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