projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Update for RDMA io engine's compatibility
[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.do_verify && td->o.verify != VERIFY_NONE)
463 buflen = (buflen + td->o.verify_interval - 1) &
464 ~(td->o.verify_interval - 1);
465
466 if (!td->o.bs_unaligned && is_power_of_2(minbs))
467 buflen = (buflen + minbs - 1) & ~(minbs - 1);
468
469 } while (!io_u_fits(td, io_u, buflen));
470
471 return buflen;
472}
473
474static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
475{
476 if (td->flags & TD_F_PROFILE_OPS) {
477 struct prof_io_ops *ops = &td->prof_io_ops;
478
479 if (ops->fill_io_u_size)
480 return ops->fill_io_u_size(td, io_u);
481 }
482
483 return __get_next_buflen(td, io_u);
484}
485
486static void set_rwmix_bytes(struct thread_data *td)
487{
488 unsigned int diff;
489
490 /*
491 * we do time or byte based switch. this is needed because
492 * buffered writes may issue a lot quicker than they complete,
493 * whereas reads do not.
494 */
495 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
496 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
497}
498
499static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
500{
501 unsigned int v;
502 unsigned long r;
503
504 if (td->o.use_os_rand) {
505 r = os_random_long(&td->rwmix_state);
506 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
507 } else {
508 r = __rand(&td->__rwmix_state);
509 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
510 }
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 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_FREE_DEF))
676 put_file_log(td, io_u->file);
677 io_u->file = NULL;
678 io_u->flags &= ~IO_U_F_FREE_DEF;
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 if (td->io_ops->flags & FIO_NOIO)
719 goto out;
720
721 set_rw_ddir(td, io_u);
722
723 /*
724 * fsync() or fdatasync() or trim etc, we are done
725 */
726 if (!ddir_rw(io_u->ddir))
727 goto out;
728
729 /*
730 * See if it's time to switch to a new zone
731 */
732 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
733 td->zone_bytes = 0;
734 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
735 io_u->file->last_pos = io_u->file->file_offset;
736 td->io_skip_bytes += td->o.zone_skip;
737 }
738
739 /*
740 * No log, let the seq/rand engine retrieve the next buflen and
741 * position.
742 */
743 if (get_next_offset(td, io_u)) {
744 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
745 return 1;
746 }
747
748 io_u->buflen = get_next_buflen(td, io_u);
749 if (!io_u->buflen) {
750 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
751 return 1;
752 }
753
754 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
755 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
756 dprint(FD_IO, " off=%llu/%lu > %llu\n",
757 (unsigned long long) io_u->offset, io_u->buflen,
758 (unsigned long long) io_u->file->real_file_size);
759 return 1;
760 }
761
762 /*
763 * mark entry before potentially trimming io_u
764 */
765 if (td_random(td) && file_randommap(td, io_u->file))
766 mark_random_map(td, io_u);
767
768out:
769 dprint_io_u(io_u, "fill_io_u");
770 td->zone_bytes += io_u->buflen;
771 return 0;
772}
773
774static void __io_u_mark_map(unsigned int *map, unsigned int nr)
775{
776 int idx = 0;
777
778 switch (nr) {
779 default:
780 idx = 6;
781 break;
782 case 33 ... 64:
783 idx = 5;
784 break;
785 case 17 ... 32:
786 idx = 4;
787 break;
788 case 9 ... 16:
789 idx = 3;
790 break;
791 case 5 ... 8:
792 idx = 2;
793 break;
794 case 1 ... 4:
795 idx = 1;
796 case 0:
797 break;
798 }
799
800 map[idx]++;
801}
802
803void io_u_mark_submit(struct thread_data *td, unsigned int nr)
804{
805 __io_u_mark_map(td->ts.io_u_submit, nr);
806 td->ts.total_submit++;
807}
808
809void io_u_mark_complete(struct thread_data *td, unsigned int nr)
810{
811 __io_u_mark_map(td->ts.io_u_complete, nr);
812 td->ts.total_complete++;
813}
814
815void io_u_mark_depth(struct thread_data *td, unsigned int nr)
816{
817 int idx = 0;
818
819 switch (td->cur_depth) {
820 default:
821 idx = 6;
822 break;
823 case 32 ... 63:
824 idx = 5;
825 break;
826 case 16 ... 31:
827 idx = 4;
828 break;
829 case 8 ... 15:
830 idx = 3;
831 break;
832 case 4 ... 7:
833 idx = 2;
834 break;
835 case 2 ... 3:
836 idx = 1;
837 case 1:
838 break;
839 }
840
841 td->ts.io_u_map[idx] += nr;
842}
843
844static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
845{
846 int idx = 0;
847
848 assert(usec < 1000);
849
850 switch (usec) {
851 case 750 ... 999:
852 idx = 9;
853 break;
854 case 500 ... 749:
855 idx = 8;
856 break;
857 case 250 ... 499:
858 idx = 7;
859 break;
860 case 100 ... 249:
861 idx = 6;
862 break;
863 case 50 ... 99:
864 idx = 5;
865 break;
866 case 20 ... 49:
867 idx = 4;
868 break;
869 case 10 ... 19:
870 idx = 3;
871 break;
872 case 4 ... 9:
873 idx = 2;
874 break;
875 case 2 ... 3:
876 idx = 1;
877 case 0 ... 1:
878 break;
879 }
880
881 assert(idx < FIO_IO_U_LAT_U_NR);
882 td->ts.io_u_lat_u[idx]++;
883}
884
885static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
886{
887 int idx = 0;
888
889 switch (msec) {
890 default:
891 idx = 11;
892 break;
893 case 1000 ... 1999:
894 idx = 10;
895 break;
896 case 750 ... 999:
897 idx = 9;
898 break;
899 case 500 ... 749:
900 idx = 8;
901 break;
902 case 250 ... 499:
903 idx = 7;
904 break;
905 case 100 ... 249:
906 idx = 6;
907 break;
908 case 50 ... 99:
909 idx = 5;
910 break;
911 case 20 ... 49:
912 idx = 4;
913 break;
914 case 10 ... 19:
915 idx = 3;
916 break;
917 case 4 ... 9:
918 idx = 2;
919 break;
920 case 2 ... 3:
921 idx = 1;
922 case 0 ... 1:
923 break;
924 }
925
926 assert(idx < FIO_IO_U_LAT_M_NR);
927 td->ts.io_u_lat_m[idx]++;
928}
929
930static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
931{
932 if (usec < 1000)
933 io_u_mark_lat_usec(td, usec);
934 else
935 io_u_mark_lat_msec(td, usec / 1000);
936}
937
938/*
939 * Get next file to service by choosing one at random
940 */
941static struct fio_file *get_next_file_rand(struct thread_data *td,
942 enum fio_file_flags goodf,
943 enum fio_file_flags badf)
944{
945 struct fio_file *f;
946 int fno;
947
948 do {
949 int opened = 0;
950 unsigned long r;
951
952 if (td->o.use_os_rand) {
953 r = os_random_long(&td->next_file_state);
954 fno = (unsigned int) ((double) td->o.nr_files
955 * (r / (OS_RAND_MAX + 1.0)));
956 } else {
957 r = __rand(&td->__next_file_state);
958 fno = (unsigned int) ((double) td->o.nr_files
959 * (r / (FRAND_MAX + 1.0)));
960 }
961
962 f = td->files[fno];
963 if (fio_file_done(f))
964 continue;
965
966 if (!fio_file_open(f)) {
967 int err;
968
969 err = td_io_open_file(td, f);
970 if (err)
971 continue;
972 opened = 1;
973 }
974
975 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
976 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
977 return f;
978 }
979 if (opened)
980 td_io_close_file(td, f);
981 } while (1);
982}
983
984/*
985 * Get next file to service by doing round robin between all available ones
986 */
987static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
988 int badf)
989{
990 unsigned int old_next_file = td->next_file;
991 struct fio_file *f;
992
993 do {
994 int opened = 0;
995
996 f = td->files[td->next_file];
997
998 td->next_file++;
999 if (td->next_file >= td->o.nr_files)
1000 td->next_file = 0;
1001
1002 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1003 if (fio_file_done(f)) {
1004 f = NULL;
1005 continue;
1006 }
1007
1008 if (!fio_file_open(f)) {
1009 int err;
1010
1011 err = td_io_open_file(td, f);
1012 if (err) {
1013 dprint(FD_FILE, "error %d on open of %s\n",
1014 err, f->file_name);
1015 f = NULL;
1016 continue;
1017 }
1018 opened = 1;
1019 }
1020
1021 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1022 f->flags);
1023 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1024 break;
1025
1026 if (opened)
1027 td_io_close_file(td, f);
1028
1029 f = NULL;
1030 } while (td->next_file != old_next_file);
1031
1032 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1033 return f;
1034}
1035
1036static struct fio_file *__get_next_file(struct thread_data *td)
1037{
1038 struct fio_file *f;
1039
1040 assert(td->o.nr_files <= td->files_index);
1041
1042 if (td->nr_done_files >= td->o.nr_files) {
1043 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1044 " nr_files=%d\n", td->nr_open_files,
1045 td->nr_done_files,
1046 td->o.nr_files);
1047 return NULL;
1048 }
1049
1050 f = td->file_service_file;
1051 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1052 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1053 goto out;
1054 if (td->file_service_left--)
1055 goto out;
1056 }
1057
1058 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1059 td->o.file_service_type == FIO_FSERVICE_SEQ)
1060 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1061 else
1062 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1063
1064 td->file_service_file = f;
1065 td->file_service_left = td->file_service_nr - 1;
1066out:
1067 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1068 return f;
1069}
1070
1071static struct fio_file *get_next_file(struct thread_data *td)
1072{
1073 if (!(td->flags & TD_F_PROFILE_OPS)) {
1074 struct prof_io_ops *ops = &td->prof_io_ops;
1075
1076 if (ops->get_next_file)
1077 return ops->get_next_file(td);
1078 }
1079
1080 return __get_next_file(td);
1081}
1082
1083static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1084{
1085 struct fio_file *f;
1086
1087 do {
1088 f = get_next_file(td);
1089 if (!f)
1090 return 1;
1091
1092 io_u->file = f;
1093 get_file(f);
1094
1095 if (!fill_io_u(td, io_u))
1096 break;
1097
1098 put_file_log(td, f);
1099 td_io_close_file(td, f);
1100 io_u->file = NULL;
1101 fio_file_set_done(f);
1102 td->nr_done_files++;
1103 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1104 td->nr_done_files, td->o.nr_files);
1105 } while (1);
1106
1107 return 0;
1108}
1109
1110
1111struct io_u *__get_io_u(struct thread_data *td)
1112{
1113 struct io_u *io_u;
1114
1115 td_io_u_lock(td);
1116
1117again:
1118 if (!io_u_rempty(&td->io_u_requeues))
1119 io_u = io_u_rpop(&td->io_u_requeues);
1120 else if (!io_u_qempty(&td->io_u_freelist))
1121 io_u = io_u_qpop(&td->io_u_freelist);
1122
1123 if (io_u) {
1124 io_u->buflen = 0;
1125 io_u->resid = 0;
1126 io_u->file = NULL;
1127 io_u->end_io = NULL;
1128 }
1129
1130 if (io_u) {
1131 assert(io_u->flags & IO_U_F_FREE);
1132 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1133 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1134 io_u->flags &= ~IO_U_F_VER_LIST;
1135
1136 io_u->error = 0;
1137 io_u->acct_ddir = -1;
1138 td->cur_depth++;
1139 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1140 } else if (td->o.verify_async) {
1141 /*
1142 * We ran out, wait for async verify threads to finish and
1143 * return one
1144 */
1145 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1146 goto again;
1147 }
1148
1149 td_io_u_unlock(td);
1150 return io_u;
1151}
1152
1153static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1154{
1155 if (!(td->flags & TD_F_TRIM_BACKLOG))
1156 return 0;
1157
1158 if (td->trim_entries) {
1159 int get_trim = 0;
1160
1161 if (td->trim_batch) {
1162 td->trim_batch--;
1163 get_trim = 1;
1164 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1165 td->last_ddir != DDIR_READ) {
1166 td->trim_batch = td->o.trim_batch;
1167 if (!td->trim_batch)
1168 td->trim_batch = td->o.trim_backlog;
1169 get_trim = 1;
1170 }
1171
1172 if (get_trim && !get_next_trim(td, io_u))
1173 return 1;
1174 }
1175
1176 return 0;
1177}
1178
1179static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1180{
1181 if (!(td->flags & TD_F_VER_BACKLOG))
1182 return 0;
1183
1184 if (td->io_hist_len) {
1185 int get_verify = 0;
1186
1187 if (td->verify_batch)
1188 get_verify = 1;
1189 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1190 td->last_ddir != DDIR_READ) {
1191 td->verify_batch = td->o.verify_batch;
1192 if (!td->verify_batch)
1193 td->verify_batch = td->o.verify_backlog;
1194 get_verify = 1;
1195 }
1196
1197 if (get_verify && !get_next_verify(td, io_u)) {
1198 td->verify_batch--;
1199 return 1;
1200 }
1201 }
1202
1203 return 0;
1204}
1205
1206/*
1207 * Fill offset and start time into the buffer content, to prevent too
1208 * easy compressible data for simple de-dupe attempts. Do this for every
1209 * 512b block in the range, since that should be the smallest block size
1210 * we can expect from a device.
1211 */
1212static void small_content_scramble(struct io_u *io_u)
1213{
1214 unsigned int i, nr_blocks = io_u->buflen / 512;
1215 uint64_t boffset;
1216 unsigned int offset;
1217 void *p, *end;
1218
1219 if (!nr_blocks)
1220 return;
1221
1222 p = io_u->xfer_buf;
1223 boffset = io_u->offset;
1224 io_u->buf_filled_len = 0;
1225
1226 for (i = 0; i < nr_blocks; i++) {
1227 /*
1228 * Fill the byte offset into a "random" start offset of
1229 * the buffer, given by the product of the usec time
1230 * and the actual offset.
1231 */
1232 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1233 offset &= ~(sizeof(uint64_t) - 1);
1234 if (offset >= 512 - sizeof(uint64_t))
1235 offset -= sizeof(uint64_t);
1236 memcpy(p + offset, &boffset, sizeof(boffset));
1237
1238 end = p + 512 - sizeof(io_u->start_time);
1239 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1240 p += 512;
1241 boffset += 512;
1242 }
1243}
1244
1245/*
1246 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1247 * etc. The returned io_u is fully ready to be prepped and submitted.
1248 */
1249struct io_u *get_io_u(struct thread_data *td)
1250{
1251 struct fio_file *f;
1252 struct io_u *io_u;
1253 int do_scramble = 0;
1254
1255 io_u = __get_io_u(td);
1256 if (!io_u) {
1257 dprint(FD_IO, "__get_io_u failed\n");
1258 return NULL;
1259 }
1260
1261 if (check_get_verify(td, io_u))
1262 goto out;
1263 if (check_get_trim(td, io_u))
1264 goto out;
1265
1266 /*
1267 * from a requeue, io_u already setup
1268 */
1269 if (io_u->file)
1270 goto out;
1271
1272 /*
1273 * If using an iolog, grab next piece if any available.
1274 */
1275 if (td->flags & TD_F_READ_IOLOG) {
1276 if (read_iolog_get(td, io_u))
1277 goto err_put;
1278 } else if (set_io_u_file(td, io_u)) {
1279 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1280 goto err_put;
1281 }
1282
1283 f = io_u->file;
1284 assert(fio_file_open(f));
1285
1286 if (ddir_rw(io_u->ddir)) {
1287 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1288 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1289 goto err_put;
1290 }
1291
1292 f->last_start = io_u->offset;
1293 f->last_pos = io_u->offset + io_u->buflen;
1294
1295 if (io_u->ddir == DDIR_WRITE) {
1296 if (td->flags & TD_F_REFILL_BUFFERS) {
1297 io_u_fill_buffer(td, io_u,
1298 io_u->xfer_buflen, io_u->xfer_buflen);
1299 } else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
1300 do_scramble = 1;
1301 if (td->flags & TD_F_VER_NONE) {
1302 populate_verify_io_u(td, io_u);
1303 do_scramble = 0;
1304 }
1305 } else if (io_u->ddir == DDIR_READ) {
1306 /*
1307 * Reset the buf_filled parameters so next time if the
1308 * buffer is used for writes it is refilled.
1309 */
1310 io_u->buf_filled_len = 0;
1311 }
1312 }
1313
1314 /*
1315 * Set io data pointers.
1316 */
1317 io_u->xfer_buf = io_u->buf;
1318 io_u->xfer_buflen = io_u->buflen;
1319
1320out:
1321 assert(io_u->file);
1322 if (!td_io_prep(td, io_u)) {
1323 if (!td->o.disable_slat)
1324 fio_gettime(&io_u->start_time, NULL);
1325 if (do_scramble)
1326 small_content_scramble(io_u);
1327 return io_u;
1328 }
1329err_put:
1330 dprint(FD_IO, "get_io_u failed\n");
1331 put_io_u(td, io_u);
1332 return NULL;
1333}
1334
1335void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1336{
1337 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1338 const char *msg[] = { "read", "write", "sync", "datasync",
1339 "sync_file_range", "wait", "trim" };
1340
1341 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1342 return;
1343
1344 log_err("fio: io_u error");
1345
1346 if (io_u->file)
1347 log_err(" on file %s", io_u->file->file_name);
1348
1349 log_err(": %s\n", strerror(io_u->error));
1350
1351 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1352 io_u->offset, io_u->xfer_buflen);
1353
1354 if (!td->error)
1355 td_verror(td, io_u->error, "io_u error");
1356}
1357
1358static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1359 struct io_completion_data *icd,
1360 const enum fio_ddir idx, unsigned int bytes)
1361{
1362 unsigned long lusec = 0;
1363
1364 if (!td->o.disable_clat || !td->o.disable_bw)
1365 lusec = utime_since(&io_u->issue_time, &icd->time);
1366
1367 if (!td->o.disable_lat) {
1368 unsigned long tusec;
1369
1370 tusec = utime_since(&io_u->start_time, &icd->time);
1371 add_lat_sample(td, idx, tusec, bytes);
1372
1373 if (td->flags & TD_F_PROFILE_OPS) {
1374 struct prof_io_ops *ops = &td->prof_io_ops;
1375
1376 if (ops->io_u_lat)
1377 icd->error = ops->io_u_lat(td, tusec);
1378 }
1379
1380 if (td->o.max_latency && tusec > td->o.max_latency) {
1381 if (!td->error)
1382 log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
1383 td_verror(td, ETIMEDOUT, "max latency exceeded");
1384 icd->error = ETIMEDOUT;
1385 }
1386 }
1387
1388 if (!td->o.disable_clat) {
1389 add_clat_sample(td, idx, lusec, bytes);
1390 io_u_mark_latency(td, lusec);
1391 }
1392
1393 if (!td->o.disable_bw)
1394 add_bw_sample(td, idx, bytes, &icd->time);
1395
1396 add_iops_sample(td, idx, &icd->time);
1397}
1398
1399static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1400{
1401 uint64_t secs, remainder, bps, bytes;
1402
1403 bytes = td->this_io_bytes[ddir];
1404 bps = td->rate_bps[ddir];
1405 secs = bytes / bps;
1406 remainder = bytes % bps;
1407 return remainder * 1000000 / bps + secs * 1000000;
1408}
1409
1410static void io_completed(struct thread_data *td, struct io_u *io_u,
1411 struct io_completion_data *icd)
1412{
1413 struct fio_file *f;
1414
1415 dprint_io_u(io_u, "io complete");
1416
1417 td_io_u_lock(td);
1418 assert(io_u->flags & IO_U_F_FLIGHT);
1419 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1420 td_io_u_unlock(td);
1421
1422 if (ddir_sync(io_u->ddir)) {
1423 td->last_was_sync = 1;
1424 f = io_u->file;
1425 if (f) {
1426 f->first_write = -1ULL;
1427 f->last_write = -1ULL;
1428 }
1429 return;
1430 }
1431
1432 td->last_was_sync = 0;
1433 td->last_ddir = io_u->ddir;
1434
1435 if (!io_u->error && ddir_rw(io_u->ddir)) {
1436 unsigned int bytes = io_u->buflen - io_u->resid;
1437 const enum fio_ddir idx = io_u->ddir;
1438 const enum fio_ddir odx = io_u->ddir ^ 1;
1439 int ret;
1440
1441 td->io_blocks[idx]++;
1442 td->this_io_blocks[idx]++;
1443 td->io_bytes[idx] += bytes;
1444
1445 if (!(io_u->flags & IO_U_F_VER_LIST))
1446 td->this_io_bytes[idx] += bytes;
1447
1448 if (idx == DDIR_WRITE) {
1449 f = io_u->file;
1450 if (f) {
1451 if (f->first_write == -1ULL ||
1452 io_u->offset < f->first_write)
1453 f->first_write = io_u->offset;
1454 if (f->last_write == -1ULL ||
1455 ((io_u->offset + bytes) > f->last_write))
1456 f->last_write = io_u->offset + bytes;
1457 }
1458 }
1459
1460 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1461 td->runstate == TD_VERIFYING)) {
1462 account_io_completion(td, io_u, icd, idx, bytes);
1463
1464 if (__should_check_rate(td, idx)) {
1465 td->rate_pending_usleep[idx] =
1466 (usec_for_io(td, idx) -
1467 utime_since_now(&td->start));
1468 }
1469 if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1470 td->rate_pending_usleep[odx] =
1471 (usec_for_io(td, odx) -
1472 utime_since_now(&td->start));
1473 }
1474
1475 if (td_write(td) && idx == DDIR_WRITE &&
1476 td->o.do_verify &&
1477 td->o.verify != VERIFY_NONE &&
1478 !td->o.experimental_verify)
1479 log_io_piece(td, io_u);
1480
1481 icd->bytes_done[idx] += bytes;
1482
1483 if (io_u->end_io) {
1484 ret = io_u->end_io(td, io_u);
1485 if (ret && !icd->error)
1486 icd->error = ret;
1487 }
1488 } else if (io_u->error) {
1489 icd->error = io_u->error;
1490 io_u_log_error(td, io_u);
1491 }
1492 if (icd->error) {
1493 enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1494 if (!td_non_fatal_error(td, eb, icd->error))
1495 return;
1496 /*
1497 * If there is a non_fatal error, then add to the error count
1498 * and clear all the errors.
1499 */
1500 update_error_count(td, icd->error);
1501 td_clear_error(td);
1502 icd->error = 0;
1503 io_u->error = 0;
1504 }
1505}
1506
1507static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1508 int nr)
1509{
1510 int ddir;
1511 if (!td->o.disable_clat || !td->o.disable_bw)
1512 fio_gettime(&icd->time, NULL);
1513
1514 icd->nr = nr;
1515
1516 icd->error = 0;
1517 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1518 icd->bytes_done[ddir] = 0;
1519}
1520
1521static void ios_completed(struct thread_data *td,
1522 struct io_completion_data *icd)
1523{
1524 struct io_u *io_u;
1525 int i;
1526
1527 for (i = 0; i < icd->nr; i++) {
1528 io_u = td->io_ops->event(td, i);
1529
1530 io_completed(td, io_u, icd);
1531
1532 if (!(io_u->flags & IO_U_F_FREE_DEF))
1533 put_io_u(td, io_u);
1534 }
1535}
1536
1537/*
1538 * Complete a single io_u for the sync engines.
1539 */
1540int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1541 uint64_t *bytes)
1542{
1543 struct io_completion_data icd;
1544
1545 init_icd(td, &icd, 1);
1546 io_completed(td, io_u, &icd);
1547
1548 if (!(io_u->flags & IO_U_F_FREE_DEF))
1549 put_io_u(td, io_u);
1550
1551 if (icd.error) {
1552 td_verror(td, icd.error, "io_u_sync_complete");
1553 return -1;
1554 }
1555
1556 if (bytes) {
1557 int ddir;
1558
1559 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1560 bytes[ddir] += icd.bytes_done[ddir];
1561 }
1562
1563 return 0;
1564}
1565
1566/*
1567 * Called to complete min_events number of io for the async engines.
1568 */
1569int io_u_queued_complete(struct thread_data *td, int min_evts,
1570 uint64_t *bytes)
1571{
1572 struct io_completion_data icd;
1573 struct timespec *tvp = NULL;
1574 int ret;
1575 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1576
1577 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1578
1579 if (!min_evts)
1580 tvp = &ts;
1581
1582 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1583 if (ret < 0) {
1584 td_verror(td, -ret, "td_io_getevents");
1585 return ret;
1586 } else if (!ret)
1587 return ret;
1588
1589 init_icd(td, &icd, ret);
1590 ios_completed(td, &icd);
1591 if (icd.error) {
1592 td_verror(td, icd.error, "io_u_queued_complete");
1593 return -1;
1594 }
1595
1596 if (bytes) {
1597 int ddir;
1598
1599 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1600 bytes[ddir] += icd.bytes_done[ddir];
1601 }
1602
1603 return 0;
1604}
1605
1606/*
1607 * Call when io_u is really queued, to update the submission latency.
1608 */
1609void io_u_queued(struct thread_data *td, struct io_u *io_u)
1610{
1611 if (!td->o.disable_slat) {
1612 unsigned long slat_time;
1613
1614 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1615 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1616 }
1617}
1618
1619void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1620 unsigned int max_bs)
1621{
1622 if (!td->o.zero_buffers) {
1623 unsigned int perc = td->o.compress_percentage;
1624
1625 if (perc) {
1626 unsigned int seg = min_write;
1627
1628 seg = min(min_write, td->o.compress_chunk);
1629 if (!seg)
1630 seg = min_write;
1631
1632 fill_random_buf_percentage(&td->buf_state, buf,
1633 perc, seg, max_bs);
1634 } else
1635 fill_random_buf(&td->buf_state, buf, max_bs);
1636 } else
1637 memset(buf, 0, max_bs);
1638}
1639
1640/*
1641 * "randomly" fill the buffer contents
1642 */
1643void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1644 unsigned int min_write, unsigned int max_bs)
1645{
1646 io_u->buf_filled_len = 0;
1647 fill_io_buffer(td, io_u->buf, min_write, max_bs);
1648}
fio - Flexible IO Tester