projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Start of ACT like benchmark profile
[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
15struct io_completion_data {
16 int nr; /* input */
17
18 int error; /* output */
19 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
20 struct timeval time; /* output */
21};
22
23/*
24 * The ->io_axmap contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
26 */
27static int random_map_free(struct fio_file *f, const uint64_t block)
28{
29 return !axmap_isset(f->io_axmap, block);
30}
31
32/*
33 * Mark a given offset as used in the map.
34 */
35static void mark_random_map(struct thread_data *td, struct io_u *io_u)
36{
37 unsigned int min_bs = td->o.rw_min_bs;
38 struct fio_file *f = io_u->file;
39 unsigned int nr_blocks;
40 uint64_t block;
41
42 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
43 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
44
45 if (!(io_u->flags & IO_U_F_BUSY_OK))
46 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
47
48 if ((nr_blocks * min_bs) < io_u->buflen)
49 io_u->buflen = nr_blocks * min_bs;
50}
51
52static uint64_t last_block(struct thread_data *td, struct fio_file *f,
53 enum fio_ddir ddir)
54{
55 uint64_t max_blocks;
56 uint64_t max_size;
57
58 assert(ddir_rw(ddir));
59
60 /*
61 * Hmm, should we make sure that ->io_size <= ->real_file_size?
62 */
63 max_size = f->io_size;
64 if (max_size > f->real_file_size)
65 max_size = f->real_file_size;
66
67 if (td->o.zone_range)
68 max_size = td->o.zone_range;
69
70 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
71 if (!max_blocks)
72 return 0;
73
74 return max_blocks;
75}
76
77struct rand_off {
78 struct flist_head list;
79 uint64_t off;
80};
81
82static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
83 enum fio_ddir ddir, uint64_t *b)
84{
85 uint64_t r, lastb;
86
87 lastb = last_block(td, f, ddir);
88 if (!lastb)
89 return 1;
90
91 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
92 uint64_t rmax;
93
94 rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
95
96 if (td->o.use_os_rand) {
97 rmax = OS_RAND_MAX;
98 r = os_random_long(&td->random_state);
99 } else {
100 rmax = FRAND_MAX;
101 r = __rand(&td->__random_state);
102 }
103
104 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
105
106 *b = (lastb - 1) * (r / ((uint64_t) rmax + 1.0));
107 } else {
108 uint64_t off = 0;
109
110 if (lfsr_next(&f->lfsr, &off, lastb))
111 return 1;
112
113 *b = off;
114 }
115
116 /*
117 * if we are not maintaining a random map, we are done.
118 */
119 if (!file_randommap(td, f))
120 goto ret;
121
122 /*
123 * calculate map offset and check if it's free
124 */
125 if (random_map_free(f, *b))
126 goto ret;
127
128 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
129 (unsigned long long) *b);
130
131 *b = axmap_next_free(f->io_axmap, *b);
132 if (*b == (uint64_t) -1ULL)
133 return 1;
134ret:
135 return 0;
136}
137
138static int __get_next_rand_offset_zipf(struct thread_data *td,
139 struct fio_file *f, enum fio_ddir ddir,
140 uint64_t *b)
141{
142 *b = zipf_next(&f->zipf);
143 return 0;
144}
145
146static int __get_next_rand_offset_pareto(struct thread_data *td,
147 struct fio_file *f, enum fio_ddir ddir,
148 uint64_t *b)
149{
150 *b = pareto_next(&f->zipf);
151 return 0;
152}
153
154static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
155{
156 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
157 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
158
159 return r1->off - r2->off;
160}
161
162static int get_off_from_method(struct thread_data *td, struct fio_file *f,
163 enum fio_ddir ddir, uint64_t *b)
164{
165 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
166 return __get_next_rand_offset(td, f, ddir, b);
167 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
168 return __get_next_rand_offset_zipf(td, f, ddir, b);
169 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
170 return __get_next_rand_offset_pareto(td, f, ddir, b);
171
172 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
173 return 1;
174}
175
176/*
177 * Sort the reads for a verify phase in batches of verifysort_nr, if
178 * specified.
179 */
180static inline int should_sort_io(struct thread_data *td)
181{
182 if (!td->o.verifysort_nr || !td->o.do_verify)
183 return 0;
184 if (!td_random(td))
185 return 0;
186 if (td->runstate != TD_VERIFYING)
187 return 0;
188 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
189 return 0;
190
191 return 1;
192}
193
194static int should_do_random(struct thread_data *td)
195{
196 unsigned int v;
197 unsigned long r;
198
199 if (td->o.perc_rand == 100)
200 return 1;
201
202 if (td->o.use_os_rand) {
203 r = os_random_long(&td->seq_rand_state);
204 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
205 } else {
206 r = __rand(&td->__seq_rand_state);
207 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
208 }
209
210 return v <= td->o.perc_rand;
211}
212
213static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
214 enum fio_ddir ddir, uint64_t *b)
215{
216 struct rand_off *r;
217 int i, ret = 1;
218
219 if (!should_sort_io(td))
220 return get_off_from_method(td, f, ddir, b);
221
222 if (!flist_empty(&td->next_rand_list)) {
223 struct rand_off *r;
224fetch:
225 r = flist_entry(td->next_rand_list.next, struct rand_off, list);
226 flist_del(&r->list);
227 *b = r->off;
228 free(r);
229 return 0;
230 }
231
232 for (i = 0; i < td->o.verifysort_nr; i++) {
233 r = malloc(sizeof(*r));
234
235 ret = get_off_from_method(td, f, ddir, &r->off);
236 if (ret) {
237 free(r);
238 break;
239 }
240
241 flist_add(&r->list, &td->next_rand_list);
242 }
243
244 if (ret && !i)
245 return ret;
246
247 assert(!flist_empty(&td->next_rand_list));
248 flist_sort(NULL, &td->next_rand_list, flist_cmp);
249 goto fetch;
250}
251
252static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
253 enum fio_ddir ddir, uint64_t *b)
254{
255 if (!get_next_rand_offset(td, f, ddir, b))
256 return 0;
257
258 if (td->o.time_based) {
259 fio_file_reset(td, f);
260 if (!get_next_rand_offset(td, f, ddir, b))
261 return 0;
262 }
263
264 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
265 f->file_name, (unsigned long long) f->last_pos,
266 (unsigned long long) f->real_file_size);
267 return 1;
268}
269
270static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
271 enum fio_ddir ddir, uint64_t *offset)
272{
273 assert(ddir_rw(ddir));
274
275 if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
276 f->last_pos = f->last_pos - f->io_size;
277
278 if (f->last_pos < f->real_file_size) {
279 uint64_t pos;
280
281 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
282 f->last_pos = f->real_file_size;
283
284 pos = f->last_pos - f->file_offset;
285 if (pos)
286 pos += td->o.ddir_seq_add;
287
288 *offset = pos;
289 return 0;
290 }
291
292 return 1;
293}
294
295static int get_next_block(struct thread_data *td, struct io_u *io_u,
296 enum fio_ddir ddir, int rw_seq)
297{
298 struct fio_file *f = io_u->file;
299 uint64_t b, offset;
300 int ret;
301
302 assert(ddir_rw(ddir));
303
304 b = offset = -1ULL;
305
306 if (rw_seq) {
307 if (td_random(td)) {
308 if (should_do_random(td))
309 ret = get_next_rand_block(td, f, ddir, &b);
310 else {
311 io_u->flags |= IO_U_F_BUSY_OK;
312 ret = get_next_seq_offset(td, f, ddir, &offset);
313 if (ret)
314 ret = get_next_rand_block(td, f, ddir, &b);
315 }
316 } else
317 ret = get_next_seq_offset(td, f, ddir, &offset);
318 } else {
319 io_u->flags |= IO_U_F_BUSY_OK;
320
321 if (td->o.rw_seq == RW_SEQ_SEQ) {
322 ret = get_next_seq_offset(td, f, ddir, &offset);
323 if (ret)
324 ret = get_next_rand_block(td, f, ddir, &b);
325 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
326 if (f->last_start != -1ULL)
327 offset = f->last_start - f->file_offset;
328 else
329 offset = 0;
330 ret = 0;
331 } else {
332 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
333 ret = 1;
334 }
335 }
336
337 if (!ret) {
338 if (offset != -1ULL)
339 io_u->offset = offset;
340 else if (b != -1ULL)
341 io_u->offset = b * td->o.ba[ddir];
342 else {
343 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
344 ret = 1;
345 }
346 }
347
348 return ret;
349}
350
351/*
352 * For random io, generate a random new block and see if it's used. Repeat
353 * until we find a free one. For sequential io, just return the end of
354 * the last io issued.
355 */
356static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
357{
358 struct fio_file *f = io_u->file;
359 enum fio_ddir ddir = io_u->ddir;
360 int rw_seq_hit = 0;
361
362 assert(ddir_rw(ddir));
363
364 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
365 rw_seq_hit = 1;
366 td->ddir_seq_nr = td->o.ddir_seq_nr;
367 }
368
369 if (get_next_block(td, io_u, ddir, rw_seq_hit))
370 return 1;
371
372 if (io_u->offset >= f->io_size) {
373 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
374 (unsigned long long) io_u->offset,
375 (unsigned long long) f->io_size);
376 return 1;
377 }
378
379 io_u->offset += f->file_offset;
380 if (io_u->offset >= f->real_file_size) {
381 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
382 (unsigned long long) io_u->offset,
383 (unsigned long long) f->real_file_size);
384 return 1;
385 }
386
387 return 0;
388}
389
390static int get_next_offset(struct thread_data *td, struct io_u *io_u)
391{
392 if (td->flags & TD_F_PROFILE_OPS) {
393 struct prof_io_ops *ops = &td->prof_io_ops;
394
395 if (ops->fill_io_u_off)
396 return ops->fill_io_u_off(td, io_u);
397 }
398
399 return __get_next_offset(td, io_u);
400}
401
402static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
403 unsigned int buflen)
404{
405 struct fio_file *f = io_u->file;
406
407 return io_u->offset + buflen <= f->io_size + get_start_offset(td);
408}
409
410static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
411{
412 const int ddir = io_u->ddir;
413 unsigned int buflen = 0;
414 unsigned int minbs, maxbs;
415 unsigned long r, rand_max;
416
417 assert(ddir_rw(ddir));
418
419 minbs = td->o.min_bs[ddir];
420 maxbs = td->o.max_bs[ddir];
421
422 if (minbs == maxbs)
423 return minbs;
424
425 /*
426 * If we can't satisfy the min block size from here, then fail
427 */
428 if (!io_u_fits(td, io_u, minbs))
429 return 0;
430
431 if (td->o.use_os_rand)
432 rand_max = OS_RAND_MAX;
433 else
434 rand_max = FRAND_MAX;
435
436 do {
437 if (td->o.use_os_rand)
438 r = os_random_long(&td->bsrange_state);
439 else
440 r = __rand(&td->__bsrange_state);
441
442 if (!td->o.bssplit_nr[ddir]) {
443 buflen = 1 + (unsigned int) ((double) maxbs *
444 (r / (rand_max + 1.0)));
445 if (buflen < minbs)
446 buflen = minbs;
447 } else {
448 long perc = 0;
449 unsigned int i;
450
451 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
452 struct bssplit *bsp = &td->o.bssplit[ddir][i];
453
454 buflen = bsp->bs;
455 perc += bsp->perc;
456 if ((r <= ((rand_max / 100L) * perc)) &&
457 io_u_fits(td, io_u, buflen))
458 break;
459 }
460 }
461
462 if (!td->o.bs_unaligned && is_power_of_2(minbs))
463 buflen = (buflen + minbs - 1) & ~(minbs - 1);
464
465 } while (!io_u_fits(td, io_u, buflen));
466
467 return buflen;
468}
469
470static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
471{
472 if (td->flags & TD_F_PROFILE_OPS) {
473 struct prof_io_ops *ops = &td->prof_io_ops;
474
475 if (ops->fill_io_u_size)
476 return ops->fill_io_u_size(td, io_u);
477 }
478
479 return __get_next_buflen(td, io_u);
480}
481
482static void set_rwmix_bytes(struct thread_data *td)
483{
484 unsigned int diff;
485
486 /*
487 * we do time or byte based switch. this is needed because
488 * buffered writes may issue a lot quicker than they complete,
489 * whereas reads do not.
490 */
491 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
492 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
493}
494
495static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
496{
497 unsigned int v;
498 unsigned long r;
499
500 if (td->o.use_os_rand) {
501 r = os_random_long(&td->rwmix_state);
502 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
503 } else {
504 r = __rand(&td->__rwmix_state);
505 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
506 }
507
508 if (v <= td->o.rwmix[DDIR_READ])
509 return DDIR_READ;
510
511 return DDIR_WRITE;
512}
513
514void io_u_quiesce(struct thread_data *td)
515{
516 /*
517 * We are going to sleep, ensure that we flush anything pending as
518 * not to skew our latency numbers.
519 *
520 * Changed to only monitor 'in flight' requests here instead of the
521 * td->cur_depth, b/c td->cur_depth does not accurately represent
522 * io's that have been actually submitted to an async engine,
523 * and cur_depth is meaningless for sync engines.
524 */
525 while (td->io_u_in_flight) {
526 int fio_unused ret;
527
528 ret = io_u_queued_complete(td, 1, NULL);
529 }
530}
531
532static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
533{
534 enum fio_ddir odir = ddir ^ 1;
535 struct timeval t;
536 long usec;
537
538 assert(ddir_rw(ddir));
539
540 if (td->rate_pending_usleep[ddir] <= 0)
541 return ddir;
542
543 /*
544 * We have too much pending sleep in this direction. See if we
545 * should switch.
546 */
547 if (td_rw(td) && td->o.rwmix[odir]) {
548 /*
549 * Other direction does not have too much pending, switch
550 */
551 if (td->rate_pending_usleep[odir] < 100000)
552 return odir;
553
554 /*
555 * Both directions have pending sleep. Sleep the minimum time
556 * and deduct from both.
557 */
558 if (td->rate_pending_usleep[ddir] <=
559 td->rate_pending_usleep[odir]) {
560 usec = td->rate_pending_usleep[ddir];
561 } else {
562 usec = td->rate_pending_usleep[odir];
563 ddir = odir;
564 }
565 } else
566 usec = td->rate_pending_usleep[ddir];
567
568 io_u_quiesce(td);
569
570 fio_gettime(&t, NULL);
571 usec_sleep(td, usec);
572 usec = utime_since_now(&t);
573
574 td->rate_pending_usleep[ddir] -= usec;
575
576 odir = ddir ^ 1;
577 if (td_rw(td) && __should_check_rate(td, odir))
578 td->rate_pending_usleep[odir] -= usec;
579
580 if (ddir_trim(ddir))
581 return ddir;
582
583 return ddir;
584}
585
586/*
587 * Return the data direction for the next io_u. If the job is a
588 * mixed read/write workload, check the rwmix cycle and switch if
589 * necessary.
590 */
591static enum fio_ddir get_rw_ddir(struct thread_data *td)
592{
593 enum fio_ddir ddir;
594
595 /*
596 * see if it's time to fsync
597 */
598 if (td->o.fsync_blocks &&
599 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
600 td->io_issues[DDIR_WRITE] && should_fsync(td))
601 return DDIR_SYNC;
602
603 /*
604 * see if it's time to fdatasync
605 */
606 if (td->o.fdatasync_blocks &&
607 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
608 td->io_issues[DDIR_WRITE] && should_fsync(td))
609 return DDIR_DATASYNC;
610
611 /*
612 * see if it's time to sync_file_range
613 */
614 if (td->sync_file_range_nr &&
615 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
616 td->io_issues[DDIR_WRITE] && should_fsync(td))
617 return DDIR_SYNC_FILE_RANGE;
618
619 if (td_rw(td)) {
620 /*
621 * Check if it's time to seed a new data direction.
622 */
623 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
624 /*
625 * Put a top limit on how many bytes we do for
626 * one data direction, to avoid overflowing the
627 * ranges too much
628 */
629 ddir = get_rand_ddir(td);
630
631 if (ddir != td->rwmix_ddir)
632 set_rwmix_bytes(td);
633
634 td->rwmix_ddir = ddir;
635 }
636 ddir = td->rwmix_ddir;
637 } else if (td_read(td))
638 ddir = DDIR_READ;
639 else if (td_write(td))
640 ddir = DDIR_WRITE;
641 else
642 ddir = DDIR_TRIM;
643
644 td->rwmix_ddir = rate_ddir(td, ddir);
645 return td->rwmix_ddir;
646}
647
648static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
649{
650 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
651
652 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
653 td->o.barrier_blocks &&
654 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
655 td->io_issues[DDIR_WRITE])
656 io_u->flags |= IO_U_F_BARRIER;
657}
658
659void put_file_log(struct thread_data *td, struct fio_file *f)
660{
661 int ret = put_file(td, f);
662
663 if (ret)
664 td_verror(td, ret, "file close");
665}
666
667void put_io_u(struct thread_data *td, struct io_u *io_u)
668{
669 td_io_u_lock(td);
670
671 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
672 put_file_log(td, io_u->file);
673 io_u->file = NULL;
674 io_u->flags &= ~IO_U_F_FREE_DEF;
675 io_u->flags |= IO_U_F_FREE;
676
677 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
678 td->cur_depth--;
679 flist_del_init(&io_u->list);
680 flist_add(&io_u->list, &td->io_u_freelist);
681 td_io_u_unlock(td);
682 td_io_u_free_notify(td);
683}
684
685void clear_io_u(struct thread_data *td, struct io_u *io_u)
686{
687 io_u->flags &= ~IO_U_F_FLIGHT;
688 put_io_u(td, io_u);
689}
690
691void requeue_io_u(struct thread_data *td, struct io_u **io_u)
692{
693 struct io_u *__io_u = *io_u;
694 enum fio_ddir ddir = acct_ddir(__io_u);
695
696 dprint(FD_IO, "requeue %p\n", __io_u);
697
698 td_io_u_lock(td);
699
700 __io_u->flags |= IO_U_F_FREE;
701 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
702 td->io_issues[ddir]--;
703
704 __io_u->flags &= ~IO_U_F_FLIGHT;
705 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
706 td->cur_depth--;
707 flist_del(&__io_u->list);
708 flist_add_tail(&__io_u->list, &td->io_u_requeues);
709 td_io_u_unlock(td);
710 *io_u = NULL;
711}
712
713static int fill_io_u(struct thread_data *td, struct io_u *io_u)
714{
715 if (td->io_ops->flags & FIO_NOIO)
716 goto out;
717
718 set_rw_ddir(td, io_u);
719
720 /*
721 * fsync() or fdatasync() or trim etc, we are done
722 */
723 if (!ddir_rw(io_u->ddir))
724 goto out;
725
726 /*
727 * See if it's time to switch to a new zone
728 */
729 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
730 td->zone_bytes = 0;
731 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
732 io_u->file->last_pos = io_u->file->file_offset;
733 td->io_skip_bytes += td->o.zone_skip;
734 }
735
736 /*
737 * No log, let the seq/rand engine retrieve the next buflen and
738 * position.
739 */
740 if (get_next_offset(td, io_u)) {
741 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
742 return 1;
743 }
744
745 io_u->buflen = get_next_buflen(td, io_u);
746 if (!io_u->buflen) {
747 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
748 return 1;
749 }
750
751 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
752 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
753 dprint(FD_IO, " off=%llu/%lu > %llu\n",
754 (unsigned long long) io_u->offset, io_u->buflen,
755 (unsigned long long) io_u->file->real_file_size);
756 return 1;
757 }
758
759 /*
760 * mark entry before potentially trimming io_u
761 */
762 if (td_random(td) && file_randommap(td, io_u->file))
763 mark_random_map(td, io_u);
764
765out:
766 dprint_io_u(io_u, "fill_io_u");
767 td->zone_bytes += io_u->buflen;
768 return 0;
769}
770
771static void __io_u_mark_map(unsigned int *map, unsigned int nr)
772{
773 int idx = 0;
774
775 switch (nr) {
776 default:
777 idx = 6;
778 break;
779 case 33 ... 64:
780 idx = 5;
781 break;
782 case 17 ... 32:
783 idx = 4;
784 break;
785 case 9 ... 16:
786 idx = 3;
787 break;
788 case 5 ... 8:
789 idx = 2;
790 break;
791 case 1 ... 4:
792 idx = 1;
793 case 0:
794 break;
795 }
796
797 map[idx]++;
798}
799
800void io_u_mark_submit(struct thread_data *td, unsigned int nr)
801{
802 __io_u_mark_map(td->ts.io_u_submit, nr);
803 td->ts.total_submit++;
804}
805
806void io_u_mark_complete(struct thread_data *td, unsigned int nr)
807{
808 __io_u_mark_map(td->ts.io_u_complete, nr);
809 td->ts.total_complete++;
810}
811
812void io_u_mark_depth(struct thread_data *td, unsigned int nr)
813{
814 int idx = 0;
815
816 switch (td->cur_depth) {
817 default:
818 idx = 6;
819 break;
820 case 32 ... 63:
821 idx = 5;
822 break;
823 case 16 ... 31:
824 idx = 4;
825 break;
826 case 8 ... 15:
827 idx = 3;
828 break;
829 case 4 ... 7:
830 idx = 2;
831 break;
832 case 2 ... 3:
833 idx = 1;
834 case 1:
835 break;
836 }
837
838 td->ts.io_u_map[idx] += nr;
839}
840
841static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
842{
843 int idx = 0;
844
845 assert(usec < 1000);
846
847 switch (usec) {
848 case 750 ... 999:
849 idx = 9;
850 break;
851 case 500 ... 749:
852 idx = 8;
853 break;
854 case 250 ... 499:
855 idx = 7;
856 break;
857 case 100 ... 249:
858 idx = 6;
859 break;
860 case 50 ... 99:
861 idx = 5;
862 break;
863 case 20 ... 49:
864 idx = 4;
865 break;
866 case 10 ... 19:
867 idx = 3;
868 break;
869 case 4 ... 9:
870 idx = 2;
871 break;
872 case 2 ... 3:
873 idx = 1;
874 case 0 ... 1:
875 break;
876 }
877
878 assert(idx < FIO_IO_U_LAT_U_NR);
879 td->ts.io_u_lat_u[idx]++;
880}
881
882static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
883{
884 int idx = 0;
885
886 switch (msec) {
887 default:
888 idx = 11;
889 break;
890 case 1000 ... 1999:
891 idx = 10;
892 break;
893 case 750 ... 999:
894 idx = 9;
895 break;
896 case 500 ... 749:
897 idx = 8;
898 break;
899 case 250 ... 499:
900 idx = 7;
901 break;
902 case 100 ... 249:
903 idx = 6;
904 break;
905 case 50 ... 99:
906 idx = 5;
907 break;
908 case 20 ... 49:
909 idx = 4;
910 break;
911 case 10 ... 19:
912 idx = 3;
913 break;
914 case 4 ... 9:
915 idx = 2;
916 break;
917 case 2 ... 3:
918 idx = 1;
919 case 0 ... 1:
920 break;
921 }
922
923 assert(idx < FIO_IO_U_LAT_M_NR);
924 td->ts.io_u_lat_m[idx]++;
925}
926
927static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
928{
929 if (usec < 1000)
930 io_u_mark_lat_usec(td, usec);
931 else
932 io_u_mark_lat_msec(td, usec / 1000);
933}
934
935/*
936 * Get next file to service by choosing one at random
937 */
938static struct fio_file *get_next_file_rand(struct thread_data *td,
939 enum fio_file_flags goodf,
940 enum fio_file_flags badf)
941{
942 struct fio_file *f;
943 int fno;
944
945 do {
946 int opened = 0;
947 unsigned long r;
948
949 if (td->o.use_os_rand) {
950 r = os_random_long(&td->next_file_state);
951 fno = (unsigned int) ((double) td->o.nr_files
952 * (r / (OS_RAND_MAX + 1.0)));
953 } else {
954 r = __rand(&td->__next_file_state);
955 fno = (unsigned int) ((double) td->o.nr_files
956 * (r / (FRAND_MAX + 1.0)));
957 }
958
959 f = td->files[fno];
960 if (fio_file_done(f))
961 continue;
962
963 if (!fio_file_open(f)) {
964 int err;
965
966 err = td_io_open_file(td, f);
967 if (err)
968 continue;
969 opened = 1;
970 }
971
972 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
973 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
974 return f;
975 }
976 if (opened)
977 td_io_close_file(td, f);
978 } while (1);
979}
980
981/*
982 * Get next file to service by doing round robin between all available ones
983 */
984static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
985 int badf)
986{
987 unsigned int old_next_file = td->next_file;
988 struct fio_file *f;
989
990 do {
991 int opened = 0;
992
993 f = td->files[td->next_file];
994
995 td->next_file++;
996 if (td->next_file >= td->o.nr_files)
997 td->next_file = 0;
998
999 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1000 if (fio_file_done(f)) {
1001 f = NULL;
1002 continue;
1003 }
1004
1005 if (!fio_file_open(f)) {
1006 int err;
1007
1008 err = td_io_open_file(td, f);
1009 if (err) {
1010 dprint(FD_FILE, "error %d on open of %s\n",
1011 err, f->file_name);
1012 f = NULL;
1013 continue;
1014 }
1015 opened = 1;
1016 }
1017
1018 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1019 f->flags);
1020 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1021 break;
1022
1023 if (opened)
1024 td_io_close_file(td, f);
1025
1026 f = NULL;
1027 } while (td->next_file != old_next_file);
1028
1029 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1030 return f;
1031}
1032
1033static struct fio_file *__get_next_file(struct thread_data *td)
1034{
1035 struct fio_file *f;
1036
1037 assert(td->o.nr_files <= td->files_index);
1038
1039 if (td->nr_done_files >= td->o.nr_files) {
1040 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1041 " nr_files=%d\n", td->nr_open_files,
1042 td->nr_done_files,
1043 td->o.nr_files);
1044 return NULL;
1045 }
1046
1047 f = td->file_service_file;
1048 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1049 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1050 goto out;
1051 if (td->file_service_left--)
1052 goto out;
1053 }
1054
1055 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1056 td->o.file_service_type == FIO_FSERVICE_SEQ)
1057 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1058 else
1059 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1060
1061 td->file_service_file = f;
1062 td->file_service_left = td->file_service_nr - 1;
1063out:
1064 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1065 return f;
1066}
1067
1068static struct fio_file *get_next_file(struct thread_data *td)
1069{
1070 if (!(td->flags & TD_F_PROFILE_OPS)) {
1071 struct prof_io_ops *ops = &td->prof_io_ops;
1072
1073 if (ops->get_next_file)
1074 return ops->get_next_file(td);
1075 }
1076
1077 return __get_next_file(td);
1078}
1079
1080static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1081{
1082 struct fio_file *f;
1083
1084 do {
1085 f = get_next_file(td);
1086 if (!f)
1087 return 1;
1088
1089 io_u->file = f;
1090 get_file(f);
1091
1092 if (!fill_io_u(td, io_u))
1093 break;
1094
1095 put_file_log(td, f);
1096 td_io_close_file(td, f);
1097 io_u->file = NULL;
1098 fio_file_set_done(f);
1099 td->nr_done_files++;
1100 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1101 td->nr_done_files, td->o.nr_files);
1102 } while (1);
1103
1104 return 0;
1105}
1106
1107
1108struct io_u *__get_io_u(struct thread_data *td)
1109{
1110 struct io_u *io_u = NULL;
1111
1112 td_io_u_lock(td);
1113
1114again:
1115 if (!flist_empty(&td->io_u_requeues))
1116 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1117 else if (!queue_full(td)) {
1118 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1119
1120 io_u->buflen = 0;
1121 io_u->resid = 0;
1122 io_u->file = NULL;
1123 io_u->end_io = NULL;
1124 }
1125
1126 if (io_u) {
1127 assert(io_u->flags & IO_U_F_FREE);
1128 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1129 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1130 io_u->flags &= ~IO_U_F_VER_LIST;
1131
1132 io_u->error = 0;
1133 io_u->acct_ddir = -1;
1134 flist_del(&io_u->list);
1135 flist_add_tail(&io_u->list, &td->io_u_busylist);
1136 td->cur_depth++;
1137 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1138 } else if (td->o.verify_async) {
1139 /*
1140 * We ran out, wait for async verify threads to finish and
1141 * return one
1142 */
1143 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1144 goto again;
1145 }
1146
1147 td_io_u_unlock(td);
1148 return io_u;
1149}
1150
1151static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1152{
1153 if (!(td->flags & TD_F_TRIM_BACKLOG))
1154 return 0;
1155
1156 if (td->trim_entries) {
1157 int get_trim = 0;
1158
1159 if (td->trim_batch) {
1160 td->trim_batch--;
1161 get_trim = 1;
1162 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1163 td->last_ddir != DDIR_READ) {
1164 td->trim_batch = td->o.trim_batch;
1165 if (!td->trim_batch)
1166 td->trim_batch = td->o.trim_backlog;
1167 get_trim = 1;
1168 }
1169
1170 if (get_trim && !get_next_trim(td, io_u))
1171 return 1;
1172 }
1173
1174 return 0;
1175}
1176
1177static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1178{
1179 if (!(td->flags & TD_F_VER_BACKLOG))
1180 return 0;
1181
1182 if (td->io_hist_len) {
1183 int get_verify = 0;
1184
1185 if (td->verify_batch)
1186 get_verify = 1;
1187 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1188 td->last_ddir != DDIR_READ) {
1189 td->verify_batch = td->o.verify_batch;
1190 if (!td->verify_batch)
1191 td->verify_batch = td->o.verify_backlog;
1192 get_verify = 1;
1193 }
1194
1195 if (get_verify && !get_next_verify(td, io_u)) {
1196 td->verify_batch--;
1197 return 1;
1198 }
1199 }
1200
1201 return 0;
1202}
1203
1204/*
1205 * Fill offset and start time into the buffer content, to prevent too
1206 * easy compressible data for simple de-dupe attempts. Do this for every
1207 * 512b block in the range, since that should be the smallest block size
1208 * we can expect from a device.
1209 */
1210static void small_content_scramble(struct io_u *io_u)
1211{
1212 unsigned int i, nr_blocks = io_u->buflen / 512;
1213 uint64_t boffset;
1214 unsigned int offset;
1215 void *p, *end;
1216
1217 if (!nr_blocks)
1218 return;
1219
1220 p = io_u->xfer_buf;
1221 boffset = io_u->offset;
1222 io_u->buf_filled_len = 0;
1223
1224 for (i = 0; i < nr_blocks; i++) {
1225 /*
1226 * Fill the byte offset into a "random" start offset of
1227 * the buffer, given by the product of the usec time
1228 * and the actual offset.
1229 */
1230 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1231 offset &= ~(sizeof(uint64_t) - 1);
1232 if (offset >= 512 - sizeof(uint64_t))
1233 offset -= sizeof(uint64_t);
1234 memcpy(p + offset, &boffset, sizeof(boffset));
1235
1236 end = p + 512 - sizeof(io_u->start_time);
1237 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1238 p += 512;
1239 boffset += 512;
1240 }
1241}
1242
1243/*
1244 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1245 * etc. The returned io_u is fully ready to be prepped and submitted.
1246 */
1247struct io_u *get_io_u(struct thread_data *td)
1248{
1249 struct fio_file *f;
1250 struct io_u *io_u;
1251 int do_scramble = 0;
1252
1253 io_u = __get_io_u(td);
1254 if (!io_u) {
1255 dprint(FD_IO, "__get_io_u failed\n");
1256 return NULL;
1257 }
1258
1259 if (check_get_verify(td, io_u))
1260 goto out;
1261 if (check_get_trim(td, io_u))
1262 goto out;
1263
1264 /*
1265 * from a requeue, io_u already setup
1266 */
1267 if (io_u->file)
1268 goto out;
1269
1270 /*
1271 * If using an iolog, grab next piece if any available.
1272 */
1273 if (td->flags & TD_F_READ_IOLOG) {
1274 if (read_iolog_get(td, io_u))
1275 goto err_put;
1276 } else if (set_io_u_file(td, io_u)) {
1277 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1278 goto err_put;
1279 }
1280
1281 f = io_u->file;
1282 assert(fio_file_open(f));
1283
1284 if (ddir_rw(io_u->ddir)) {
1285 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1286 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1287 goto err_put;
1288 }
1289
1290 f->last_start = io_u->offset;
1291 f->last_pos = io_u->offset + io_u->buflen;
1292
1293 if (io_u->ddir == DDIR_WRITE) {
1294 if (td->flags & TD_F_REFILL_BUFFERS) {
1295 io_u_fill_buffer(td, io_u,
1296 io_u->xfer_buflen, io_u->xfer_buflen);
1297 } else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
1298 do_scramble = 1;
1299 if (td->flags & TD_F_VER_NONE) {
1300 populate_verify_io_u(td, io_u);
1301 do_scramble = 0;
1302 }
1303 } else if (io_u->ddir == DDIR_READ) {
1304 /*
1305 * Reset the buf_filled parameters so next time if the
1306 * buffer is used for writes it is refilled.
1307 */
1308 io_u->buf_filled_len = 0;
1309 }
1310 }
1311
1312 /*
1313 * Set io data pointers.
1314 */
1315 io_u->xfer_buf = io_u->buf;
1316 io_u->xfer_buflen = io_u->buflen;
1317
1318out:
1319 assert(io_u->file);
1320 if (!td_io_prep(td, io_u)) {
1321 if (!td->o.disable_slat)
1322 fio_gettime(&io_u->start_time, NULL);
1323 if (do_scramble)
1324 small_content_scramble(io_u);
1325 return io_u;
1326 }
1327err_put:
1328 dprint(FD_IO, "get_io_u failed\n");
1329 put_io_u(td, io_u);
1330 return NULL;
1331}
1332
1333void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1334{
1335 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1336 const char *msg[] = { "read", "write", "sync", "datasync",
1337 "sync_file_range", "wait", "trim" };
1338
1339 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1340 return;
1341
1342 log_err("fio: io_u error");
1343
1344 if (io_u->file)
1345 log_err(" on file %s", io_u->file->file_name);
1346
1347 log_err(": %s\n", strerror(io_u->error));
1348
1349 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1350 io_u->offset, io_u->xfer_buflen);
1351
1352 if (!td->error)
1353 td_verror(td, io_u->error, "io_u error");
1354}
1355
1356static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1357 struct io_completion_data *icd,
1358 const enum fio_ddir idx, unsigned int bytes)
1359{
1360 unsigned long lusec = 0;
1361
1362 if (!td->o.disable_clat || !td->o.disable_bw)
1363 lusec = utime_since(&io_u->issue_time, &icd->time);
1364
1365 if (!td->o.disable_lat) {
1366 unsigned long tusec;
1367
1368 tusec = utime_since(&io_u->start_time, &icd->time);
1369 add_lat_sample(td, idx, tusec, bytes);
1370
1371 if (td->flags & TD_F_PROFILE_OPS) {
1372 struct prof_io_ops *ops = &td->prof_io_ops;
1373
1374 if (ops->io_u_lat)
1375 icd->error = ops->io_u_lat(td, tusec);
1376 }
1377
1378 if (td->o.max_latency && tusec > td->o.max_latency) {
1379 if (!td->error)
1380 log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
1381 td_verror(td, ETIMEDOUT, "max latency exceeded");
1382 icd->error = ETIMEDOUT;
1383 }
1384 }
1385
1386 if (!td->o.disable_clat) {
1387 add_clat_sample(td, idx, lusec, bytes);
1388 io_u_mark_latency(td, lusec);
1389 }
1390
1391 if (!td->o.disable_bw)
1392 add_bw_sample(td, idx, bytes, &icd->time);
1393
1394 add_iops_sample(td, idx, &icd->time);
1395}
1396
1397static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1398{
1399 uint64_t secs, remainder, bps, bytes;
1400
1401 bytes = td->this_io_bytes[ddir];
1402 bps = td->rate_bps[ddir];
1403 secs = bytes / bps;
1404 remainder = bytes % bps;
1405 return remainder * 1000000 / bps + secs * 1000000;
1406}
1407
1408static void io_completed(struct thread_data *td, struct io_u *io_u,
1409 struct io_completion_data *icd)
1410{
1411 struct fio_file *f;
1412
1413 dprint_io_u(io_u, "io complete");
1414
1415 td_io_u_lock(td);
1416 assert(io_u->flags & IO_U_F_FLIGHT);
1417 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1418 td_io_u_unlock(td);
1419
1420 if (ddir_sync(io_u->ddir)) {
1421 td->last_was_sync = 1;
1422 f = io_u->file;
1423 if (f) {
1424 f->first_write = -1ULL;
1425 f->last_write = -1ULL;
1426 }
1427 return;
1428 }
1429
1430 td->last_was_sync = 0;
1431 td->last_ddir = io_u->ddir;
1432
1433 if (!io_u->error && ddir_rw(io_u->ddir)) {
1434 unsigned int bytes = io_u->buflen - io_u->resid;
1435 const enum fio_ddir idx = io_u->ddir;
1436 const enum fio_ddir odx = io_u->ddir ^ 1;
1437 int ret;
1438
1439 td->io_blocks[idx]++;
1440 td->this_io_blocks[idx]++;
1441 td->io_bytes[idx] += bytes;
1442
1443 if (!(io_u->flags & IO_U_F_VER_LIST))
1444 td->this_io_bytes[idx] += bytes;
1445
1446 if (idx == DDIR_WRITE) {
1447 f = io_u->file;
1448 if (f) {
1449 if (f->first_write == -1ULL ||
1450 io_u->offset < f->first_write)
1451 f->first_write = io_u->offset;
1452 if (f->last_write == -1ULL ||
1453 ((io_u->offset + bytes) > f->last_write))
1454 f->last_write = io_u->offset + bytes;
1455 }
1456 }
1457
1458 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1459 td->runstate == TD_VERIFYING)) {
1460 account_io_completion(td, io_u, icd, idx, bytes);
1461
1462 if (__should_check_rate(td, idx)) {
1463 td->rate_pending_usleep[idx] =
1464 (usec_for_io(td, idx) -
1465 utime_since_now(&td->start));
1466 }
1467 if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1468 td->rate_pending_usleep[odx] =
1469 (usec_for_io(td, odx) -
1470 utime_since_now(&td->start));
1471 }
1472
1473 if (td_write(td) && idx == DDIR_WRITE &&
1474 td->o.do_verify &&
1475 td->o.verify != VERIFY_NONE &&
1476 !td->o.experimental_verify)
1477 log_io_piece(td, io_u);
1478
1479 icd->bytes_done[idx] += bytes;
1480
1481 if (io_u->end_io) {
1482 ret = io_u->end_io(td, io_u);
1483 if (ret && !icd->error)
1484 icd->error = ret;
1485 }
1486 } else if (io_u->error) {
1487 icd->error = io_u->error;
1488 io_u_log_error(td, io_u);
1489 }
1490 if (icd->error) {
1491 enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1492 if (!td_non_fatal_error(td, eb, icd->error))
1493 return;
1494 /*
1495 * If there is a non_fatal error, then add to the error count
1496 * and clear all the errors.
1497 */
1498 update_error_count(td, icd->error);
1499 td_clear_error(td);
1500 icd->error = 0;
1501 io_u->error = 0;
1502 }
1503}
1504
1505static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1506 int nr)
1507{
1508 int ddir;
1509 if (!td->o.disable_clat || !td->o.disable_bw)
1510 fio_gettime(&icd->time, NULL);
1511
1512 icd->nr = nr;
1513
1514 icd->error = 0;
1515 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1516 icd->bytes_done[ddir] = 0;
1517}
1518
1519static void ios_completed(struct thread_data *td,
1520 struct io_completion_data *icd)
1521{
1522 struct io_u *io_u;
1523 int i;
1524
1525 for (i = 0; i < icd->nr; i++) {
1526 io_u = td->io_ops->event(td, i);
1527
1528 io_completed(td, io_u, icd);
1529
1530 if (!(io_u->flags & IO_U_F_FREE_DEF))
1531 put_io_u(td, io_u);
1532 }
1533}
1534
1535/*
1536 * Complete a single io_u for the sync engines.
1537 */
1538int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1539 uint64_t *bytes)
1540{
1541 struct io_completion_data icd;
1542
1543 init_icd(td, &icd, 1);
1544 io_completed(td, io_u, &icd);
1545
1546 if (!(io_u->flags & IO_U_F_FREE_DEF))
1547 put_io_u(td, io_u);
1548
1549 if (icd.error) {
1550 td_verror(td, icd.error, "io_u_sync_complete");
1551 return -1;
1552 }
1553
1554 if (bytes) {
1555 int ddir;
1556
1557 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1558 bytes[ddir] += icd.bytes_done[ddir];
1559 }
1560
1561 return 0;
1562}
1563
1564/*
1565 * Called to complete min_events number of io for the async engines.
1566 */
1567int io_u_queued_complete(struct thread_data *td, int min_evts,
1568 uint64_t *bytes)
1569{
1570 struct io_completion_data icd;
1571 struct timespec *tvp = NULL;
1572 int ret;
1573 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1574
1575 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1576
1577 if (!min_evts)
1578 tvp = &ts;
1579
1580 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1581 if (ret < 0) {
1582 td_verror(td, -ret, "td_io_getevents");
1583 return ret;
1584 } else if (!ret)
1585 return ret;
1586
1587 init_icd(td, &icd, ret);
1588 ios_completed(td, &icd);
1589 if (icd.error) {
1590 td_verror(td, icd.error, "io_u_queued_complete");
1591 return -1;
1592 }
1593
1594 if (bytes) {
1595 int ddir;
1596
1597 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1598 bytes[ddir] += icd.bytes_done[ddir];
1599 }
1600
1601 return 0;
1602}
1603
1604/*
1605 * Call when io_u is really queued, to update the submission latency.
1606 */
1607void io_u_queued(struct thread_data *td, struct io_u *io_u)
1608{
1609 if (!td->o.disable_slat) {
1610 unsigned long slat_time;
1611
1612 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1613 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1614 }
1615}
1616
1617void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1618 unsigned int max_bs)
1619{
1620 if (!td->o.zero_buffers) {
1621 unsigned int perc = td->o.compress_percentage;
1622
1623 if (perc) {
1624 unsigned int seg = min_write;
1625
1626 seg = min(min_write, td->o.compress_chunk);
1627 if (!seg)
1628 seg = min_write;
1629
1630 fill_random_buf_percentage(&td->buf_state, buf,
1631 perc, seg, max_bs);
1632 } else
1633 fill_random_buf(&td->buf_state, buf, max_bs);
1634 } else
1635 memset(buf, 0, max_bs);
1636}
1637
1638/*
1639 * "randomly" fill the buffer contents
1640 */
1641void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1642 unsigned int min_write, unsigned int max_bs)
1643{
1644 io_u->buf_filled_len = 0;
1645 fill_io_buffer(td, io_u->buf, min_write, max_bs);
1646}
fio - Flexible IO Tester