projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Remove use of OS provided random functions
[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 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
72 if (!max_blocks)
73 return 0;
74
75 return max_blocks;
76}
77
78struct rand_off {
79 struct flist_head list;
80 uint64_t off;
81};
82
83static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
84 enum fio_ddir ddir, uint64_t *b)
85{
86 uint64_t r, lastb;
87
88 lastb = last_block(td, f, ddir);
89 if (!lastb)
90 return 1;
91
92 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
93 r = __rand(&td->__random_state);
94
95 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
96
97 *b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0));
98 } else {
99 uint64_t off = 0;
100
101 if (lfsr_next(&f->lfsr, &off, lastb))
102 return 1;
103
104 *b = off;
105 }
106
107 /*
108 * if we are not maintaining a random map, we are done.
109 */
110 if (!file_randommap(td, f))
111 goto ret;
112
113 /*
114 * calculate map offset and check if it's free
115 */
116 if (random_map_free(f, *b))
117 goto ret;
118
119 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
120 (unsigned long long) *b);
121
122 *b = axmap_next_free(f->io_axmap, *b);
123 if (*b == (uint64_t) -1ULL)
124 return 1;
125ret:
126 return 0;
127}
128
129static int __get_next_rand_offset_zipf(struct thread_data *td,
130 struct fio_file *f, enum fio_ddir ddir,
131 uint64_t *b)
132{
133 *b = zipf_next(&f->zipf);
134 return 0;
135}
136
137static int __get_next_rand_offset_pareto(struct thread_data *td,
138 struct fio_file *f, enum fio_ddir ddir,
139 uint64_t *b)
140{
141 *b = pareto_next(&f->zipf);
142 return 0;
143}
144
145static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
146{
147 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
148 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
149
150 return r1->off - r2->off;
151}
152
153static int get_off_from_method(struct thread_data *td, struct fio_file *f,
154 enum fio_ddir ddir, uint64_t *b)
155{
156 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
157 return __get_next_rand_offset(td, f, ddir, b);
158 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
159 return __get_next_rand_offset_zipf(td, f, ddir, b);
160 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
161 return __get_next_rand_offset_pareto(td, f, ddir, b);
162
163 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
164 return 1;
165}
166
167/*
168 * Sort the reads for a verify phase in batches of verifysort_nr, if
169 * specified.
170 */
171static inline int should_sort_io(struct thread_data *td)
172{
173 if (!td->o.verifysort_nr || !td->o.do_verify)
174 return 0;
175 if (!td_random(td))
176 return 0;
177 if (td->runstate != TD_VERIFYING)
178 return 0;
179 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
180 return 0;
181
182 return 1;
183}
184
185static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
186{
187 unsigned int v;
188 unsigned long r;
189
190 if (td->o.perc_rand[ddir] == 100)
191 return 1;
192
193 r = __rand(&td->__seq_rand_state[ddir]);
194 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
195
196 return v <= td->o.perc_rand[ddir];
197}
198
199static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
200 enum fio_ddir ddir, uint64_t *b)
201{
202 struct rand_off *r;
203 int i, ret = 1;
204
205 if (!should_sort_io(td))
206 return get_off_from_method(td, f, ddir, b);
207
208 if (!flist_empty(&td->next_rand_list)) {
209 struct rand_off *r;
210fetch:
211 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
212 flist_del(&r->list);
213 *b = r->off;
214 free(r);
215 return 0;
216 }
217
218 for (i = 0; i < td->o.verifysort_nr; i++) {
219 r = malloc(sizeof(*r));
220
221 ret = get_off_from_method(td, f, ddir, &r->off);
222 if (ret) {
223 free(r);
224 break;
225 }
226
227 flist_add(&r->list, &td->next_rand_list);
228 }
229
230 if (ret && !i)
231 return ret;
232
233 assert(!flist_empty(&td->next_rand_list));
234 flist_sort(NULL, &td->next_rand_list, flist_cmp);
235 goto fetch;
236}
237
238static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
239 enum fio_ddir ddir, uint64_t *b)
240{
241 if (!get_next_rand_offset(td, f, ddir, b))
242 return 0;
243
244 if (td->o.time_based) {
245 fio_file_reset(td, f);
246 if (!get_next_rand_offset(td, f, ddir, b))
247 return 0;
248 }
249
250 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
251 f->file_name, (unsigned long long) f->last_pos,
252 (unsigned long long) f->real_file_size);
253 return 1;
254}
255
256static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
257 enum fio_ddir ddir, uint64_t *offset)
258{
259 struct thread_options *o = &td->o;
260
261 assert(ddir_rw(ddir));
262
263 if (f->last_pos >= f->io_size + get_start_offset(td, f) &&
264 o->time_based)
265 f->last_pos = f->last_pos - f->io_size;
266
267 if (f->last_pos < f->real_file_size) {
268 uint64_t pos;
269
270 if (f->last_pos == f->file_offset && o->ddir_seq_add < 0)
271 f->last_pos = f->real_file_size;
272
273 pos = f->last_pos - f->file_offset;
274 if (pos && o->ddir_seq_add) {
275 pos += o->ddir_seq_add;
276
277 /*
278 * If we reach beyond the end of the file
279 * with holed IO, wrap around to the
280 * beginning again.
281 */
282 if (pos >= f->real_file_size)
283 pos = f->file_offset;
284 }
285
286 *offset = pos;
287 return 0;
288 }
289
290 return 1;
291}
292
293static int get_next_block(struct thread_data *td, struct io_u *io_u,
294 enum fio_ddir ddir, int rw_seq,
295 unsigned int *is_random)
296{
297 struct fio_file *f = io_u->file;
298 uint64_t b, offset;
299 int ret;
300
301 assert(ddir_rw(ddir));
302
303 b = offset = -1ULL;
304
305 if (rw_seq) {
306 if (td_random(td)) {
307 if (should_do_random(td, ddir)) {
308 ret = get_next_rand_block(td, f, ddir, &b);
309 *is_random = 1;
310 } else {
311 *is_random = 0;
312 io_u->flags |= IO_U_F_BUSY_OK;
313 ret = get_next_seq_offset(td, f, ddir, &offset);
314 if (ret)
315 ret = get_next_rand_block(td, f, ddir, &b);
316 }
317 } else {
318 *is_random = 0;
319 ret = get_next_seq_offset(td, f, ddir, &offset);
320 }
321 } else {
322 io_u->flags |= IO_U_F_BUSY_OK;
323 *is_random = 0;
324
325 if (td->o.rw_seq == RW_SEQ_SEQ) {
326 ret = get_next_seq_offset(td, f, ddir, &offset);
327 if (ret) {
328 ret = get_next_rand_block(td, f, ddir, &b);
329 *is_random = 0;
330 }
331 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
332 if (f->last_start != -1ULL)
333 offset = f->last_start - f->file_offset;
334 else
335 offset = 0;
336 ret = 0;
337 } else {
338 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
339 ret = 1;
340 }
341 }
342
343 if (!ret) {
344 if (offset != -1ULL)
345 io_u->offset = offset;
346 else if (b != -1ULL)
347 io_u->offset = b * td->o.ba[ddir];
348 else {
349 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
350 ret = 1;
351 }
352 }
353
354 return ret;
355}
356
357/*
358 * For random io, generate a random new block and see if it's used. Repeat
359 * until we find a free one. For sequential io, just return the end of
360 * the last io issued.
361 */
362static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
363 unsigned int *is_random)
364{
365 struct fio_file *f = io_u->file;
366 enum fio_ddir ddir = io_u->ddir;
367 int rw_seq_hit = 0;
368
369 assert(ddir_rw(ddir));
370
371 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
372 rw_seq_hit = 1;
373 td->ddir_seq_nr = td->o.ddir_seq_nr;
374 }
375
376 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
377 return 1;
378
379 if (io_u->offset >= f->io_size) {
380 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
381 (unsigned long long) io_u->offset,
382 (unsigned long long) f->io_size);
383 return 1;
384 }
385
386 io_u->offset += f->file_offset;
387 if (io_u->offset >= f->real_file_size) {
388 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
389 (unsigned long long) io_u->offset,
390 (unsigned long long) f->real_file_size);
391 return 1;
392 }
393
394 return 0;
395}
396
397static int get_next_offset(struct thread_data *td, struct io_u *io_u,
398 unsigned int *is_random)
399{
400 if (td->flags & TD_F_PROFILE_OPS) {
401 struct prof_io_ops *ops = &td->prof_io_ops;
402
403 if (ops->fill_io_u_off)
404 return ops->fill_io_u_off(td, io_u, is_random);
405 }
406
407 return __get_next_offset(td, io_u, is_random);
408}
409
410static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
411 unsigned int buflen)
412{
413 struct fio_file *f = io_u->file;
414
415 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
416}
417
418static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
419 unsigned int is_random)
420{
421 int ddir = io_u->ddir;
422 unsigned int buflen = 0;
423 unsigned int minbs, maxbs;
424 unsigned long r;
425
426 assert(ddir_rw(ddir));
427
428 if (td->o.bs_is_seq_rand)
429 ddir = is_random ? DDIR_WRITE: DDIR_READ;
430
431 minbs = td->o.min_bs[ddir];
432 maxbs = td->o.max_bs[ddir];
433
434 if (minbs == maxbs)
435 return minbs;
436
437 /*
438 * If we can't satisfy the min block size from here, then fail
439 */
440 if (!io_u_fits(td, io_u, minbs))
441 return 0;
442
443 do {
444 r = __rand(&td->__bsrange_state);
445
446 if (!td->o.bssplit_nr[ddir]) {
447 buflen = 1 + (unsigned int) ((double) maxbs *
448 (r / (FRAND_MAX + 1.0)));
449 if (buflen < minbs)
450 buflen = minbs;
451 } else {
452 long perc = 0;
453 unsigned int i;
454
455 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
456 struct bssplit *bsp = &td->o.bssplit[ddir][i];
457
458 buflen = bsp->bs;
459 perc += bsp->perc;
460 if ((r <= ((FRAND_MAX / 100L) * perc)) &&
461 io_u_fits(td, io_u, buflen))
462 break;
463 }
464 }
465
466 if (td->o.do_verify && td->o.verify != VERIFY_NONE)
467 buflen = (buflen + td->o.verify_interval - 1) &
468 ~(td->o.verify_interval - 1);
469
470 if (!td->o.bs_unaligned && is_power_of_2(minbs))
471 buflen = (buflen + minbs - 1) & ~(minbs - 1);
472
473 } while (!io_u_fits(td, io_u, buflen));
474
475 return buflen;
476}
477
478static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
479 unsigned int is_random)
480{
481 if (td->flags & TD_F_PROFILE_OPS) {
482 struct prof_io_ops *ops = &td->prof_io_ops;
483
484 if (ops->fill_io_u_size)
485 return ops->fill_io_u_size(td, io_u, is_random);
486 }
487
488 return __get_next_buflen(td, io_u, is_random);
489}
490
491static void set_rwmix_bytes(struct thread_data *td)
492{
493 unsigned int diff;
494
495 /*
496 * we do time or byte based switch. this is needed because
497 * buffered writes may issue a lot quicker than they complete,
498 * whereas reads do not.
499 */
500 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
501 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
502}
503
504static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
505{
506 unsigned int v;
507 unsigned long r;
508
509 r = __rand(&td->__rwmix_state);
510 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
511
512 if (v <= td->o.rwmix[DDIR_READ])
513 return DDIR_READ;
514
515 return DDIR_WRITE;
516}
517
518void io_u_quiesce(struct thread_data *td)
519{
520 /*
521 * We are going to sleep, ensure that we flush anything pending as
522 * not to skew our latency numbers.
523 *
524 * Changed to only monitor 'in flight' requests here instead of the
525 * td->cur_depth, b/c td->cur_depth does not accurately represent
526 * io's that have been actually submitted to an async engine,
527 * and cur_depth is meaningless for sync engines.
528 */
529 while (td->io_u_in_flight) {
530 int fio_unused ret;
531
532 ret = io_u_queued_complete(td, 1, NULL);
533 }
534}
535
536static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
537{
538 enum fio_ddir odir = ddir ^ 1;
539 struct timeval t;
540 long usec;
541
542 assert(ddir_rw(ddir));
543
544 if (td->rate_pending_usleep[ddir] <= 0)
545 return ddir;
546
547 /*
548 * We have too much pending sleep in this direction. See if we
549 * should switch.
550 */
551 if (td_rw(td) && td->o.rwmix[odir]) {
552 /*
553 * Other direction does not have too much pending, switch
554 */
555 if (td->rate_pending_usleep[odir] < 100000)
556 return odir;
557
558 /*
559 * Both directions have pending sleep. Sleep the minimum time
560 * and deduct from both.
561 */
562 if (td->rate_pending_usleep[ddir] <=
563 td->rate_pending_usleep[odir]) {
564 usec = td->rate_pending_usleep[ddir];
565 } else {
566 usec = td->rate_pending_usleep[odir];
567 ddir = odir;
568 }
569 } else
570 usec = td->rate_pending_usleep[ddir];
571
572 io_u_quiesce(td);
573
574 fio_gettime(&t, NULL);
575 usec_sleep(td, usec);
576 usec = utime_since_now(&t);
577
578 td->rate_pending_usleep[ddir] -= usec;
579
580 odir = ddir ^ 1;
581 if (td_rw(td) && __should_check_rate(td, odir))
582 td->rate_pending_usleep[odir] -= usec;
583
584 if (ddir_trim(ddir))
585 return ddir;
586
587 return ddir;
588}
589
590/*
591 * Return the data direction for the next io_u. If the job is a
592 * mixed read/write workload, check the rwmix cycle and switch if
593 * necessary.
594 */
595static enum fio_ddir get_rw_ddir(struct thread_data *td)
596{
597 enum fio_ddir ddir;
598
599 /*
600 * see if it's time to fsync
601 */
602 if (td->o.fsync_blocks &&
603 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
604 td->io_issues[DDIR_WRITE] && should_fsync(td))
605 return DDIR_SYNC;
606
607 /*
608 * see if it's time to fdatasync
609 */
610 if (td->o.fdatasync_blocks &&
611 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
612 td->io_issues[DDIR_WRITE] && should_fsync(td))
613 return DDIR_DATASYNC;
614
615 /*
616 * see if it's time to sync_file_range
617 */
618 if (td->sync_file_range_nr &&
619 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
620 td->io_issues[DDIR_WRITE] && should_fsync(td))
621 return DDIR_SYNC_FILE_RANGE;
622
623 if (td_rw(td)) {
624 /*
625 * Check if it's time to seed a new data direction.
626 */
627 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
628 /*
629 * Put a top limit on how many bytes we do for
630 * one data direction, to avoid overflowing the
631 * ranges too much
632 */
633 ddir = get_rand_ddir(td);
634
635 if (ddir != td->rwmix_ddir)
636 set_rwmix_bytes(td);
637
638 td->rwmix_ddir = ddir;
639 }
640 ddir = td->rwmix_ddir;
641 } else if (td_read(td))
642 ddir = DDIR_READ;
643 else if (td_write(td))
644 ddir = DDIR_WRITE;
645 else
646 ddir = DDIR_TRIM;
647
648 td->rwmix_ddir = rate_ddir(td, ddir);
649 return td->rwmix_ddir;
650}
651
652static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
653{
654 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
655
656 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
657 td->o.barrier_blocks &&
658 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
659 td->io_issues[DDIR_WRITE])
660 io_u->flags |= IO_U_F_BARRIER;
661}
662
663void put_file_log(struct thread_data *td, struct fio_file *f)
664{
665 unsigned int ret = put_file(td, f);
666
667 if (ret)
668 td_verror(td, ret, "file close");
669}
670
671void put_io_u(struct thread_data *td, struct io_u *io_u)
672{
673 td_io_u_lock(td);
674
675 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
676 put_file_log(td, io_u->file);
677
678 io_u->file = NULL;
679 io_u->flags |= IO_U_F_FREE;
680
681 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
682 td->cur_depth--;
683 io_u_qpush(&td->io_u_freelist, io_u);
684 td_io_u_unlock(td);
685 td_io_u_free_notify(td);
686}
687
688void clear_io_u(struct thread_data *td, struct io_u *io_u)
689{
690 io_u->flags &= ~IO_U_F_FLIGHT;
691 put_io_u(td, io_u);
692}
693
694void requeue_io_u(struct thread_data *td, struct io_u **io_u)
695{
696 struct io_u *__io_u = *io_u;
697 enum fio_ddir ddir = acct_ddir(__io_u);
698
699 dprint(FD_IO, "requeue %p\n", __io_u);
700
701 td_io_u_lock(td);
702
703 __io_u->flags |= IO_U_F_FREE;
704 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
705 td->io_issues[ddir]--;
706
707 __io_u->flags &= ~IO_U_F_FLIGHT;
708 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
709 td->cur_depth--;
710
711 io_u_rpush(&td->io_u_requeues, __io_u);
712 td_io_u_unlock(td);
713 *io_u = NULL;
714}
715
716static int fill_io_u(struct thread_data *td, struct io_u *io_u)
717{
718 unsigned int is_random;
719
720 if (td->io_ops->flags & FIO_NOIO)
721 goto out;
722
723 set_rw_ddir(td, io_u);
724
725 /*
726 * fsync() or fdatasync() or trim etc, we are done
727 */
728 if (!ddir_rw(io_u->ddir))
729 goto out;
730
731 /*
732 * See if it's time to switch to a new zone
733 */
734 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
735 struct fio_file *f = io_u->file;
736
737 td->zone_bytes = 0;
738 f->file_offset += td->o.zone_range + td->o.zone_skip;
739
740 /*
741 * Wrap from the beginning, if we exceed the file size
742 */
743 if (f->file_offset >= f->real_file_size)
744 f->file_offset = f->real_file_size - f->file_offset;
745 f->last_pos = f->file_offset;
746 td->io_skip_bytes += td->o.zone_skip;
747 }
748
749 /*
750 * No log, let the seq/rand engine retrieve the next buflen and
751 * position.
752 */
753 if (get_next_offset(td, io_u, &is_random)) {
754 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
755 return 1;
756 }
757
758 io_u->buflen = get_next_buflen(td, io_u, is_random);
759 if (!io_u->buflen) {
760 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
761 return 1;
762 }
763
764 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
765 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
766 dprint(FD_IO, " off=%llu/%lu > %llu\n",
767 (unsigned long long) io_u->offset, io_u->buflen,
768 (unsigned long long) io_u->file->real_file_size);
769 return 1;
770 }
771
772 /*
773 * mark entry before potentially trimming io_u
774 */
775 if (td_random(td) && file_randommap(td, io_u->file))
776 mark_random_map(td, io_u);
777
778out:
779 dprint_io_u(io_u, "fill_io_u");
780 td->zone_bytes += io_u->buflen;
781 return 0;
782}
783
784static void __io_u_mark_map(unsigned int *map, unsigned int nr)
785{
786 int idx = 0;
787
788 switch (nr) {
789 default:
790 idx = 6;
791 break;
792 case 33 ... 64:
793 idx = 5;
794 break;
795 case 17 ... 32:
796 idx = 4;
797 break;
798 case 9 ... 16:
799 idx = 3;
800 break;
801 case 5 ... 8:
802 idx = 2;
803 break;
804 case 1 ... 4:
805 idx = 1;
806 case 0:
807 break;
808 }
809
810 map[idx]++;
811}
812
813void io_u_mark_submit(struct thread_data *td, unsigned int nr)
814{
815 __io_u_mark_map(td->ts.io_u_submit, nr);
816 td->ts.total_submit++;
817}
818
819void io_u_mark_complete(struct thread_data *td, unsigned int nr)
820{
821 __io_u_mark_map(td->ts.io_u_complete, nr);
822 td->ts.total_complete++;
823}
824
825void io_u_mark_depth(struct thread_data *td, unsigned int nr)
826{
827 int idx = 0;
828
829 switch (td->cur_depth) {
830 default:
831 idx = 6;
832 break;
833 case 32 ... 63:
834 idx = 5;
835 break;
836 case 16 ... 31:
837 idx = 4;
838 break;
839 case 8 ... 15:
840 idx = 3;
841 break;
842 case 4 ... 7:
843 idx = 2;
844 break;
845 case 2 ... 3:
846 idx = 1;
847 case 1:
848 break;
849 }
850
851 td->ts.io_u_map[idx] += nr;
852}
853
854static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
855{
856 int idx = 0;
857
858 assert(usec < 1000);
859
860 switch (usec) {
861 case 750 ... 999:
862 idx = 9;
863 break;
864 case 500 ... 749:
865 idx = 8;
866 break;
867 case 250 ... 499:
868 idx = 7;
869 break;
870 case 100 ... 249:
871 idx = 6;
872 break;
873 case 50 ... 99:
874 idx = 5;
875 break;
876 case 20 ... 49:
877 idx = 4;
878 break;
879 case 10 ... 19:
880 idx = 3;
881 break;
882 case 4 ... 9:
883 idx = 2;
884 break;
885 case 2 ... 3:
886 idx = 1;
887 case 0 ... 1:
888 break;
889 }
890
891 assert(idx < FIO_IO_U_LAT_U_NR);
892 td->ts.io_u_lat_u[idx]++;
893}
894
895static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
896{
897 int idx = 0;
898
899 switch (msec) {
900 default:
901 idx = 11;
902 break;
903 case 1000 ... 1999:
904 idx = 10;
905 break;
906 case 750 ... 999:
907 idx = 9;
908 break;
909 case 500 ... 749:
910 idx = 8;
911 break;
912 case 250 ... 499:
913 idx = 7;
914 break;
915 case 100 ... 249:
916 idx = 6;
917 break;
918 case 50 ... 99:
919 idx = 5;
920 break;
921 case 20 ... 49:
922 idx = 4;
923 break;
924 case 10 ... 19:
925 idx = 3;
926 break;
927 case 4 ... 9:
928 idx = 2;
929 break;
930 case 2 ... 3:
931 idx = 1;
932 case 0 ... 1:
933 break;
934 }
935
936 assert(idx < FIO_IO_U_LAT_M_NR);
937 td->ts.io_u_lat_m[idx]++;
938}
939
940static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
941{
942 if (usec < 1000)
943 io_u_mark_lat_usec(td, usec);
944 else
945 io_u_mark_lat_msec(td, usec / 1000);
946}
947
948/*
949 * Get next file to service by choosing one at random
950 */
951static struct fio_file *get_next_file_rand(struct thread_data *td,
952 enum fio_file_flags goodf,
953 enum fio_file_flags badf)
954{
955 struct fio_file *f;
956 int fno;
957
958 do {
959 int opened = 0;
960 unsigned long r;
961
962 r = __rand(&td->__next_file_state);
963 fno = (unsigned int) ((double) td->o.nr_files
964 * (r / (FRAND_MAX + 1.0)));
965
966 f = td->files[fno];
967 if (fio_file_done(f))
968 continue;
969
970 if (!fio_file_open(f)) {
971 int err;
972
973 if (td->nr_open_files >= td->o.open_files)
974 return ERR_PTR(-EBUSY);
975
976 err = td_io_open_file(td, f);
977 if (err)
978 continue;
979 opened = 1;
980 }
981
982 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
983 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
984 return f;
985 }
986 if (opened)
987 td_io_close_file(td, f);
988 } while (1);
989}
990
991/*
992 * Get next file to service by doing round robin between all available ones
993 */
994static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
995 int badf)
996{
997 unsigned int old_next_file = td->next_file;
998 struct fio_file *f;
999
1000 do {
1001 int opened = 0;
1002
1003 f = td->files[td->next_file];
1004
1005 td->next_file++;
1006 if (td->next_file >= td->o.nr_files)
1007 td->next_file = 0;
1008
1009 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1010 if (fio_file_done(f)) {
1011 f = NULL;
1012 continue;
1013 }
1014
1015 if (!fio_file_open(f)) {
1016 int err;
1017
1018 if (td->nr_open_files >= td->o.open_files)
1019 return ERR_PTR(-EBUSY);
1020
1021 err = td_io_open_file(td, f);
1022 if (err) {
1023 dprint(FD_FILE, "error %d on open of %s\n",
1024 err, f->file_name);
1025 f = NULL;
1026 continue;
1027 }
1028 opened = 1;
1029 }
1030
1031 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1032 f->flags);
1033 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1034 break;
1035
1036 if (opened)
1037 td_io_close_file(td, f);
1038
1039 f = NULL;
1040 } while (td->next_file != old_next_file);
1041
1042 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1043 return f;
1044}
1045
1046static struct fio_file *__get_next_file(struct thread_data *td)
1047{
1048 struct fio_file *f;
1049
1050 assert(td->o.nr_files <= td->files_index);
1051
1052 if (td->nr_done_files >= td->o.nr_files) {
1053 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1054 " nr_files=%d\n", td->nr_open_files,
1055 td->nr_done_files,
1056 td->o.nr_files);
1057 return NULL;
1058 }
1059
1060 f = td->file_service_file;
1061 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1062 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1063 goto out;
1064 if (td->file_service_left--)
1065 goto out;
1066 }
1067
1068 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1069 td->o.file_service_type == FIO_FSERVICE_SEQ)
1070 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1071 else
1072 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1073
1074 if (IS_ERR(f))
1075 return f;
1076
1077 td->file_service_file = f;
1078 td->file_service_left = td->file_service_nr - 1;
1079out:
1080 if (f)
1081 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1082 else
1083 dprint(FD_FILE, "get_next_file: NULL\n");
1084 return f;
1085}
1086
1087static struct fio_file *get_next_file(struct thread_data *td)
1088{
1089 if (td->flags & TD_F_PROFILE_OPS) {
1090 struct prof_io_ops *ops = &td->prof_io_ops;
1091
1092 if (ops->get_next_file)
1093 return ops->get_next_file(td);
1094 }
1095
1096 return __get_next_file(td);
1097}
1098
1099static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1100{
1101 struct fio_file *f;
1102
1103 do {
1104 f = get_next_file(td);
1105 if (IS_ERR_OR_NULL(f))
1106 return PTR_ERR(f);
1107
1108 io_u->file = f;
1109 get_file(f);
1110
1111 if (!fill_io_u(td, io_u))
1112 break;
1113
1114 put_file_log(td, f);
1115 td_io_close_file(td, f);
1116 io_u->file = NULL;
1117 fio_file_set_done(f);
1118 td->nr_done_files++;
1119 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1120 td->nr_done_files, td->o.nr_files);
1121 } while (1);
1122
1123 return 0;
1124}
1125
1126static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1127 unsigned long tusec, unsigned long max_usec)
1128{
1129 if (!td->error)
1130 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1131 td_verror(td, ETIMEDOUT, "max latency exceeded");
1132 icd->error = ETIMEDOUT;
1133}
1134
1135static void lat_new_cycle(struct thread_data *td)
1136{
1137 fio_gettime(&td->latency_ts, NULL);
1138 td->latency_ios = ddir_rw_sum(td->io_blocks);
1139 td->latency_failed = 0;
1140}
1141
1142/*
1143 * We had an IO outside the latency target. Reduce the queue depth. If we
1144 * are at QD=1, then it's time to give up.
1145 */
1146static int __lat_target_failed(struct thread_data *td)
1147{
1148 if (td->latency_qd == 1)
1149 return 1;
1150
1151 td->latency_qd_high = td->latency_qd;
1152
1153 if (td->latency_qd == td->latency_qd_low)
1154 td->latency_qd_low--;
1155
1156 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1157
1158 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1159
1160 /*
1161 * When we ramp QD down, quiesce existing IO to prevent
1162 * a storm of ramp downs due to pending higher depth.
1163 */
1164 io_u_quiesce(td);
1165 lat_new_cycle(td);
1166 return 0;
1167}
1168
1169static int lat_target_failed(struct thread_data *td)
1170{
1171 if (td->o.latency_percentile.u.f == 100.0)
1172 return __lat_target_failed(td);
1173
1174 td->latency_failed++;
1175 return 0;
1176}
1177
1178void lat_target_init(struct thread_data *td)
1179{
1180 td->latency_end_run = 0;
1181
1182 if (td->o.latency_target) {
1183 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1184 fio_gettime(&td->latency_ts, NULL);
1185 td->latency_qd = 1;
1186 td->latency_qd_high = td->o.iodepth;
1187 td->latency_qd_low = 1;
1188 td->latency_ios = ddir_rw_sum(td->io_blocks);
1189 } else
1190 td->latency_qd = td->o.iodepth;
1191}
1192
1193void lat_target_reset(struct thread_data *td)
1194{
1195 if (!td->latency_end_run)
1196 lat_target_init(td);
1197}
1198
1199static void lat_target_success(struct thread_data *td)
1200{
1201 const unsigned int qd = td->latency_qd;
1202 struct thread_options *o = &td->o;
1203
1204 td->latency_qd_low = td->latency_qd;
1205
1206 /*
1207 * If we haven't failed yet, we double up to a failing value instead
1208 * of bisecting from highest possible queue depth. If we have set
1209 * a limit other than td->o.iodepth, bisect between that.
1210 */
1211 if (td->latency_qd_high != o->iodepth)
1212 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1213 else
1214 td->latency_qd *= 2;
1215
1216 if (td->latency_qd > o->iodepth)
1217 td->latency_qd = o->iodepth;
1218
1219 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1220
1221 /*
1222 * Same as last one, we are done. Let it run a latency cycle, so
1223 * we get only the results from the targeted depth.
1224 */
1225 if (td->latency_qd == qd) {
1226 if (td->latency_end_run) {
1227 dprint(FD_RATE, "We are done\n");
1228 td->done = 1;
1229 } else {
1230 dprint(FD_RATE, "Quiesce and final run\n");
1231 io_u_quiesce(td);
1232 td->latency_end_run = 1;
1233 reset_all_stats(td);
1234 reset_io_stats(td);
1235 }
1236 }
1237
1238 lat_new_cycle(td);
1239}
1240
1241/*
1242 * Check if we can bump the queue depth
1243 */
1244void lat_target_check(struct thread_data *td)
1245{
1246 uint64_t usec_window;
1247 uint64_t ios;
1248 double success_ios;
1249
1250 usec_window = utime_since_now(&td->latency_ts);
1251 if (usec_window < td->o.latency_window)
1252 return;
1253
1254 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1255 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1256 success_ios *= 100.0;
1257
1258 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1259
1260 if (success_ios >= td->o.latency_percentile.u.f)
1261 lat_target_success(td);
1262 else
1263 __lat_target_failed(td);
1264}
1265
1266/*
1267 * If latency target is enabled, we might be ramping up or down and not
1268 * using the full queue depth available.
1269 */
1270int queue_full(const struct thread_data *td)
1271{
1272 const int qempty = io_u_qempty(&td->io_u_freelist);
1273
1274 if (qempty)
1275 return 1;
1276 if (!td->o.latency_target)
1277 return 0;
1278
1279 return td->cur_depth >= td->latency_qd;
1280}
1281
1282struct io_u *__get_io_u(struct thread_data *td)
1283{
1284 struct io_u *io_u = NULL;
1285
1286 td_io_u_lock(td);
1287
1288again:
1289 if (!io_u_rempty(&td->io_u_requeues))
1290 io_u = io_u_rpop(&td->io_u_requeues);
1291 else if (!queue_full(td)) {
1292 io_u = io_u_qpop(&td->io_u_freelist);
1293
1294 io_u->file = NULL;
1295 io_u->buflen = 0;
1296 io_u->resid = 0;
1297 io_u->end_io = NULL;
1298 }
1299
1300 if (io_u) {
1301 assert(io_u->flags & IO_U_F_FREE);
1302 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1303 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1304 IO_U_F_VER_LIST);
1305
1306 io_u->error = 0;
1307 io_u->acct_ddir = -1;
1308 td->cur_depth++;
1309 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1310 io_u->ipo = NULL;
1311 } else if (td->o.verify_async) {
1312 /*
1313 * We ran out, wait for async verify threads to finish and
1314 * return one
1315 */
1316 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1317 goto again;
1318 }
1319
1320 td_io_u_unlock(td);
1321 return io_u;
1322}
1323
1324static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1325{
1326 if (!(td->flags & TD_F_TRIM_BACKLOG))
1327 return 0;
1328
1329 if (td->trim_entries) {
1330 int get_trim = 0;
1331
1332 if (td->trim_batch) {
1333 td->trim_batch--;
1334 get_trim = 1;
1335 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1336 td->last_ddir != DDIR_READ) {
1337 td->trim_batch = td->o.trim_batch;
1338 if (!td->trim_batch)
1339 td->trim_batch = td->o.trim_backlog;
1340 get_trim = 1;
1341 }
1342
1343 if (get_trim && !get_next_trim(td, io_u))
1344 return 1;
1345 }
1346
1347 return 0;
1348}
1349
1350static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1351{
1352 if (!(td->flags & TD_F_VER_BACKLOG))
1353 return 0;
1354
1355 if (td->io_hist_len) {
1356 int get_verify = 0;
1357
1358 if (td->verify_batch)
1359 get_verify = 1;
1360 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1361 td->last_ddir != DDIR_READ) {
1362 td->verify_batch = td->o.verify_batch;
1363 if (!td->verify_batch)
1364 td->verify_batch = td->o.verify_backlog;
1365 get_verify = 1;
1366 }
1367
1368 if (get_verify && !get_next_verify(td, io_u)) {
1369 td->verify_batch--;
1370 return 1;
1371 }
1372 }
1373
1374 return 0;
1375}
1376
1377/*
1378 * Fill offset and start time into the buffer content, to prevent too
1379 * easy compressible data for simple de-dupe attempts. Do this for every
1380 * 512b block in the range, since that should be the smallest block size
1381 * we can expect from a device.
1382 */
1383static void small_content_scramble(struct io_u *io_u)
1384{
1385 unsigned int i, nr_blocks = io_u->buflen / 512;
1386 uint64_t boffset;
1387 unsigned int offset;
1388 void *p, *end;
1389
1390 if (!nr_blocks)
1391 return;
1392
1393 p = io_u->xfer_buf;
1394 boffset = io_u->offset;
1395 io_u->buf_filled_len = 0;
1396
1397 for (i = 0; i < nr_blocks; i++) {
1398 /*
1399 * Fill the byte offset into a "random" start offset of
1400 * the buffer, given by the product of the usec time
1401 * and the actual offset.
1402 */
1403 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1404 offset &= ~(sizeof(uint64_t) - 1);
1405 if (offset >= 512 - sizeof(uint64_t))
1406 offset -= sizeof(uint64_t);
1407 memcpy(p + offset, &boffset, sizeof(boffset));
1408
1409 end = p + 512 - sizeof(io_u->start_time);
1410 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1411 p += 512;
1412 boffset += 512;
1413 }
1414}
1415
1416/*
1417 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1418 * etc. The returned io_u is fully ready to be prepped and submitted.
1419 */
1420struct io_u *get_io_u(struct thread_data *td)
1421{
1422 struct fio_file *f;
1423 struct io_u *io_u;
1424 int do_scramble = 0;
1425 long ret = 0;
1426
1427 io_u = __get_io_u(td);
1428 if (!io_u) {
1429 dprint(FD_IO, "__get_io_u failed\n");
1430 return NULL;
1431 }
1432
1433 if (check_get_verify(td, io_u))
1434 goto out;
1435 if (check_get_trim(td, io_u))
1436 goto out;
1437
1438 /*
1439 * from a requeue, io_u already setup
1440 */
1441 if (io_u->file)
1442 goto out;
1443
1444 /*
1445 * If using an iolog, grab next piece if any available.
1446 */
1447 if (td->flags & TD_F_READ_IOLOG) {
1448 if (read_iolog_get(td, io_u))
1449 goto err_put;
1450 } else if (set_io_u_file(td, io_u)) {
1451 ret = -EBUSY;
1452 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1453 goto err_put;
1454 }
1455
1456 f = io_u->file;
1457 if (!f) {
1458 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1459 goto err_put;
1460 }
1461
1462 assert(fio_file_open(f));
1463
1464 if (ddir_rw(io_u->ddir)) {
1465 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1466 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1467 goto err_put;
1468 }
1469
1470 f->last_start = io_u->offset;
1471 f->last_pos = io_u->offset + io_u->buflen;
1472
1473 if (io_u->ddir == DDIR_WRITE) {
1474 if (td->flags & TD_F_REFILL_BUFFERS) {
1475 io_u_fill_buffer(td, io_u,
1476 td->o.min_bs[DDIR_WRITE],
1477 io_u->xfer_buflen);
1478 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1479 !(td->flags & TD_F_COMPRESS))
1480 do_scramble = 1;
1481 if (td->flags & TD_F_VER_NONE) {
1482 populate_verify_io_u(td, io_u);
1483 do_scramble = 0;
1484 }
1485 } else if (io_u->ddir == DDIR_READ) {
1486 /*
1487 * Reset the buf_filled parameters so next time if the
1488 * buffer is used for writes it is refilled.
1489 */
1490 io_u->buf_filled_len = 0;
1491 }
1492 }
1493
1494 /*
1495 * Set io data pointers.
1496 */
1497 io_u->xfer_buf = io_u->buf;
1498 io_u->xfer_buflen = io_u->buflen;
1499
1500out:
1501 assert(io_u->file);
1502 if (!td_io_prep(td, io_u)) {
1503 if (!td->o.disable_slat)
1504 fio_gettime(&io_u->start_time, NULL);
1505 if (do_scramble)
1506 small_content_scramble(io_u);
1507 return io_u;
1508 }
1509err_put:
1510 dprint(FD_IO, "get_io_u failed\n");
1511 put_io_u(td, io_u);
1512 return ERR_PTR(ret);
1513}
1514
1515void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1516{
1517 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1518
1519 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1520 return;
1521
1522 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1523 io_u->file ? " on file " : "",
1524 io_u->file ? io_u->file->file_name : "",
1525 strerror(io_u->error),
1526 io_ddir_name(io_u->ddir),
1527 io_u->offset, io_u->xfer_buflen);
1528
1529 if (!td->error)
1530 td_verror(td, io_u->error, "io_u error");
1531}
1532
1533static inline int gtod_reduce(struct thread_data *td)
1534{
1535 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1536 && td->o.disable_bw;
1537}
1538
1539static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1540 struct io_completion_data *icd,
1541 const enum fio_ddir idx, unsigned int bytes)
1542{
1543 unsigned long lusec = 0;
1544
1545 if (!gtod_reduce(td))
1546 lusec = utime_since(&io_u->issue_time, &icd->time);
1547
1548 if (!td->o.disable_lat) {
1549 unsigned long tusec;
1550
1551 tusec = utime_since(&io_u->start_time, &icd->time);
1552 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1553
1554 if (td->flags & TD_F_PROFILE_OPS) {
1555 struct prof_io_ops *ops = &td->prof_io_ops;
1556
1557 if (ops->io_u_lat)
1558 icd->error = ops->io_u_lat(td, tusec);
1559 }
1560
1561 if (td->o.max_latency && tusec > td->o.max_latency)
1562 lat_fatal(td, icd, tusec, td->o.max_latency);
1563 if (td->o.latency_target && tusec > td->o.latency_target) {
1564 if (lat_target_failed(td))
1565 lat_fatal(td, icd, tusec, td->o.latency_target);
1566 }
1567 }
1568
1569 if (!td->o.disable_clat) {
1570 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1571 io_u_mark_latency(td, lusec);
1572 }
1573
1574 if (!td->o.disable_bw)
1575 add_bw_sample(td, idx, bytes, &icd->time);
1576
1577 if (!gtod_reduce(td))
1578 add_iops_sample(td, idx, bytes, &icd->time);
1579}
1580
1581static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1582{
1583 uint64_t secs, remainder, bps, bytes;
1584
1585 bytes = td->this_io_bytes[ddir];
1586 bps = td->rate_bps[ddir];
1587 secs = bytes / bps;
1588 remainder = bytes % bps;
1589 return remainder * 1000000 / bps + secs * 1000000;
1590}
1591
1592static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1593 struct io_completion_data *icd)
1594{
1595 struct io_u *io_u = *io_u_ptr;
1596 enum fio_ddir ddir = io_u->ddir;
1597 struct fio_file *f = io_u->file;
1598
1599 dprint_io_u(io_u, "io complete");
1600
1601 td_io_u_lock(td);
1602 assert(io_u->flags & IO_U_F_FLIGHT);
1603 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1604
1605 /*
1606 * Mark IO ok to verify
1607 */
1608 if (io_u->ipo) {
1609 /*
1610 * Remove errored entry from the verification list
1611 */
1612 if (io_u->error)
1613 unlog_io_piece(td, io_u);
1614 else {
1615 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1616 write_barrier();
1617 }
1618 }
1619
1620 td_io_u_unlock(td);
1621
1622 if (ddir_sync(ddir)) {
1623 td->last_was_sync = 1;
1624 if (f) {
1625 f->first_write = -1ULL;
1626 f->last_write = -1ULL;
1627 }
1628 return;
1629 }
1630
1631 td->last_was_sync = 0;
1632 td->last_ddir = ddir;
1633
1634 if (!io_u->error && ddir_rw(ddir)) {
1635 unsigned int bytes = io_u->buflen - io_u->resid;
1636 const enum fio_ddir oddir = ddir ^ 1;
1637 int ret;
1638
1639 td->io_blocks[ddir]++;
1640 td->this_io_blocks[ddir]++;
1641 td->io_bytes[ddir] += bytes;
1642
1643 if (!(io_u->flags & IO_U_F_VER_LIST))
1644 td->this_io_bytes[ddir] += bytes;
1645
1646 if (ddir == DDIR_WRITE && f) {
1647 if (f->first_write == -1ULL ||
1648 io_u->offset < f->first_write)
1649 f->first_write = io_u->offset;
1650 if (f->last_write == -1ULL ||
1651 ((io_u->offset + bytes) > f->last_write))
1652 f->last_write = io_u->offset + bytes;
1653 }
1654
1655 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1656 td->runstate == TD_VERIFYING)) {
1657 account_io_completion(td, io_u, icd, ddir, bytes);
1658
1659 if (__should_check_rate(td, ddir)) {
1660 td->rate_pending_usleep[ddir] =
1661 (usec_for_io(td, ddir) -
1662 utime_since_now(&td->start));
1663 }
1664 if (ddir != DDIR_TRIM &&
1665 __should_check_rate(td, oddir)) {
1666 td->rate_pending_usleep[oddir] =
1667 (usec_for_io(td, oddir) -
1668 utime_since_now(&td->start));
1669 }
1670 }
1671
1672 icd->bytes_done[ddir] += bytes;
1673
1674 if (io_u->end_io) {
1675 ret = io_u->end_io(td, io_u_ptr);
1676 io_u = *io_u_ptr;
1677 if (ret && !icd->error)
1678 icd->error = ret;
1679 }
1680 } else if (io_u->error) {
1681 icd->error = io_u->error;
1682 io_u_log_error(td, io_u);
1683 }
1684 if (icd->error) {
1685 enum error_type_bit eb = td_error_type(ddir, icd->error);
1686
1687 if (!td_non_fatal_error(td, eb, icd->error))
1688 return;
1689
1690 /*
1691 * If there is a non_fatal error, then add to the error count
1692 * and clear all the errors.
1693 */
1694 update_error_count(td, icd->error);
1695 td_clear_error(td);
1696 icd->error = 0;
1697 if (io_u)
1698 io_u->error = 0;
1699 }
1700}
1701
1702static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1703 int nr)
1704{
1705 int ddir;
1706
1707 if (!gtod_reduce(td))
1708 fio_gettime(&icd->time, NULL);
1709
1710 icd->nr = nr;
1711
1712 icd->error = 0;
1713 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1714 icd->bytes_done[ddir] = 0;
1715}
1716
1717static void ios_completed(struct thread_data *td,
1718 struct io_completion_data *icd)
1719{
1720 struct io_u *io_u;
1721 int i;
1722
1723 for (i = 0; i < icd->nr; i++) {
1724 io_u = td->io_ops->event(td, i);
1725
1726 io_completed(td, &io_u, icd);
1727
1728 if (io_u)
1729 put_io_u(td, io_u);
1730 }
1731}
1732
1733/*
1734 * Complete a single io_u for the sync engines.
1735 */
1736int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1737 uint64_t *bytes)
1738{
1739 struct io_completion_data icd;
1740
1741 init_icd(td, &icd, 1);
1742 io_completed(td, &io_u, &icd);
1743
1744 if (io_u)
1745 put_io_u(td, io_u);
1746
1747 if (icd.error) {
1748 td_verror(td, icd.error, "io_u_sync_complete");
1749 return -1;
1750 }
1751
1752 if (bytes) {
1753 int ddir;
1754
1755 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1756 bytes[ddir] += icd.bytes_done[ddir];
1757 }
1758
1759 return 0;
1760}
1761
1762/*
1763 * Called to complete min_events number of io for the async engines.
1764 */
1765int io_u_queued_complete(struct thread_data *td, int min_evts,
1766 uint64_t *bytes)
1767{
1768 struct io_completion_data icd;
1769 struct timespec *tvp = NULL;
1770 int ret;
1771 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1772
1773 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1774
1775 if (!min_evts)
1776 tvp = &ts;
1777 else if (min_evts > td->cur_depth)
1778 min_evts = td->cur_depth;
1779
1780 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1781 if (ret < 0) {
1782 td_verror(td, -ret, "td_io_getevents");
1783 return ret;
1784 } else if (!ret)
1785 return ret;
1786
1787 init_icd(td, &icd, ret);
1788 ios_completed(td, &icd);
1789 if (icd.error) {
1790 td_verror(td, icd.error, "io_u_queued_complete");
1791 return -1;
1792 }
1793
1794 if (bytes) {
1795 int ddir;
1796
1797 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1798 bytes[ddir] += icd.bytes_done[ddir];
1799 }
1800
1801 return 0;
1802}
1803
1804/*
1805 * Call when io_u is really queued, to update the submission latency.
1806 */
1807void io_u_queued(struct thread_data *td, struct io_u *io_u)
1808{
1809 if (!td->o.disable_slat) {
1810 unsigned long slat_time;
1811
1812 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1813 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1814 io_u->offset);
1815 }
1816}
1817
1818/*
1819 * See if we should reuse the last seed, if dedupe is enabled
1820 */
1821static struct frand_state *get_buf_state(struct thread_data *td)
1822{
1823 unsigned int v;
1824 unsigned long r;
1825
1826 if (!td->o.dedupe_percentage)
1827 return &td->buf_state;
1828 else if (td->o.dedupe_percentage == 100)
1829 return &td->buf_state_prev;
1830
1831 r = __rand(&td->dedupe_state);
1832 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
1833
1834 if (v <= td->o.dedupe_percentage)
1835 return &td->buf_state_prev;
1836
1837 return &td->buf_state;
1838}
1839
1840static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1841{
1842 if (rs == &td->buf_state)
1843 frand_copy(&td->buf_state_prev, rs);
1844}
1845
1846void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1847 unsigned int max_bs)
1848{
1849 if (td->o.buffer_pattern_bytes)
1850 fill_buffer_pattern(td, buf, max_bs);
1851 else if (!td->o.zero_buffers) {
1852 unsigned int perc = td->o.compress_percentage;
1853 struct frand_state *rs;
1854 unsigned int left = max_bs;
1855
1856 do {
1857 rs = get_buf_state(td);
1858
1859 min_write = min(min_write, left);
1860
1861 if (perc) {
1862 unsigned int seg = min_write;
1863
1864 seg = min(min_write, td->o.compress_chunk);
1865 if (!seg)
1866 seg = min_write;
1867
1868 fill_random_buf_percentage(rs, buf, perc, seg,
1869 min_write);
1870 } else
1871 fill_random_buf(rs, buf, min_write);
1872
1873 buf += min_write;
1874 left -= min_write;
1875 save_buf_state(td, rs);
1876 } while (left);
1877 } else
1878 memset(buf, 0, max_bs);
1879}
1880
1881/*
1882 * "randomly" fill the buffer contents
1883 */
1884void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1885 unsigned int min_write, unsigned int max_bs)
1886{
1887 io_u->buf_filled_len = 0;
1888 fill_io_buffer(td, io_u->buf, min_write, max_bs);
1889}
fio - Flexible IO Tester