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