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