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