projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
sg: don't return success if non bdev/char is used
[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
16struct io_completion_data {
17 int nr; /* input */
18
19 int error; /* output */
20 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
21 struct timeval time; /* output */
22};
23
24/*
25 * The ->io_axmap contains a map of blocks we have or have not done io
26 * to yet. Used to make sure we cover the entire range in a fair fashion.
27 */
28static int random_map_free(struct fio_file *f, const uint64_t block)
29{
30 return !axmap_isset(f->io_axmap, block);
31}
32
33/*
34 * Mark a given offset as used in the map.
35 */
36static void mark_random_map(struct thread_data *td, struct io_u *io_u)
37{
38 unsigned int min_bs = td->o.rw_min_bs;
39 struct fio_file *f = io_u->file;
40 unsigned int nr_blocks;
41 uint64_t block;
42
43 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
44 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
45
46 if (!(io_u->flags & IO_U_F_BUSY_OK))
47 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
48
49 if ((nr_blocks * min_bs) < io_u->buflen)
50 io_u->buflen = nr_blocks * min_bs;
51}
52
53static uint64_t last_block(struct thread_data *td, struct fio_file *f,
54 enum fio_ddir ddir)
55{
56 uint64_t max_blocks;
57 uint64_t max_size;
58
59 assert(ddir_rw(ddir));
60
61 /*
62 * Hmm, should we make sure that ->io_size <= ->real_file_size?
63 */
64 max_size = f->io_size;
65 if (max_size > f->real_file_size)
66 max_size = f->real_file_size;
67
68 if (td->o.zone_range)
69 max_size = td->o.zone_range;
70
71 if (td->o.min_bs[ddir] > td->o.ba[ddir])
72 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
73
74 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
75 if (!max_blocks)
76 return 0;
77
78 return max_blocks;
79}
80
81struct rand_off {
82 struct flist_head list;
83 uint64_t off;
84};
85
86static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
87 enum fio_ddir ddir, uint64_t *b)
88{
89 uint64_t r;
90
91 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
92 uint64_t lastb;
93
94 lastb = last_block(td, f, ddir);
95 if (!lastb)
96 return 1;
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 / ((uint64_t) FRAND_MAX + 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 flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
153{
154 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
155 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
156
157 return r1->off - r2->off;
158}
159
160static int get_off_from_method(struct thread_data *td, struct fio_file *f,
161 enum fio_ddir ddir, uint64_t *b)
162{
163 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
164 return __get_next_rand_offset(td, f, ddir, b);
165 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
166 return __get_next_rand_offset_zipf(td, f, ddir, b);
167 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
168 return __get_next_rand_offset_pareto(td, f, ddir, b);
169
170 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
171 return 1;
172}
173
174/*
175 * Sort the reads for a verify phase in batches of verifysort_nr, if
176 * specified.
177 */
178static inline int should_sort_io(struct thread_data *td)
179{
180 if (!td->o.verifysort_nr || !td->o.do_verify)
181 return 0;
182 if (!td_random(td))
183 return 0;
184 if (td->runstate != TD_VERIFYING)
185 return 0;
186 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
187 return 0;
188
189 return 1;
190}
191
192static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
193{
194 unsigned int v;
195 unsigned long r;
196
197 if (td->o.perc_rand[ddir] == 100)
198 return 1;
199
200 r = __rand(&td->seq_rand_state[ddir]);
201 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
202
203 return v <= td->o.perc_rand[ddir];
204}
205
206static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
207 enum fio_ddir ddir, uint64_t *b)
208{
209 struct rand_off *r;
210 int i, ret = 1;
211
212 if (!should_sort_io(td))
213 return get_off_from_method(td, f, ddir, b);
214
215 if (!flist_empty(&td->next_rand_list)) {
216fetch:
217 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
218 flist_del(&r->list);
219 *b = r->off;
220 free(r);
221 return 0;
222 }
223
224 for (i = 0; i < td->o.verifysort_nr; i++) {
225 r = malloc(sizeof(*r));
226
227 ret = get_off_from_method(td, f, ddir, &r->off);
228 if (ret) {
229 free(r);
230 break;
231 }
232
233 flist_add(&r->list, &td->next_rand_list);
234 }
235
236 if (ret && !i)
237 return ret;
238
239 assert(!flist_empty(&td->next_rand_list));
240 flist_sort(NULL, &td->next_rand_list, flist_cmp);
241 goto fetch;
242}
243
244static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
245 enum fio_ddir ddir, uint64_t *b)
246{
247 if (!get_next_rand_offset(td, f, ddir, b))
248 return 0;
249
250 if (td->o.time_based) {
251 fio_file_reset(td, f);
252 if (!get_next_rand_offset(td, f, ddir, b))
253 return 0;
254 }
255
256 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
257 f->file_name, (unsigned long long) f->last_pos[ddir],
258 (unsigned long long) f->real_file_size);
259 return 1;
260}
261
262static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
263 enum fio_ddir ddir, uint64_t *offset)
264{
265 struct thread_options *o = &td->o;
266
267 assert(ddir_rw(ddir));
268
269 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
270 o->time_based)
271 f->last_pos[ddir] = f->last_pos[ddir] - f->io_size;
272
273 if (f->last_pos[ddir] < f->real_file_size) {
274 uint64_t pos;
275
276 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0)
277 f->last_pos[ddir] = f->real_file_size;
278
279 pos = f->last_pos[ddir] - f->file_offset;
280 if (pos && o->ddir_seq_add) {
281 pos += o->ddir_seq_add;
282
283 /*
284 * If we reach beyond the end of the file
285 * with holed IO, wrap around to the
286 * beginning again.
287 */
288 if (pos >= f->real_file_size)
289 pos = f->file_offset;
290 }
291
292 *offset = pos;
293 return 0;
294 }
295
296 return 1;
297}
298
299static int get_next_block(struct thread_data *td, struct io_u *io_u,
300 enum fio_ddir ddir, int rw_seq,
301 unsigned int *is_random)
302{
303 struct fio_file *f = io_u->file;
304 uint64_t b, offset;
305 int ret;
306
307 assert(ddir_rw(ddir));
308
309 b = offset = -1ULL;
310
311 if (rw_seq) {
312 if (td_random(td)) {
313 if (should_do_random(td, ddir)) {
314 ret = get_next_rand_block(td, f, ddir, &b);
315 *is_random = 1;
316 } else {
317 *is_random = 0;
318 io_u->flags |= IO_U_F_BUSY_OK;
319 ret = get_next_seq_offset(td, f, ddir, &offset);
320 if (ret)
321 ret = get_next_rand_block(td, f, ddir, &b);
322 }
323 } else {
324 *is_random = 0;
325 ret = get_next_seq_offset(td, f, ddir, &offset);
326 }
327 } else {
328 io_u->flags |= IO_U_F_BUSY_OK;
329 *is_random = 0;
330
331 if (td->o.rw_seq == RW_SEQ_SEQ) {
332 ret = get_next_seq_offset(td, f, ddir, &offset);
333 if (ret) {
334 ret = get_next_rand_block(td, f, ddir, &b);
335 *is_random = 0;
336 }
337 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
338 if (f->last_start[ddir] != -1ULL)
339 offset = f->last_start[ddir] - f->file_offset;
340 else
341 offset = 0;
342 ret = 0;
343 } else {
344 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
345 ret = 1;
346 }
347 }
348
349 if (!ret) {
350 if (offset != -1ULL)
351 io_u->offset = offset;
352 else if (b != -1ULL)
353 io_u->offset = b * td->o.ba[ddir];
354 else {
355 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
356 ret = 1;
357 }
358 }
359
360 return ret;
361}
362
363/*
364 * For random io, generate a random new block and see if it's used. Repeat
365 * until we find a free one. For sequential io, just return the end of
366 * the last io issued.
367 */
368static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
369 unsigned int *is_random)
370{
371 struct fio_file *f = io_u->file;
372 enum fio_ddir ddir = io_u->ddir;
373 int rw_seq_hit = 0;
374
375 assert(ddir_rw(ddir));
376
377 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
378 rw_seq_hit = 1;
379 td->ddir_seq_nr = td->o.ddir_seq_nr;
380 }
381
382 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
383 return 1;
384
385 if (io_u->offset >= f->io_size) {
386 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
387 (unsigned long long) io_u->offset,
388 (unsigned long long) f->io_size);
389 return 1;
390 }
391
392 io_u->offset += f->file_offset;
393 if (io_u->offset >= f->real_file_size) {
394 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
395 (unsigned long long) io_u->offset,
396 (unsigned long long) f->real_file_size);
397 return 1;
398 }
399
400 return 0;
401}
402
403static int get_next_offset(struct thread_data *td, struct io_u *io_u,
404 unsigned int *is_random)
405{
406 if (td->flags & TD_F_PROFILE_OPS) {
407 struct prof_io_ops *ops = &td->prof_io_ops;
408
409 if (ops->fill_io_u_off)
410 return ops->fill_io_u_off(td, io_u, is_random);
411 }
412
413 return __get_next_offset(td, io_u, is_random);
414}
415
416static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
417 unsigned int buflen)
418{
419 struct fio_file *f = io_u->file;
420
421 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
422}
423
424static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
425 unsigned int is_random)
426{
427 int ddir = io_u->ddir;
428 unsigned int buflen = 0;
429 unsigned int minbs, maxbs;
430 unsigned long r;
431
432 assert(ddir_rw(ddir));
433
434 if (td->o.bs_is_seq_rand)
435 ddir = is_random ? DDIR_WRITE: DDIR_READ;
436
437 minbs = td->o.min_bs[ddir];
438 maxbs = td->o.max_bs[ddir];
439
440 if (minbs == maxbs)
441 return minbs;
442
443 /*
444 * If we can't satisfy the min block size from here, then fail
445 */
446 if (!io_u_fits(td, io_u, minbs))
447 return 0;
448
449 do {
450 r = __rand(&td->bsrange_state);
451
452 if (!td->o.bssplit_nr[ddir]) {
453 buflen = 1 + (unsigned int) ((double) maxbs *
454 (r / (FRAND_MAX + 1.0)));
455 if (buflen < minbs)
456 buflen = minbs;
457 } else {
458 long perc = 0;
459 unsigned int i;
460
461 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
462 struct bssplit *bsp = &td->o.bssplit[ddir][i];
463
464 buflen = bsp->bs;
465 perc += bsp->perc;
466 if ((r <= ((FRAND_MAX / 100L) * perc)) &&
467 io_u_fits(td, io_u, buflen))
468 break;
469 }
470 }
471
472 if (td->o.do_verify && td->o.verify != VERIFY_NONE)
473 buflen = (buflen + td->o.verify_interval - 1) &
474 ~(td->o.verify_interval - 1);
475
476 if (!td->o.bs_unaligned && is_power_of_2(minbs))
477 buflen = (buflen + minbs - 1) & ~(minbs - 1);
478
479 } while (!io_u_fits(td, io_u, buflen));
480
481 return buflen;
482}
483
484static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
485 unsigned int is_random)
486{
487 if (td->flags & TD_F_PROFILE_OPS) {
488 struct prof_io_ops *ops = &td->prof_io_ops;
489
490 if (ops->fill_io_u_size)
491 return ops->fill_io_u_size(td, io_u, is_random);
492 }
493
494 return __get_next_buflen(td, io_u, is_random);
495}
496
497static void set_rwmix_bytes(struct thread_data *td)
498{
499 unsigned int diff;
500
501 /*
502 * we do time or byte based switch. this is needed because
503 * buffered writes may issue a lot quicker than they complete,
504 * whereas reads do not.
505 */
506 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
507 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
508}
509
510static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
511{
512 unsigned int v;
513 unsigned long r;
514
515 r = __rand(&td->rwmix_state);
516 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
517
518 if (v <= td->o.rwmix[DDIR_READ])
519 return DDIR_READ;
520
521 return DDIR_WRITE;
522}
523
524void io_u_quiesce(struct thread_data *td)
525{
526 /*
527 * We are going to sleep, ensure that we flush anything pending as
528 * not to skew our latency numbers.
529 *
530 * Changed to only monitor 'in flight' requests here instead of the
531 * td->cur_depth, b/c td->cur_depth does not accurately represent
532 * io's that have been actually submitted to an async engine,
533 * and cur_depth is meaningless for sync engines.
534 */
535 if (td->io_u_queued || td->cur_depth) {
536 int fio_unused ret;
537
538 ret = td_io_commit(td);
539 }
540
541 while (td->io_u_in_flight) {
542 int fio_unused ret;
543
544 ret = io_u_queued_complete(td, 1, NULL);
545 }
546}
547
548static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
549{
550 enum fio_ddir odir = ddir ^ 1;
551 long usec;
552
553 assert(ddir_rw(ddir));
554
555 if (td->rate_pending_usleep[ddir] <= 0)
556 return ddir;
557
558 /*
559 * We have too much pending sleep in this direction. See if we
560 * should switch.
561 */
562 if (td_rw(td) && td->o.rwmix[odir]) {
563 /*
564 * Other direction does not have too much pending, switch
565 */
566 if (td->rate_pending_usleep[odir] < 100000)
567 return odir;
568
569 /*
570 * Both directions have pending sleep. Sleep the minimum time
571 * and deduct from both.
572 */
573 if (td->rate_pending_usleep[ddir] <=
574 td->rate_pending_usleep[odir]) {
575 usec = td->rate_pending_usleep[ddir];
576 } else {
577 usec = td->rate_pending_usleep[odir];
578 ddir = odir;
579 }
580 } else
581 usec = td->rate_pending_usleep[ddir];
582
583 io_u_quiesce(td);
584
585 usec = usec_sleep(td, usec);
586
587 td->rate_pending_usleep[ddir] -= usec;
588
589 odir = ddir ^ 1;
590 if (td_rw(td) && __should_check_rate(td, odir))
591 td->rate_pending_usleep[odir] -= usec;
592
593 return ddir;
594}
595
596/*
597 * Return the data direction for the next io_u. If the job is a
598 * mixed read/write workload, check the rwmix cycle and switch if
599 * necessary.
600 */
601static enum fio_ddir get_rw_ddir(struct thread_data *td)
602{
603 enum fio_ddir ddir;
604
605 /*
606 * see if it's time to fsync
607 */
608 if (td->o.fsync_blocks &&
609 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
610 td->io_issues[DDIR_WRITE] && should_fsync(td))
611 return DDIR_SYNC;
612
613 /*
614 * see if it's time to fdatasync
615 */
616 if (td->o.fdatasync_blocks &&
617 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
618 td->io_issues[DDIR_WRITE] && should_fsync(td))
619 return DDIR_DATASYNC;
620
621 /*
622 * see if it's time to sync_file_range
623 */
624 if (td->sync_file_range_nr &&
625 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
626 td->io_issues[DDIR_WRITE] && should_fsync(td))
627 return DDIR_SYNC_FILE_RANGE;
628
629 if (td_rw(td)) {
630 /*
631 * Check if it's time to seed a new data direction.
632 */
633 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
634 /*
635 * Put a top limit on how many bytes we do for
636 * one data direction, to avoid overflowing the
637 * ranges too much
638 */
639 ddir = get_rand_ddir(td);
640
641 if (ddir != td->rwmix_ddir)
642 set_rwmix_bytes(td);
643
644 td->rwmix_ddir = ddir;
645 }
646 ddir = td->rwmix_ddir;
647 } else if (td_read(td))
648 ddir = DDIR_READ;
649 else if (td_write(td))
650 ddir = DDIR_WRITE;
651 else
652 ddir = DDIR_TRIM;
653
654 td->rwmix_ddir = rate_ddir(td, ddir);
655 return td->rwmix_ddir;
656}
657
658static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
659{
660 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
661
662 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
663 td->o.barrier_blocks &&
664 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
665 td->io_issues[DDIR_WRITE])
666 io_u->flags |= IO_U_F_BARRIER;
667}
668
669void put_file_log(struct thread_data *td, struct fio_file *f)
670{
671 unsigned int ret = put_file(td, f);
672
673 if (ret)
674 td_verror(td, ret, "file close");
675}
676
677void put_io_u(struct thread_data *td, struct io_u *io_u)
678{
679 td_io_u_lock(td);
680
681 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
682 put_file_log(td, io_u->file);
683
684 io_u->file = NULL;
685 io_u->flags |= IO_U_F_FREE;
686
687 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
688 td->cur_depth--;
689 io_u_qpush(&td->io_u_freelist, io_u);
690 td_io_u_unlock(td);
691 td_io_u_free_notify(td);
692}
693
694void clear_io_u(struct thread_data *td, struct io_u *io_u)
695{
696 io_u->flags &= ~IO_U_F_FLIGHT;
697 put_io_u(td, io_u);
698}
699
700void requeue_io_u(struct thread_data *td, struct io_u **io_u)
701{
702 struct io_u *__io_u = *io_u;
703 enum fio_ddir ddir = acct_ddir(__io_u);
704
705 dprint(FD_IO, "requeue %p\n", __io_u);
706
707 td_io_u_lock(td);
708
709 __io_u->flags |= IO_U_F_FREE;
710 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
711 td->io_issues[ddir]--;
712
713 __io_u->flags &= ~IO_U_F_FLIGHT;
714 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
715 td->cur_depth--;
716
717 io_u_rpush(&td->io_u_requeues, __io_u);
718 td_io_u_unlock(td);
719 *io_u = NULL;
720}
721
722static int fill_io_u(struct thread_data *td, struct io_u *io_u)
723{
724 unsigned int is_random;
725
726 if (td->io_ops->flags & FIO_NOIO)
727 goto out;
728
729 set_rw_ddir(td, io_u);
730
731 /*
732 * fsync() or fdatasync() or trim etc, we are done
733 */
734 if (!ddir_rw(io_u->ddir))
735 goto out;
736
737 /*
738 * See if it's time to switch to a new zone
739 */
740 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
741 struct fio_file *f = io_u->file;
742
743 td->zone_bytes = 0;
744 f->file_offset += td->o.zone_range + td->o.zone_skip;
745
746 /*
747 * Wrap from the beginning, if we exceed the file size
748 */
749 if (f->file_offset >= f->real_file_size)
750 f->file_offset = f->real_file_size - f->file_offset;
751 f->last_pos[io_u->ddir] = f->file_offset;
752 td->io_skip_bytes += td->o.zone_skip;
753 }
754
755 /*
756 * No log, let the seq/rand engine retrieve the next buflen and
757 * position.
758 */
759 if (get_next_offset(td, io_u, &is_random)) {
760 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
761 return 1;
762 }
763
764 io_u->buflen = get_next_buflen(td, io_u, is_random);
765 if (!io_u->buflen) {
766 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
767 return 1;
768 }
769
770 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
771 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
772 dprint(FD_IO, " off=%llu/%lu > %llu\n",
773 (unsigned long long) io_u->offset, io_u->buflen,
774 (unsigned long long) io_u->file->real_file_size);
775 return 1;
776 }
777
778 /*
779 * mark entry before potentially trimming io_u
780 */
781 if (td_random(td) && file_randommap(td, io_u->file))
782 mark_random_map(td, io_u);
783
784out:
785 dprint_io_u(io_u, "fill_io_u");
786 td->zone_bytes += io_u->buflen;
787 return 0;
788}
789
790static void __io_u_mark_map(unsigned int *map, unsigned int nr)
791{
792 int idx = 0;
793
794 switch (nr) {
795 default:
796 idx = 6;
797 break;
798 case 33 ... 64:
799 idx = 5;
800 break;
801 case 17 ... 32:
802 idx = 4;
803 break;
804 case 9 ... 16:
805 idx = 3;
806 break;
807 case 5 ... 8:
808 idx = 2;
809 break;
810 case 1 ... 4:
811 idx = 1;
812 case 0:
813 break;
814 }
815
816 map[idx]++;
817}
818
819void io_u_mark_submit(struct thread_data *td, unsigned int nr)
820{
821 __io_u_mark_map(td->ts.io_u_submit, nr);
822 td->ts.total_submit++;
823}
824
825void io_u_mark_complete(struct thread_data *td, unsigned int nr)
826{
827 __io_u_mark_map(td->ts.io_u_complete, nr);
828 td->ts.total_complete++;
829}
830
831void io_u_mark_depth(struct thread_data *td, unsigned int nr)
832{
833 int idx = 0;
834
835 switch (td->cur_depth) {
836 default:
837 idx = 6;
838 break;
839 case 32 ... 63:
840 idx = 5;
841 break;
842 case 16 ... 31:
843 idx = 4;
844 break;
845 case 8 ... 15:
846 idx = 3;
847 break;
848 case 4 ... 7:
849 idx = 2;
850 break;
851 case 2 ... 3:
852 idx = 1;
853 case 1:
854 break;
855 }
856
857 td->ts.io_u_map[idx] += nr;
858}
859
860static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
861{
862 int idx = 0;
863
864 assert(usec < 1000);
865
866 switch (usec) {
867 case 750 ... 999:
868 idx = 9;
869 break;
870 case 500 ... 749:
871 idx = 8;
872 break;
873 case 250 ... 499:
874 idx = 7;
875 break;
876 case 100 ... 249:
877 idx = 6;
878 break;
879 case 50 ... 99:
880 idx = 5;
881 break;
882 case 20 ... 49:
883 idx = 4;
884 break;
885 case 10 ... 19:
886 idx = 3;
887 break;
888 case 4 ... 9:
889 idx = 2;
890 break;
891 case 2 ... 3:
892 idx = 1;
893 case 0 ... 1:
894 break;
895 }
896
897 assert(idx < FIO_IO_U_LAT_U_NR);
898 td->ts.io_u_lat_u[idx]++;
899}
900
901static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
902{
903 int idx = 0;
904
905 switch (msec) {
906 default:
907 idx = 11;
908 break;
909 case 1000 ... 1999:
910 idx = 10;
911 break;
912 case 750 ... 999:
913 idx = 9;
914 break;
915 case 500 ... 749:
916 idx = 8;
917 break;
918 case 250 ... 499:
919 idx = 7;
920 break;
921 case 100 ... 249:
922 idx = 6;
923 break;
924 case 50 ... 99:
925 idx = 5;
926 break;
927 case 20 ... 49:
928 idx = 4;
929 break;
930 case 10 ... 19:
931 idx = 3;
932 break;
933 case 4 ... 9:
934 idx = 2;
935 break;
936 case 2 ... 3:
937 idx = 1;
938 case 0 ... 1:
939 break;
940 }
941
942 assert(idx < FIO_IO_U_LAT_M_NR);
943 td->ts.io_u_lat_m[idx]++;
944}
945
946static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
947{
948 if (usec < 1000)
949 io_u_mark_lat_usec(td, usec);
950 else
951 io_u_mark_lat_msec(td, usec / 1000);
952}
953
954/*
955 * Get next file to service by choosing one at random
956 */
957static struct fio_file *get_next_file_rand(struct thread_data *td,
958 enum fio_file_flags goodf,
959 enum fio_file_flags badf)
960{
961 struct fio_file *f;
962 int fno;
963
964 do {
965 int opened = 0;
966 unsigned long r;
967
968 r = __rand(&td->next_file_state);
969 fno = (unsigned int) ((double) td->o.nr_files
970 * (r / (FRAND_MAX + 1.0)));
971
972 f = td->files[fno];
973 if (fio_file_done(f))
974 continue;
975
976 if (!fio_file_open(f)) {
977 int err;
978
979 if (td->nr_open_files >= td->o.open_files)
980 return ERR_PTR(-EBUSY);
981
982 err = td_io_open_file(td, f);
983 if (err)
984 continue;
985 opened = 1;
986 }
987
988 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
989 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
990 return f;
991 }
992 if (opened)
993 td_io_close_file(td, f);
994 } while (1);
995}
996
997/*
998 * Get next file to service by doing round robin between all available ones
999 */
1000static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1001 int badf)
1002{
1003 unsigned int old_next_file = td->next_file;
1004 struct fio_file *f;
1005
1006 do {
1007 int opened = 0;
1008
1009 f = td->files[td->next_file];
1010
1011 td->next_file++;
1012 if (td->next_file >= td->o.nr_files)
1013 td->next_file = 0;
1014
1015 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1016 if (fio_file_done(f)) {
1017 f = NULL;
1018 continue;
1019 }
1020
1021 if (!fio_file_open(f)) {
1022 int err;
1023
1024 if (td->nr_open_files >= td->o.open_files)
1025 return ERR_PTR(-EBUSY);
1026
1027 err = td_io_open_file(td, f);
1028 if (err) {
1029 dprint(FD_FILE, "error %d on open of %s\n",
1030 err, f->file_name);
1031 f = NULL;
1032 continue;
1033 }
1034 opened = 1;
1035 }
1036
1037 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1038 f->flags);
1039 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1040 break;
1041
1042 if (opened)
1043 td_io_close_file(td, f);
1044
1045 f = NULL;
1046 } while (td->next_file != old_next_file);
1047
1048 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1049 return f;
1050}
1051
1052static struct fio_file *__get_next_file(struct thread_data *td)
1053{
1054 struct fio_file *f;
1055
1056 assert(td->o.nr_files <= td->files_index);
1057
1058 if (td->nr_done_files >= td->o.nr_files) {
1059 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1060 " nr_files=%d\n", td->nr_open_files,
1061 td->nr_done_files,
1062 td->o.nr_files);
1063 return NULL;
1064 }
1065
1066 f = td->file_service_file;
1067 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1068 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1069 goto out;
1070 if (td->file_service_left--)
1071 goto out;
1072 }
1073
1074 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1075 td->o.file_service_type == FIO_FSERVICE_SEQ)
1076 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1077 else
1078 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1079
1080 if (IS_ERR(f))
1081 return f;
1082
1083 td->file_service_file = f;
1084 td->file_service_left = td->file_service_nr - 1;
1085out:
1086 if (f)
1087 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1088 else
1089 dprint(FD_FILE, "get_next_file: NULL\n");
1090 return f;
1091}
1092
1093static struct fio_file *get_next_file(struct thread_data *td)
1094{
1095 if (td->flags & TD_F_PROFILE_OPS) {
1096 struct prof_io_ops *ops = &td->prof_io_ops;
1097
1098 if (ops->get_next_file)
1099 return ops->get_next_file(td);
1100 }
1101
1102 return __get_next_file(td);
1103}
1104
1105static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1106{
1107 struct fio_file *f;
1108
1109 do {
1110 f = get_next_file(td);
1111 if (IS_ERR_OR_NULL(f))
1112 return PTR_ERR(f);
1113
1114 io_u->file = f;
1115 get_file(f);
1116
1117 if (!fill_io_u(td, io_u))
1118 break;
1119
1120 put_file_log(td, f);
1121 td_io_close_file(td, f);
1122 io_u->file = NULL;
1123 fio_file_set_done(f);
1124 td->nr_done_files++;
1125 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1126 td->nr_done_files, td->o.nr_files);
1127 } while (1);
1128
1129 return 0;
1130}
1131
1132static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1133 unsigned long tusec, unsigned long max_usec)
1134{
1135 if (!td->error)
1136 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1137 td_verror(td, ETIMEDOUT, "max latency exceeded");
1138 icd->error = ETIMEDOUT;
1139}
1140
1141static void lat_new_cycle(struct thread_data *td)
1142{
1143 fio_gettime(&td->latency_ts, NULL);
1144 td->latency_ios = ddir_rw_sum(td->io_blocks);
1145 td->latency_failed = 0;
1146}
1147
1148/*
1149 * We had an IO outside the latency target. Reduce the queue depth. If we
1150 * are at QD=1, then it's time to give up.
1151 */
1152static int __lat_target_failed(struct thread_data *td)
1153{
1154 if (td->latency_qd == 1)
1155 return 1;
1156
1157 td->latency_qd_high = td->latency_qd;
1158
1159 if (td->latency_qd == td->latency_qd_low)
1160 td->latency_qd_low--;
1161
1162 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1163
1164 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1165
1166 /*
1167 * When we ramp QD down, quiesce existing IO to prevent
1168 * a storm of ramp downs due to pending higher depth.
1169 */
1170 io_u_quiesce(td);
1171 lat_new_cycle(td);
1172 return 0;
1173}
1174
1175static int lat_target_failed(struct thread_data *td)
1176{
1177 if (td->o.latency_percentile.u.f == 100.0)
1178 return __lat_target_failed(td);
1179
1180 td->latency_failed++;
1181 return 0;
1182}
1183
1184void lat_target_init(struct thread_data *td)
1185{
1186 td->latency_end_run = 0;
1187
1188 if (td->o.latency_target) {
1189 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1190 fio_gettime(&td->latency_ts, NULL);
1191 td->latency_qd = 1;
1192 td->latency_qd_high = td->o.iodepth;
1193 td->latency_qd_low = 1;
1194 td->latency_ios = ddir_rw_sum(td->io_blocks);
1195 } else
1196 td->latency_qd = td->o.iodepth;
1197}
1198
1199void lat_target_reset(struct thread_data *td)
1200{
1201 if (!td->latency_end_run)
1202 lat_target_init(td);
1203}
1204
1205static void lat_target_success(struct thread_data *td)
1206{
1207 const unsigned int qd = td->latency_qd;
1208 struct thread_options *o = &td->o;
1209
1210 td->latency_qd_low = td->latency_qd;
1211
1212 /*
1213 * If we haven't failed yet, we double up to a failing value instead
1214 * of bisecting from highest possible queue depth. If we have set
1215 * a limit other than td->o.iodepth, bisect between that.
1216 */
1217 if (td->latency_qd_high != o->iodepth)
1218 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1219 else
1220 td->latency_qd *= 2;
1221
1222 if (td->latency_qd > o->iodepth)
1223 td->latency_qd = o->iodepth;
1224
1225 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1226
1227 /*
1228 * Same as last one, we are done. Let it run a latency cycle, so
1229 * we get only the results from the targeted depth.
1230 */
1231 if (td->latency_qd == qd) {
1232 if (td->latency_end_run) {
1233 dprint(FD_RATE, "We are done\n");
1234 td->done = 1;
1235 } else {
1236 dprint(FD_RATE, "Quiesce and final run\n");
1237 io_u_quiesce(td);
1238 td->latency_end_run = 1;
1239 reset_all_stats(td);
1240 reset_io_stats(td);
1241 }
1242 }
1243
1244 lat_new_cycle(td);
1245}
1246
1247/*
1248 * Check if we can bump the queue depth
1249 */
1250void lat_target_check(struct thread_data *td)
1251{
1252 uint64_t usec_window;
1253 uint64_t ios;
1254 double success_ios;
1255
1256 usec_window = utime_since_now(&td->latency_ts);
1257 if (usec_window < td->o.latency_window)
1258 return;
1259
1260 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1261 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1262 success_ios *= 100.0;
1263
1264 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1265
1266 if (success_ios >= td->o.latency_percentile.u.f)
1267 lat_target_success(td);
1268 else
1269 __lat_target_failed(td);
1270}
1271
1272/*
1273 * If latency target is enabled, we might be ramping up or down and not
1274 * using the full queue depth available.
1275 */
1276int queue_full(const struct thread_data *td)
1277{
1278 const int qempty = io_u_qempty(&td->io_u_freelist);
1279
1280 if (qempty)
1281 return 1;
1282 if (!td->o.latency_target)
1283 return 0;
1284
1285 return td->cur_depth >= td->latency_qd;
1286}
1287
1288struct io_u *__get_io_u(struct thread_data *td)
1289{
1290 struct io_u *io_u = NULL;
1291
1292 if (td->stop_io)
1293 return NULL;
1294
1295 td_io_u_lock(td);
1296
1297again:
1298 if (!io_u_rempty(&td->io_u_requeues))
1299 io_u = io_u_rpop(&td->io_u_requeues);
1300 else if (!queue_full(td)) {
1301 io_u = io_u_qpop(&td->io_u_freelist);
1302
1303 io_u->file = NULL;
1304 io_u->buflen = 0;
1305 io_u->resid = 0;
1306 io_u->end_io = NULL;
1307 }
1308
1309 if (io_u) {
1310 assert(io_u->flags & IO_U_F_FREE);
1311 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1312 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1313 IO_U_F_VER_LIST);
1314
1315 io_u->error = 0;
1316 io_u->acct_ddir = -1;
1317 td->cur_depth++;
1318 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1319 io_u->ipo = NULL;
1320 } else if (td->o.verify_async) {
1321 /*
1322 * We ran out, wait for async verify threads to finish and
1323 * return one
1324 */
1325 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1326 goto again;
1327 }
1328
1329 td_io_u_unlock(td);
1330 return io_u;
1331}
1332
1333static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1334{
1335 if (!(td->flags & TD_F_TRIM_BACKLOG))
1336 return 0;
1337
1338 if (td->trim_entries) {
1339 int get_trim = 0;
1340
1341 if (td->trim_batch) {
1342 td->trim_batch--;
1343 get_trim = 1;
1344 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1345 td->last_ddir != DDIR_READ) {
1346 td->trim_batch = td->o.trim_batch;
1347 if (!td->trim_batch)
1348 td->trim_batch = td->o.trim_backlog;
1349 get_trim = 1;
1350 }
1351
1352 if (get_trim && !get_next_trim(td, io_u))
1353 return 1;
1354 }
1355
1356 return 0;
1357}
1358
1359static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1360{
1361 if (!(td->flags & TD_F_VER_BACKLOG))
1362 return 0;
1363
1364 if (td->io_hist_len) {
1365 int get_verify = 0;
1366
1367 if (td->verify_batch)
1368 get_verify = 1;
1369 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1370 td->last_ddir != DDIR_READ) {
1371 td->verify_batch = td->o.verify_batch;
1372 if (!td->verify_batch)
1373 td->verify_batch = td->o.verify_backlog;
1374 get_verify = 1;
1375 }
1376
1377 if (get_verify && !get_next_verify(td, io_u)) {
1378 td->verify_batch--;
1379 return 1;
1380 }
1381 }
1382
1383 return 0;
1384}
1385
1386/*
1387 * Fill offset and start time into the buffer content, to prevent too
1388 * easy compressible data for simple de-dupe attempts. Do this for every
1389 * 512b block in the range, since that should be the smallest block size
1390 * we can expect from a device.
1391 */
1392static void small_content_scramble(struct io_u *io_u)
1393{
1394 unsigned int i, nr_blocks = io_u->buflen / 512;
1395 uint64_t boffset;
1396 unsigned int offset;
1397 void *p, *end;
1398
1399 if (!nr_blocks)
1400 return;
1401
1402 p = io_u->xfer_buf;
1403 boffset = io_u->offset;
1404 io_u->buf_filled_len = 0;
1405
1406 for (i = 0; i < nr_blocks; i++) {
1407 /*
1408 * Fill the byte offset into a "random" start offset of
1409 * the buffer, given by the product of the usec time
1410 * and the actual offset.
1411 */
1412 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1413 offset &= ~(sizeof(uint64_t) - 1);
1414 if (offset >= 512 - sizeof(uint64_t))
1415 offset -= sizeof(uint64_t);
1416 memcpy(p + offset, &boffset, sizeof(boffset));
1417
1418 end = p + 512 - sizeof(io_u->start_time);
1419 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1420 p += 512;
1421 boffset += 512;
1422 }
1423}
1424
1425/*
1426 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1427 * etc. The returned io_u is fully ready to be prepped and submitted.
1428 */
1429struct io_u *get_io_u(struct thread_data *td)
1430{
1431 struct fio_file *f;
1432 struct io_u *io_u;
1433 int do_scramble = 0;
1434 long ret = 0;
1435
1436 io_u = __get_io_u(td);
1437 if (!io_u) {
1438 dprint(FD_IO, "__get_io_u failed\n");
1439 return NULL;
1440 }
1441
1442 if (check_get_verify(td, io_u))
1443 goto out;
1444 if (check_get_trim(td, io_u))
1445 goto out;
1446
1447 /*
1448 * from a requeue, io_u already setup
1449 */
1450 if (io_u->file)
1451 goto out;
1452
1453 /*
1454 * If using an iolog, grab next piece if any available.
1455 */
1456 if (td->flags & TD_F_READ_IOLOG) {
1457 if (read_iolog_get(td, io_u))
1458 goto err_put;
1459 } else if (set_io_u_file(td, io_u)) {
1460 ret = -EBUSY;
1461 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1462 goto err_put;
1463 }
1464
1465 f = io_u->file;
1466 if (!f) {
1467 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1468 goto err_put;
1469 }
1470
1471 assert(fio_file_open(f));
1472
1473 if (ddir_rw(io_u->ddir)) {
1474 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1475 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1476 goto err_put;
1477 }
1478
1479 f->last_start[io_u->ddir] = io_u->offset;
1480 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1481
1482 if (io_u->ddir == DDIR_WRITE) {
1483 if (td->flags & TD_F_REFILL_BUFFERS) {
1484 io_u_fill_buffer(td, io_u,
1485 td->o.min_bs[DDIR_WRITE],
1486 io_u->xfer_buflen);
1487 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1488 !(td->flags & TD_F_COMPRESS))
1489 do_scramble = 1;
1490 if (td->flags & TD_F_VER_NONE) {
1491 populate_verify_io_u(td, io_u);
1492 do_scramble = 0;
1493 }
1494 } else if (io_u->ddir == DDIR_READ) {
1495 /*
1496 * Reset the buf_filled parameters so next time if the
1497 * buffer is used for writes it is refilled.
1498 */
1499 io_u->buf_filled_len = 0;
1500 }
1501 }
1502
1503 /*
1504 * Set io data pointers.
1505 */
1506 io_u->xfer_buf = io_u->buf;
1507 io_u->xfer_buflen = io_u->buflen;
1508
1509out:
1510 assert(io_u->file);
1511 if (!td_io_prep(td, io_u)) {
1512 if (!td->o.disable_slat)
1513 fio_gettime(&io_u->start_time, NULL);
1514 if (do_scramble)
1515 small_content_scramble(io_u);
1516 return io_u;
1517 }
1518err_put:
1519 dprint(FD_IO, "get_io_u failed\n");
1520 put_io_u(td, io_u);
1521 return ERR_PTR(ret);
1522}
1523
1524void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1525{
1526 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1527
1528 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1529 return;
1530
1531 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1532 io_u->file ? " on file " : "",
1533 io_u->file ? io_u->file->file_name : "",
1534 strerror(io_u->error),
1535 io_ddir_name(io_u->ddir),
1536 io_u->offset, io_u->xfer_buflen);
1537
1538 if (!td->error)
1539 td_verror(td, io_u->error, "io_u error");
1540}
1541
1542static inline int gtod_reduce(struct thread_data *td)
1543{
1544 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1545 && td->o.disable_bw;
1546}
1547
1548static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1549 struct io_completion_data *icd,
1550 const enum fio_ddir idx, unsigned int bytes)
1551{
1552 unsigned long lusec = 0;
1553
1554 if (!gtod_reduce(td))
1555 lusec = utime_since(&io_u->issue_time, &icd->time);
1556
1557 if (!td->o.disable_lat) {
1558 unsigned long tusec;
1559
1560 tusec = utime_since(&io_u->start_time, &icd->time);
1561 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1562
1563 if (td->flags & TD_F_PROFILE_OPS) {
1564 struct prof_io_ops *ops = &td->prof_io_ops;
1565
1566 if (ops->io_u_lat)
1567 icd->error = ops->io_u_lat(td, tusec);
1568 }
1569
1570 if (td->o.max_latency && tusec > td->o.max_latency)
1571 lat_fatal(td, icd, tusec, td->o.max_latency);
1572 if (td->o.latency_target && tusec > td->o.latency_target) {
1573 if (lat_target_failed(td))
1574 lat_fatal(td, icd, tusec, td->o.latency_target);
1575 }
1576 }
1577
1578 if (!td->o.disable_clat) {
1579 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1580 io_u_mark_latency(td, lusec);
1581 }
1582
1583 if (!td->o.disable_bw)
1584 add_bw_sample(td, idx, bytes, &icd->time);
1585
1586 if (!gtod_reduce(td))
1587 add_iops_sample(td, idx, bytes, &icd->time);
1588}
1589
1590static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1591{
1592 uint64_t secs, remainder, bps, bytes;
1593
1594 bytes = td->this_io_bytes[ddir];
1595 bps = td->rate_bps[ddir];
1596 secs = bytes / bps;
1597 remainder = bytes % bps;
1598 return remainder * 1000000 / bps + secs * 1000000;
1599}
1600
1601static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1602 struct io_completion_data *icd)
1603{
1604 struct io_u *io_u = *io_u_ptr;
1605 enum fio_ddir ddir = io_u->ddir;
1606 struct fio_file *f = io_u->file;
1607
1608 dprint_io_u(io_u, "io complete");
1609
1610 td_io_u_lock(td);
1611 assert(io_u->flags & IO_U_F_FLIGHT);
1612 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1613
1614 /*
1615 * Mark IO ok to verify
1616 */
1617 if (io_u->ipo) {
1618 /*
1619 * Remove errored entry from the verification list
1620 */
1621 if (io_u->error)
1622 unlog_io_piece(td, io_u);
1623 else {
1624 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1625 write_barrier();
1626 }
1627 }
1628
1629 td_io_u_unlock(td);
1630
1631 if (ddir_sync(ddir)) {
1632 td->last_was_sync = 1;
1633 if (f) {
1634 f->first_write = -1ULL;
1635 f->last_write = -1ULL;
1636 }
1637 return;
1638 }
1639
1640 td->last_was_sync = 0;
1641 td->last_ddir = ddir;
1642
1643 if (!io_u->error && ddir_rw(ddir)) {
1644 unsigned int bytes = io_u->buflen - io_u->resid;
1645 const enum fio_ddir oddir = ddir ^ 1;
1646 int ret;
1647
1648 td->io_blocks[ddir]++;
1649 td->this_io_blocks[ddir]++;
1650 td->io_bytes[ddir] += bytes;
1651
1652 if (!(io_u->flags & IO_U_F_VER_LIST))
1653 td->this_io_bytes[ddir] += bytes;
1654
1655 if (ddir == DDIR_WRITE) {
1656 if (f) {
1657 if (f->first_write == -1ULL ||
1658 io_u->offset < f->first_write)
1659 f->first_write = io_u->offset;
1660 if (f->last_write == -1ULL ||
1661 ((io_u->offset + bytes) > f->last_write))
1662 f->last_write = io_u->offset + bytes;
1663 }
1664 if (td->last_write_comp) {
1665 int idx = td->last_write_idx++;
1666
1667 td->last_write_comp[idx] = io_u->offset;
1668 if (td->last_write_idx == td->o.iodepth)
1669 td->last_write_idx = 0;
1670 }
1671 }
1672
1673 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1674 td->runstate == TD_VERIFYING)) {
1675 account_io_completion(td, io_u, icd, ddir, bytes);
1676
1677 if (__should_check_rate(td, ddir)) {
1678 td->rate_pending_usleep[ddir] =
1679 (usec_for_io(td, ddir) -
1680 utime_since_now(&td->start));
1681 }
1682 if (ddir != DDIR_TRIM &&
1683 __should_check_rate(td, oddir)) {
1684 td->rate_pending_usleep[oddir] =
1685 (usec_for_io(td, oddir) -
1686 utime_since_now(&td->start));
1687 }
1688 }
1689
1690 icd->bytes_done[ddir] += bytes;
1691
1692 if (io_u->end_io) {
1693 ret = io_u->end_io(td, io_u_ptr);
1694 io_u = *io_u_ptr;
1695 if (ret && !icd->error)
1696 icd->error = ret;
1697 }
1698 } else if (io_u->error) {
1699 icd->error = io_u->error;
1700 io_u_log_error(td, io_u);
1701 }
1702 if (icd->error) {
1703 enum error_type_bit eb = td_error_type(ddir, icd->error);
1704
1705 if (!td_non_fatal_error(td, eb, icd->error))
1706 return;
1707
1708 /*
1709 * If there is a non_fatal error, then add to the error count
1710 * and clear all the errors.
1711 */
1712 update_error_count(td, icd->error);
1713 td_clear_error(td);
1714 icd->error = 0;
1715 if (io_u)
1716 io_u->error = 0;
1717 }
1718}
1719
1720static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1721 int nr)
1722{
1723 int ddir;
1724
1725 if (!gtod_reduce(td))
1726 fio_gettime(&icd->time, NULL);
1727
1728 icd->nr = nr;
1729
1730 icd->error = 0;
1731 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1732 icd->bytes_done[ddir] = 0;
1733}
1734
1735static void ios_completed(struct thread_data *td,
1736 struct io_completion_data *icd)
1737{
1738 struct io_u *io_u;
1739 int i;
1740
1741 for (i = 0; i < icd->nr; i++) {
1742 io_u = td->io_ops->event(td, i);
1743
1744 io_completed(td, &io_u, icd);
1745
1746 if (io_u)
1747 put_io_u(td, io_u);
1748 }
1749}
1750
1751/*
1752 * Complete a single io_u for the sync engines.
1753 */
1754int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1755 uint64_t *bytes)
1756{
1757 struct io_completion_data icd;
1758
1759 init_icd(td, &icd, 1);
1760 io_completed(td, &io_u, &icd);
1761
1762 if (io_u)
1763 put_io_u(td, io_u);
1764
1765 if (icd.error) {
1766 td_verror(td, icd.error, "io_u_sync_complete");
1767 return -1;
1768 }
1769
1770 if (bytes) {
1771 int ddir;
1772
1773 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1774 bytes[ddir] += icd.bytes_done[ddir];
1775 }
1776
1777 return 0;
1778}
1779
1780/*
1781 * Called to complete min_events number of io for the async engines.
1782 */
1783int io_u_queued_complete(struct thread_data *td, int min_evts,
1784 uint64_t *bytes)
1785{
1786 struct io_completion_data icd;
1787 struct timespec *tvp = NULL;
1788 int ret;
1789 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1790
1791 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1792
1793 if (!min_evts)
1794 tvp = &ts;
1795 else if (min_evts > td->cur_depth)
1796 min_evts = td->cur_depth;
1797
1798 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1799 if (ret < 0) {
1800 td_verror(td, -ret, "td_io_getevents");
1801 return ret;
1802 } else if (!ret)
1803 return ret;
1804
1805 init_icd(td, &icd, ret);
1806 ios_completed(td, &icd);
1807 if (icd.error) {
1808 td_verror(td, icd.error, "io_u_queued_complete");
1809 return -1;
1810 }
1811
1812 if (bytes) {
1813 int ddir;
1814
1815 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1816 bytes[ddir] += icd.bytes_done[ddir];
1817 }
1818
1819 return 0;
1820}
1821
1822/*
1823 * Call when io_u is really queued, to update the submission latency.
1824 */
1825void io_u_queued(struct thread_data *td, struct io_u *io_u)
1826{
1827 if (!td->o.disable_slat) {
1828 unsigned long slat_time;
1829
1830 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1831 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1832 io_u->offset);
1833 }
1834}
1835
1836/*
1837 * See if we should reuse the last seed, if dedupe is enabled
1838 */
1839static struct frand_state *get_buf_state(struct thread_data *td)
1840{
1841 unsigned int v;
1842 unsigned long r;
1843
1844 if (!td->o.dedupe_percentage)
1845 return &td->buf_state;
1846 else if (td->o.dedupe_percentage == 100)
1847 return &td->buf_state_prev;
1848
1849 r = __rand(&td->dedupe_state);
1850 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
1851
1852 if (v <= td->o.dedupe_percentage)
1853 return &td->buf_state_prev;
1854
1855 return &td->buf_state;
1856}
1857
1858static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1859{
1860 if (rs == &td->buf_state)
1861 frand_copy(&td->buf_state_prev, rs);
1862}
1863
1864void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1865 unsigned int max_bs)
1866{
1867 struct thread_options *o = &td->o;
1868
1869 if (o->compress_percentage || o->dedupe_percentage) {
1870 unsigned int perc = td->o.compress_percentage;
1871 struct frand_state *rs;
1872 unsigned int left = max_bs;
1873
1874 do {
1875 rs = get_buf_state(td);
1876
1877 min_write = min(min_write, left);
1878
1879 if (perc) {
1880 unsigned int seg = min_write;
1881
1882 seg = min(min_write, td->o.compress_chunk);
1883 if (!seg)
1884 seg = min_write;
1885
1886 fill_random_buf_percentage(rs, buf, perc, seg,
1887 min_write, o->buffer_pattern,
1888 o->buffer_pattern_bytes);
1889 } else
1890 fill_random_buf(rs, buf, min_write);
1891
1892 buf += min_write;
1893 left -= min_write;
1894 save_buf_state(td, rs);
1895 } while (left);
1896 } else if (o->buffer_pattern_bytes)
1897 fill_buffer_pattern(td, buf, max_bs);
1898 else if (o->zero_buffers)
1899 memset(buf, 0, max_bs);
1900 else
1901 fill_random_buf(get_buf_state(td), buf, max_bs);
1902}
1903
1904/*
1905 * "randomly" fill the buffer contents
1906 */
1907void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1908 unsigned int min_write, unsigned int max_bs)
1909{
1910 io_u->buf_filled_len = 0;
1911 fill_io_buffer(td, io_u->buf, min_write, max_bs);
1912}
fio - Flexible IO Tester