projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Merge branch 'manpage_typos' of https://github.com/dirtyharrycallahan/fio
[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 unsigned int v;
167
168 lastb = last_block(td, f, ddir);
169 if (!lastb)
170 return 1;
171
172 if (!td->o.zone_split_nr[ddir]) {
173bail:
174 return __get_next_rand_offset(td, f, ddir, b, lastb);
175 }
176
177 /*
178 * Generate a value, v, between 1 and 100, both inclusive
179 */
180 v = rand32_between(&td->zone_state, 1, 100);
181
182 /*
183 * Find our generated table. 'send' is the end block of this zone,
184 * 'stotal' is our start offset.
185 */
186 zsi = &td->zone_state_index[ddir][v - 1];
187 stotal = zsi->size_prev / td->o.ba[ddir];
188 send = zsi->size / td->o.ba[ddir];
189
190 /*
191 * Should never happen
192 */
193 if (send == -1U) {
194 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
195 log_err("fio: bug in zoned generation\n");
196 goto bail;
197 } else if (send > lastb) {
198 /*
199 * This happens if the user specifies ranges that exceed
200 * the file/device size. We can't handle that gracefully,
201 * so error and exit.
202 */
203 log_err("fio: zoned_abs sizes exceed file size\n");
204 return 1;
205 }
206
207 /*
208 * Generate index from 0..send-stotal
209 */
210 if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1)
211 return 1;
212
213 *b += stotal;
214 return 0;
215}
216
217static int __get_next_rand_offset_zoned(struct thread_data *td,
218 struct fio_file *f, enum fio_ddir ddir,
219 uint64_t *b)
220{
221 unsigned int v, send, stotal;
222 uint64_t offset, lastb;
223 struct zone_split_index *zsi;
224
225 lastb = last_block(td, f, ddir);
226 if (!lastb)
227 return 1;
228
229 if (!td->o.zone_split_nr[ddir]) {
230bail:
231 return __get_next_rand_offset(td, f, ddir, b, lastb);
232 }
233
234 /*
235 * Generate a value, v, between 1 and 100, both inclusive
236 */
237 v = rand32_between(&td->zone_state, 1, 100);
238
239 zsi = &td->zone_state_index[ddir][v - 1];
240 stotal = zsi->size_perc_prev;
241 send = zsi->size_perc;
242
243 /*
244 * Should never happen
245 */
246 if (send == -1U) {
247 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
248 log_err("fio: bug in zoned generation\n");
249 goto bail;
250 }
251
252 /*
253 * 'send' is some percentage below or equal to 100 that
254 * marks the end of the current IO range. 'stotal' marks
255 * the start, in percent.
256 */
257 if (stotal)
258 offset = stotal * lastb / 100ULL;
259 else
260 offset = 0;
261
262 lastb = lastb * (send - stotal) / 100ULL;
263
264 /*
265 * Generate index from 0..send-of-lastb
266 */
267 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
268 return 1;
269
270 /*
271 * Add our start offset, if any
272 */
273 if (offset)
274 *b += offset;
275
276 return 0;
277}
278
279static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
280{
281 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
282 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
283
284 return r1->off - r2->off;
285}
286
287static int get_off_from_method(struct thread_data *td, struct fio_file *f,
288 enum fio_ddir ddir, uint64_t *b)
289{
290 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
291 uint64_t lastb;
292
293 lastb = last_block(td, f, ddir);
294 if (!lastb)
295 return 1;
296
297 return __get_next_rand_offset(td, f, ddir, b, lastb);
298 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
299 return __get_next_rand_offset_zipf(td, f, ddir, b);
300 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
301 return __get_next_rand_offset_pareto(td, f, ddir, b);
302 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
303 return __get_next_rand_offset_gauss(td, f, ddir, b);
304 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
305 return __get_next_rand_offset_zoned(td, f, ddir, b);
306 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
307 return __get_next_rand_offset_zoned_abs(td, f, ddir, b);
308
309 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
310 return 1;
311}
312
313/*
314 * Sort the reads for a verify phase in batches of verifysort_nr, if
315 * specified.
316 */
317static inline bool should_sort_io(struct thread_data *td)
318{
319 if (!td->o.verifysort_nr || !td->o.do_verify)
320 return false;
321 if (!td_random(td))
322 return false;
323 if (td->runstate != TD_VERIFYING)
324 return false;
325 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
326 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
327 return false;
328
329 return true;
330}
331
332static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
333{
334 unsigned int v;
335
336 if (td->o.perc_rand[ddir] == 100)
337 return true;
338
339 v = rand32_between(&td->seq_rand_state[ddir], 1, 100);
340
341 return v <= td->o.perc_rand[ddir];
342}
343
344static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
345 enum fio_ddir ddir, uint64_t *b)
346{
347 struct rand_off *r;
348 int i, ret = 1;
349
350 if (!should_sort_io(td))
351 return get_off_from_method(td, f, ddir, b);
352
353 if (!flist_empty(&td->next_rand_list)) {
354fetch:
355 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
356 flist_del(&r->list);
357 *b = r->off;
358 free(r);
359 return 0;
360 }
361
362 for (i = 0; i < td->o.verifysort_nr; i++) {
363 r = malloc(sizeof(*r));
364
365 ret = get_off_from_method(td, f, ddir, &r->off);
366 if (ret) {
367 free(r);
368 break;
369 }
370
371 flist_add(&r->list, &td->next_rand_list);
372 }
373
374 if (ret && !i)
375 return ret;
376
377 assert(!flist_empty(&td->next_rand_list));
378 flist_sort(NULL, &td->next_rand_list, flist_cmp);
379 goto fetch;
380}
381
382static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f)
383{
384 struct thread_options *o = &td->o;
385
386 if (o->invalidate_cache && !o->odirect) {
387 int fio_unused ret;
388
389 ret = file_invalidate_cache(td, f);
390 }
391}
392
393static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
394 enum fio_ddir ddir, uint64_t *b)
395{
396 if (!get_next_rand_offset(td, f, ddir, b))
397 return 0;
398
399 if (td->o.time_based ||
400 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
401 fio_file_reset(td, f);
402 if (!get_next_rand_offset(td, f, ddir, b))
403 return 0;
404 loop_cache_invalidate(td, f);
405 }
406
407 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
408 f->file_name, (unsigned long long) f->last_pos[ddir],
409 (unsigned long long) f->real_file_size);
410 return 1;
411}
412
413static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
414 enum fio_ddir ddir, uint64_t *offset)
415{
416 struct thread_options *o = &td->o;
417
418 assert(ddir_rw(ddir));
419
420 /*
421 * If we reach the end for a time based run, reset us back to 0
422 * and invalidate the cache, if we need to.
423 */
424 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
425 o->time_based) {
426 f->last_pos[ddir] = f->file_offset;
427 loop_cache_invalidate(td, f);
428 }
429
430 if (f->last_pos[ddir] < f->real_file_size) {
431 uint64_t pos;
432
433 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) {
434 if (f->real_file_size > f->io_size)
435 f->last_pos[ddir] = f->io_size;
436 else
437 f->last_pos[ddir] = f->real_file_size;
438 }
439
440 pos = f->last_pos[ddir] - f->file_offset;
441 if (pos && o->ddir_seq_add) {
442 pos += o->ddir_seq_add;
443
444 /*
445 * If we reach beyond the end of the file
446 * with holed IO, wrap around to the
447 * beginning again. If we're doing backwards IO,
448 * wrap to the end.
449 */
450 if (pos >= f->real_file_size) {
451 if (o->ddir_seq_add > 0)
452 pos = f->file_offset;
453 else {
454 if (f->real_file_size > f->io_size)
455 pos = f->io_size;
456 else
457 pos = f->real_file_size;
458
459 pos += o->ddir_seq_add;
460 }
461 }
462 }
463
464 *offset = pos;
465 return 0;
466 }
467
468 return 1;
469}
470
471static int get_next_block(struct thread_data *td, struct io_u *io_u,
472 enum fio_ddir ddir, int rw_seq,
473 unsigned int *is_random)
474{
475 struct fio_file *f = io_u->file;
476 uint64_t b, offset;
477 int ret;
478
479 assert(ddir_rw(ddir));
480
481 b = offset = -1ULL;
482
483 if (rw_seq) {
484 if (td_random(td)) {
485 if (should_do_random(td, ddir)) {
486 ret = get_next_rand_block(td, f, ddir, &b);
487 *is_random = 1;
488 } else {
489 *is_random = 0;
490 io_u_set(td, io_u, IO_U_F_BUSY_OK);
491 ret = get_next_seq_offset(td, f, ddir, &offset);
492 if (ret)
493 ret = get_next_rand_block(td, f, ddir, &b);
494 }
495 } else {
496 *is_random = 0;
497 ret = get_next_seq_offset(td, f, ddir, &offset);
498 }
499 } else {
500 io_u_set(td, io_u, IO_U_F_BUSY_OK);
501 *is_random = 0;
502
503 if (td->o.rw_seq == RW_SEQ_SEQ) {
504 ret = get_next_seq_offset(td, f, ddir, &offset);
505 if (ret) {
506 ret = get_next_rand_block(td, f, ddir, &b);
507 *is_random = 0;
508 }
509 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
510 if (f->last_start[ddir] != -1ULL)
511 offset = f->last_start[ddir] - f->file_offset;
512 else
513 offset = 0;
514 ret = 0;
515 } else {
516 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
517 ret = 1;
518 }
519 }
520
521 if (!ret) {
522 if (offset != -1ULL)
523 io_u->offset = offset;
524 else if (b != -1ULL)
525 io_u->offset = b * td->o.ba[ddir];
526 else {
527 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
528 ret = 1;
529 }
530 }
531
532 return ret;
533}
534
535/*
536 * For random io, generate a random new block and see if it's used. Repeat
537 * until we find a free one. For sequential io, just return the end of
538 * the last io issued.
539 */
540static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
541 unsigned int *is_random)
542{
543 struct fio_file *f = io_u->file;
544 enum fio_ddir ddir = io_u->ddir;
545 int rw_seq_hit = 0;
546
547 assert(ddir_rw(ddir));
548
549 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
550 rw_seq_hit = 1;
551 td->ddir_seq_nr = td->o.ddir_seq_nr;
552 }
553
554 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
555 return 1;
556
557 if (io_u->offset >= f->io_size) {
558 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
559 (unsigned long long) io_u->offset,
560 (unsigned long long) f->io_size);
561 return 1;
562 }
563
564 io_u->offset += f->file_offset;
565 if (io_u->offset >= f->real_file_size) {
566 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
567 (unsigned long long) io_u->offset,
568 (unsigned long long) f->real_file_size);
569 return 1;
570 }
571
572 return 0;
573}
574
575static int get_next_offset(struct thread_data *td, struct io_u *io_u,
576 unsigned int *is_random)
577{
578 if (td->flags & TD_F_PROFILE_OPS) {
579 struct prof_io_ops *ops = &td->prof_io_ops;
580
581 if (ops->fill_io_u_off)
582 return ops->fill_io_u_off(td, io_u, is_random);
583 }
584
585 return __get_next_offset(td, io_u, is_random);
586}
587
588static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
589 unsigned int buflen)
590{
591 struct fio_file *f = io_u->file;
592
593 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
594}
595
596static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
597 unsigned int is_random)
598{
599 int ddir = io_u->ddir;
600 unsigned int buflen = 0;
601 unsigned int minbs, maxbs;
602 uint64_t frand_max, r;
603 bool power_2;
604
605 assert(ddir_rw(ddir));
606
607 if (td->o.bs_is_seq_rand)
608 ddir = is_random ? DDIR_WRITE: DDIR_READ;
609
610 minbs = td->o.min_bs[ddir];
611 maxbs = td->o.max_bs[ddir];
612
613 if (minbs == maxbs)
614 return minbs;
615
616 /*
617 * If we can't satisfy the min block size from here, then fail
618 */
619 if (!io_u_fits(td, io_u, minbs))
620 return 0;
621
622 frand_max = rand_max(&td->bsrange_state[ddir]);
623 do {
624 r = __rand(&td->bsrange_state[ddir]);
625
626 if (!td->o.bssplit_nr[ddir]) {
627 buflen = 1 + (unsigned int) ((double) maxbs *
628 (r / (frand_max + 1.0)));
629 if (buflen < minbs)
630 buflen = minbs;
631 } else {
632 long long perc = 0;
633 unsigned int i;
634
635 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
636 struct bssplit *bsp = &td->o.bssplit[ddir][i];
637
638 buflen = bsp->bs;
639 perc += bsp->perc;
640 if (!perc)
641 break;
642 if ((r / perc <= frand_max / 100ULL) &&
643 io_u_fits(td, io_u, buflen))
644 break;
645 }
646 }
647
648 power_2 = is_power_of_2(minbs);
649 if (!td->o.bs_unaligned && power_2)
650 buflen &= ~(minbs - 1);
651 else if (!td->o.bs_unaligned && !power_2)
652 buflen -= buflen % minbs;
653 } while (!io_u_fits(td, io_u, buflen));
654
655 return buflen;
656}
657
658static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
659 unsigned int is_random)
660{
661 if (td->flags & TD_F_PROFILE_OPS) {
662 struct prof_io_ops *ops = &td->prof_io_ops;
663
664 if (ops->fill_io_u_size)
665 return ops->fill_io_u_size(td, io_u, is_random);
666 }
667
668 return __get_next_buflen(td, io_u, is_random);
669}
670
671static void set_rwmix_bytes(struct thread_data *td)
672{
673 unsigned int diff;
674
675 /*
676 * we do time or byte based switch. this is needed because
677 * buffered writes may issue a lot quicker than they complete,
678 * whereas reads do not.
679 */
680 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
681 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
682}
683
684static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
685{
686 unsigned int v;
687
688 v = rand32_between(&td->rwmix_state, 1, 100);
689
690 if (v <= td->o.rwmix[DDIR_READ])
691 return DDIR_READ;
692
693 return DDIR_WRITE;
694}
695
696int io_u_quiesce(struct thread_data *td)
697{
698 int completed = 0;
699
700 /*
701 * We are going to sleep, ensure that we flush anything pending as
702 * not to skew our latency numbers.
703 *
704 * Changed to only monitor 'in flight' requests here instead of the
705 * td->cur_depth, b/c td->cur_depth does not accurately represent
706 * io's that have been actually submitted to an async engine,
707 * and cur_depth is meaningless for sync engines.
708 */
709 if (td->io_u_queued || td->cur_depth) {
710 int fio_unused ret;
711
712 ret = td_io_commit(td);
713 }
714
715 while (td->io_u_in_flight) {
716 int ret;
717
718 ret = io_u_queued_complete(td, 1);
719 if (ret > 0)
720 completed += ret;
721 }
722
723 if (td->flags & TD_F_REGROW_LOGS)
724 regrow_logs(td);
725
726 return completed;
727}
728
729static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
730{
731 enum fio_ddir odir = ddir ^ 1;
732 uint64_t usec;
733 uint64_t now;
734
735 assert(ddir_rw(ddir));
736 now = utime_since_now(&td->start);
737
738 /*
739 * if rate_next_io_time is in the past, need to catch up to rate
740 */
741 if (td->rate_next_io_time[ddir] <= now)
742 return ddir;
743
744 /*
745 * We are ahead of rate in this direction. See if we
746 * should switch.
747 */
748 if (td_rw(td) && td->o.rwmix[odir]) {
749 /*
750 * Other direction is behind rate, switch
751 */
752 if (td->rate_next_io_time[odir] <= now)
753 return odir;
754
755 /*
756 * Both directions are ahead of rate. sleep the min,
757 * switch if necessary
758 */
759 if (td->rate_next_io_time[ddir] <=
760 td->rate_next_io_time[odir]) {
761 usec = td->rate_next_io_time[ddir] - now;
762 } else {
763 usec = td->rate_next_io_time[odir] - now;
764 ddir = odir;
765 }
766 } else
767 usec = td->rate_next_io_time[ddir] - now;
768
769 if (td->o.io_submit_mode == IO_MODE_INLINE)
770 io_u_quiesce(td);
771
772 usec_sleep(td, usec);
773 return ddir;
774}
775
776/*
777 * Return the data direction for the next io_u. If the job is a
778 * mixed read/write workload, check the rwmix cycle and switch if
779 * necessary.
780 */
781static enum fio_ddir get_rw_ddir(struct thread_data *td)
782{
783 enum fio_ddir ddir;
784
785 /*
786 * See if it's time to fsync/fdatasync/sync_file_range first,
787 * and if not then move on to check regular I/Os.
788 */
789 if (should_fsync(td)) {
790 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
791 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
792 return DDIR_SYNC;
793
794 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
795 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
796 return DDIR_DATASYNC;
797
798 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
799 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
800 return DDIR_SYNC_FILE_RANGE;
801 }
802
803 if (td_rw(td)) {
804 /*
805 * Check if it's time to seed a new data direction.
806 */
807 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
808 /*
809 * Put a top limit on how many bytes we do for
810 * one data direction, to avoid overflowing the
811 * ranges too much
812 */
813 ddir = get_rand_ddir(td);
814
815 if (ddir != td->rwmix_ddir)
816 set_rwmix_bytes(td);
817
818 td->rwmix_ddir = ddir;
819 }
820 ddir = td->rwmix_ddir;
821 } else if (td_read(td))
822 ddir = DDIR_READ;
823 else if (td_write(td))
824 ddir = DDIR_WRITE;
825 else if (td_trim(td))
826 ddir = DDIR_TRIM;
827 else
828 ddir = DDIR_INVAL;
829
830 td->rwmix_ddir = rate_ddir(td, ddir);
831 return td->rwmix_ddir;
832}
833
834static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
835{
836 enum fio_ddir ddir = get_rw_ddir(td);
837
838 if (td_trimwrite(td)) {
839 struct fio_file *f = io_u->file;
840 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
841 ddir = DDIR_TRIM;
842 else
843 ddir = DDIR_WRITE;
844 }
845
846 io_u->ddir = io_u->acct_ddir = ddir;
847
848 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
849 td->o.barrier_blocks &&
850 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
851 td->io_issues[DDIR_WRITE])
852 io_u_set(td, io_u, IO_U_F_BARRIER);
853}
854
855void put_file_log(struct thread_data *td, struct fio_file *f)
856{
857 unsigned int ret = put_file(td, f);
858
859 if (ret)
860 td_verror(td, ret, "file close");
861}
862
863void put_io_u(struct thread_data *td, struct io_u *io_u)
864{
865 if (td->parent)
866 td = td->parent;
867
868 td_io_u_lock(td);
869
870 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
871 put_file_log(td, io_u->file);
872
873 io_u->file = NULL;
874 io_u_set(td, io_u, IO_U_F_FREE);
875
876 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
877 td->cur_depth--;
878 assert(!(td->flags & TD_F_CHILD));
879 }
880 io_u_qpush(&td->io_u_freelist, io_u);
881 td_io_u_unlock(td);
882 td_io_u_free_notify(td);
883}
884
885void clear_io_u(struct thread_data *td, struct io_u *io_u)
886{
887 io_u_clear(td, io_u, IO_U_F_FLIGHT);
888 put_io_u(td, io_u);
889}
890
891void requeue_io_u(struct thread_data *td, struct io_u **io_u)
892{
893 struct io_u *__io_u = *io_u;
894 enum fio_ddir ddir = acct_ddir(__io_u);
895
896 dprint(FD_IO, "requeue %p\n", __io_u);
897
898 if (td->parent)
899 td = td->parent;
900
901 td_io_u_lock(td);
902
903 io_u_set(td, __io_u, IO_U_F_FREE);
904 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
905 td->io_issues[ddir]--;
906
907 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
908 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
909 td->cur_depth--;
910 assert(!(td->flags & TD_F_CHILD));
911 }
912
913 io_u_rpush(&td->io_u_requeues, __io_u);
914 td_io_u_unlock(td);
915 td_io_u_free_notify(td);
916 *io_u = NULL;
917}
918
919static void __fill_io_u_zone(struct thread_data *td, struct io_u *io_u)
920{
921 struct fio_file *f = io_u->file;
922
923 /*
924 * See if it's time to switch to a new zone
925 */
926 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
927 td->zone_bytes = 0;
928 f->file_offset += td->o.zone_range + td->o.zone_skip;
929
930 /*
931 * Wrap from the beginning, if we exceed the file size
932 */
933 if (f->file_offset >= f->real_file_size)
934 f->file_offset = f->real_file_size - f->file_offset;
935 f->last_pos[io_u->ddir] = f->file_offset;
936 td->io_skip_bytes += td->o.zone_skip;
937 }
938
939 /*
940 * If zone_size > zone_range, then maintain the same zone until
941 * zone_bytes >= zone_size.
942 */
943 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
944 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
945 f->file_offset, f->last_pos[io_u->ddir]);
946 f->last_pos[io_u->ddir] = f->file_offset;
947 }
948
949 /*
950 * For random: if 'norandommap' is not set and zone_size > zone_range,
951 * map needs to be reset as it's done with zone_range everytime.
952 */
953 if ((td->zone_bytes % td->o.zone_range) == 0) {
954 fio_file_reset(td, f);
955 }
956}
957
958static int fill_io_u(struct thread_data *td, struct io_u *io_u)
959{
960 unsigned int is_random;
961
962 if (td_ioengine_flagged(td, FIO_NOIO))
963 goto out;
964
965 set_rw_ddir(td, io_u);
966
967 /*
968 * fsync() or fdatasync() or trim etc, we are done
969 */
970 if (!ddir_rw(io_u->ddir))
971 goto out;
972
973 /*
974 * When file is zoned zone_range is always positive
975 */
976 if (td->o.zone_range) {
977 __fill_io_u_zone(td, io_u);
978 }
979
980 /*
981 * No log, let the seq/rand engine retrieve the next buflen and
982 * position.
983 */
984 if (get_next_offset(td, io_u, &is_random)) {
985 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
986 return 1;
987 }
988
989 io_u->buflen = get_next_buflen(td, io_u, is_random);
990 if (!io_u->buflen) {
991 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
992 return 1;
993 }
994
995 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
996 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%lx exceeds file size=0x%llx\n",
997 io_u,
998 (unsigned long long) io_u->offset, io_u->buflen,
999 (unsigned long long) io_u->file->real_file_size);
1000 return 1;
1001 }
1002
1003 /*
1004 * mark entry before potentially trimming io_u
1005 */
1006 if (td_random(td) && file_randommap(td, io_u->file))
1007 mark_random_map(td, io_u);
1008
1009out:
1010 dprint_io_u(io_u, "fill");
1011 td->zone_bytes += io_u->buflen;
1012 return 0;
1013}
1014
1015static void __io_u_mark_map(uint64_t *map, unsigned int nr)
1016{
1017 int idx = 0;
1018
1019 switch (nr) {
1020 default:
1021 idx = 6;
1022 break;
1023 case 33 ... 64:
1024 idx = 5;
1025 break;
1026 case 17 ... 32:
1027 idx = 4;
1028 break;
1029 case 9 ... 16:
1030 idx = 3;
1031 break;
1032 case 5 ... 8:
1033 idx = 2;
1034 break;
1035 case 1 ... 4:
1036 idx = 1;
1037 case 0:
1038 break;
1039 }
1040
1041 map[idx]++;
1042}
1043
1044void io_u_mark_submit(struct thread_data *td, unsigned int nr)
1045{
1046 __io_u_mark_map(td->ts.io_u_submit, nr);
1047 td->ts.total_submit++;
1048}
1049
1050void io_u_mark_complete(struct thread_data *td, unsigned int nr)
1051{
1052 __io_u_mark_map(td->ts.io_u_complete, nr);
1053 td->ts.total_complete++;
1054}
1055
1056void io_u_mark_depth(struct thread_data *td, unsigned int nr)
1057{
1058 int idx = 0;
1059
1060 switch (td->cur_depth) {
1061 default:
1062 idx = 6;
1063 break;
1064 case 32 ... 63:
1065 idx = 5;
1066 break;
1067 case 16 ... 31:
1068 idx = 4;
1069 break;
1070 case 8 ... 15:
1071 idx = 3;
1072 break;
1073 case 4 ... 7:
1074 idx = 2;
1075 break;
1076 case 2 ... 3:
1077 idx = 1;
1078 case 1:
1079 break;
1080 }
1081
1082 td->ts.io_u_map[idx] += nr;
1083}
1084
1085static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1086{
1087 int idx = 0;
1088
1089 assert(nsec < 1000);
1090
1091 switch (nsec) {
1092 case 750 ... 999:
1093 idx = 9;
1094 break;
1095 case 500 ... 749:
1096 idx = 8;
1097 break;
1098 case 250 ... 499:
1099 idx = 7;
1100 break;
1101 case 100 ... 249:
1102 idx = 6;
1103 break;
1104 case 50 ... 99:
1105 idx = 5;
1106 break;
1107 case 20 ... 49:
1108 idx = 4;
1109 break;
1110 case 10 ... 19:
1111 idx = 3;
1112 break;
1113 case 4 ... 9:
1114 idx = 2;
1115 break;
1116 case 2 ... 3:
1117 idx = 1;
1118 case 0 ... 1:
1119 break;
1120 }
1121
1122 assert(idx < FIO_IO_U_LAT_N_NR);
1123 td->ts.io_u_lat_n[idx]++;
1124}
1125
1126static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1127{
1128 int idx = 0;
1129
1130 assert(usec < 1000 && usec >= 1);
1131
1132 switch (usec) {
1133 case 750 ... 999:
1134 idx = 9;
1135 break;
1136 case 500 ... 749:
1137 idx = 8;
1138 break;
1139 case 250 ... 499:
1140 idx = 7;
1141 break;
1142 case 100 ... 249:
1143 idx = 6;
1144 break;
1145 case 50 ... 99:
1146 idx = 5;
1147 break;
1148 case 20 ... 49:
1149 idx = 4;
1150 break;
1151 case 10 ... 19:
1152 idx = 3;
1153 break;
1154 case 4 ... 9:
1155 idx = 2;
1156 break;
1157 case 2 ... 3:
1158 idx = 1;
1159 case 0 ... 1:
1160 break;
1161 }
1162
1163 assert(idx < FIO_IO_U_LAT_U_NR);
1164 td->ts.io_u_lat_u[idx]++;
1165}
1166
1167static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1168{
1169 int idx = 0;
1170
1171 assert(msec >= 1);
1172
1173 switch (msec) {
1174 default:
1175 idx = 11;
1176 break;
1177 case 1000 ... 1999:
1178 idx = 10;
1179 break;
1180 case 750 ... 999:
1181 idx = 9;
1182 break;
1183 case 500 ... 749:
1184 idx = 8;
1185 break;
1186 case 250 ... 499:
1187 idx = 7;
1188 break;
1189 case 100 ... 249:
1190 idx = 6;
1191 break;
1192 case 50 ... 99:
1193 idx = 5;
1194 break;
1195 case 20 ... 49:
1196 idx = 4;
1197 break;
1198 case 10 ... 19:
1199 idx = 3;
1200 break;
1201 case 4 ... 9:
1202 idx = 2;
1203 break;
1204 case 2 ... 3:
1205 idx = 1;
1206 case 0 ... 1:
1207 break;
1208 }
1209
1210 assert(idx < FIO_IO_U_LAT_M_NR);
1211 td->ts.io_u_lat_m[idx]++;
1212}
1213
1214static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1215{
1216 if (nsec < 1000)
1217 io_u_mark_lat_nsec(td, nsec);
1218 else if (nsec < 1000000)
1219 io_u_mark_lat_usec(td, nsec / 1000);
1220 else
1221 io_u_mark_lat_msec(td, nsec / 1000000);
1222}
1223
1224static unsigned int __get_next_fileno_rand(struct thread_data *td)
1225{
1226 unsigned long fileno;
1227
1228 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1229 uint64_t frand_max = rand_max(&td->next_file_state);
1230 unsigned long r;
1231
1232 r = __rand(&td->next_file_state);
1233 return (unsigned int) ((double) td->o.nr_files
1234 * (r / (frand_max + 1.0)));
1235 }
1236
1237 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1238 fileno = zipf_next(&td->next_file_zipf);
1239 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1240 fileno = pareto_next(&td->next_file_zipf);
1241 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1242 fileno = gauss_next(&td->next_file_gauss);
1243 else {
1244 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1245 assert(0);
1246 return 0;
1247 }
1248
1249 return fileno >> FIO_FSERVICE_SHIFT;
1250}
1251
1252/*
1253 * Get next file to service by choosing one at random
1254 */
1255static struct fio_file *get_next_file_rand(struct thread_data *td,
1256 enum fio_file_flags goodf,
1257 enum fio_file_flags badf)
1258{
1259 struct fio_file *f;
1260 int fno;
1261
1262 do {
1263 int opened = 0;
1264
1265 fno = __get_next_fileno_rand(td);
1266
1267 f = td->files[fno];
1268 if (fio_file_done(f))
1269 continue;
1270
1271 if (!fio_file_open(f)) {
1272 int err;
1273
1274 if (td->nr_open_files >= td->o.open_files)
1275 return ERR_PTR(-EBUSY);
1276
1277 err = td_io_open_file(td, f);
1278 if (err)
1279 continue;
1280 opened = 1;
1281 }
1282
1283 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1284 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1285 return f;
1286 }
1287 if (opened)
1288 td_io_close_file(td, f);
1289 } while (1);
1290}
1291
1292/*
1293 * Get next file to service by doing round robin between all available ones
1294 */
1295static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1296 int badf)
1297{
1298 unsigned int old_next_file = td->next_file;
1299 struct fio_file *f;
1300
1301 do {
1302 int opened = 0;
1303
1304 f = td->files[td->next_file];
1305
1306 td->next_file++;
1307 if (td->next_file >= td->o.nr_files)
1308 td->next_file = 0;
1309
1310 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1311 if (fio_file_done(f)) {
1312 f = NULL;
1313 continue;
1314 }
1315
1316 if (!fio_file_open(f)) {
1317 int err;
1318
1319 if (td->nr_open_files >= td->o.open_files)
1320 return ERR_PTR(-EBUSY);
1321
1322 err = td_io_open_file(td, f);
1323 if (err) {
1324 dprint(FD_FILE, "error %d on open of %s\n",
1325 err, f->file_name);
1326 f = NULL;
1327 continue;
1328 }
1329 opened = 1;
1330 }
1331
1332 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1333 f->flags);
1334 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1335 break;
1336
1337 if (opened)
1338 td_io_close_file(td, f);
1339
1340 f = NULL;
1341 } while (td->next_file != old_next_file);
1342
1343 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1344 return f;
1345}
1346
1347static struct fio_file *__get_next_file(struct thread_data *td)
1348{
1349 struct fio_file *f;
1350
1351 assert(td->o.nr_files <= td->files_index);
1352
1353 if (td->nr_done_files >= td->o.nr_files) {
1354 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1355 " nr_files=%d\n", td->nr_open_files,
1356 td->nr_done_files,
1357 td->o.nr_files);
1358 return NULL;
1359 }
1360
1361 f = td->file_service_file;
1362 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1363 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1364 goto out;
1365 if (td->file_service_left--)
1366 goto out;
1367 }
1368
1369 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1370 td->o.file_service_type == FIO_FSERVICE_SEQ)
1371 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1372 else
1373 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1374
1375 if (IS_ERR(f))
1376 return f;
1377
1378 td->file_service_file = f;
1379 td->file_service_left = td->file_service_nr - 1;
1380out:
1381 if (f)
1382 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1383 else
1384 dprint(FD_FILE, "get_next_file: NULL\n");
1385 return f;
1386}
1387
1388static struct fio_file *get_next_file(struct thread_data *td)
1389{
1390 if (td->flags & TD_F_PROFILE_OPS) {
1391 struct prof_io_ops *ops = &td->prof_io_ops;
1392
1393 if (ops->get_next_file)
1394 return ops->get_next_file(td);
1395 }
1396
1397 return __get_next_file(td);
1398}
1399
1400static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1401{
1402 struct fio_file *f;
1403
1404 do {
1405 f = get_next_file(td);
1406 if (IS_ERR_OR_NULL(f))
1407 return PTR_ERR(f);
1408
1409 io_u->file = f;
1410 get_file(f);
1411
1412 if (!fill_io_u(td, io_u))
1413 break;
1414
1415 put_file_log(td, f);
1416 td_io_close_file(td, f);
1417 io_u->file = NULL;
1418 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1419 fio_file_reset(td, f);
1420 else {
1421 fio_file_set_done(f);
1422 td->nr_done_files++;
1423 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1424 td->nr_done_files, td->o.nr_files);
1425 }
1426 } while (1);
1427
1428 return 0;
1429}
1430
1431static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1432 unsigned long long tnsec, unsigned long long max_nsec)
1433{
1434 if (!td->error)
1435 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
1436 td_verror(td, ETIMEDOUT, "max latency exceeded");
1437 icd->error = ETIMEDOUT;
1438}
1439
1440static void lat_new_cycle(struct thread_data *td)
1441{
1442 fio_gettime(&td->latency_ts, NULL);
1443 td->latency_ios = ddir_rw_sum(td->io_blocks);
1444 td->latency_failed = 0;
1445}
1446
1447/*
1448 * We had an IO outside the latency target. Reduce the queue depth. If we
1449 * are at QD=1, then it's time to give up.
1450 */
1451static bool __lat_target_failed(struct thread_data *td)
1452{
1453 if (td->latency_qd == 1)
1454 return true;
1455
1456 td->latency_qd_high = td->latency_qd;
1457
1458 if (td->latency_qd == td->latency_qd_low)
1459 td->latency_qd_low--;
1460
1461 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1462
1463 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1464
1465 /*
1466 * When we ramp QD down, quiesce existing IO to prevent
1467 * a storm of ramp downs due to pending higher depth.
1468 */
1469 io_u_quiesce(td);
1470 lat_new_cycle(td);
1471 return false;
1472}
1473
1474static bool lat_target_failed(struct thread_data *td)
1475{
1476 if (td->o.latency_percentile.u.f == 100.0)
1477 return __lat_target_failed(td);
1478
1479 td->latency_failed++;
1480 return false;
1481}
1482
1483void lat_target_init(struct thread_data *td)
1484{
1485 td->latency_end_run = 0;
1486
1487 if (td->o.latency_target) {
1488 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1489 fio_gettime(&td->latency_ts, NULL);
1490 td->latency_qd = 1;
1491 td->latency_qd_high = td->o.iodepth;
1492 td->latency_qd_low = 1;
1493 td->latency_ios = ddir_rw_sum(td->io_blocks);
1494 } else
1495 td->latency_qd = td->o.iodepth;
1496}
1497
1498void lat_target_reset(struct thread_data *td)
1499{
1500 if (!td->latency_end_run)
1501 lat_target_init(td);
1502}
1503
1504static void lat_target_success(struct thread_data *td)
1505{
1506 const unsigned int qd = td->latency_qd;
1507 struct thread_options *o = &td->o;
1508
1509 td->latency_qd_low = td->latency_qd;
1510
1511 /*
1512 * If we haven't failed yet, we double up to a failing value instead
1513 * of bisecting from highest possible queue depth. If we have set
1514 * a limit other than td->o.iodepth, bisect between that.
1515 */
1516 if (td->latency_qd_high != o->iodepth)
1517 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1518 else
1519 td->latency_qd *= 2;
1520
1521 if (td->latency_qd > o->iodepth)
1522 td->latency_qd = o->iodepth;
1523
1524 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1525
1526 /*
1527 * Same as last one, we are done. Let it run a latency cycle, so
1528 * we get only the results from the targeted depth.
1529 */
1530 if (td->latency_qd == qd) {
1531 if (td->latency_end_run) {
1532 dprint(FD_RATE, "We are done\n");
1533 td->done = 1;
1534 } else {
1535 dprint(FD_RATE, "Quiesce and final run\n");
1536 io_u_quiesce(td);
1537 td->latency_end_run = 1;
1538 reset_all_stats(td);
1539 reset_io_stats(td);
1540 }
1541 }
1542
1543 lat_new_cycle(td);
1544}
1545
1546/*
1547 * Check if we can bump the queue depth
1548 */
1549void lat_target_check(struct thread_data *td)
1550{
1551 uint64_t usec_window;
1552 uint64_t ios;
1553 double success_ios;
1554
1555 usec_window = utime_since_now(&td->latency_ts);
1556 if (usec_window < td->o.latency_window)
1557 return;
1558
1559 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1560 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1561 success_ios *= 100.0;
1562
1563 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1564
1565 if (success_ios >= td->o.latency_percentile.u.f)
1566 lat_target_success(td);
1567 else
1568 __lat_target_failed(td);
1569}
1570
1571/*
1572 * If latency target is enabled, we might be ramping up or down and not
1573 * using the full queue depth available.
1574 */
1575bool queue_full(const struct thread_data *td)
1576{
1577 const int qempty = io_u_qempty(&td->io_u_freelist);
1578
1579 if (qempty)
1580 return true;
1581 if (!td->o.latency_target)
1582 return false;
1583
1584 return td->cur_depth >= td->latency_qd;
1585}
1586
1587struct io_u *__get_io_u(struct thread_data *td)
1588{
1589 struct io_u *io_u = NULL;
1590
1591 if (td->stop_io)
1592 return NULL;
1593
1594 td_io_u_lock(td);
1595
1596again:
1597 if (!io_u_rempty(&td->io_u_requeues))
1598 io_u = io_u_rpop(&td->io_u_requeues);
1599 else if (!queue_full(td)) {
1600 io_u = io_u_qpop(&td->io_u_freelist);
1601
1602 io_u->file = NULL;
1603 io_u->buflen = 0;
1604 io_u->resid = 0;
1605 io_u->end_io = NULL;
1606 }
1607
1608 if (io_u) {
1609 assert(io_u->flags & IO_U_F_FREE);
1610 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1611 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1612 IO_U_F_VER_LIST);
1613
1614 io_u->error = 0;
1615 io_u->acct_ddir = -1;
1616 td->cur_depth++;
1617 assert(!(td->flags & TD_F_CHILD));
1618 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1619 io_u->ipo = NULL;
1620 } else if (td_async_processing(td)) {
1621 /*
1622 * We ran out, wait for async verify threads to finish and
1623 * return one
1624 */
1625 assert(!(td->flags & TD_F_CHILD));
1626 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1627 goto again;
1628 }
1629
1630 td_io_u_unlock(td);
1631 return io_u;
1632}
1633
1634static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1635{
1636 if (!(td->flags & TD_F_TRIM_BACKLOG))
1637 return false;
1638 if (!td->trim_entries)
1639 return false;
1640
1641 if (td->trim_batch) {
1642 td->trim_batch--;
1643 if (get_next_trim(td, io_u))
1644 return true;
1645 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1646 td->last_ddir != DDIR_READ) {
1647 td->trim_batch = td->o.trim_batch;
1648 if (!td->trim_batch)
1649 td->trim_batch = td->o.trim_backlog;
1650 if (get_next_trim(td, io_u))
1651 return true;
1652 }
1653
1654 return false;
1655}
1656
1657static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1658{
1659 if (!(td->flags & TD_F_VER_BACKLOG))
1660 return false;
1661
1662 if (td->io_hist_len) {
1663 int get_verify = 0;
1664
1665 if (td->verify_batch)
1666 get_verify = 1;
1667 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1668 td->last_ddir != DDIR_READ) {
1669 td->verify_batch = td->o.verify_batch;
1670 if (!td->verify_batch)
1671 td->verify_batch = td->o.verify_backlog;
1672 get_verify = 1;
1673 }
1674
1675 if (get_verify && !get_next_verify(td, io_u)) {
1676 td->verify_batch--;
1677 return true;
1678 }
1679 }
1680
1681 return false;
1682}
1683
1684/*
1685 * Fill offset and start time into the buffer content, to prevent too
1686 * easy compressible data for simple de-dupe attempts. Do this for every
1687 * 512b block in the range, since that should be the smallest block size
1688 * we can expect from a device.
1689 */
1690static void small_content_scramble(struct io_u *io_u)
1691{
1692 unsigned int i, nr_blocks = io_u->buflen >> 9;
1693 unsigned int offset;
1694 uint64_t boffset, *iptr;
1695 char *p;
1696
1697 if (!nr_blocks)
1698 return;
1699
1700 p = io_u->xfer_buf;
1701 boffset = io_u->offset;
1702
1703 if (io_u->buf_filled_len)
1704 io_u->buf_filled_len = 0;
1705
1706 /*
1707 * Generate random index between 0..7. We do chunks of 512b, if
1708 * we assume a cacheline is 64 bytes, then we have 8 of those.
1709 * Scramble content within the blocks in the same cacheline to
1710 * speed things up.
1711 */
1712 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1713
1714 for (i = 0; i < nr_blocks; i++) {
1715 /*
1716 * Fill offset into start of cacheline, time into end
1717 * of cacheline
1718 */
1719 iptr = (void *) p + (offset << 6);
1720 *iptr = boffset;
1721
1722 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1723 iptr[0] = io_u->start_time.tv_sec;
1724 iptr[1] = io_u->start_time.tv_nsec;
1725
1726 p += 512;
1727 boffset += 512;
1728 }
1729}
1730
1731/*
1732 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1733 * etc. The returned io_u is fully ready to be prepped and submitted.
1734 */
1735struct io_u *get_io_u(struct thread_data *td)
1736{
1737 struct fio_file *f;
1738 struct io_u *io_u;
1739 int do_scramble = 0;
1740 long ret = 0;
1741
1742 io_u = __get_io_u(td);
1743 if (!io_u) {
1744 dprint(FD_IO, "__get_io_u failed\n");
1745 return NULL;
1746 }
1747
1748 if (check_get_verify(td, io_u))
1749 goto out;
1750 if (check_get_trim(td, io_u))
1751 goto out;
1752
1753 /*
1754 * from a requeue, io_u already setup
1755 */
1756 if (io_u->file)
1757 goto out;
1758
1759 /*
1760 * If using an iolog, grab next piece if any available.
1761 */
1762 if (td->flags & TD_F_READ_IOLOG) {
1763 if (read_iolog_get(td, io_u))
1764 goto err_put;
1765 } else if (set_io_u_file(td, io_u)) {
1766 ret = -EBUSY;
1767 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1768 goto err_put;
1769 }
1770
1771 f = io_u->file;
1772 if (!f) {
1773 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1774 goto err_put;
1775 }
1776
1777 assert(fio_file_open(f));
1778
1779 if (ddir_rw(io_u->ddir)) {
1780 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1781 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1782 goto err_put;
1783 }
1784
1785 f->last_start[io_u->ddir] = io_u->offset;
1786 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1787
1788 if (io_u->ddir == DDIR_WRITE) {
1789 if (td->flags & TD_F_REFILL_BUFFERS) {
1790 io_u_fill_buffer(td, io_u,
1791 td->o.min_bs[DDIR_WRITE],
1792 io_u->buflen);
1793 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1794 !(td->flags & TD_F_COMPRESS))
1795 do_scramble = 1;
1796 if (td->flags & TD_F_VER_NONE) {
1797 populate_verify_io_u(td, io_u);
1798 do_scramble = 0;
1799 }
1800 } else if (io_u->ddir == DDIR_READ) {
1801 /*
1802 * Reset the buf_filled parameters so next time if the
1803 * buffer is used for writes it is refilled.
1804 */
1805 io_u->buf_filled_len = 0;
1806 }
1807 }
1808
1809 /*
1810 * Set io data pointers.
1811 */
1812 io_u->xfer_buf = io_u->buf;
1813 io_u->xfer_buflen = io_u->buflen;
1814
1815out:
1816 assert(io_u->file);
1817 if (!td_io_prep(td, io_u)) {
1818 if (!td->o.disable_lat)
1819 fio_gettime(&io_u->start_time, NULL);
1820
1821 if (do_scramble)
1822 small_content_scramble(io_u);
1823
1824 return io_u;
1825 }
1826err_put:
1827 dprint(FD_IO, "get_io_u failed\n");
1828 put_io_u(td, io_u);
1829 return ERR_PTR(ret);
1830}
1831
1832static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1833{
1834 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1835
1836 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1837 return;
1838
1839 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1840 io_u->file ? " on file " : "",
1841 io_u->file ? io_u->file->file_name : "",
1842 strerror(io_u->error),
1843 io_ddir_name(io_u->ddir),
1844 io_u->offset, io_u->xfer_buflen);
1845
1846 if (td->io_ops->errdetails) {
1847 char *err = td->io_ops->errdetails(io_u);
1848
1849 log_err("fio: %s\n", err);
1850 free(err);
1851 }
1852
1853 if (!td->error)
1854 td_verror(td, io_u->error, "io_u error");
1855}
1856
1857void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1858{
1859 __io_u_log_error(td, io_u);
1860 if (td->parent)
1861 __io_u_log_error(td->parent, io_u);
1862}
1863
1864static inline bool gtod_reduce(struct thread_data *td)
1865{
1866 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1867 || td->o.gtod_reduce;
1868}
1869
1870static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1871 struct io_completion_data *icd,
1872 const enum fio_ddir idx, unsigned int bytes)
1873{
1874 const int no_reduce = !gtod_reduce(td);
1875 unsigned long long llnsec = 0;
1876
1877 if (td->parent)
1878 td = td->parent;
1879
1880 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1881 return;
1882
1883 if (no_reduce)
1884 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1885
1886 if (!td->o.disable_lat) {
1887 unsigned long long tnsec;
1888
1889 tnsec = ntime_since(&io_u->start_time, &icd->time);
1890 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
1891
1892 if (td->flags & TD_F_PROFILE_OPS) {
1893 struct prof_io_ops *ops = &td->prof_io_ops;
1894
1895 if (ops->io_u_lat)
1896 icd->error = ops->io_u_lat(td, tnsec);
1897 }
1898
1899 if (td->o.max_latency && tnsec > td->o.max_latency)
1900 lat_fatal(td, icd, tnsec, td->o.max_latency);
1901 if (td->o.latency_target && tnsec > td->o.latency_target) {
1902 if (lat_target_failed(td))
1903 lat_fatal(td, icd, tnsec, td->o.latency_target);
1904 }
1905 }
1906
1907 if (ddir_rw(idx)) {
1908 if (!td->o.disable_clat) {
1909 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1910 io_u_mark_latency(td, llnsec);
1911 }
1912
1913 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1914 add_bw_sample(td, io_u, bytes, llnsec);
1915
1916 if (no_reduce && per_unit_log(td->iops_log))
1917 add_iops_sample(td, io_u, bytes);
1918 } else if (ddir_sync(idx) && !td->o.disable_clat)
1919 add_sync_clat_sample(&td->ts, llnsec);
1920
1921 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1922 uint32_t *info = io_u_block_info(td, io_u);
1923 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1924 if (io_u->ddir == DDIR_TRIM) {
1925 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1926 BLOCK_INFO_TRIMS(*info) + 1);
1927 } else if (io_u->ddir == DDIR_WRITE) {
1928 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1929 *info);
1930 }
1931 }
1932 }
1933}
1934
1935static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1936 uint64_t offset, unsigned int bytes)
1937{
1938 int idx;
1939
1940 if (!f)
1941 return;
1942
1943 if (f->first_write == -1ULL || offset < f->first_write)
1944 f->first_write = offset;
1945 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1946 f->last_write = offset + bytes;
1947
1948 if (!f->last_write_comp)
1949 return;
1950
1951 idx = f->last_write_idx++;
1952 f->last_write_comp[idx] = offset;
1953 if (f->last_write_idx == td->o.iodepth)
1954 f->last_write_idx = 0;
1955}
1956
1957static bool should_account(struct thread_data *td)
1958{
1959 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1960 td->runstate == TD_VERIFYING);
1961}
1962
1963static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1964 struct io_completion_data *icd)
1965{
1966 struct io_u *io_u = *io_u_ptr;
1967 enum fio_ddir ddir = io_u->ddir;
1968 struct fio_file *f = io_u->file;
1969
1970 dprint_io_u(io_u, "complete");
1971
1972 assert(io_u->flags & IO_U_F_FLIGHT);
1973 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1974
1975 /*
1976 * Mark IO ok to verify
1977 */
1978 if (io_u->ipo) {
1979 /*
1980 * Remove errored entry from the verification list
1981 */
1982 if (io_u->error)
1983 unlog_io_piece(td, io_u);
1984 else {
1985 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1986 write_barrier();
1987 }
1988 }
1989
1990 if (ddir_sync(ddir)) {
1991 td->last_was_sync = true;
1992 if (f) {
1993 f->first_write = -1ULL;
1994 f->last_write = -1ULL;
1995 }
1996 if (should_account(td))
1997 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
1998 return;
1999 }
2000
2001 td->last_was_sync = false;
2002 td->last_ddir = ddir;
2003
2004 if (!io_u->error && ddir_rw(ddir)) {
2005 unsigned int bytes = io_u->buflen - io_u->resid;
2006 int ret;
2007
2008 td->io_blocks[ddir]++;
2009 td->io_bytes[ddir] += bytes;
2010
2011 if (!(io_u->flags & IO_U_F_VER_LIST)) {
2012 td->this_io_blocks[ddir]++;
2013 td->this_io_bytes[ddir] += bytes;
2014 }
2015
2016 if (ddir == DDIR_WRITE)
2017 file_log_write_comp(td, f, io_u->offset, bytes);
2018
2019 if (should_account(td))
2020 account_io_completion(td, io_u, icd, ddir, bytes);
2021
2022 icd->bytes_done[ddir] += bytes;
2023
2024 if (io_u->end_io) {
2025 ret = io_u->end_io(td, io_u_ptr);
2026 io_u = *io_u_ptr;
2027 if (ret && !icd->error)
2028 icd->error = ret;
2029 }
2030 } else if (io_u->error) {
2031 icd->error = io_u->error;
2032 io_u_log_error(td, io_u);
2033 }
2034 if (icd->error) {
2035 enum error_type_bit eb = td_error_type(ddir, icd->error);
2036
2037 if (!td_non_fatal_error(td, eb, icd->error))
2038 return;
2039
2040 /*
2041 * If there is a non_fatal error, then add to the error count
2042 * and clear all the errors.
2043 */
2044 update_error_count(td, icd->error);
2045 td_clear_error(td);
2046 icd->error = 0;
2047 if (io_u)
2048 io_u->error = 0;
2049 }
2050}
2051
2052static void init_icd(struct thread_data *td, struct io_completion_data *icd,
2053 int nr)
2054{
2055 int ddir;
2056
2057 if (!gtod_reduce(td))
2058 fio_gettime(&icd->time, NULL);
2059
2060 icd->nr = nr;
2061
2062 icd->error = 0;
2063 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2064 icd->bytes_done[ddir] = 0;
2065}
2066
2067static void ios_completed(struct thread_data *td,
2068 struct io_completion_data *icd)
2069{
2070 struct io_u *io_u;
2071 int i;
2072
2073 for (i = 0; i < icd->nr; i++) {
2074 io_u = td->io_ops->event(td, i);
2075
2076 io_completed(td, &io_u, icd);
2077
2078 if (io_u)
2079 put_io_u(td, io_u);
2080 }
2081}
2082
2083/*
2084 * Complete a single io_u for the sync engines.
2085 */
2086int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
2087{
2088 struct io_completion_data icd;
2089 int ddir;
2090
2091 init_icd(td, &icd, 1);
2092 io_completed(td, &io_u, &icd);
2093
2094 if (io_u)
2095 put_io_u(td, io_u);
2096
2097 if (icd.error) {
2098 td_verror(td, icd.error, "io_u_sync_complete");
2099 return -1;
2100 }
2101
2102 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2103 td->bytes_done[ddir] += icd.bytes_done[ddir];
2104
2105 return 0;
2106}
2107
2108/*
2109 * Called to complete min_events number of io for the async engines.
2110 */
2111int io_u_queued_complete(struct thread_data *td, int min_evts)
2112{
2113 struct io_completion_data icd;
2114 struct timespec *tvp = NULL;
2115 int ret, ddir;
2116 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2117
2118 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2119
2120 if (!min_evts)
2121 tvp = &ts;
2122 else if (min_evts > td->cur_depth)
2123 min_evts = td->cur_depth;
2124
2125 /* No worries, td_io_getevents fixes min and max if they are
2126 * set incorrectly */
2127 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2128 if (ret < 0) {
2129 td_verror(td, -ret, "td_io_getevents");
2130 return ret;
2131 } else if (!ret)
2132 return ret;
2133
2134 init_icd(td, &icd, ret);
2135 ios_completed(td, &icd);
2136 if (icd.error) {
2137 td_verror(td, icd.error, "io_u_queued_complete");
2138 return -1;
2139 }
2140
2141 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2142 td->bytes_done[ddir] += icd.bytes_done[ddir];
2143
2144 return ret;
2145}
2146
2147/*
2148 * Call when io_u is really queued, to update the submission latency.
2149 */
2150void io_u_queued(struct thread_data *td, struct io_u *io_u)
2151{
2152 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2153 unsigned long slat_time;
2154
2155 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2156
2157 if (td->parent)
2158 td = td->parent;
2159
2160 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2161 io_u->offset);
2162 }
2163}
2164
2165/*
2166 * See if we should reuse the last seed, if dedupe is enabled
2167 */
2168static struct frand_state *get_buf_state(struct thread_data *td)
2169{
2170 unsigned int v;
2171
2172 if (!td->o.dedupe_percentage)
2173 return &td->buf_state;
2174 else if (td->o.dedupe_percentage == 100) {
2175 frand_copy(&td->buf_state_prev, &td->buf_state);
2176 return &td->buf_state;
2177 }
2178
2179 v = rand32_between(&td->dedupe_state, 1, 100);
2180
2181 if (v <= td->o.dedupe_percentage)
2182 return &td->buf_state_prev;
2183
2184 return &td->buf_state;
2185}
2186
2187static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2188{
2189 if (td->o.dedupe_percentage == 100)
2190 frand_copy(rs, &td->buf_state_prev);
2191 else if (rs == &td->buf_state)
2192 frand_copy(&td->buf_state_prev, rs);
2193}
2194
2195void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2196 unsigned int max_bs)
2197{
2198 struct thread_options *o = &td->o;
2199
2200 if (o->mem_type == MEM_CUDA_MALLOC)
2201 return;
2202
2203 if (o->compress_percentage || o->dedupe_percentage) {
2204 unsigned int perc = td->o.compress_percentage;
2205 struct frand_state *rs;
2206 unsigned int left = max_bs;
2207 unsigned int this_write;
2208
2209 do {
2210 rs = get_buf_state(td);
2211
2212 min_write = min(min_write, left);
2213
2214 if (perc) {
2215 this_write = min_not_zero(min_write,
2216 td->o.compress_chunk);
2217
2218 fill_random_buf_percentage(rs, buf, perc,
2219 this_write, this_write,
2220 o->buffer_pattern,
2221 o->buffer_pattern_bytes);
2222 } else {
2223 fill_random_buf(rs, buf, min_write);
2224 this_write = min_write;
2225 }
2226
2227 buf += this_write;
2228 left -= this_write;
2229 save_buf_state(td, rs);
2230 } while (left);
2231 } else if (o->buffer_pattern_bytes)
2232 fill_buffer_pattern(td, buf, max_bs);
2233 else if (o->zero_buffers)
2234 memset(buf, 0, max_bs);
2235 else
2236 fill_random_buf(get_buf_state(td), buf, max_bs);
2237}
2238
2239/*
2240 * "randomly" fill the buffer contents
2241 */
2242void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2243 unsigned int min_write, unsigned int max_bs)
2244{
2245 io_u->buf_filled_len = 0;
2246 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2247}
2248
2249static int do_sync_file_range(const struct thread_data *td,
2250 struct fio_file *f)
2251{
2252 off64_t offset, nbytes;
2253
2254 offset = f->first_write;
2255 nbytes = f->last_write - f->first_write;
2256
2257 if (!nbytes)
2258 return 0;
2259
2260 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2261}
2262
2263int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2264{
2265 int ret;
2266
2267 if (io_u->ddir == DDIR_SYNC) {
2268 ret = fsync(io_u->file->fd);
2269 } else if (io_u->ddir == DDIR_DATASYNC) {
2270#ifdef CONFIG_FDATASYNC
2271 ret = fdatasync(io_u->file->fd);
2272#else
2273 ret = io_u->xfer_buflen;
2274 io_u->error = EINVAL;
2275#endif
2276 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2277 ret = do_sync_file_range(td, io_u->file);
2278 else {
2279 ret = io_u->xfer_buflen;
2280 io_u->error = EINVAL;
2281 }
2282
2283 if (ret < 0)
2284 io_u->error = errno;
2285
2286 return ret;
2287}
2288
2289int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2290{
2291#ifndef FIO_HAVE_TRIM
2292 io_u->error = EINVAL;
2293 return 0;
2294#else
2295 struct fio_file *f = io_u->file;
2296 int ret;
2297
2298 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2299 if (!ret)
2300 return io_u->xfer_buflen;
2301
2302 io_u->error = ret;
2303 return 0;
2304#endif
2305}
fio - Flexible IO Tester