projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Add verify trim support
[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
14struct io_completion_data {
15 int nr; /* input */
16
17 int error; /* output */
18 unsigned long bytes_done[2]; /* output */
19 struct timeval time; /* output */
20};
21
22/*
23 * The ->file_map[] contains a map of blocks we have or have not done io
24 * to yet. Used to make sure we cover the entire range in a fair fashion.
25 */
26static int random_map_free(struct fio_file *f, const unsigned long long block)
27{
28 unsigned int idx = RAND_MAP_IDX(f, block);
29 unsigned int bit = RAND_MAP_BIT(f, block);
30
31 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
32
33 return (f->file_map[idx] & (1 << bit)) == 0;
34}
35
36/*
37 * Mark a given offset as used in the map.
38 */
39static void mark_random_map(struct thread_data *td, struct io_u *io_u)
40{
41 unsigned int min_bs = td->o.rw_min_bs;
42 struct fio_file *f = io_u->file;
43 unsigned long long block;
44 unsigned int blocks, nr_blocks;
45 int busy_check;
46
47 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
48 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
49 blocks = 0;
50 busy_check = !(io_u->flags & IO_U_F_BUSY_OK);
51
52 while (nr_blocks) {
53 unsigned int this_blocks, mask;
54 unsigned int idx, bit;
55
56 /*
57 * If we have a mixed random workload, we may
58 * encounter blocks we already did IO to.
59 */
60 if (!busy_check) {
61 blocks = nr_blocks;
62 break;
63 }
64 if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block))
65 break;
66
67 idx = RAND_MAP_IDX(f, block);
68 bit = RAND_MAP_BIT(f, block);
69
70 fio_assert(td, idx < f->num_maps);
71
72 this_blocks = nr_blocks;
73 if (this_blocks + bit > BLOCKS_PER_MAP)
74 this_blocks = BLOCKS_PER_MAP - bit;
75
76 do {
77 if (this_blocks == BLOCKS_PER_MAP)
78 mask = -1U;
79 else
80 mask = ((1U << this_blocks) - 1) << bit;
81
82 if (!(f->file_map[idx] & mask))
83 break;
84
85 this_blocks--;
86 } while (this_blocks);
87
88 if (!this_blocks)
89 break;
90
91 f->file_map[idx] |= mask;
92 nr_blocks -= this_blocks;
93 blocks += this_blocks;
94 block += this_blocks;
95 }
96
97 if ((blocks * min_bs) < io_u->buflen)
98 io_u->buflen = blocks * min_bs;
99}
100
101static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
102 enum fio_ddir ddir)
103{
104 unsigned long long max_blocks;
105 unsigned long long max_size;
106
107 assert(ddir_rw(ddir));
108
109 /*
110 * Hmm, should we make sure that ->io_size <= ->real_file_size?
111 */
112 max_size = f->io_size;
113 if (max_size > f->real_file_size)
114 max_size = f->real_file_size;
115
116 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
117 if (!max_blocks)
118 return 0;
119
120 return max_blocks;
121}
122
123/*
124 * Return the next free block in the map.
125 */
126static int get_next_free_block(struct thread_data *td, struct fio_file *f,
127 enum fio_ddir ddir, unsigned long long *b)
128{
129 unsigned long long min_bs = td->o.rw_min_bs;
130 int i;
131
132 i = f->last_free_lookup;
133 *b = (i * BLOCKS_PER_MAP);
134 while ((*b) * min_bs < f->real_file_size &&
135 (*b) * min_bs < f->io_size) {
136 if (f->file_map[i] != (unsigned int) -1) {
137 *b += ffz(f->file_map[i]);
138 if (*b > last_block(td, f, ddir))
139 break;
140 f->last_free_lookup = i;
141 return 0;
142 }
143
144 *b += BLOCKS_PER_MAP;
145 i++;
146 }
147
148 dprint(FD_IO, "failed finding a free block\n");
149 return 1;
150}
151
152static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
153 enum fio_ddir ddir, unsigned long long *b)
154{
155 unsigned long long r;
156 int loops = 5;
157
158 do {
159 r = os_random_long(&td->random_state);
160 dprint(FD_RANDOM, "off rand %llu\n", r);
161 *b = (last_block(td, f, ddir) - 1)
162 * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
163
164 /*
165 * if we are not maintaining a random map, we are done.
166 */
167 if (!file_randommap(td, f))
168 return 0;
169
170 /*
171 * calculate map offset and check if it's free
172 */
173 if (random_map_free(f, *b))
174 return 0;
175
176 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
177 *b);
178 } while (--loops);
179
180 /*
181 * we get here, if we didn't suceed in looking up a block. generate
182 * a random start offset into the filemap, and find the first free
183 * block from there.
184 */
185 loops = 10;
186 do {
187 f->last_free_lookup = (f->num_maps - 1) *
188 (r / (OS_RAND_MAX + 1.0));
189 if (!get_next_free_block(td, f, ddir, b))
190 return 0;
191
192 r = os_random_long(&td->random_state);
193 } while (--loops);
194
195 /*
196 * that didn't work either, try exhaustive search from the start
197 */
198 f->last_free_lookup = 0;
199 return get_next_free_block(td, f, ddir, b);
200}
201
202static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
203 enum fio_ddir ddir, unsigned long long *b)
204{
205 if (get_next_rand_offset(td, f, ddir, b)) {
206 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
207 f->file_name, f->last_pos, f->real_file_size);
208 return 1;
209 }
210
211 return 0;
212}
213
214static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
215 enum fio_ddir ddir, unsigned long long *b)
216{
217 assert(ddir_rw(ddir));
218
219 if (f->last_pos < f->real_file_size) {
220 *b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
221 return 0;
222 }
223
224 return 1;
225}
226
227static int get_next_block(struct thread_data *td, struct io_u *io_u,
228 enum fio_ddir ddir, int rw_seq, unsigned long long *b)
229{
230 struct fio_file *f = io_u->file;
231 int ret;
232
233 assert(ddir_rw(ddir));
234
235 if (rw_seq) {
236 if (td_random(td))
237 ret = get_next_rand_block(td, f, ddir, b);
238 else
239 ret = get_next_seq_block(td, f, ddir, b);
240 } else {
241 io_u->flags |= IO_U_F_BUSY_OK;
242
243 if (td->o.rw_seq == RW_SEQ_SEQ) {
244 ret = get_next_seq_block(td, f, ddir, b);
245 if (ret)
246 ret = get_next_rand_block(td, f, ddir, b);
247 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
248 if (f->last_start != -1ULL)
249 *b = (f->last_start - f->file_offset) / td->o.min_bs[ddir];
250 else
251 *b = 0;
252 ret = 0;
253 } else {
254 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
255 ret = 1;
256 }
257 }
258
259 return ret;
260}
261
262/*
263 * For random io, generate a random new block and see if it's used. Repeat
264 * until we find a free one. For sequential io, just return the end of
265 * the last io issued.
266 */
267static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
268{
269 struct fio_file *f = io_u->file;
270 unsigned long long b;
271 enum fio_ddir ddir = io_u->ddir;
272 int rw_seq_hit = 0;
273
274 assert(ddir_rw(ddir));
275
276 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
277 rw_seq_hit = 1;
278 td->ddir_seq_nr = td->o.ddir_seq_nr;
279 }
280
281 if (get_next_block(td, io_u, ddir, rw_seq_hit, &b)) {
282 printf("fail\n");
283 return 1;
284 }
285
286 io_u->offset = b * td->o.ba[ddir];
287 if (io_u->offset >= f->io_size) {
288 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
289 io_u->offset, f->io_size);
290 return 1;
291 }
292
293 io_u->offset += f->file_offset;
294 if (io_u->offset >= f->real_file_size) {
295 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
296 io_u->offset, f->real_file_size);
297 return 1;
298 }
299
300 return 0;
301}
302
303static int get_next_offset(struct thread_data *td, struct io_u *io_u)
304{
305 struct prof_io_ops *ops = &td->prof_io_ops;
306
307 if (ops->fill_io_u_off)
308 return ops->fill_io_u_off(td, io_u);
309
310 return __get_next_offset(td, io_u);
311}
312
313static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
314{
315 const int ddir = io_u->ddir;
316 unsigned int uninitialized_var(buflen);
317 unsigned int minbs, maxbs;
318 long r;
319
320 assert(ddir_rw(ddir));
321
322 minbs = td->o.min_bs[ddir];
323 maxbs = td->o.max_bs[ddir];
324
325 if (minbs == maxbs)
326 buflen = minbs;
327 else {
328 r = os_random_long(&td->bsrange_state);
329 if (!td->o.bssplit_nr[ddir]) {
330 buflen = 1 + (unsigned int) ((double) maxbs *
331 (r / (OS_RAND_MAX + 1.0)));
332 if (buflen < minbs)
333 buflen = minbs;
334 } else {
335 long perc = 0;
336 unsigned int i;
337
338 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
339 struct bssplit *bsp = &td->o.bssplit[ddir][i];
340
341 buflen = bsp->bs;
342 perc += bsp->perc;
343 if (r <= ((OS_RAND_MAX / 100L) * perc))
344 break;
345 }
346 }
347 if (!td->o.bs_unaligned && is_power_of_2(minbs))
348 buflen = (buflen + minbs - 1) & ~(minbs - 1);
349 }
350
351 if (io_u->offset + buflen > io_u->file->real_file_size) {
352 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
353 minbs, ddir);
354 buflen = minbs;
355 }
356
357 return buflen;
358}
359
360static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
361{
362 struct prof_io_ops *ops = &td->prof_io_ops;
363
364 if (ops->fill_io_u_size)
365 return ops->fill_io_u_size(td, io_u);
366
367 return __get_next_buflen(td, io_u);
368}
369
370static void set_rwmix_bytes(struct thread_data *td)
371{
372 unsigned int diff;
373
374 /*
375 * we do time or byte based switch. this is needed because
376 * buffered writes may issue a lot quicker than they complete,
377 * whereas reads do not.
378 */
379 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
380 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
381}
382
383static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
384{
385 unsigned int v;
386 long r;
387
388 r = os_random_long(&td->rwmix_state);
389 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
390 if (v <= td->o.rwmix[DDIR_READ])
391 return DDIR_READ;
392
393 return DDIR_WRITE;
394}
395
396static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
397{
398 enum fio_ddir odir = ddir ^ 1;
399 struct timeval t;
400 long usec;
401
402 assert(ddir_rw(ddir));
403
404 if (td->rate_pending_usleep[ddir] <= 0)
405 return ddir;
406
407 /*
408 * We have too much pending sleep in this direction. See if we
409 * should switch.
410 */
411 if (td_rw(td)) {
412 /*
413 * Other direction does not have too much pending, switch
414 */
415 if (td->rate_pending_usleep[odir] < 100000)
416 return odir;
417
418 /*
419 * Both directions have pending sleep. Sleep the minimum time
420 * and deduct from both.
421 */
422 if (td->rate_pending_usleep[ddir] <=
423 td->rate_pending_usleep[odir]) {
424 usec = td->rate_pending_usleep[ddir];
425 } else {
426 usec = td->rate_pending_usleep[odir];
427 ddir = odir;
428 }
429 } else
430 usec = td->rate_pending_usleep[ddir];
431
432 fio_gettime(&t, NULL);
433 usec_sleep(td, usec);
434 usec = utime_since_now(&t);
435
436 td->rate_pending_usleep[ddir] -= usec;
437
438 odir = ddir ^ 1;
439 if (td_rw(td) && __should_check_rate(td, odir))
440 td->rate_pending_usleep[odir] -= usec;
441
442 return ddir;
443}
444
445/*
446 * Return the data direction for the next io_u. If the job is a
447 * mixed read/write workload, check the rwmix cycle and switch if
448 * necessary.
449 */
450static enum fio_ddir get_rw_ddir(struct thread_data *td)
451{
452 enum fio_ddir ddir;
453
454 /*
455 * see if it's time to fsync
456 */
457 if (td->o.fsync_blocks &&
458 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
459 td->io_issues[DDIR_WRITE] && should_fsync(td))
460 return DDIR_SYNC;
461
462 /*
463 * see if it's time to fdatasync
464 */
465 if (td->o.fdatasync_blocks &&
466 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
467 td->io_issues[DDIR_WRITE] && should_fsync(td))
468 return DDIR_DATASYNC;
469
470 /*
471 * see if it's time to sync_file_range
472 */
473 if (td->sync_file_range_nr &&
474 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
475 td->io_issues[DDIR_WRITE] && should_fsync(td))
476 return DDIR_SYNC_FILE_RANGE;
477
478 if (td_rw(td)) {
479 /*
480 * Check if it's time to seed a new data direction.
481 */
482 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
483 /*
484 * Put a top limit on how many bytes we do for
485 * one data direction, to avoid overflowing the
486 * ranges too much
487 */
488 ddir = get_rand_ddir(td);
489
490 if (ddir != td->rwmix_ddir)
491 set_rwmix_bytes(td);
492
493 td->rwmix_ddir = ddir;
494 }
495 ddir = td->rwmix_ddir;
496 } else if (td_read(td))
497 ddir = DDIR_READ;
498 else
499 ddir = DDIR_WRITE;
500
501 td->rwmix_ddir = rate_ddir(td, ddir);
502 return td->rwmix_ddir;
503}
504
505void put_file_log(struct thread_data *td, struct fio_file *f)
506{
507 int ret = put_file(td, f);
508
509 if (ret)
510 td_verror(td, ret, "file close");
511}
512
513void put_io_u(struct thread_data *td, struct io_u *io_u)
514{
515 td_io_u_lock(td);
516
517 io_u->flags |= IO_U_F_FREE;
518 io_u->flags &= ~IO_U_F_FREE_DEF;
519
520 if (io_u->file)
521 put_file_log(td, io_u->file);
522
523 io_u->file = NULL;
524 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
525 td->cur_depth--;
526 flist_del_init(&io_u->list);
527 flist_add(&io_u->list, &td->io_u_freelist);
528 td_io_u_unlock(td);
529 td_io_u_free_notify(td);
530}
531
532void clear_io_u(struct thread_data *td, struct io_u *io_u)
533{
534 io_u->flags &= ~IO_U_F_FLIGHT;
535 put_io_u(td, io_u);
536}
537
538void requeue_io_u(struct thread_data *td, struct io_u **io_u)
539{
540 struct io_u *__io_u = *io_u;
541
542 dprint(FD_IO, "requeue %p\n", __io_u);
543
544 td_io_u_lock(td);
545
546 __io_u->flags |= IO_U_F_FREE;
547 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
548 td->io_issues[__io_u->ddir]--;
549
550 __io_u->flags &= ~IO_U_F_FLIGHT;
551 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
552 td->cur_depth--;
553 flist_del(&__io_u->list);
554 flist_add_tail(&__io_u->list, &td->io_u_requeues);
555 td_io_u_unlock(td);
556 *io_u = NULL;
557}
558
559static int fill_io_u(struct thread_data *td, struct io_u *io_u)
560{
561 if (td->io_ops->flags & FIO_NOIO)
562 goto out;
563
564 io_u->ddir = get_rw_ddir(td);
565
566 /*
567 * fsync() or fdatasync() or trim etc, we are done
568 */
569 if (!ddir_rw(io_u->ddir))
570 goto out;
571
572 /*
573 * See if it's time to switch to a new zone
574 */
575 if (td->zone_bytes >= td->o.zone_size) {
576 td->zone_bytes = 0;
577 io_u->file->last_pos += td->o.zone_skip;
578 td->io_skip_bytes += td->o.zone_skip;
579 }
580
581 /*
582 * No log, let the seq/rand engine retrieve the next buflen and
583 * position.
584 */
585 if (get_next_offset(td, io_u)) {
586 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
587 return 1;
588 }
589
590 io_u->buflen = get_next_buflen(td, io_u);
591 if (!io_u->buflen) {
592 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
593 return 1;
594 }
595
596 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
597 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
598 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
599 io_u->buflen, io_u->file->real_file_size);
600 return 1;
601 }
602
603 /*
604 * mark entry before potentially trimming io_u
605 */
606 if (td_random(td) && file_randommap(td, io_u->file))
607 mark_random_map(td, io_u);
608
609 /*
610 * If using a write iolog, store this entry.
611 */
612out:
613 dprint_io_u(io_u, "fill_io_u");
614 td->zone_bytes += io_u->buflen;
615 log_io_u(td, io_u);
616 return 0;
617}
618
619static void __io_u_mark_map(unsigned int *map, unsigned int nr)
620{
621 int index = 0;
622
623 switch (nr) {
624 default:
625 index = 6;
626 break;
627 case 33 ... 64:
628 index = 5;
629 break;
630 case 17 ... 32:
631 index = 4;
632 break;
633 case 9 ... 16:
634 index = 3;
635 break;
636 case 5 ... 8:
637 index = 2;
638 break;
639 case 1 ... 4:
640 index = 1;
641 case 0:
642 break;
643 }
644
645 map[index]++;
646}
647
648void io_u_mark_submit(struct thread_data *td, unsigned int nr)
649{
650 __io_u_mark_map(td->ts.io_u_submit, nr);
651 td->ts.total_submit++;
652}
653
654void io_u_mark_complete(struct thread_data *td, unsigned int nr)
655{
656 __io_u_mark_map(td->ts.io_u_complete, nr);
657 td->ts.total_complete++;
658}
659
660void io_u_mark_depth(struct thread_data *td, unsigned int nr)
661{
662 int index = 0;
663
664 switch (td->cur_depth) {
665 default:
666 index = 6;
667 break;
668 case 32 ... 63:
669 index = 5;
670 break;
671 case 16 ... 31:
672 index = 4;
673 break;
674 case 8 ... 15:
675 index = 3;
676 break;
677 case 4 ... 7:
678 index = 2;
679 break;
680 case 2 ... 3:
681 index = 1;
682 case 1:
683 break;
684 }
685
686 td->ts.io_u_map[index] += nr;
687}
688
689static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
690{
691 int index = 0;
692
693 assert(usec < 1000);
694
695 switch (usec) {
696 case 750 ... 999:
697 index = 9;
698 break;
699 case 500 ... 749:
700 index = 8;
701 break;
702 case 250 ... 499:
703 index = 7;
704 break;
705 case 100 ... 249:
706 index = 6;
707 break;
708 case 50 ... 99:
709 index = 5;
710 break;
711 case 20 ... 49:
712 index = 4;
713 break;
714 case 10 ... 19:
715 index = 3;
716 break;
717 case 4 ... 9:
718 index = 2;
719 break;
720 case 2 ... 3:
721 index = 1;
722 case 0 ... 1:
723 break;
724 }
725
726 assert(index < FIO_IO_U_LAT_U_NR);
727 td->ts.io_u_lat_u[index]++;
728}
729
730static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
731{
732 int index = 0;
733
734 switch (msec) {
735 default:
736 index = 11;
737 break;
738 case 1000 ... 1999:
739 index = 10;
740 break;
741 case 750 ... 999:
742 index = 9;
743 break;
744 case 500 ... 749:
745 index = 8;
746 break;
747 case 250 ... 499:
748 index = 7;
749 break;
750 case 100 ... 249:
751 index = 6;
752 break;
753 case 50 ... 99:
754 index = 5;
755 break;
756 case 20 ... 49:
757 index = 4;
758 break;
759 case 10 ... 19:
760 index = 3;
761 break;
762 case 4 ... 9:
763 index = 2;
764 break;
765 case 2 ... 3:
766 index = 1;
767 case 0 ... 1:
768 break;
769 }
770
771 assert(index < FIO_IO_U_LAT_M_NR);
772 td->ts.io_u_lat_m[index]++;
773}
774
775static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
776{
777 if (usec < 1000)
778 io_u_mark_lat_usec(td, usec);
779 else
780 io_u_mark_lat_msec(td, usec / 1000);
781}
782
783/*
784 * Get next file to service by choosing one at random
785 */
786static struct fio_file *get_next_file_rand(struct thread_data *td,
787 enum fio_file_flags goodf,
788 enum fio_file_flags badf)
789{
790 struct fio_file *f;
791 int fno;
792
793 do {
794 long r = os_random_long(&td->next_file_state);
795 int opened = 0;
796
797 fno = (unsigned int) ((double) td->o.nr_files
798 * (r / (OS_RAND_MAX + 1.0)));
799 f = td->files[fno];
800 if (fio_file_done(f))
801 continue;
802
803 if (!fio_file_open(f)) {
804 int err;
805
806 err = td_io_open_file(td, f);
807 if (err)
808 continue;
809 opened = 1;
810 }
811
812 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
813 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
814 return f;
815 }
816 if (opened)
817 td_io_close_file(td, f);
818 } while (1);
819}
820
821/*
822 * Get next file to service by doing round robin between all available ones
823 */
824static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
825 int badf)
826{
827 unsigned int old_next_file = td->next_file;
828 struct fio_file *f;
829
830 do {
831 int opened = 0;
832
833 f = td->files[td->next_file];
834
835 td->next_file++;
836 if (td->next_file >= td->o.nr_files)
837 td->next_file = 0;
838
839 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
840 if (fio_file_done(f)) {
841 f = NULL;
842 continue;
843 }
844
845 if (!fio_file_open(f)) {
846 int err;
847
848 err = td_io_open_file(td, f);
849 if (err) {
850 dprint(FD_FILE, "error %d on open of %s\n",
851 err, f->file_name);
852 f = NULL;
853 continue;
854 }
855 opened = 1;
856 }
857
858 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
859 f->flags);
860 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
861 break;
862
863 if (opened)
864 td_io_close_file(td, f);
865
866 f = NULL;
867 } while (td->next_file != old_next_file);
868
869 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
870 return f;
871}
872
873static struct fio_file *__get_next_file(struct thread_data *td)
874{
875 struct fio_file *f;
876
877 assert(td->o.nr_files <= td->files_index);
878
879 if (td->nr_done_files >= td->o.nr_files) {
880 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
881 " nr_files=%d\n", td->nr_open_files,
882 td->nr_done_files,
883 td->o.nr_files);
884 return NULL;
885 }
886
887 f = td->file_service_file;
888 if (f && fio_file_open(f) && !fio_file_closing(f)) {
889 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
890 goto out;
891 if (td->file_service_left--)
892 goto out;
893 }
894
895 if (td->o.file_service_type == FIO_FSERVICE_RR ||
896 td->o.file_service_type == FIO_FSERVICE_SEQ)
897 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
898 else
899 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
900
901 td->file_service_file = f;
902 td->file_service_left = td->file_service_nr - 1;
903out:
904 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
905 return f;
906}
907
908static struct fio_file *get_next_file(struct thread_data *td)
909{
910 struct prof_io_ops *ops = &td->prof_io_ops;
911
912 if (ops->get_next_file)
913 return ops->get_next_file(td);
914
915 return __get_next_file(td);
916}
917
918static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
919{
920 struct fio_file *f;
921
922 do {
923 f = get_next_file(td);
924 if (!f)
925 return 1;
926
927 io_u->file = f;
928 get_file(f);
929
930 if (!fill_io_u(td, io_u))
931 break;
932
933 put_file_log(td, f);
934 td_io_close_file(td, f);
935 io_u->file = NULL;
936 fio_file_set_done(f);
937 td->nr_done_files++;
938 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
939 td->nr_done_files, td->o.nr_files);
940 } while (1);
941
942 return 0;
943}
944
945
946struct io_u *__get_io_u(struct thread_data *td)
947{
948 struct io_u *io_u = NULL;
949
950 td_io_u_lock(td);
951
952again:
953 if (!flist_empty(&td->io_u_requeues))
954 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
955 else if (!queue_full(td)) {
956 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
957
958 io_u->buflen = 0;
959 io_u->resid = 0;
960 io_u->file = NULL;
961 io_u->end_io = NULL;
962 }
963
964 if (io_u) {
965 assert(io_u->flags & IO_U_F_FREE);
966 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
967
968 io_u->error = 0;
969 flist_del(&io_u->list);
970 flist_add(&io_u->list, &td->io_u_busylist);
971 td->cur_depth++;
972 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
973 } else if (td->o.verify_async) {
974 /*
975 * We ran out, wait for async verify threads to finish and
976 * return one
977 */
978 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
979 goto again;
980 }
981
982 td_io_u_unlock(td);
983 return io_u;
984}
985
986static int check_get_trim(struct thread_data *td, struct io_u *io_u)
987{
988 if (td->o.trim_backlog && td->trim_entries) {
989 int get_trim = 0;
990
991 if (td->trim_batch) {
992 td->trim_batch--;
993 get_trim = 1;
994 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
995 td->last_ddir != DDIR_READ) {
996 td->trim_batch = td->o.trim_batch;
997 if (!td->trim_batch)
998 td->trim_batch = td->o.trim_backlog;
999 get_trim = 1;
1000 }
1001
1002 if (get_trim && !get_next_trim(td, io_u))
1003 return 1;
1004 }
1005
1006 return 0;
1007}
1008
1009static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1010{
1011 if (td->o.verify_backlog && td->io_hist_len) {
1012 int get_verify = 0;
1013
1014 if (td->verify_batch) {
1015 td->verify_batch--;
1016 get_verify = 1;
1017 } else if (!(td->io_hist_len % td->o.verify_backlog) &&
1018 td->last_ddir != DDIR_READ) {
1019 td->verify_batch = td->o.verify_batch;
1020 if (!td->verify_batch)
1021 td->verify_batch = td->o.verify_backlog;
1022 get_verify = 1;
1023 }
1024
1025 if (get_verify && !get_next_verify(td, io_u))
1026 return 1;
1027 }
1028
1029 return 0;
1030}
1031
1032/*
1033 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1034 * etc. The returned io_u is fully ready to be prepped and submitted.
1035 */
1036struct io_u *get_io_u(struct thread_data *td)
1037{
1038 struct fio_file *f;
1039 struct io_u *io_u;
1040
1041 io_u = __get_io_u(td);
1042 if (!io_u) {
1043 dprint(FD_IO, "__get_io_u failed\n");
1044 return NULL;
1045 }
1046
1047 if (check_get_verify(td, io_u))
1048 goto out;
1049 if (check_get_trim(td, io_u))
1050 goto out;
1051
1052 /*
1053 * from a requeue, io_u already setup
1054 */
1055 if (io_u->file)
1056 goto out;
1057
1058 /*
1059 * If using an iolog, grab next piece if any available.
1060 */
1061 if (td->o.read_iolog_file) {
1062 if (read_iolog_get(td, io_u))
1063 goto err_put;
1064 } else if (set_io_u_file(td, io_u)) {
1065 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1066 goto err_put;
1067 }
1068
1069 f = io_u->file;
1070 assert(fio_file_open(f));
1071
1072 if (ddir_rw(io_u->ddir)) {
1073 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1074 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1075 goto err_put;
1076 }
1077
1078 f->last_start = io_u->offset;
1079 f->last_pos = io_u->offset + io_u->buflen;
1080
1081 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
1082 populate_verify_io_u(td, io_u);
1083 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
1084 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
1085 else if (io_u->ddir == DDIR_READ) {
1086 /*
1087 * Reset the buf_filled parameters so next time if the
1088 * buffer is used for writes it is refilled.
1089 */
1090 io_u->buf_filled_len = 0;
1091 }
1092 }
1093
1094 /*
1095 * Set io data pointers.
1096 */
1097 io_u->xfer_buf = io_u->buf;
1098 io_u->xfer_buflen = io_u->buflen;
1099
1100out:
1101 assert(io_u->file);
1102 if (!td_io_prep(td, io_u)) {
1103 if (!td->o.disable_slat)
1104 fio_gettime(&io_u->start_time, NULL);
1105 return io_u;
1106 }
1107err_put:
1108 dprint(FD_IO, "get_io_u failed\n");
1109 put_io_u(td, io_u);
1110 return NULL;
1111}
1112
1113void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1114{
1115 const char *msg[] = { "read", "write", "sync", "datasync",
1116 "sync_file_range", "wait", "trim" };
1117
1118
1119
1120 log_err("fio: io_u error");
1121
1122 if (io_u->file)
1123 log_err(" on file %s", io_u->file->file_name);
1124
1125 log_err(": %s\n", strerror(io_u->error));
1126
1127 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1128 io_u->offset, io_u->xfer_buflen);
1129
1130 if (!td->error)
1131 td_verror(td, io_u->error, "io_u error");
1132}
1133
1134static void io_completed(struct thread_data *td, struct io_u *io_u,
1135 struct io_completion_data *icd)
1136{
1137 /*
1138 * Older gcc's are too dumb to realize that usec is always used
1139 * initialized, silence that warning.
1140 */
1141 unsigned long uninitialized_var(usec);
1142 struct fio_file *f;
1143
1144 dprint_io_u(io_u, "io complete");
1145
1146 td_io_u_lock(td);
1147 assert(io_u->flags & IO_U_F_FLIGHT);
1148 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1149 td_io_u_unlock(td);
1150
1151 if (ddir_sync(io_u->ddir)) {
1152 td->last_was_sync = 1;
1153 f = io_u->file;
1154 if (f) {
1155 f->first_write = -1ULL;
1156 f->last_write = -1ULL;
1157 }
1158 return;
1159 }
1160
1161 td->last_was_sync = 0;
1162 td->last_ddir = io_u->ddir;
1163
1164 if (!io_u->error && ddir_rw(io_u->ddir)) {
1165 unsigned int bytes = io_u->buflen - io_u->resid;
1166 const enum fio_ddir idx = io_u->ddir;
1167 const enum fio_ddir odx = io_u->ddir ^ 1;
1168 int ret;
1169
1170 td->io_blocks[idx]++;
1171 td->io_bytes[idx] += bytes;
1172 td->this_io_bytes[idx] += bytes;
1173
1174 if (idx == DDIR_WRITE) {
1175 f = io_u->file;
1176 if (f) {
1177 if (f->first_write == -1ULL ||
1178 io_u->offset < f->first_write)
1179 f->first_write = io_u->offset;
1180 if (f->last_write == -1ULL ||
1181 ((io_u->offset + bytes) > f->last_write))
1182 f->last_write = io_u->offset + bytes;
1183 }
1184 }
1185
1186 if (ramp_time_over(td)) {
1187 unsigned long uninitialized_var(lusec);
1188
1189 if (!td->o.disable_clat || !td->o.disable_bw)
1190 lusec = utime_since(&io_u->issue_time,
1191 &icd->time);
1192 if (!td->o.disable_lat) {
1193 unsigned long tusec;
1194
1195 tusec = utime_since(&io_u->start_time,
1196 &icd->time);
1197 add_lat_sample(td, idx, tusec, bytes);
1198 }
1199 if (!td->o.disable_clat) {
1200 add_clat_sample(td, idx, lusec, bytes);
1201 io_u_mark_latency(td, lusec);
1202 }
1203 if (!td->o.disable_bw)
1204 add_bw_sample(td, idx, bytes, &icd->time);
1205 if (__should_check_rate(td, idx)) {
1206 td->rate_pending_usleep[idx] =
1207 ((td->this_io_bytes[idx] *
1208 td->rate_nsec_cycle[idx]) / 1000 -
1209 utime_since_now(&td->start));
1210 }
1211 if (__should_check_rate(td, idx ^ 1))
1212 td->rate_pending_usleep[odx] =
1213 ((td->this_io_bytes[odx] *
1214 td->rate_nsec_cycle[odx]) / 1000 -
1215 utime_since_now(&td->start));
1216 }
1217
1218 if (td_write(td) && idx == DDIR_WRITE &&
1219 td->o.do_verify &&
1220 td->o.verify != VERIFY_NONE)
1221 log_io_piece(td, io_u);
1222
1223 icd->bytes_done[idx] += bytes;
1224
1225 if (io_u->end_io) {
1226 ret = io_u->end_io(td, io_u);
1227 if (ret && !icd->error)
1228 icd->error = ret;
1229 }
1230 } else if (io_u->error) {
1231 icd->error = io_u->error;
1232 io_u_log_error(td, io_u);
1233 }
1234 if (td->o.continue_on_error && icd->error &&
1235 td_non_fatal_error(icd->error)) {
1236 /*
1237 * If there is a non_fatal error, then add to the error count
1238 * and clear all the errors.
1239 */
1240 update_error_count(td, icd->error);
1241 td_clear_error(td);
1242 icd->error = 0;
1243 io_u->error = 0;
1244 }
1245}
1246
1247static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1248 int nr)
1249{
1250 if (!td->o.disable_clat || !td->o.disable_bw)
1251 fio_gettime(&icd->time, NULL);
1252
1253 icd->nr = nr;
1254
1255 icd->error = 0;
1256 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1257}
1258
1259static void ios_completed(struct thread_data *td,
1260 struct io_completion_data *icd)
1261{
1262 struct io_u *io_u;
1263 int i;
1264
1265 for (i = 0; i < icd->nr; i++) {
1266 io_u = td->io_ops->event(td, i);
1267
1268 io_completed(td, io_u, icd);
1269
1270 if (!(io_u->flags & IO_U_F_FREE_DEF))
1271 put_io_u(td, io_u);
1272 }
1273}
1274
1275/*
1276 * Complete a single io_u for the sync engines.
1277 */
1278int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1279 unsigned long *bytes)
1280{
1281 struct io_completion_data icd;
1282
1283 init_icd(td, &icd, 1);
1284 io_completed(td, io_u, &icd);
1285
1286 if (!(io_u->flags & IO_U_F_FREE_DEF))
1287 put_io_u(td, io_u);
1288
1289 if (icd.error) {
1290 td_verror(td, icd.error, "io_u_sync_complete");
1291 return -1;
1292 }
1293
1294 if (bytes) {
1295 bytes[0] += icd.bytes_done[0];
1296 bytes[1] += icd.bytes_done[1];
1297 }
1298
1299 return 0;
1300}
1301
1302/*
1303 * Called to complete min_events number of io for the async engines.
1304 */
1305int io_u_queued_complete(struct thread_data *td, int min_evts,
1306 unsigned long *bytes)
1307{
1308 struct io_completion_data icd;
1309 struct timespec *tvp = NULL;
1310 int ret;
1311 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1312
1313 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1314
1315 if (!min_evts)
1316 tvp = &ts;
1317
1318 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1319 if (ret < 0) {
1320 td_verror(td, -ret, "td_io_getevents");
1321 return ret;
1322 } else if (!ret)
1323 return ret;
1324
1325 init_icd(td, &icd, ret);
1326 ios_completed(td, &icd);
1327 if (icd.error) {
1328 td_verror(td, icd.error, "io_u_queued_complete");
1329 return -1;
1330 }
1331
1332 if (bytes) {
1333 bytes[0] += icd.bytes_done[0];
1334 bytes[1] += icd.bytes_done[1];
1335 }
1336
1337 return 0;
1338}
1339
1340/*
1341 * Call when io_u is really queued, to update the submission latency.
1342 */
1343void io_u_queued(struct thread_data *td, struct io_u *io_u)
1344{
1345 if (!td->o.disable_slat) {
1346 unsigned long slat_time;
1347
1348 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1349 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1350 }
1351}
1352
1353/*
1354 * "randomly" fill the buffer contents
1355 */
1356void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1357 unsigned int max_bs)
1358{
1359 if (!td->o.zero_buffers)
1360 fill_random_buf(io_u->buf, max_bs);
1361 else
1362 memset(io_u->buf, 0, max_bs);
1363}
fio - Flexible IO Tester