projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Fix end-of-job slowdown for random IO with a random map
[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] & (1UL << 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 idx, bit;
54 unsigned long mask, this_blocks;
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 = -1UL;
79 else
80 mask = ((1UL << 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 block, min_bs = td->o.rw_min_bs, lastb;
130 int i;
131
132 lastb = last_block(td, f, ddir);
133 if (!lastb)
134 return 1;
135
136 i = f->last_free_lookup;
137 block = i * BLOCKS_PER_MAP;
138 while (block * min_bs < f->real_file_size &&
139 block * min_bs < f->io_size) {
140 if (f->file_map[i] != -1UL) {
141 block += ffz(f->file_map[i]);
142 if (block > lastb)
143 break;
144 f->last_free_lookup = i;
145 *b = block;
146 return 0;
147 }
148
149 block += BLOCKS_PER_MAP;
150 i++;
151 }
152
153 dprint(FD_IO, "failed finding a free block\n");
154 return 1;
155}
156
157static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
158 enum fio_ddir ddir, unsigned long long *b)
159{
160 unsigned long long r, lastb;
161 int loops = 5;
162
163 lastb = last_block(td, f, ddir);
164 if (!lastb)
165 return 1;
166
167 if (f->failed_rands >= 200)
168 goto ffz;
169
170 do {
171 r = os_random_long(&td->random_state);
172 dprint(FD_RANDOM, "off rand %llu\n", r);
173 *b = (lastb - 1) * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
174
175 /*
176 * if we are not maintaining a random map, we are done.
177 */
178 if (!file_randommap(td, f))
179 goto ret_good;
180
181 /*
182 * calculate map offset and check if it's free
183 */
184 if (random_map_free(f, *b))
185 goto ret_good;
186
187 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
188 *b);
189 } while (--loops);
190
191 if (!f->failed_rands++)
192 f->last_free_lookup = 0;
193
194 /*
195 * we get here, if we didn't suceed in looking up a block. generate
196 * a random start offset into the filemap, and find the first free
197 * block from there.
198 */
199 loops = 10;
200 do {
201 f->last_free_lookup = (f->num_maps - 1) *
202 (r / (OS_RAND_MAX + 1.0));
203 if (!get_next_free_block(td, f, ddir, b))
204 goto ret;
205
206 r = os_random_long(&td->random_state);
207 } while (--loops);
208
209 /*
210 * that didn't work either, try exhaustive search from the start
211 */
212 f->last_free_lookup = 0;
213ffz:
214 if (!get_next_free_block(td, f, ddir, b))
215 return 0;
216 f->last_free_lookup = 0;
217 return get_next_free_block(td, f, ddir, b);
218ret_good:
219 f->failed_rands = 0;
220ret:
221 return 0;
222}
223
224static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
225 enum fio_ddir ddir, unsigned long long *b)
226{
227 if (get_next_rand_offset(td, f, ddir, b)) {
228 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
229 f->file_name, f->last_pos, f->real_file_size);
230 return 1;
231 }
232
233 return 0;
234}
235
236static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
237 enum fio_ddir ddir, unsigned long long *b)
238{
239 assert(ddir_rw(ddir));
240
241 if (f->last_pos < f->real_file_size) {
242 *b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
243 return 0;
244 }
245
246 return 1;
247}
248
249static int get_next_block(struct thread_data *td, struct io_u *io_u,
250 enum fio_ddir ddir, int rw_seq, unsigned long long *b)
251{
252 struct fio_file *f = io_u->file;
253 int ret;
254
255 assert(ddir_rw(ddir));
256
257 if (rw_seq) {
258 if (td_random(td))
259 ret = get_next_rand_block(td, f, ddir, b);
260 else
261 ret = get_next_seq_block(td, f, ddir, b);
262 } else {
263 io_u->flags |= IO_U_F_BUSY_OK;
264
265 if (td->o.rw_seq == RW_SEQ_SEQ) {
266 ret = get_next_seq_block(td, f, ddir, b);
267 if (ret)
268 ret = get_next_rand_block(td, f, ddir, b);
269 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
270 if (f->last_start != -1ULL)
271 *b = (f->last_start - f->file_offset)
272 / td->o.min_bs[ddir];
273 else
274 *b = 0;
275 ret = 0;
276 } else {
277 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
278 ret = 1;
279 }
280 }
281
282 return ret;
283}
284
285/*
286 * For random io, generate a random new block and see if it's used. Repeat
287 * until we find a free one. For sequential io, just return the end of
288 * the last io issued.
289 */
290static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
291{
292 struct fio_file *f = io_u->file;
293 unsigned long long b;
294 enum fio_ddir ddir = io_u->ddir;
295 int rw_seq_hit = 0;
296
297 assert(ddir_rw(ddir));
298
299 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
300 rw_seq_hit = 1;
301 td->ddir_seq_nr = td->o.ddir_seq_nr;
302 }
303
304 if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
305 return 1;
306
307 io_u->offset = b * td->o.ba[ddir];
308 if (io_u->offset >= f->io_size) {
309 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
310 io_u->offset, f->io_size);
311 return 1;
312 }
313
314 io_u->offset += f->file_offset;
315 if (io_u->offset >= f->real_file_size) {
316 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
317 io_u->offset, f->real_file_size);
318 return 1;
319 }
320
321 return 0;
322}
323
324static int get_next_offset(struct thread_data *td, struct io_u *io_u)
325{
326 struct prof_io_ops *ops = &td->prof_io_ops;
327
328 if (ops->fill_io_u_off)
329 return ops->fill_io_u_off(td, io_u);
330
331 return __get_next_offset(td, io_u);
332}
333
334static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
335{
336 const int ddir = io_u->ddir;
337 unsigned int uninitialized_var(buflen);
338 unsigned int minbs, maxbs;
339 long r;
340
341 assert(ddir_rw(ddir));
342
343 minbs = td->o.min_bs[ddir];
344 maxbs = td->o.max_bs[ddir];
345
346 if (minbs == maxbs)
347 buflen = minbs;
348 else {
349 r = os_random_long(&td->bsrange_state);
350 if (!td->o.bssplit_nr[ddir]) {
351 buflen = 1 + (unsigned int) ((double) maxbs *
352 (r / (OS_RAND_MAX + 1.0)));
353 if (buflen < minbs)
354 buflen = minbs;
355 } else {
356 long perc = 0;
357 unsigned int i;
358
359 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
360 struct bssplit *bsp = &td->o.bssplit[ddir][i];
361
362 buflen = bsp->bs;
363 perc += bsp->perc;
364 if (r <= ((OS_RAND_MAX / 100L) * perc))
365 break;
366 }
367 }
368 if (!td->o.bs_unaligned && is_power_of_2(minbs))
369 buflen = (buflen + minbs - 1) & ~(minbs - 1);
370 }
371
372 if (io_u->offset + buflen > io_u->file->real_file_size) {
373 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
374 minbs, ddir);
375 buflen = minbs;
376 }
377
378 return buflen;
379}
380
381static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
382{
383 struct prof_io_ops *ops = &td->prof_io_ops;
384
385 if (ops->fill_io_u_size)
386 return ops->fill_io_u_size(td, io_u);
387
388 return __get_next_buflen(td, io_u);
389}
390
391static void set_rwmix_bytes(struct thread_data *td)
392{
393 unsigned int diff;
394
395 /*
396 * we do time or byte based switch. this is needed because
397 * buffered writes may issue a lot quicker than they complete,
398 * whereas reads do not.
399 */
400 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
401 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
402}
403
404static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
405{
406 unsigned int v;
407 long r;
408
409 r = os_random_long(&td->rwmix_state);
410 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
411 if (v <= td->o.rwmix[DDIR_READ])
412 return DDIR_READ;
413
414 return DDIR_WRITE;
415}
416
417static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
418{
419 enum fio_ddir odir = ddir ^ 1;
420 struct timeval t;
421 long usec;
422
423 assert(ddir_rw(ddir));
424
425 if (td->rate_pending_usleep[ddir] <= 0)
426 return ddir;
427
428 /*
429 * We have too much pending sleep in this direction. See if we
430 * should switch.
431 */
432 if (td_rw(td)) {
433 /*
434 * Other direction does not have too much pending, switch
435 */
436 if (td->rate_pending_usleep[odir] < 100000)
437 return odir;
438
439 /*
440 * Both directions have pending sleep. Sleep the minimum time
441 * and deduct from both.
442 */
443 if (td->rate_pending_usleep[ddir] <=
444 td->rate_pending_usleep[odir]) {
445 usec = td->rate_pending_usleep[ddir];
446 } else {
447 usec = td->rate_pending_usleep[odir];
448 ddir = odir;
449 }
450 } else
451 usec = td->rate_pending_usleep[ddir];
452
453 fio_gettime(&t, NULL);
454 usec_sleep(td, usec);
455 usec = utime_since_now(&t);
456
457 td->rate_pending_usleep[ddir] -= usec;
458
459 odir = ddir ^ 1;
460 if (td_rw(td) && __should_check_rate(td, odir))
461 td->rate_pending_usleep[odir] -= usec;
462
463 return ddir;
464}
465
466/*
467 * Return the data direction for the next io_u. If the job is a
468 * mixed read/write workload, check the rwmix cycle and switch if
469 * necessary.
470 */
471static enum fio_ddir get_rw_ddir(struct thread_data *td)
472{
473 enum fio_ddir ddir;
474
475 /*
476 * see if it's time to fsync
477 */
478 if (td->o.fsync_blocks &&
479 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
480 td->io_issues[DDIR_WRITE] && should_fsync(td))
481 return DDIR_SYNC;
482
483 /*
484 * see if it's time to fdatasync
485 */
486 if (td->o.fdatasync_blocks &&
487 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
488 td->io_issues[DDIR_WRITE] && should_fsync(td))
489 return DDIR_DATASYNC;
490
491 /*
492 * see if it's time to sync_file_range
493 */
494 if (td->sync_file_range_nr &&
495 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
496 td->io_issues[DDIR_WRITE] && should_fsync(td))
497 return DDIR_SYNC_FILE_RANGE;
498
499 if (td_rw(td)) {
500 /*
501 * Check if it's time to seed a new data direction.
502 */
503 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
504 /*
505 * Put a top limit on how many bytes we do for
506 * one data direction, to avoid overflowing the
507 * ranges too much
508 */
509 ddir = get_rand_ddir(td);
510
511 if (ddir != td->rwmix_ddir)
512 set_rwmix_bytes(td);
513
514 td->rwmix_ddir = ddir;
515 }
516 ddir = td->rwmix_ddir;
517 } else if (td_read(td))
518 ddir = DDIR_READ;
519 else
520 ddir = DDIR_WRITE;
521
522 td->rwmix_ddir = rate_ddir(td, ddir);
523 return td->rwmix_ddir;
524}
525
526static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
527{
528 io_u->ddir = get_rw_ddir(td);
529
530 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
531 td->o.barrier_blocks &&
532 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
533 td->io_issues[DDIR_WRITE])
534 io_u->flags |= IO_U_F_BARRIER;
535}
536
537void put_file_log(struct thread_data *td, struct fio_file *f)
538{
539 int ret = put_file(td, f);
540
541 if (ret)
542 td_verror(td, ret, "file close");
543}
544
545void put_io_u(struct thread_data *td, struct io_u *io_u)
546{
547 td_io_u_lock(td);
548
549 io_u->flags |= IO_U_F_FREE;
550 io_u->flags &= ~IO_U_F_FREE_DEF;
551
552 if (io_u->file)
553 put_file_log(td, io_u->file);
554
555 io_u->file = NULL;
556 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
557 td->cur_depth--;
558 flist_del_init(&io_u->list);
559 flist_add(&io_u->list, &td->io_u_freelist);
560 td_io_u_unlock(td);
561 td_io_u_free_notify(td);
562}
563
564void clear_io_u(struct thread_data *td, struct io_u *io_u)
565{
566 io_u->flags &= ~IO_U_F_FLIGHT;
567 put_io_u(td, io_u);
568}
569
570void requeue_io_u(struct thread_data *td, struct io_u **io_u)
571{
572 struct io_u *__io_u = *io_u;
573
574 dprint(FD_IO, "requeue %p\n", __io_u);
575
576 td_io_u_lock(td);
577
578 __io_u->flags |= IO_U_F_FREE;
579 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
580 td->io_issues[__io_u->ddir]--;
581
582 __io_u->flags &= ~IO_U_F_FLIGHT;
583 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
584 td->cur_depth--;
585 flist_del(&__io_u->list);
586 flist_add_tail(&__io_u->list, &td->io_u_requeues);
587 td_io_u_unlock(td);
588 *io_u = NULL;
589}
590
591static int fill_io_u(struct thread_data *td, struct io_u *io_u)
592{
593 if (td->io_ops->flags & FIO_NOIO)
594 goto out;
595
596 set_rw_ddir(td, io_u);
597
598 /*
599 * fsync() or fdatasync() or trim etc, we are done
600 */
601 if (!ddir_rw(io_u->ddir))
602 goto out;
603
604 /*
605 * See if it's time to switch to a new zone
606 */
607 if (td->zone_bytes >= td->o.zone_size) {
608 td->zone_bytes = 0;
609 io_u->file->last_pos += td->o.zone_skip;
610 td->io_skip_bytes += td->o.zone_skip;
611 }
612
613 /*
614 * No log, let the seq/rand engine retrieve the next buflen and
615 * position.
616 */
617 if (get_next_offset(td, io_u)) {
618 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
619 return 1;
620 }
621
622 io_u->buflen = get_next_buflen(td, io_u);
623 if (!io_u->buflen) {
624 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
625 return 1;
626 }
627
628 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
629 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
630 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
631 io_u->buflen, io_u->file->real_file_size);
632 return 1;
633 }
634
635 /*
636 * mark entry before potentially trimming io_u
637 */
638 if (td_random(td) && file_randommap(td, io_u->file))
639 mark_random_map(td, io_u);
640
641 /*
642 * If using a write iolog, store this entry.
643 */
644out:
645 dprint_io_u(io_u, "fill_io_u");
646 td->zone_bytes += io_u->buflen;
647 log_io_u(td, io_u);
648 return 0;
649}
650
651static void __io_u_mark_map(unsigned int *map, unsigned int nr)
652{
653 int idx = 0;
654
655 switch (nr) {
656 default:
657 idx = 6;
658 break;
659 case 33 ... 64:
660 idx = 5;
661 break;
662 case 17 ... 32:
663 idx = 4;
664 break;
665 case 9 ... 16:
666 idx = 3;
667 break;
668 case 5 ... 8:
669 idx = 2;
670 break;
671 case 1 ... 4:
672 idx = 1;
673 case 0:
674 break;
675 }
676
677 map[idx]++;
678}
679
680void io_u_mark_submit(struct thread_data *td, unsigned int nr)
681{
682 __io_u_mark_map(td->ts.io_u_submit, nr);
683 td->ts.total_submit++;
684}
685
686void io_u_mark_complete(struct thread_data *td, unsigned int nr)
687{
688 __io_u_mark_map(td->ts.io_u_complete, nr);
689 td->ts.total_complete++;
690}
691
692void io_u_mark_depth(struct thread_data *td, unsigned int nr)
693{
694 int idx = 0;
695
696 switch (td->cur_depth) {
697 default:
698 idx = 6;
699 break;
700 case 32 ... 63:
701 idx = 5;
702 break;
703 case 16 ... 31:
704 idx = 4;
705 break;
706 case 8 ... 15:
707 idx = 3;
708 break;
709 case 4 ... 7:
710 idx = 2;
711 break;
712 case 2 ... 3:
713 idx = 1;
714 case 1:
715 break;
716 }
717
718 td->ts.io_u_map[idx] += nr;
719}
720
721static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
722{
723 int idx = 0;
724
725 assert(usec < 1000);
726
727 switch (usec) {
728 case 750 ... 999:
729 idx = 9;
730 break;
731 case 500 ... 749:
732 idx = 8;
733 break;
734 case 250 ... 499:
735 idx = 7;
736 break;
737 case 100 ... 249:
738 idx = 6;
739 break;
740 case 50 ... 99:
741 idx = 5;
742 break;
743 case 20 ... 49:
744 idx = 4;
745 break;
746 case 10 ... 19:
747 idx = 3;
748 break;
749 case 4 ... 9:
750 idx = 2;
751 break;
752 case 2 ... 3:
753 idx = 1;
754 case 0 ... 1:
755 break;
756 }
757
758 assert(idx < FIO_IO_U_LAT_U_NR);
759 td->ts.io_u_lat_u[idx]++;
760}
761
762static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
763{
764 int idx = 0;
765
766 switch (msec) {
767 default:
768 idx = 11;
769 break;
770 case 1000 ... 1999:
771 idx = 10;
772 break;
773 case 750 ... 999:
774 idx = 9;
775 break;
776 case 500 ... 749:
777 idx = 8;
778 break;
779 case 250 ... 499:
780 idx = 7;
781 break;
782 case 100 ... 249:
783 idx = 6;
784 break;
785 case 50 ... 99:
786 idx = 5;
787 break;
788 case 20 ... 49:
789 idx = 4;
790 break;
791 case 10 ... 19:
792 idx = 3;
793 break;
794 case 4 ... 9:
795 idx = 2;
796 break;
797 case 2 ... 3:
798 idx = 1;
799 case 0 ... 1:
800 break;
801 }
802
803 assert(idx < FIO_IO_U_LAT_M_NR);
804 td->ts.io_u_lat_m[idx]++;
805}
806
807static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
808{
809 if (usec < 1000)
810 io_u_mark_lat_usec(td, usec);
811 else
812 io_u_mark_lat_msec(td, usec / 1000);
813}
814
815/*
816 * Get next file to service by choosing one at random
817 */
818static struct fio_file *get_next_file_rand(struct thread_data *td,
819 enum fio_file_flags goodf,
820 enum fio_file_flags badf)
821{
822 struct fio_file *f;
823 int fno;
824
825 do {
826 long r = os_random_long(&td->next_file_state);
827 int opened = 0;
828
829 fno = (unsigned int) ((double) td->o.nr_files
830 * (r / (OS_RAND_MAX + 1.0)));
831 f = td->files[fno];
832 if (fio_file_done(f))
833 continue;
834
835 if (!fio_file_open(f)) {
836 int err;
837
838 err = td_io_open_file(td, f);
839 if (err)
840 continue;
841 opened = 1;
842 }
843
844 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
845 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
846 return f;
847 }
848 if (opened)
849 td_io_close_file(td, f);
850 } while (1);
851}
852
853/*
854 * Get next file to service by doing round robin between all available ones
855 */
856static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
857 int badf)
858{
859 unsigned int old_next_file = td->next_file;
860 struct fio_file *f;
861
862 do {
863 int opened = 0;
864
865 f = td->files[td->next_file];
866
867 td->next_file++;
868 if (td->next_file >= td->o.nr_files)
869 td->next_file = 0;
870
871 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
872 if (fio_file_done(f)) {
873 f = NULL;
874 continue;
875 }
876
877 if (!fio_file_open(f)) {
878 int err;
879
880 err = td_io_open_file(td, f);
881 if (err) {
882 dprint(FD_FILE, "error %d on open of %s\n",
883 err, f->file_name);
884 f = NULL;
885 continue;
886 }
887 opened = 1;
888 }
889
890 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
891 f->flags);
892 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
893 break;
894
895 if (opened)
896 td_io_close_file(td, f);
897
898 f = NULL;
899 } while (td->next_file != old_next_file);
900
901 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
902 return f;
903}
904
905static struct fio_file *__get_next_file(struct thread_data *td)
906{
907 struct fio_file *f;
908
909 assert(td->o.nr_files <= td->files_index);
910
911 if (td->nr_done_files >= td->o.nr_files) {
912 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
913 " nr_files=%d\n", td->nr_open_files,
914 td->nr_done_files,
915 td->o.nr_files);
916 return NULL;
917 }
918
919 f = td->file_service_file;
920 if (f && fio_file_open(f) && !fio_file_closing(f)) {
921 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
922 goto out;
923 if (td->file_service_left--)
924 goto out;
925 }
926
927 if (td->o.file_service_type == FIO_FSERVICE_RR ||
928 td->o.file_service_type == FIO_FSERVICE_SEQ)
929 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
930 else
931 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
932
933 td->file_service_file = f;
934 td->file_service_left = td->file_service_nr - 1;
935out:
936 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
937 return f;
938}
939
940static struct fio_file *get_next_file(struct thread_data *td)
941{
942 struct prof_io_ops *ops = &td->prof_io_ops;
943
944 if (ops->get_next_file)
945 return ops->get_next_file(td);
946
947 return __get_next_file(td);
948}
949
950static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
951{
952 struct fio_file *f;
953
954 do {
955 f = get_next_file(td);
956 if (!f)
957 return 1;
958
959 io_u->file = f;
960 get_file(f);
961
962 if (!fill_io_u(td, io_u))
963 break;
964
965 put_file_log(td, f);
966 td_io_close_file(td, f);
967 io_u->file = NULL;
968 fio_file_set_done(f);
969 td->nr_done_files++;
970 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
971 td->nr_done_files, td->o.nr_files);
972 } while (1);
973
974 return 0;
975}
976
977
978struct io_u *__get_io_u(struct thread_data *td)
979{
980 struct io_u *io_u = NULL;
981
982 td_io_u_lock(td);
983
984again:
985 if (!flist_empty(&td->io_u_requeues))
986 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
987 else if (!queue_full(td)) {
988 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
989
990 io_u->buflen = 0;
991 io_u->resid = 0;
992 io_u->file = NULL;
993 io_u->end_io = NULL;
994 }
995
996 if (io_u) {
997 assert(io_u->flags & IO_U_F_FREE);
998 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
999 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1000
1001 io_u->error = 0;
1002 flist_del(&io_u->list);
1003 flist_add(&io_u->list, &td->io_u_busylist);
1004 td->cur_depth++;
1005 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1006 } else if (td->o.verify_async) {
1007 /*
1008 * We ran out, wait for async verify threads to finish and
1009 * return one
1010 */
1011 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1012 goto again;
1013 }
1014
1015 td_io_u_unlock(td);
1016 return io_u;
1017}
1018
1019static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1020{
1021 if (td->o.trim_backlog && td->trim_entries) {
1022 int get_trim = 0;
1023
1024 if (td->trim_batch) {
1025 td->trim_batch--;
1026 get_trim = 1;
1027 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1028 td->last_ddir != DDIR_READ) {
1029 td->trim_batch = td->o.trim_batch;
1030 if (!td->trim_batch)
1031 td->trim_batch = td->o.trim_backlog;
1032 get_trim = 1;
1033 }
1034
1035 if (get_trim && !get_next_trim(td, io_u))
1036 return 1;
1037 }
1038
1039 return 0;
1040}
1041
1042static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1043{
1044 if (td->o.verify_backlog && td->io_hist_len) {
1045 int get_verify = 0;
1046
1047 if (td->verify_batch) {
1048 td->verify_batch--;
1049 get_verify = 1;
1050 } else if (!(td->io_hist_len % td->o.verify_backlog) &&
1051 td->last_ddir != DDIR_READ) {
1052 td->verify_batch = td->o.verify_batch;
1053 if (!td->verify_batch)
1054 td->verify_batch = td->o.verify_backlog;
1055 get_verify = 1;
1056 }
1057
1058 if (get_verify && !get_next_verify(td, io_u))
1059 return 1;
1060 }
1061
1062 return 0;
1063}
1064
1065/*
1066 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1067 * etc. The returned io_u is fully ready to be prepped and submitted.
1068 */
1069struct io_u *get_io_u(struct thread_data *td)
1070{
1071 struct fio_file *f;
1072 struct io_u *io_u;
1073
1074 io_u = __get_io_u(td);
1075 if (!io_u) {
1076 dprint(FD_IO, "__get_io_u failed\n");
1077 return NULL;
1078 }
1079
1080 if (check_get_verify(td, io_u))
1081 goto out;
1082 if (check_get_trim(td, io_u))
1083 goto out;
1084
1085 /*
1086 * from a requeue, io_u already setup
1087 */
1088 if (io_u->file)
1089 goto out;
1090
1091 /*
1092 * If using an iolog, grab next piece if any available.
1093 */
1094 if (td->o.read_iolog_file) {
1095 if (read_iolog_get(td, io_u))
1096 goto err_put;
1097 } else if (set_io_u_file(td, io_u)) {
1098 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1099 goto err_put;
1100 }
1101
1102 f = io_u->file;
1103 assert(fio_file_open(f));
1104
1105 if (ddir_rw(io_u->ddir)) {
1106 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1107 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1108 goto err_put;
1109 }
1110
1111 f->last_start = io_u->offset;
1112 f->last_pos = io_u->offset + io_u->buflen;
1113
1114 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
1115 populate_verify_io_u(td, io_u);
1116 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
1117 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
1118 else if (io_u->ddir == DDIR_READ) {
1119 /*
1120 * Reset the buf_filled parameters so next time if the
1121 * buffer is used for writes it is refilled.
1122 */
1123 io_u->buf_filled_len = 0;
1124 }
1125 }
1126
1127 /*
1128 * Set io data pointers.
1129 */
1130 io_u->xfer_buf = io_u->buf;
1131 io_u->xfer_buflen = io_u->buflen;
1132
1133out:
1134 assert(io_u->file);
1135 if (!td_io_prep(td, io_u)) {
1136 if (!td->o.disable_slat)
1137 fio_gettime(&io_u->start_time, NULL);
1138 return io_u;
1139 }
1140err_put:
1141 dprint(FD_IO, "get_io_u failed\n");
1142 put_io_u(td, io_u);
1143 return NULL;
1144}
1145
1146void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1147{
1148 const char *msg[] = { "read", "write", "sync", "datasync",
1149 "sync_file_range", "wait", "trim" };
1150
1151
1152
1153 log_err("fio: io_u error");
1154
1155 if (io_u->file)
1156 log_err(" on file %s", io_u->file->file_name);
1157
1158 log_err(": %s\n", strerror(io_u->error));
1159
1160 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1161 io_u->offset, io_u->xfer_buflen);
1162
1163 if (!td->error)
1164 td_verror(td, io_u->error, "io_u error");
1165}
1166
1167static void io_completed(struct thread_data *td, struct io_u *io_u,
1168 struct io_completion_data *icd)
1169{
1170 /*
1171 * Older gcc's are too dumb to realize that usec is always used
1172 * initialized, silence that warning.
1173 */
1174 unsigned long uninitialized_var(usec);
1175 struct fio_file *f;
1176
1177 dprint_io_u(io_u, "io complete");
1178
1179 td_io_u_lock(td);
1180 assert(io_u->flags & IO_U_F_FLIGHT);
1181 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1182 td_io_u_unlock(td);
1183
1184 if (ddir_sync(io_u->ddir)) {
1185 td->last_was_sync = 1;
1186 f = io_u->file;
1187 if (f) {
1188 f->first_write = -1ULL;
1189 f->last_write = -1ULL;
1190 }
1191 return;
1192 }
1193
1194 td->last_was_sync = 0;
1195 td->last_ddir = io_u->ddir;
1196
1197 if (!io_u->error && ddir_rw(io_u->ddir)) {
1198 unsigned int bytes = io_u->buflen - io_u->resid;
1199 const enum fio_ddir idx = io_u->ddir;
1200 const enum fio_ddir odx = io_u->ddir ^ 1;
1201 int ret;
1202
1203 td->io_blocks[idx]++;
1204 td->io_bytes[idx] += bytes;
1205 td->this_io_bytes[idx] += bytes;
1206
1207 if (idx == DDIR_WRITE) {
1208 f = io_u->file;
1209 if (f) {
1210 if (f->first_write == -1ULL ||
1211 io_u->offset < f->first_write)
1212 f->first_write = io_u->offset;
1213 if (f->last_write == -1ULL ||
1214 ((io_u->offset + bytes) > f->last_write))
1215 f->last_write = io_u->offset + bytes;
1216 }
1217 }
1218
1219 if (ramp_time_over(td)) {
1220 unsigned long uninitialized_var(lusec);
1221
1222 if (!td->o.disable_clat || !td->o.disable_bw)
1223 lusec = utime_since(&io_u->issue_time,
1224 &icd->time);
1225 if (!td->o.disable_lat) {
1226 unsigned long tusec;
1227
1228 tusec = utime_since(&io_u->start_time,
1229 &icd->time);
1230 add_lat_sample(td, idx, tusec, bytes);
1231 }
1232 if (!td->o.disable_clat) {
1233 add_clat_sample(td, idx, lusec, bytes);
1234 io_u_mark_latency(td, lusec);
1235 }
1236 if (!td->o.disable_bw)
1237 add_bw_sample(td, idx, bytes, &icd->time);
1238 if (__should_check_rate(td, idx)) {
1239 td->rate_pending_usleep[idx] =
1240 ((td->this_io_bytes[idx] *
1241 td->rate_nsec_cycle[idx]) / 1000 -
1242 utime_since_now(&td->start));
1243 }
1244 if (__should_check_rate(td, idx ^ 1))
1245 td->rate_pending_usleep[odx] =
1246 ((td->this_io_bytes[odx] *
1247 td->rate_nsec_cycle[odx]) / 1000 -
1248 utime_since_now(&td->start));
1249 }
1250
1251 if (td_write(td) && idx == DDIR_WRITE &&
1252 td->o.do_verify &&
1253 td->o.verify != VERIFY_NONE)
1254 log_io_piece(td, io_u);
1255
1256 icd->bytes_done[idx] += bytes;
1257
1258 if (io_u->end_io) {
1259 ret = io_u->end_io(td, io_u);
1260 if (ret && !icd->error)
1261 icd->error = ret;
1262 }
1263 } else if (io_u->error) {
1264 icd->error = io_u->error;
1265 io_u_log_error(td, io_u);
1266 }
1267 if (td->o.continue_on_error && icd->error &&
1268 td_non_fatal_error(icd->error)) {
1269 /*
1270 * If there is a non_fatal error, then add to the error count
1271 * and clear all the errors.
1272 */
1273 update_error_count(td, icd->error);
1274 td_clear_error(td);
1275 icd->error = 0;
1276 io_u->error = 0;
1277 }
1278}
1279
1280static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1281 int nr)
1282{
1283 if (!td->o.disable_clat || !td->o.disable_bw)
1284 fio_gettime(&icd->time, NULL);
1285
1286 icd->nr = nr;
1287
1288 icd->error = 0;
1289 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1290}
1291
1292static void ios_completed(struct thread_data *td,
1293 struct io_completion_data *icd)
1294{
1295 struct io_u *io_u;
1296 int i;
1297
1298 for (i = 0; i < icd->nr; i++) {
1299 io_u = td->io_ops->event(td, i);
1300
1301 io_completed(td, io_u, icd);
1302
1303 if (!(io_u->flags & IO_U_F_FREE_DEF))
1304 put_io_u(td, io_u);
1305 }
1306}
1307
1308/*
1309 * Complete a single io_u for the sync engines.
1310 */
1311int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1312 unsigned long *bytes)
1313{
1314 struct io_completion_data icd;
1315
1316 init_icd(td, &icd, 1);
1317 io_completed(td, io_u, &icd);
1318
1319 if (!(io_u->flags & IO_U_F_FREE_DEF))
1320 put_io_u(td, io_u);
1321
1322 if (icd.error) {
1323 td_verror(td, icd.error, "io_u_sync_complete");
1324 return -1;
1325 }
1326
1327 if (bytes) {
1328 bytes[0] += icd.bytes_done[0];
1329 bytes[1] += icd.bytes_done[1];
1330 }
1331
1332 return 0;
1333}
1334
1335/*
1336 * Called to complete min_events number of io for the async engines.
1337 */
1338int io_u_queued_complete(struct thread_data *td, int min_evts,
1339 unsigned long *bytes)
1340{
1341 struct io_completion_data icd;
1342 struct timespec *tvp = NULL;
1343 int ret;
1344 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1345
1346 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1347
1348 if (!min_evts)
1349 tvp = &ts;
1350
1351 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1352 if (ret < 0) {
1353 td_verror(td, -ret, "td_io_getevents");
1354 return ret;
1355 } else if (!ret)
1356 return ret;
1357
1358 init_icd(td, &icd, ret);
1359 ios_completed(td, &icd);
1360 if (icd.error) {
1361 td_verror(td, icd.error, "io_u_queued_complete");
1362 return -1;
1363 }
1364
1365 if (bytes) {
1366 bytes[0] += icd.bytes_done[0];
1367 bytes[1] += icd.bytes_done[1];
1368 }
1369
1370 return 0;
1371}
1372
1373/*
1374 * Call when io_u is really queued, to update the submission latency.
1375 */
1376void io_u_queued(struct thread_data *td, struct io_u *io_u)
1377{
1378 if (!td->o.disable_slat) {
1379 unsigned long slat_time;
1380
1381 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1382 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1383 }
1384}
1385
1386/*
1387 * "randomly" fill the buffer contents
1388 */
1389void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1390 unsigned int max_bs)
1391{
1392 io_u->buf_filled_len = 0;
1393
1394 if (!td->o.zero_buffers)
1395 fill_random_buf(io_u->buf, max_bs);
1396 else
1397 memset(io_u->buf, 0, max_bs);
1398}
fio - Flexible IO Tester