{
605 int ddir = io_u->ddir;
606 unsigned int buflen = 0;
607 unsigned int minbs, maxbs;
608 uint64_t frand_max, r;
609 bool power_2;
610
611 assert(ddir_rw(ddir));
612
613 if (td->o.bs_is_seq_rand)
614 ddir = is_random ? DDIR_WRITE: DDIR_READ;
615
616 minbs = td->o.min_bs[ddir];
617 maxbs = td->o.max_bs[ddir];
618
619 if (minbs == maxbs)
620 return minbs;
621
622 /*
623 * If we can't satisfy the min block size from here, then fail
624 */
625 if (!io_u_fits(td, io_u, minbs))
626 return 0;
627
628 frand_max = rand_max(&td->bsrange_state[ddir]);
629 do {
630 r = __rand(&td->bsrange_state[ddir]);
631
632 if (!td->o.bssplit_nr[ddir]) {
633 buflen = 1 + (unsigned int) ((double) maxbs *
634 (r / (frand_max + 1.0)));
635 if (buflen < minbs)
636 buflen = minbs;
637 } else {
638 long long perc = 0;
639 unsigned int i;
640
641 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
642 struct bssplit *bsp = &td->o.bssplit[ddir][i];
643
644 buflen = bsp->bs;
645 perc += bsp->perc;
646 if (!perc)
647 break;
648 if ((r / perc <= frand_max / 100ULL) &&
649 io_u_fits(td, io_u, buflen))
650 break;
651 }
652 }
653
654 power_2 = is_power_of_2(minbs);
655 if (!td->o.bs_unaligned && power_2)
656 buflen &= ~(minbs - 1);
657 else if (!td->o.bs_unaligned && !power_2)
658 buflen -= buflen % minbs;
659 } while (!io_u_fits(td, io_u, buflen));
660
661 return buflen;
662}
663
664static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
665 unsigned int is_random)
666{
667 if (td->flags & TD_F_PROFILE_OPS) {
668 struct prof_io_ops *ops = &td->prof_io_ops;
669
670 if (ops->fill_io_u_size)
671 return ops->fill_io_u_size(td, io_u, is_random);
672 }
673
674 return __get_next_buflen(td, io_u, is_random);
675}
676
677static void set_rwmix_bytes(struct thread_data *td)
678{
679 unsigned int diff;
680
681 /*
682 * we do time or byte based switch. this is needed because
683 * buffered writes may issue a lot quicker than they complete,
684 * whereas reads do not.
685 */
686 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
687 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
688}
689
690static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
691{
692 unsigned int v;
693
694 v = rand32_between(&td->rwmix_state, 1, 100);
695
696 if (v <= td->o.rwmix[DDIR_READ])
697 return DDIR_READ;
698
699 return DDIR_WRITE;
700}
701
702int io_u_quiesce(struct thread_data *td)
703{
704 int completed = 0;
705
706 /*
707 * We are going to sleep, ensure that we flush anything pending as
708 * not to skew our latency numbers.
709 *
710 * Changed to only monitor 'in flight' requests here instead of the
711 * td->cur_depth, b/c td->cur_depth does not accurately represent
712 * io's that have been actually submitted to an async engine,
713 * and cur_depth is meaningless for sync engines.
714 */
715 if (td->io_u_queued || td->cur_depth) {
716 int fio_unused ret;
717
718 ret = td_io_commit(td);
719 }
720
721 while (td->io_u_in_flight) {
722 int ret;
723
724 ret = io_u_queued_complete(td, 1);
725 if (ret > 0)
726 completed += ret;
727 }
728
729 if (td->flags & TD_F_REGROW_LOGS)
730 regrow_logs(td);
731
732 return completed;
733}
734
735static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
736{
737 enum fio_ddir odir = ddir ^ 1;
738 uint64_t usec;
739 uint64_t now;
740
741 assert(ddir_rw(ddir));
742 now = utime_since_now(&td->start);
743
744 /*
745 * if rate_next_io_time is in the past, need to catch up to rate
746 */
747 if (td->rate_next_io_time[ddir] <= now)
748 return ddir;
749
750 /*
751 * We are ahead of rate in this direction. See if we
752 * should switch.
753 */
754 if (td_rw(td) && td->o.rwmix[odir]) {
755 /*
756 * Other direction is behind rate, switch
757 */
758 if (td->rate_next_io_time[odir] <= now)
759 return odir;
760
761 /*
762 * Both directions are ahead of rate. sleep the min,
763 * switch if necessary
764 */
765 if (td->rate_next_io_time[ddir] <=
766 td->rate_next_io_time[odir]) {
767 usec = td->rate_next_io_time[ddir] - now;
768 } else {
769 usec = td->rate_next_io_time[odir] - now;
770 ddir = odir;
771 }
772 } else
773 usec = td->rate_next_io_time[ddir] - now;
774
775 if (td->o.io_submit_mode == IO_MODE_INLINE)
776 io_u_quiesce(td);
777
778 usec_sleep(td, usec);
779 return ddir;
780}
781
782/*
783 * Return the data direction for the next io_u. If the job is a
784 * mixed read/write workload, check the rwmix cycle and switch if
785 * necessary.
786 */
787static enum fio_ddir get_rw_ddir(struct thread_data *td)
788{
789 enum fio_ddir ddir;
790
791 /*
792 * See if it's time to fsync/fdatasync/sync_file_range first,
793 * and if not then move on to check regular I/Os.
794 */
795 if (should_fsync(td)) {
796 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
797 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
798 return DDIR_SYNC;
799
800 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
801 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
802 return DDIR_DATASYNC;
803
804 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
805 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
806 return DDIR_SYNC_FILE_RANGE;
807 }
808
809 if (td_rw(td)) {
810 /*
811 * Check if it's time to seed a new data direction.
812 */
813 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
814 /*
815 * Put a top limit on how many bytes we do for
816 * one data direction, to avoid overflowing the
817 * ranges too much
818 */
819 ddir = get_rand_ddir(td);
820
821 if (ddir != td->rwmix_ddir)
822 set_rwmix_bytes(td);
823
824 td->rwmix_ddir = ddir;
825 }
826 ddir = td->rwmix_ddir;
827 } else if (td_read(td))
828 ddir = DDIR_READ;
829 else if (td_write(td))
830 ddir = DDIR_WRITE;
831 else if (td_trim(td))
832 ddir = DDIR_TRIM;
833 else
834 ddir = DDIR_INVAL;
835
836 td->rwmix_ddir = rate_ddir(td, ddir);
837 return td->rwmix_ddir;
838}
839
840static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
841{
842 enum fio_ddir ddir = get_rw_ddir(td);
843
844 if (td_trimwrite(td)) {
845 struct fio_file *f = io_u->file;
846 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
847 ddir = DDIR_TRIM;
848 else
849 ddir = DDIR_WRITE;
850 }
851
852 io_u->ddir = io_u->acct_ddir = ddir;
853
854 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
855 td->o.barrier_blocks &&
856 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
857 td->io_issues[DDIR_WRITE])
858 io_u_set(td, io_u, IO_U_F_BARRIER);
859}
860
861void put_file_log(struct thread_data *td, struct fio_file *f)
862{
863 unsigned int ret = put_file(td, f);
864
865 if (ret)
866 td_verror(td, ret, "file close");
867}
868
869void put_io_u(struct thread_data *td, struct io_u *io_u)
870{
871 if (td->parent)
872 td = td->parent;
873
874 td_io_u_lock(td);
875
876 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
877 put_file_log(td, io_u->file);
878
879 io_u->file = NULL;
880 io_u_set(td, io_u, IO_U_F_FREE);
881
882 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
883 td->cur_depth--;
884 assert(!(td->flags & TD_F_CHILD));
885 }
886 io_u_qpush(&td->io_u_freelist, io_u);
887 td_io_u_unlock(td);
888 td_io_u_free_notify(td);
889}
890
891void clear_io_u(struct thread_data *td, struct io_u *io_u)
892{
893 io_u_clear(td, io_u, IO_U_F_FLIGHT);
894 put_io_u(td, io_u);
895}
896
897void requeue_io_u(struct thread_data *td, struct io_u **io_u)
898{
899 struct io_u *__io_u = *io_u;
900 enum fio_ddir ddir = acct_ddir(__io_u);
901
902 dprint(FD_IO, "requeue %p\n", __io_u);
903
904 if (td->parent)
905 td = td->parent;
906
907 td_io_u_lock(td);
908
909 io_u_set(td, __io_u, IO_U_F_FREE);
910 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
911 td->io_issues[ddir]--;
912
913 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
914 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
915 td->cur_depth--;
916 assert(!(td->flags & TD_F_CHILD));
917 }
918
919 io_u_rpush(&td->io_u_requeues, __io_u);
920 td_io_u_unlock(td);
921 td_io_u_free_notify(td);
922 *io_u = NULL;
923}
924
925static void __fill_io_u_zone(struct thread_data *td, struct io_u *io_u)
926{
927 struct fio_file *f = io_u->file;
928
929 /*
930 * See if it's time to switch to a new zone
931 */
932 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
933 td->zone_bytes = 0;
934 f->file_offset += td->o.zone_range + td->o.zone_skip;
935
936 /*
937 * Wrap from the beginning, if we exceed the file size
938 */
939 if (f->file_offset >= f->real_file_size)
940 f->file_offset = f->real_file_size - f->file_offset;
941 f->last_pos[io_u->ddir] = f->file_offset;
942 td->io_skip_bytes += td->o.zone_skip;
943 }
944
945 /*
946 * If zone_size > zone_range, then maintain the same zone until
947 * zone_bytes >= zone_size.
948 */
949 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
950 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
951 f->file_offset, f->last_pos[io_u->ddir]);
952 f->last_pos[io_u->ddir] = f->file_offset;
953 }
954
955 /*
956 * For random: if 'norandommap' is not set and zone_size > zone_range,
957 * map needs to be reset as it's done with zone_range everytime.
958 */
959 if ((td->zone_bytes % td->o.zone_range) == 0) {
960 fio_file_reset(td, f);
961 }
962}
963
964static int fill_io_u(struct thread_data *td, struct io_u *io_u)
965{
966 unsigned int is_random;
967
968 if (td_ioengine_flagged(td, FIO_NOIO))
969 goto out;
970
971 set_rw_ddir(td, io_u);
972
973 /*
974 * fsync() or fdatasync() or trim etc, we are done
975 */
976 if (!ddir_rw(io_u->ddir))
977 goto out;
978
979 /*
980 * When file is zoned zone_range is always positive
981 */
982 if (td->o.zone_range) {
983 __fill_io_u_zone(td, io_u);
984 }
985
986 /*
987 * No log, let the seq/rand engine retrieve the next buflen and
988 * position.
989 */
990 if (get_next_offset(td, io_u, &is_random)) {
991 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
992 return 1;
993 }
994
995 io_u->buflen = get_next_buflen(td, io_u, is_random);
996 if (!io_u->buflen) {
997 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
998 return 1;
999 }
1000
1001 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
1002 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%lx exceeds file size=0x%llx\n",
1003 io_u,
1004 (unsigned long long) io_u->offset, io_u->buflen,
1005 (unsigned long long) io_u->file->real_file_size);
1006 return 1;
1007 }
1008
1009 /*
1010 * mark entry before potentially trimming io_u
1011 */
1012 if (td_random(td) && file_randommap(td, io_u->file))
1013 mark_random_map(td, io_u);
1014
1015out:
1016 dprint_io_u(io_u, "fill");
1017 td->zone_bytes += io_u->buflen;
1018 return 0;
1019}
1020
1021static void __io_u_mark_map(unsigned int *map, unsigned int nr)
1022{
1023 int idx = 0;
1024
1025 switch (nr) {
1026 default:
1027 idx = 6;
1028 break;
1029 case 33 ... 64:
1030 idx = 5;
1031 break;
1032 case 17 ... 32:
1033 idx = 4;
1034 break;
1035 case 9 ... 16:
1036 idx = 3;
1037 break;
1038 case 5 ... 8:
1039 idx = 2;
1040 break;
1041 case 1 ... 4:
1042 idx = 1;
1043 case 0:
1044 break;
1045 }
1046
1047 map[idx]++;
1048}
1049
1050void io_u_mark_submit(struct thread_data *td, unsigned int nr)
1051{
1052 __io_u_mark_map(td->ts.io_u_submit, nr);
1053 td->ts.total_submit++;
1054}
1055
1056void io_u_mark_complete(struct thread_data *td, unsigned int nr)
1057{
1058 __io_u_mark_map(td->ts.io_u_complete, nr);
1059 td->ts.total_complete++;
1060}
1061
1062void io_u_mark_depth(struct thread_data *td, unsigned int nr)
1063{
1064 int idx = 0;
1065
1066 switch (td->cur_depth) {
1067 default:
1068 idx = 6;
1069 break;
1070 case 32 ... 63:
1071 idx = 5;
1072 break;
1073 case 16 ... 31:
1074 idx = 4;
1075 break;
1076 case 8 ... 15:
1077 idx = 3;
1078 break;
1079 case 4 ... 7:
1080 idx = 2;
1081 break;
1082 case 2 ... 3:
1083 idx = 1;
1084 case 1:
1085 break;
1086 }
1087
1088 td->ts.io_u_map[idx] += nr;
1089}
1090
1091static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1092{
1093 int idx = 0;
1094
1095 assert(nsec < 1000);
1096
1097 switch (nsec) {
1098 case 750 ... 999:
1099 idx = 9;
1100 break;
1101 case 500 ... 749:
1102 idx = 8;
1103 break;
1104 case 250 ... 499:
1105 idx = 7;
1106 break;
1107 case 100 ... 249:
1108 idx = 6;
1109 break;
1110 case 50 ... 99:
1111 idx = 5;
1112 break;
1113 case 20 ... 49:
1114 idx = 4;
1115 break;
1116 case 10 ... 19:
1117 idx = 3;
1118 break;
1119 case 4 ... 9:
1120 idx = 2;
1121 break;
1122 case 2 ... 3:
1123 idx = 1;
1124 case 0 ... 1:
1125 break;
1126 }
1127
1128 assert(idx < FIO_IO_U_LAT_N_NR);
1129 td->ts.io_u_lat_n[idx]++;
1130}
1131
1132static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1133{
1134 int idx = 0;
1135
1136 assert(usec < 1000 && usec >= 1);
1137
1138 switch (usec) {
1139 case 750 ... 999:
1140 idx = 9;
1141 break;
1142 case 500 ... 749:
1143 idx = 8;
1144 break;
1145 case 250 ... 499:
1146 idx = 7;
1147 break;
1148 case 100 ... 249:
1149 idx = 6;
1150 break;
1151 case 50 ... 99:
1152 idx = 5;
1153 break;
1154 case 20 ... 49:
1155 idx = 4;
1156 break;
1157 case 10 ... 19:
1158 idx = 3;
1159 break;
1160 case 4 ... 9:
1161 idx = 2;
1162 break;
1163 case 2 ... 3:
1164 idx = 1;
1165 case 0 ... 1:
1166 break;
1167 }
1168
1169 assert(idx < FIO_IO_U_LAT_U_NR);
1170 td->ts.io_u_lat_u[idx]++;
1171}
1172
1173static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1174{
1175 int idx = 0;
1176
1177 assert(msec >= 1);
1178
1179 switch (msec) {
1180 default:
1181 idx = 11;
1182 break;
1183 case 1000 ... 1999:
1184 idx = 10;
1185 break;
1186 case 750 ... 999:
1187 idx = 9;
1188 break;
1189 case 500 ... 749:
1190 idx = 8;
1191 break;
1192 case 250 ... 499:
1193 idx = 7;
1194 break;
1195 case 100 ... 249:
1196 idx = 6;
1197 break;
1198 case 50 ... 99:
1199 idx = 5;
1200 break;
1201 case 20 ... 49:
1202 idx = 4;
1203 break;
1204 case 10 ... 19:
1205 idx = 3;
1206 break;
1207 case 4 ... 9:
1208 idx = 2;
1209 break;
1210 case 2 ... 3:
1211 idx = 1;
1212 case 0 ... 1:
1213 break;
1214 }
1215
1216 assert(idx < FIO_IO_U_LAT_M_NR);
1217 td->ts.io_u_lat_m[idx]++;
1218}
1219
1220static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1221{
1222 if (nsec < 1000)
1223 io_u_mark_lat_nsec(td, nsec);
1224 else if (nsec < 1000000)
1225 io_u_mark_lat_usec(td, nsec / 1000);
1226 else
1227 io_u_mark_lat_msec(td, nsec / 1000000);
1228}
1229
1230static unsigned int __get_next_fileno_rand(struct thread_data *td)
1231{
1232 unsigned long fileno;
1233
1234 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1235 uint64_t frand_max = rand_max(&td->next_file_state);
1236 unsigned long r;
1237
1238 r = __rand(&td->next_file_state);
1239 return (unsigned int) ((double) td->o.nr_files
1240 * (r / (frand_max + 1.0)));
1241 }
1242
1243 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1244 fileno = zipf_next(&td->next_file_zipf);
1245 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1246 fileno = pareto_next(&td->next_file_zipf);
1247 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1248 fileno = gauss_next(&td->next_file_gauss);
1249 else {
1250 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1251 assert(0);
1252 return 0;
1253 }
1254
1255 return fileno >> FIO_FSERVICE_SHIFT;
1256}
1257
1258/*
1259 * Get next file to service by choosing one at random
1260 */
1261static struct fio_file *get_next_file_rand(struct thread_data *td,
1262 enum fio_file_flags goodf,
1263 enum fio_file_flags badf)
1264{
1265 struct fio_file *f;
1266 int fno;
1267
1268 do {
1269 int opened = 0;
1270
1271 fno = __get_next_fileno_rand(td);
1272
1273 f = td->files[fno];
1274 if (fio_file_done(f))
1275 continue;
1276
1277 if (!fio_file_open(f)) {
1278 int err;
1279
1280 if (td->nr_open_files >= td->o.open_files)
1281 return ERR_PTR(-EBUSY);
1282
1283 err = td_io_open_file(td, f);
1284 if (err)
1285 continue;
1286 opened = 1;
1287 }
1288
1289 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1290 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1291 return f;
1292 }
1293 if (opened)
1294 td_io_close_file(td, f);
1295 } while (1);
1296}
1297
1298/*
1299 * Get next file to service by doing round robin between all available ones
1300 */
1301static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1302 int badf)
1303{
1304 unsigned int old_next_file = td->next_file;
1305 struct fio_file *f;
1306
1307 do {
1308 int opened = 0;
1309
1310 f = td->files[td->next_file];
1311
1312 td->next_file++;
1313 if (td->next_file >= td->o.nr_files)
1314 td->next_file = 0;
1315
1316 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1317 if (fio_file_done(f)) {
1318 f = NULL;
1319 continue;
1320 }
1321
1322 if (!fio_file_open(f)) {
1323 int err;
1324
1325 if (td->nr_open_files >= td->o.open_files)
1326 return ERR_PTR(-EBUSY);
1327
1328 err = td_io_open_file(td, f);
1329 if (err) {
1330 dprint(FD_FILE, "error %d on open of %s\n",
1331 err, f->file_name);
1332 f = NULL;
1333 continue;
1334 }
1335 opened = 1;
1336 }
1337
1338 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1339 f->flags);
1340 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1341 break;
1342
1343 if (opened)
1344 td_io_close_file(td, f);
1345
1346 f = NULL;
1347 } while (td->next_file != old_next_file);
1348
1349 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1350 return f;
1351}
1352
1353static struct fio_file *__get_next_file(struct thread_data *td)
1354{
1355 struct fio_file *f;
1356
1357 assert(td->o.nr_files <= td->files_index);
1358
1359 if (td->nr_done_files >= td->o.nr_files) {
1360 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1361 " nr_files=%d\n", td->nr_open_files,
1362 td->nr_done_files,
1363 td->o.nr_files);
1364 return NULL;
1365 }
1366
1367 f = td->file_service_file;
1368 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1369 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1370 goto out;
1371 if (td->file_service_left--)
1372 goto out;
1373 }
1374
1375 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1376 td->o.file_service_type == FIO_FSERVICE_SEQ)
1377 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1378 else
1379 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1380
1381 if (IS_ERR(f))
1382 return f;
1383
1384 td->file_service_file = f;
1385 td->file_service_left = td->file_service_nr - 1;
1386out:
1387 if (f)
1388 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1389 else
1390 dprint(FD_FILE, "get_next_file: NULL\n");
1391 return f;
1392}
1393
1394static struct fio_file *get_next_file(struct thread_data *td)
1395{
1396 if (td->flags & TD_F_PROFILE_OPS) {
1397 struct prof_io_ops *ops = &td->prof_io_ops;
1398
1399 if (ops->get_next_file)
1400 return ops->get_next_file(td);
1401 }
1402
1403 return __get_next_file(td);
1404}
1405
1406static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1407{
1408 struct fio_file *f;
1409
1410 do {
1411 f = get_next_file(td);
1412 if (IS_ERR_OR_NULL(f))
1413 return PTR_ERR(f);
1414
1415 io_u->file = f;
1416 get_file(f);
1417
1418 if (!fill_io_u(td, io_u))
1419 break;
1420
1421 put_file_log(td, f);
1422 td_io_close_file(td, f);
1423 io_u->file = NULL;
1424 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1425 fio_file_reset(td, f);
1426 else {
1427 fio_file_set_done(f);
1428 td->nr_done_files++;
1429 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1430 td->nr_done_files, td->o.nr_files);
1431 }
1432 } while (1);
1433
1434 return 0;
1435}
1436
1437static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1438 unsigned long long tnsec, unsigned long long max_nsec)
1439{
1440 if (!td->error)
1441 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
1442 td_verror(td, ETIMEDOUT, "max latency exceeded");
1443 icd->error = ETIMEDOUT;
1444}
1445
1446static void lat_new_cycle(struct thread_data *td)
1447{
1448 fio_gettime(&td->latency_ts, NULL);
1449 td->latency_ios = ddir_rw_sum(td->io_blocks);
1450 td->latency_failed = 0;
1451}
1452
1453/*
1454 * We had an IO outside the latency target. Reduce the queue depth. If we
1455 * are at QD=1, then it's time to give up.
1456 */
1457static bool __lat_target_failed(struct thread_data *td)
1458{
1459 if (td->latency_qd == 1)
1460 return true;
1461
1462 td->latency_qd_high = td->latency_qd;
1463
1464 if (td->latency_qd == td->latency_qd_low)
1465 td->latency_qd_low--;
1466
1467 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1468
1469 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1470
1471 /*
1472 * When we ramp QD down, quiesce existing IO to prevent
1473 * a storm of ramp downs due to pending higher depth.
1474 */
1475 io_u_quiesce(td);
1476 lat_new_cycle(td);
1477 return false;
1478}
1479
1480static bool lat_target_failed(struct thread_data *td)
1481{
1482 if (td->o.latency_percentile.u.f == 100.0)
1483 return __lat_target_failed(td);
1484
1485 td->latency_failed++;
1486 return false;
1487}
1488
1489void lat_target_init(struct thread_data *td)
1490{
1491 td->latency_end_run = 0;
1492
1493 if (td->o.latency_target) {
1494 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1495 fio_gettime(&td->latency_ts, NULL);
1496 td->latency_qd = 1;
1497 td->latency_qd_high = td->o.iodepth;
1498 td->latency_qd_low = 1;
1499 td->latency_ios = ddir_rw_sum(td->io_blocks);
1500 } else
1501 td->latency_qd = td->o.iodepth;
1502}
1503
1504void lat_target_reset(struct thread_data *td)
1505{
1506 if (!td->latency_end_run)
1507 lat_target_init(td);
1508}
1509
1510static void lat_target_success(struct thread_data *td)
1511{
1512 const unsigned int qd = td->latency_qd;
1513 struct thread_options *o = &td->o;
1514
1515 td->latency_qd_low = td->latency_qd;
1516
1517 /*
1518 * If we haven't failed yet, we double up to a failing value instead
1519 * of bisecting from highest possible queue depth. If we have set
1520 * a limit other than td->o.iodepth, bisect between that.
1521 */
1522 if (td->latency_qd_high != o->iodepth)
1523 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1524 else
1525 td->latency_qd *= 2;
1526
1527 if (td->latency_qd > o->iodepth)
1528 td->latency_qd = o->iodepth;
1529
1530 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1531
1532 /*
1533 * Same as last one, we are done. Let it run a latency cycle, so
1534 * we get only the results from the targeted depth.
1535 */
1536 if (td->latency_qd == qd) {
1537 if (td->latency_end_run) {
1538 dprint(FD_RATE, "We are done\n");
1539 td->done = 1;
1540 } else {
1541 dprint(FD_RATE, "Quiesce and final run\n");
1542 io_u_quiesce(td);
1543 td->latency_end_run = 1;
1544 reset_all_stats(td);
1545 reset_io_stats(td);
1546 }
1547 }
1548
1549 lat_new_cycle(td);
1550}
1551
1552/*
1553 * Check if we can bump the queue depth
1554 */
1555void lat_target_check(struct thread_data *td)
1556{
1557 uint64_t usec_window;
1558 uint64_t ios;
1559 double success_ios;
1560
1561 usec_window = utime_since_now(&td->latency_ts);
1562 if (usec_window < td->o.latency_window)
1563 return;
1564
1565 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1566 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1567 success_ios *= 100.0;
1568
1569 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1570
1571 if (success_ios >= td->o.latency_percentile.u.f)
1572 lat_target_success(td);
1573 else
1574 __lat_target_failed(td);
1575}
1576
1577/*
1578 * If latency target is enabled, we might be ramping up or down and not
1579 * using the full queue depth available.
1580 */
1581bool queue_full(const struct thread_data *td)
1582{
1583 const int qempty = io_u_qempty(&td->io_u_freelist);
1584
1585 if (qempty)
1586 return true;
1587 if (!td->o.latency_target)
1588 return false;
1589
1590 return td->cur_depth >= td->latency_qd;
1591}
1592
1593struct io_u *__get_io_u(struct thread_data *td)
1594{
1595 struct io_u *io_u = NULL;
1596
1597 if (td->stop_io)
1598 return NULL;
1599
1600 td_io_u_lock(td);
1601
1602again:
1603 if (!io_u_rempty(&td->io_u_requeues))
1604 io_u = io_u_rpop(&td->io_u_requeues);
1605 else if (!queue_full(td)) {
1606 io_u = io_u_qpop(&td->io_u_freelist);
1607
1608 io_u->file = NULL;
1609 io_u->buflen = 0;
1610 io_u->resid = 0;
1611 io_u->end_io = NULL;
1612 }
1613
1614 if (io_u) {
1615 assert(io_u->flags & IO_U_F_FREE);
1616 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1617 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1618 IO_U_F_VER_LIST);
1619
1620 io_u->error = 0;
1621 io_u->acct_ddir = -1;
1622 td->cur_depth++;
1623 assert(!(td->flags & TD_F_CHILD));
1624 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1625 io_u->ipo = NULL;
1626 } else if (td_async_processing(td)) {
1627 /*
1628 * We ran out, wait for async verify threads to finish and
1629 * return one
1630 */
1631 assert(!(td->flags & TD_F_CHILD));
1632 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1633 goto again;
1634 }
1635
1636 td_io_u_unlock(td);
1637 return io_u;
1638}
1639
1640static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1641{
1642 if (!(td->flags & TD_F_TRIM_BACKLOG))
1643 return false;
1644 if (!td->trim_entries)
1645 return false;
1646
1647 if (td->trim_batch) {
1648 td->trim_batch--;
1649 if (get_next_trim(td, io_u))
1650 return true;
1651 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1652 td->last_ddir != DDIR_READ) {
1653 td->trim_batch = td->o.trim_batch;
1654 if (!td->trim_batch)
1655 td->trim_batch = td->o.trim_backlog;
1656 if (get_next_trim(td, io_u))
1657 return true;
1658 }
1659
1660 return false;
1661}
1662
1663static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1664{
1665 if (!(td->flags & TD_F_VER_BACKLOG))
1666 return false;
1667
1668 if (td->io_hist_len) {
1669 int get_verify = 0;
1670
1671 if (td->verify_batch)
1672 get_verify = 1;
1673 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1674 td->last_ddir != DDIR_READ) {
1675 td->verify_batch = td->o.verify_batch;
1676 if (!td->verify_batch)
1677 td->verify_batch = td->o.verify_backlog;
1678 get_verify = 1;
1679 }
1680
1681 if (get_verify && !get_next_verify(td, io_u)) {
1682 td->verify_batch--;
1683 return true;
1684 }
1685 }
1686
1687 return false;
1688}
1689
1690/*
1691 * Fill offset and start time into the buffer content, to prevent too
1692 * easy compressible data for simple de-dupe attempts. Do this for every
1693 * 512b block in the range, since that should be the smallest block size
1694 * we can expect from a device.
1695 */
1696static void small_content_scramble(struct io_u *io_u)
1697{
1698 unsigned int i, nr_blocks = io_u->buflen >> 9;
1699 unsigned int offset;
1700 uint64_t boffset, *iptr;
1701 char *p;
1702
1703 if (!nr_blocks)
1704 return;
1705
1706 p = io_u->xfer_buf;
1707 boffset = io_u->offset;
1708
1709 if (io_u->buf_filled_len)
1710 io_u->buf_filled_len = 0;
1711
1712 /*
1713 * Generate random index between 0..7. We do chunks of 512b, if
1714 * we assume a cacheline is 64 bytes, then we have 8 of those.
1715 * Scramble content within the blocks in the same cacheline to
1716 * speed things up.
1717 */
1718 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1719
1720 for (i = 0; i < nr_blocks; i++) {
1721 /*
1722 * Fill offset into start of cacheline, time into end
1723 * of cacheline
1724 */
1725 iptr = (void *) p + (offset << 6);
1726 *iptr = boffset;
1727
1728 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1729 iptr[0] = io_u->start_time.tv_sec;
1730 iptr[1] = io_u->start_time.tv_nsec;
1731
1732 p += 512;
1733 boffset += 512;
1734 }
1735}
1736
1737/*
1738 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1739 * etc. The returned io_u is fully ready to be prepped and submitted.
1740 */
1741struct io_u *get_io_u(struct thread_data *td)
1742{
1743 struct fio_file *f;
1744 struct io_u *io_u;
1745 int do_scramble = 0;
1746 long ret = 0;
1747
1748 io_u = __get_io_u(td);
1749 if (!io_u) {
1750 dprint(FD_IO, "__get_io_u failed\n");
1751 return NULL;
1752 }
1753
1754 if (check_get_verify(td, io_u))
1755 goto out;
1756 if (check_get_trim(td, io_u))
1757 goto out;
1758
1759 /*
1760 * from a requeue, io_u already setup
1761 */
1762 if (io_u->file)
1763 goto out;
1764
1765 /*
1766 * If using an iolog, grab next piece if any available.
1767 */
1768 if (td->flags & TD_F_READ_IOLOG) {
1769 if (read_iolog_get(td, io_u))
1770 goto err_put;
1771 } else if (set_io_u_file(td, io_u)) {
1772 ret = -EBUSY;
1773 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1774 goto err_put;
1775 }
1776
1777 f = io_u->file;
1778 if (!f) {
1779 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1780 goto err_put;
1781 }
1782
1783 assert(fio_file_open(f));
1784
1785 if (ddir_rw(io_u->ddir)) {
1786 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1787 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1788 goto err_put;
1789 }
1790
1791 f->last_start[io_u->ddir] = io_u->offset;
1792 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1793
1794 if (io_u->ddir == DDIR_WRITE) {
1795 if (td->flags & TD_F_REFILL_BUFFERS) {
1796 io_u_fill_buffer(td, io_u,
1797 td->o.min_bs[DDIR_WRITE],
1798 io_u->buflen);
1799 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1800 !(td->flags & TD_F_COMPRESS))
1801 do_scramble = 1;
1802 if (td->flags & TD_F_VER_NONE) {
1803 populate_verify_io_u(td, io_u);
1804 do_scramble = 0;
1805 }
1806 } else if (io_u->ddir == DDIR_READ) {
1807 /*
1808 * Reset the buf_filled parameters so next time if the
1809 * buffer is used for writes it is refilled.
1810 */
1811 io_u->buf_filled_len = 0;
1812 }
1813 }
1814
1815 /*
1816 * Set io data pointers.
1817 */
1818 io_u->xfer_buf = io_u->buf;
1819 io_u->xfer_buflen = io_u->buflen;
1820
1821out:
1822 assert(io_u->file);
1823 if (!td_io_prep(td, io_u)) {
1824 if (!td->o.disable_lat)
1825 fio_gettime(&io_u->start_time, NULL);
1826
1827 if (do_scramble)
1828 small_content_scramble(io_u);
1829
1830 return io_u;
1831 }
1832err_put:
1833 dprint(FD_IO, "get_io_u failed\n");
1834 put_io_u(td, io_u);
1835 return ERR_PTR(ret);
1836}
1837
1838static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1839{
1840 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1841
1842 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1843 return;
1844
1845 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1846 io_u->file ? " on file " : "",
1847 io_u->file ? io_u->file->file_name : "",
1848 strerror(io_u->error),
1849 io_ddir_name(io_u->ddir),
1850 io_u->offset, io_u->xfer_buflen);
1851
1852 if (td->io_ops->errdetails) {
1853 char *err = td->io_ops->errdetails(io_u);
1854
1855 log_err("fio: %s\n", err);
1856 free(err);
1857 }
1858
1859 if (!td->error)
1860 td_verror(td, io_u->error, "io_u error");
1861}
1862
1863void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1864{
1865 __io_u_log_error(td, io_u);
1866 if (td->parent)
1867 __io_u_log_error(td->parent, io_u);
1868}
1869
1870static inline bool gtod_reduce(struct thread_data *td)
1871{
1872 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1873 || td->o.gtod_reduce;
1874}
1875
1876static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1877 struct io_completion_data *icd,
1878 const enum fio_ddir idx, unsigned int bytes)
1879{
1880 const int no_reduce = !gtod_reduce(td);
1881 unsigned long long llnsec = 0;
1882
1883 if (td->parent)
1884 td = td->parent;
1885
1886 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1887 return;
1888
1889 if (no_reduce)
1890 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1891
1892 if (!td->o.disable_lat) {
1893 unsigned long long tnsec;
1894
1895 tnsec = ntime_since(&io_u->start_time, &icd->time);
1896 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
1897
1898 if (td->flags & TD_F_PROFILE_OPS) {
1899 struct prof_io_ops *ops = &td->prof_io_ops;
1900
1901 if (ops->io_u_lat)
1902 icd->error = ops->io_u_lat(td, tnsec);
1903 }
1904
1905 if (td->o.max_latency && tnsec > td->o.max_latency)
1906 lat_fatal(td, icd, tnsec, td->o.max_latency);
1907 if (td->o.latency_target && tnsec > td->o.latency_target) {
1908 if (lat_target_failed(td))
1909 lat_fatal(td, icd, tnsec, td->o.latency_target);
1910 }
1911 }
1912
1913 if (ddir_rw(idx)) {
1914 if (!td->o.disable_clat) {
1915 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1916 io_u_mark_latency(td, llnsec);
1917 }
1918
1919 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1920 add_bw_sample(td, io_u, bytes, llnsec);
1921
1922 if (no_reduce && per_unit_log(td->iops_log))
1923 add_iops_sample(td, io_u, bytes);
1924 }
1925
1926 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1927 uint32_t *info = io_u_block_info(td, io_u);
1928 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1929 if (io_u->ddir == DDIR_TRIM) {
1930 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1931 BLOCK_INFO_TRIMS(*info) + 1);
1932 } else if (io_u->ddir == DDIR_WRITE) {
1933 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1934 *info);
1935 }
1936 }
1937 }
1938}
1939
1940static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1941 uint64_t offset, unsigned int bytes)
1942{
1943 int idx;
1944
1945 if (!f)
1946 return;
1947
1948 if (f->first_write == -1ULL || offset < f->first_write)
1949 f->first_write = offset;
1950 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1951 f->last_write = offset + bytes;
1952
1953 if (!f->last_write_comp)
1954 return;
1955
1956 idx = f->last_write_idx++;
1957 f->last_write_comp[idx] = offset;
1958 if (f->last_write_idx == td->o.iodepth)
1959 f->last_write_idx = 0;
1960}
1961
1962static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1963 struct io_completion_data *icd)
1964{
1965 struct io_u *io_u = *io_u_ptr;
1966 enum fio_ddir ddir = io_u->ddir;
1967 struct fio_file *f = io_u->file;
1968
1969 dprint_io_u(io_u, "complete");
1970
1971 assert(io_u->flags & IO_U_F_FLIGHT);
1972 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1973
1974 /*
1975 * Mark IO ok to verify
1976 */
1977 if (io_u->ipo) {
1978 /*
1979 * Remove errored entry from the verification list
1980 */
1981 if (io_u->error)
1982 unlog_io_piece(td, io_u);
1983 else {
1984 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1985 write_barrier();
1986 }
1987 }
1988
1989 if (ddir_sync(ddir)) {
1990 td->last_was_sync = 1;
1991 if (f) {
1992 f->first_write = -1ULL;
1993 f->last_write = -1ULL;
1994 }
1995 return;
1996 }
1997
1998 td->last_was_sync = 0;
1999 td->last_ddir = ddir;
2000
2001 if (!io_u->error && ddir_rw(ddir)) {
2002 unsigned int bytes = io_u->buflen - io_u->resid;
2003 int ret;
2004
2005 td->io_blocks[ddir]++;
2006 td->io_bytes[ddir] += bytes;
2007
2008 if (!(io_u->flags & IO_U_F_VER_LIST)) {
2009 td->this_io_blocks[ddir]++;
2010 td->this_io_bytes[ddir] += bytes;
2011 }
2012
2013 if (ddir == DDIR_WRITE)
2014 file_log_write_comp(td, f, io_u->offset, bytes);
2015
2016 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
2017 td->runstate == TD_VERIFYING))
2018 account_io_completion(td, io_u, icd, ddir, bytes);
2019
2020 icd->bytes_done[ddir] += bytes;
2021
2022 if (io_u->end_io) {
2023 ret = io_u->end_io(td, io_u_ptr);
2024 io_u = *io_u_ptr;
2025 if (ret && !icd->error)
2026 icd->error = ret;
2027 }
2028 } else if (io_u->error) {
2029 icd->error = io_u->error;
2030 io_u_log_error(td, io_u);
2031 }
2032 if (icd->error) {
2033 enum error_type_bit eb = td_error_type(ddir, icd->error);
2034
2035 if (!td_non_fatal_error(td, eb, icd->error))
2036 return;
2037
2038 /*
2039 * If there is a non_fatal error, then add to the error count
2040 * and clear all the errors.
2041 */
2042 update_error_count(td, icd->error);
2043 td_clear_error(td);
2044 icd->error = 0;
2045 if (io_u)
2046 io_u->error = 0;
2047 }
2048}
2049
2050static void init_icd(struct thread_data *td, struct io_completion_data *icd,
2051 int nr)
2052{
2053 int ddir;
2054
2055 if (!gtod_reduce(td))
2056 fio_gettime(&icd->time, NULL);
2057
2058 icd->nr = nr;
2059
2060 icd->error = 0;
2061 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2062 icd->bytes_done[ddir] = 0;
2063}
2064
2065static void ios_completed(struct thread_data *td,
2066 struct io_completion_data *icd)
2067{
2068 struct io_u *io_u;
2069 int i;
2070
2071 for (i = 0; i < icd->nr; i++) {
2072 io_u = td->io_ops->event(td, i);
2073
2074 io_completed(td, &io_u, icd);
2075
2076 if (io_u)
2077 put_io_u(td, io_u);
2078 }
2079}
2080
2081/*
2082 * Complete a single io_u for the sync engines.
2083 */
2084int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
2085{
2086 struct io_completion_data icd;
2087 int ddir;
2088
2089 init_icd(td, &icd, 1);
2090 io_completed(td, &io_u, &icd);
2091
2092 if (io_u)
2093 put_io_u(td, io_u);
2094
2095 if (icd.error) {
2096 td_verror(td, icd.error, "io_u_sync_complete");
2097 return -1;
2098 }
2099
2100 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2101 td->bytes_done[ddir] += icd.bytes_done[ddir];
2102
2103 return 0;
2104}
2105
2106/*
2107 * Called to complete min_events number of io for the async engines.
2108 */
2109int io_u_queued_complete(struct thread_data *td, int min_evts)
2110{
2111 struct io_completion_data icd;
2112 struct timespec *tvp = NULL;
2113 int ret, ddir;
2114 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2115
2116 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2117
2118 if (!min_evts)
2119 tvp = &ts;
2120 else if (min_evts > td->cur_depth)
2121 min_evts = td->cur_depth;
2122
2123 /* No worries, td_io_getevents fixes min and max if they are
2124 * set incorrectly */
2125 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2126 if (ret < 0) {
2127 td_verror(td, -ret, "td_io_getevents");
2128 return ret;
2129 } else if (!ret)
2130 return ret;
2131
2132 init_icd(td, &icd, ret);
2133 ios_completed(td, &icd);
2134 if (icd.error) {
2135 td_verror(td, icd.error, "io_u_queued_complete");
2136 return -1;
2137 }
2138
2139 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2140 td->bytes_done[ddir] += icd.bytes_done[ddir];
2141
2142 return ret;
2143}
2144
2145/*
2146 * Call when io_u is really queued, to update the submission latency.
2147 */
2148void io_u_queued(struct thread_data *td, struct io_u *io_u)
2149{
2150 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2151 unsigned long slat_time;
2152
2153 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2154
2155 if (td->parent)
2156 td = td->parent;
2157
2158 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2159 io_u->offset);
2160 }
2161}
2162
2163/*
2164 * See if we should reuse the last seed, if dedupe is enabled
2165 */
2166static struct frand_state *get_buf_state(struct thread_data *td)
2167{
2168 unsigned int v;
2169
2170 if (!td->o.dedupe_percentage)
2171 return &td->buf_state;
2172 else if (td->o.dedupe_percentage == 100) {
2173 frand_copy(&td->buf_state_prev, &td->buf_state);
2174 return &td->buf_state;
2175 }
2176
2177 v = rand32_between(&td->dedupe_state, 1, 100);
2178
2179 if (v <= td->o.dedupe_percentage)
2180 return &td->buf_state_prev;
2181
2182 return &td->buf_state;
2183}
2184
2185static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2186{
2187 if (td->o.dedupe_percentage == 100)
2188 frand_copy(rs, &td->buf_state_prev);
2189 else if (rs == &td->buf_state)
2190 frand_copy(&td->buf_state_prev, rs);
2191}
2192
2193void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2194 unsigned int max_bs)
2195{
2196 struct thread_options *o = &td->o;
2197
2198 if (o->mem_type == MEM_CUDA_MALLOC)
2199 return;
2200
2201 if (o->compress_percentage || o->dedupe_percentage) {
2202 unsigned int perc = td->o.compress_percentage;
2203 struct frand_state *rs;
2204 unsigned int left = max_bs;
2205 unsigned int this_write;
2206
2207 do {
2208 rs = get_buf_state(td);
2209
2210 min_write = min(min_write, left);
2211
2212 if (perc) {
2213 this_write = min_not_zero(min_write,
2214 td->o.compress_chunk);
2215
2216 fill_random_buf_percentage(rs, buf, perc,
2217 this_write, this_write,
2218 o->buffer_pattern,
2219 o->buffer_pattern_bytes);
2220 } else {
2221 fill_random_buf(rs, buf, min_write);
2222 this_write = min_write;
2223 }
2224
2225 buf += this_write;
2226 left -= this_write;
2227 save_buf_state(td, rs);
2228 } while (left);
2229 } else if (o->buffer_pattern_bytes)
2230 fill_buffer_pattern(td, buf, max_bs);
2231 else if (o->zero_buffers)
2232 memset(buf, 0, max_bs);
2233 else
2234 fill_random_buf(get_buf_state(td), buf, max_bs);
2235}
2236
2237/*
2238 * "randomly" fill the buffer contents
2239 */
2240void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2241 unsigned int min_write, unsigned int max_bs)
2242{
2243 io_u->buf_filled_len = 0;
2244 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2245}
2246
2247static int do_sync_file_range(const struct thread_data *td,
2248 struct fio_file *f)
2249{
2250 off64_t offset, nbytes;
2251
2252 offset = f->first_write;
2253 nbytes = f->last_write - f->first_write;
2254
2255 if (!nbytes)
2256 return 0;
2257
2258 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2259}
2260
2261int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2262{
2263 int ret;
2264
2265 if (io_u->ddir == DDIR_SYNC) {
2266 ret = fsync(io_u->file->fd);
2267 } else if (io_u->ddir == DDIR_DATASYNC) {
2268#ifdef CONFIG_FDATASYNC
2269 ret = fdatasync(io_u->file->fd);
2270#else
2271 ret = io_u->xfer_buflen;
2272 io_u->error = EINVAL;
2273#endif
2274 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2275 ret = do_sync_file_range(td, io_u->file);
2276 else {
2277 ret = io_u->xfer_buflen;
2278 io_u->error = EINVAL;
2279 }
2280
2281 if (ret < 0)
2282 io_u->error = errno;
2283
2284 return ret;
2285}
2286
2287int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2288{
2289#ifndef FIO_HAVE_TRIM
2290 io_u->error = EINVAL;
2291 return 0;
2292#else
2293 struct fio_file *f = io_u->file;
2294 int ret;
2295
2296 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2297 if (!ret)
2298 return io_u->xfer_buflen;
2299
2300 io_u->error = ret;
2301 return 0;
2302#endif
2303}
fio - Flexible IO Tester