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