projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Fix problem with iodepth_batch_complete=0 and accounting
[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 if (td->o.zone_range)
117 max_size = td->o.zone_range;
118
119 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
120 if (!max_blocks)
121 return 0;
122
123 return max_blocks;
124}
125
126/*
127 * Return the next free block in the map.
128 */
129static int get_next_free_block(struct thread_data *td, struct fio_file *f,
130 enum fio_ddir ddir, unsigned long long *b)
131{
132 unsigned long long block, min_bs = td->o.rw_min_bs, lastb;
133 int i;
134
135 lastb = last_block(td, f, ddir);
136 if (!lastb)
137 return 1;
138
139 i = f->last_free_lookup;
140 block = i * BLOCKS_PER_MAP;
141 while (block * min_bs < f->real_file_size &&
142 block * min_bs < f->io_size) {
143 if (f->file_map[i] != -1UL) {
144 block += ffz(f->file_map[i]);
145 if (block > lastb)
146 break;
147 f->last_free_lookup = i;
148 *b = block;
149 return 0;
150 }
151
152 block += BLOCKS_PER_MAP;
153 i++;
154 }
155
156 dprint(FD_IO, "failed finding a free block\n");
157 return 1;
158}
159
160static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
161 enum fio_ddir ddir, unsigned long long *b)
162{
163 unsigned long long rmax, r, lastb;
164 int loops = 5;
165
166 lastb = last_block(td, f, ddir);
167 if (!lastb)
168 return 1;
169
170 if (f->failed_rands >= 200)
171 goto ffz;
172
173 rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
174 do {
175 if (td->o.use_os_rand)
176 r = os_random_long(&td->random_state);
177 else
178 r = __rand(&td->__random_state);
179
180 *b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0));
181
182 dprint(FD_RANDOM, "off rand %llu\n", r);
183
184
185 /*
186 * if we are not maintaining a random map, we are done.
187 */
188 if (!file_randommap(td, f))
189 goto ret_good;
190
191 /*
192 * calculate map offset and check if it's free
193 */
194 if (random_map_free(f, *b))
195 goto ret_good;
196
197 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
198 *b);
199 } while (--loops);
200
201 if (!f->failed_rands++)
202 f->last_free_lookup = 0;
203
204 /*
205 * we get here, if we didn't suceed in looking up a block. generate
206 * a random start offset into the filemap, and find the first free
207 * block from there.
208 */
209 loops = 10;
210 do {
211 f->last_free_lookup = (f->num_maps - 1) *
212 (r / ((unsigned long long) rmax + 1.0));
213 if (!get_next_free_block(td, f, ddir, b))
214 goto ret;
215
216 if (td->o.use_os_rand)
217 r = os_random_long(&td->random_state);
218 else
219 r = __rand(&td->__random_state);
220 } while (--loops);
221
222 /*
223 * that didn't work either, try exhaustive search from the start
224 */
225 f->last_free_lookup = 0;
226ffz:
227 if (!get_next_free_block(td, f, ddir, b))
228 return 0;
229 f->last_free_lookup = 0;
230 return get_next_free_block(td, f, ddir, b);
231ret_good:
232 f->failed_rands = 0;
233ret:
234 return 0;
235}
236
237static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
238 enum fio_ddir ddir, unsigned long long *b)
239{
240 if (get_next_rand_offset(td, f, ddir, b)) {
241 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
242 f->file_name, f->last_pos, f->real_file_size);
243 return 1;
244 }
245
246 return 0;
247}
248
249static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
250 enum fio_ddir ddir, unsigned long long *b)
251{
252 assert(ddir_rw(ddir));
253
254 if (f->last_pos < f->real_file_size) {
255 unsigned long long pos;
256
257 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
258 f->last_pos = f->real_file_size;
259
260 pos = f->last_pos - f->file_offset;
261 if (pos)
262 pos += td->o.ddir_seq_add;
263
264 *b = pos / td->o.min_bs[ddir];
265 return 0;
266 }
267
268 return 1;
269}
270
271static int get_next_block(struct thread_data *td, struct io_u *io_u,
272 enum fio_ddir ddir, int rw_seq, unsigned long long *b)
273{
274 struct fio_file *f = io_u->file;
275 int ret;
276
277 assert(ddir_rw(ddir));
278
279 if (rw_seq) {
280 if (td_random(td))
281 ret = get_next_rand_block(td, f, ddir, b);
282 else
283 ret = get_next_seq_block(td, f, ddir, b);
284 } else {
285 io_u->flags |= IO_U_F_BUSY_OK;
286
287 if (td->o.rw_seq == RW_SEQ_SEQ) {
288 ret = get_next_seq_block(td, f, ddir, b);
289 if (ret)
290 ret = get_next_rand_block(td, f, ddir, b);
291 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
292 if (f->last_start != -1ULL)
293 *b = (f->last_start - f->file_offset)
294 / td->o.min_bs[ddir];
295 else
296 *b = 0;
297 ret = 0;
298 } else {
299 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
300 ret = 1;
301 }
302 }
303
304 return ret;
305}
306
307/*
308 * For random io, generate a random new block and see if it's used. Repeat
309 * until we find a free one. For sequential io, just return the end of
310 * the last io issued.
311 */
312static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
313{
314 struct fio_file *f = io_u->file;
315 unsigned long long b;
316 enum fio_ddir ddir = io_u->ddir;
317 int rw_seq_hit = 0;
318
319 assert(ddir_rw(ddir));
320
321 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
322 rw_seq_hit = 1;
323 td->ddir_seq_nr = td->o.ddir_seq_nr;
324 }
325
326 if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
327 return 1;
328
329 io_u->offset = b * td->o.ba[ddir];
330 if (io_u->offset >= f->io_size) {
331 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
332 io_u->offset, f->io_size);
333 return 1;
334 }
335
336 io_u->offset += f->file_offset;
337 if (io_u->offset >= f->real_file_size) {
338 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
339 io_u->offset, f->real_file_size);
340 return 1;
341 }
342
343 return 0;
344}
345
346static int get_next_offset(struct thread_data *td, struct io_u *io_u)
347{
348 struct prof_io_ops *ops = &td->prof_io_ops;
349
350 if (ops->fill_io_u_off)
351 return ops->fill_io_u_off(td, io_u);
352
353 return __get_next_offset(td, io_u);
354}
355
356static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
357 unsigned int buflen)
358{
359 struct fio_file *f = io_u->file;
360
361 return io_u->offset + buflen <= f->io_size + td->o.start_offset;
362}
363
364static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
365{
366 const int ddir = io_u->ddir;
367 unsigned int uninitialized_var(buflen);
368 unsigned int minbs, maxbs;
369 unsigned long r, rand_max;
370
371 assert(ddir_rw(ddir));
372
373 minbs = td->o.min_bs[ddir];
374 maxbs = td->o.max_bs[ddir];
375
376 if (minbs == maxbs)
377 return minbs;
378
379 /*
380 * If we can't satisfy the min block size from here, then fail
381 */
382 if (!io_u_fits(td, io_u, minbs))
383 return 0;
384
385 if (td->o.use_os_rand)
386 rand_max = OS_RAND_MAX;
387 else
388 rand_max = FRAND_MAX;
389
390 do {
391 if (td->o.use_os_rand)
392 r = os_random_long(&td->bsrange_state);
393 else
394 r = __rand(&td->__bsrange_state);
395
396 if (!td->o.bssplit_nr[ddir]) {
397 buflen = 1 + (unsigned int) ((double) maxbs *
398 (r / (rand_max + 1.0)));
399 if (buflen < minbs)
400 buflen = minbs;
401 } else {
402 long perc = 0;
403 unsigned int i;
404
405 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
406 struct bssplit *bsp = &td->o.bssplit[ddir][i];
407
408 buflen = bsp->bs;
409 perc += bsp->perc;
410 if ((r <= ((rand_max / 100L) * perc)) &&
411 io_u_fits(td, io_u, buflen))
412 break;
413 }
414 }
415
416 if (!td->o.bs_unaligned && is_power_of_2(minbs))
417 buflen = (buflen + minbs - 1) & ~(minbs - 1);
418
419 } while (!io_u_fits(td, io_u, buflen));
420
421 return buflen;
422}
423
424static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
425{
426 struct prof_io_ops *ops = &td->prof_io_ops;
427
428 if (ops->fill_io_u_size)
429 return ops->fill_io_u_size(td, io_u);
430
431 return __get_next_buflen(td, io_u);
432}
433
434static void set_rwmix_bytes(struct thread_data *td)
435{
436 unsigned int diff;
437
438 /*
439 * we do time or byte based switch. this is needed because
440 * buffered writes may issue a lot quicker than they complete,
441 * whereas reads do not.
442 */
443 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
444 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
445}
446
447static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
448{
449 unsigned int v;
450 unsigned long r;
451
452 if (td->o.use_os_rand) {
453 r = os_random_long(&td->rwmix_state);
454 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
455 } else {
456 r = __rand(&td->__rwmix_state);
457 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
458 }
459
460 if (v <= td->o.rwmix[DDIR_READ])
461 return DDIR_READ;
462
463 return DDIR_WRITE;
464}
465
466static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
467{
468 enum fio_ddir odir = ddir ^ 1;
469 struct timeval t;
470 long usec;
471
472 assert(ddir_rw(ddir));
473
474 if (td->rate_pending_usleep[ddir] <= 0)
475 return ddir;
476
477 /*
478 * We have too much pending sleep in this direction. See if we
479 * should switch.
480 */
481 if (td_rw(td)) {
482 /*
483 * Other direction does not have too much pending, switch
484 */
485 if (td->rate_pending_usleep[odir] < 100000)
486 return odir;
487
488 /*
489 * Both directions have pending sleep. Sleep the minimum time
490 * and deduct from both.
491 */
492 if (td->rate_pending_usleep[ddir] <=
493 td->rate_pending_usleep[odir]) {
494 usec = td->rate_pending_usleep[ddir];
495 } else {
496 usec = td->rate_pending_usleep[odir];
497 ddir = odir;
498 }
499 } else
500 usec = td->rate_pending_usleep[ddir];
501
502 /*
503 * We are going to sleep, ensure that we flush anything pending as
504 * not to skew our latency numbers.
505 *
506 * Changed to only monitor 'in flight' requests here instead of the
507 * td->cur_depth, b/c td->cur_depth does not accurately represent
508 * io's that have been actually submitted to an async engine,
509 * and cur_depth is meaningless for sync engines.
510 */
511 if (td->io_u_in_flight) {
512 int fio_unused ret;
513
514 ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
515 }
516
517 fio_gettime(&t, NULL);
518 usec_sleep(td, usec);
519 usec = utime_since_now(&t);
520
521 td->rate_pending_usleep[ddir] -= usec;
522
523 odir = ddir ^ 1;
524 if (td_rw(td) && __should_check_rate(td, odir))
525 td->rate_pending_usleep[odir] -= usec;
526
527 return ddir;
528}
529
530/*
531 * Return the data direction for the next io_u. If the job is a
532 * mixed read/write workload, check the rwmix cycle and switch if
533 * necessary.
534 */
535static enum fio_ddir get_rw_ddir(struct thread_data *td)
536{
537 enum fio_ddir ddir;
538
539 /*
540 * see if it's time to fsync
541 */
542 if (td->o.fsync_blocks &&
543 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
544 td->io_issues[DDIR_WRITE] && should_fsync(td))
545 return DDIR_SYNC;
546
547 /*
548 * see if it's time to fdatasync
549 */
550 if (td->o.fdatasync_blocks &&
551 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
552 td->io_issues[DDIR_WRITE] && should_fsync(td))
553 return DDIR_DATASYNC;
554
555 /*
556 * see if it's time to sync_file_range
557 */
558 if (td->sync_file_range_nr &&
559 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
560 td->io_issues[DDIR_WRITE] && should_fsync(td))
561 return DDIR_SYNC_FILE_RANGE;
562
563 if (td_rw(td)) {
564 /*
565 * Check if it's time to seed a new data direction.
566 */
567 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
568 /*
569 * Put a top limit on how many bytes we do for
570 * one data direction, to avoid overflowing the
571 * ranges too much
572 */
573 ddir = get_rand_ddir(td);
574
575 if (ddir != td->rwmix_ddir)
576 set_rwmix_bytes(td);
577
578 td->rwmix_ddir = ddir;
579 }
580 ddir = td->rwmix_ddir;
581 } else if (td_read(td))
582 ddir = DDIR_READ;
583 else
584 ddir = DDIR_WRITE;
585
586 td->rwmix_ddir = rate_ddir(td, ddir);
587 return td->rwmix_ddir;
588}
589
590static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
591{
592 io_u->ddir = get_rw_ddir(td);
593
594 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
595 td->o.barrier_blocks &&
596 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
597 td->io_issues[DDIR_WRITE])
598 io_u->flags |= IO_U_F_BARRIER;
599}
600
601void put_file_log(struct thread_data *td, struct fio_file *f)
602{
603 int ret = put_file(td, f);
604
605 if (ret)
606 td_verror(td, ret, "file close");
607}
608
609void put_io_u(struct thread_data *td, struct io_u *io_u)
610{
611 td_io_u_lock(td);
612
613 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
614 put_file_log(td, io_u->file);
615 io_u->file = NULL;
616 io_u->flags &= ~IO_U_F_FREE_DEF;
617 io_u->flags |= IO_U_F_FREE;
618
619 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
620 td->cur_depth--;
621 flist_del_init(&io_u->list);
622 flist_add(&io_u->list, &td->io_u_freelist);
623 td_io_u_unlock(td);
624 td_io_u_free_notify(td);
625}
626
627void clear_io_u(struct thread_data *td, struct io_u *io_u)
628{
629 io_u->flags &= ~IO_U_F_FLIGHT;
630 put_io_u(td, io_u);
631}
632
633void requeue_io_u(struct thread_data *td, struct io_u **io_u)
634{
635 struct io_u *__io_u = *io_u;
636
637 dprint(FD_IO, "requeue %p\n", __io_u);
638
639 td_io_u_lock(td);
640
641 __io_u->flags |= IO_U_F_FREE;
642 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
643 td->io_issues[__io_u->ddir]--;
644
645 __io_u->flags &= ~IO_U_F_FLIGHT;
646 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
647 td->cur_depth--;
648 flist_del(&__io_u->list);
649 flist_add_tail(&__io_u->list, &td->io_u_requeues);
650 td_io_u_unlock(td);
651 *io_u = NULL;
652}
653
654static int fill_io_u(struct thread_data *td, struct io_u *io_u)
655{
656 if (td->io_ops->flags & FIO_NOIO)
657 goto out;
658
659 set_rw_ddir(td, io_u);
660
661 /*
662 * fsync() or fdatasync() or trim etc, we are done
663 */
664 if (!ddir_rw(io_u->ddir))
665 goto out;
666
667 /*
668 * See if it's time to switch to a new zone
669 */
670 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
671 td->zone_bytes = 0;
672 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
673 io_u->file->last_pos = io_u->file->file_offset;
674 td->io_skip_bytes += td->o.zone_skip;
675 }
676
677 /*
678 * No log, let the seq/rand engine retrieve the next buflen and
679 * position.
680 */
681 if (get_next_offset(td, io_u)) {
682 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
683 return 1;
684 }
685
686 io_u->buflen = get_next_buflen(td, io_u);
687 if (!io_u->buflen) {
688 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
689 return 1;
690 }
691
692 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
693 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
694 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
695 io_u->buflen, io_u->file->real_file_size);
696 return 1;
697 }
698
699 /*
700 * mark entry before potentially trimming io_u
701 */
702 if (td_random(td) && file_randommap(td, io_u->file))
703 mark_random_map(td, io_u);
704
705 /*
706 * If using a write iolog, store this entry.
707 */
708out:
709 dprint_io_u(io_u, "fill_io_u");
710 td->zone_bytes += io_u->buflen;
711 log_io_u(td, io_u);
712 return 0;
713}
714
715static void __io_u_mark_map(unsigned int *map, unsigned int nr)
716{
717 int idx = 0;
718
719 switch (nr) {
720 default:
721 idx = 6;
722 break;
723 case 33 ... 64:
724 idx = 5;
725 break;
726 case 17 ... 32:
727 idx = 4;
728 break;
729 case 9 ... 16:
730 idx = 3;
731 break;
732 case 5 ... 8:
733 idx = 2;
734 break;
735 case 1 ... 4:
736 idx = 1;
737 case 0:
738 break;
739 }
740
741 map[idx]++;
742}
743
744void io_u_mark_submit(struct thread_data *td, unsigned int nr)
745{
746 __io_u_mark_map(td->ts.io_u_submit, nr);
747 td->ts.total_submit++;
748}
749
750void io_u_mark_complete(struct thread_data *td, unsigned int nr)
751{
752 __io_u_mark_map(td->ts.io_u_complete, nr);
753 td->ts.total_complete++;
754}
755
756void io_u_mark_depth(struct thread_data *td, unsigned int nr)
757{
758 int idx = 0;
759
760 switch (td->cur_depth) {
761 default:
762 idx = 6;
763 break;
764 case 32 ... 63:
765 idx = 5;
766 break;
767 case 16 ... 31:
768 idx = 4;
769 break;
770 case 8 ... 15:
771 idx = 3;
772 break;
773 case 4 ... 7:
774 idx = 2;
775 break;
776 case 2 ... 3:
777 idx = 1;
778 case 1:
779 break;
780 }
781
782 td->ts.io_u_map[idx] += nr;
783}
784
785static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
786{
787 int idx = 0;
788
789 assert(usec < 1000);
790
791 switch (usec) {
792 case 750 ... 999:
793 idx = 9;
794 break;
795 case 500 ... 749:
796 idx = 8;
797 break;
798 case 250 ... 499:
799 idx = 7;
800 break;
801 case 100 ... 249:
802 idx = 6;
803 break;
804 case 50 ... 99:
805 idx = 5;
806 break;
807 case 20 ... 49:
808 idx = 4;
809 break;
810 case 10 ... 19:
811 idx = 3;
812 break;
813 case 4 ... 9:
814 idx = 2;
815 break;
816 case 2 ... 3:
817 idx = 1;
818 case 0 ... 1:
819 break;
820 }
821
822 assert(idx < FIO_IO_U_LAT_U_NR);
823 td->ts.io_u_lat_u[idx]++;
824}
825
826static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
827{
828 int idx = 0;
829
830 switch (msec) {
831 default:
832 idx = 11;
833 break;
834 case 1000 ... 1999:
835 idx = 10;
836 break;
837 case 750 ... 999:
838 idx = 9;
839 break;
840 case 500 ... 749:
841 idx = 8;
842 break;
843 case 250 ... 499:
844 idx = 7;
845 break;
846 case 100 ... 249:
847 idx = 6;
848 break;
849 case 50 ... 99:
850 idx = 5;
851 break;
852 case 20 ... 49:
853 idx = 4;
854 break;
855 case 10 ... 19:
856 idx = 3;
857 break;
858 case 4 ... 9:
859 idx = 2;
860 break;
861 case 2 ... 3:
862 idx = 1;
863 case 0 ... 1:
864 break;
865 }
866
867 assert(idx < FIO_IO_U_LAT_M_NR);
868 td->ts.io_u_lat_m[idx]++;
869}
870
871static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
872{
873 if (usec < 1000)
874 io_u_mark_lat_usec(td, usec);
875 else
876 io_u_mark_lat_msec(td, usec / 1000);
877}
878
879/*
880 * Get next file to service by choosing one at random
881 */
882static struct fio_file *get_next_file_rand(struct thread_data *td,
883 enum fio_file_flags goodf,
884 enum fio_file_flags badf)
885{
886 struct fio_file *f;
887 int fno;
888
889 do {
890 int opened = 0;
891 unsigned long r;
892
893 if (td->o.use_os_rand) {
894 r = os_random_long(&td->next_file_state);
895 fno = (unsigned int) ((double) td->o.nr_files
896 * (r / (OS_RAND_MAX + 1.0)));
897 } else {
898 r = __rand(&td->__next_file_state);
899 fno = (unsigned int) ((double) td->o.nr_files
900 * (r / (FRAND_MAX + 1.0)));
901 }
902
903 f = td->files[fno];
904 if (fio_file_done(f))
905 continue;
906
907 if (!fio_file_open(f)) {
908 int err;
909
910 err = td_io_open_file(td, f);
911 if (err)
912 continue;
913 opened = 1;
914 }
915
916 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
917 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
918 return f;
919 }
920 if (opened)
921 td_io_close_file(td, f);
922 } while (1);
923}
924
925/*
926 * Get next file to service by doing round robin between all available ones
927 */
928static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
929 int badf)
930{
931 unsigned int old_next_file = td->next_file;
932 struct fio_file *f;
933
934 do {
935 int opened = 0;
936
937 f = td->files[td->next_file];
938
939 td->next_file++;
940 if (td->next_file >= td->o.nr_files)
941 td->next_file = 0;
942
943 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
944 if (fio_file_done(f)) {
945 f = NULL;
946 continue;
947 }
948
949 if (!fio_file_open(f)) {
950 int err;
951
952 err = td_io_open_file(td, f);
953 if (err) {
954 dprint(FD_FILE, "error %d on open of %s\n",
955 err, f->file_name);
956 f = NULL;
957 continue;
958 }
959 opened = 1;
960 }
961
962 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
963 f->flags);
964 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
965 break;
966
967 if (opened)
968 td_io_close_file(td, f);
969
970 f = NULL;
971 } while (td->next_file != old_next_file);
972
973 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
974 return f;
975}
976
977static struct fio_file *__get_next_file(struct thread_data *td)
978{
979 struct fio_file *f;
980
981 assert(td->o.nr_files <= td->files_index);
982
983 if (td->nr_done_files >= td->o.nr_files) {
984 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
985 " nr_files=%d\n", td->nr_open_files,
986 td->nr_done_files,
987 td->o.nr_files);
988 return NULL;
989 }
990
991 f = td->file_service_file;
992 if (f && fio_file_open(f) && !fio_file_closing(f)) {
993 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
994 goto out;
995 if (td->file_service_left--)
996 goto out;
997 }
998
999 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1000 td->o.file_service_type == FIO_FSERVICE_SEQ)
1001 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1002 else
1003 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1004
1005 td->file_service_file = f;
1006 td->file_service_left = td->file_service_nr - 1;
1007out:
1008 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1009 return f;
1010}
1011
1012static struct fio_file *get_next_file(struct thread_data *td)
1013{
1014 struct prof_io_ops *ops = &td->prof_io_ops;
1015
1016 if (ops->get_next_file)
1017 return ops->get_next_file(td);
1018
1019 return __get_next_file(td);
1020}
1021
1022static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1023{
1024 struct fio_file *f;
1025
1026 do {
1027 f = get_next_file(td);
1028 if (!f)
1029 return 1;
1030
1031 io_u->file = f;
1032 get_file(f);
1033
1034 if (!fill_io_u(td, io_u))
1035 break;
1036
1037 put_file_log(td, f);
1038 td_io_close_file(td, f);
1039 io_u->file = NULL;
1040 fio_file_set_done(f);
1041 td->nr_done_files++;
1042 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1043 td->nr_done_files, td->o.nr_files);
1044 } while (1);
1045
1046 return 0;
1047}
1048
1049
1050struct io_u *__get_io_u(struct thread_data *td)
1051{
1052 struct io_u *io_u = NULL;
1053
1054 td_io_u_lock(td);
1055
1056again:
1057 if (!flist_empty(&td->io_u_requeues))
1058 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1059 else if (!queue_full(td)) {
1060 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1061
1062 io_u->buflen = 0;
1063 io_u->resid = 0;
1064 io_u->file = NULL;
1065 io_u->end_io = NULL;
1066 }
1067
1068 if (io_u) {
1069 assert(io_u->flags & IO_U_F_FREE);
1070 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1071 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1072 io_u->flags &= ~IO_U_F_VER_LIST;
1073
1074 io_u->error = 0;
1075 flist_del(&io_u->list);
1076 flist_add(&io_u->list, &td->io_u_busylist);
1077 td->cur_depth++;
1078 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1079 } else if (td->o.verify_async) {
1080 /*
1081 * We ran out, wait for async verify threads to finish and
1082 * return one
1083 */
1084 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1085 goto again;
1086 }
1087
1088 td_io_u_unlock(td);
1089 return io_u;
1090}
1091
1092static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1093{
1094 if (td->o.trim_backlog && td->trim_entries) {
1095 int get_trim = 0;
1096
1097 if (td->trim_batch) {
1098 td->trim_batch--;
1099 get_trim = 1;
1100 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1101 td->last_ddir != DDIR_READ) {
1102 td->trim_batch = td->o.trim_batch;
1103 if (!td->trim_batch)
1104 td->trim_batch = td->o.trim_backlog;
1105 get_trim = 1;
1106 }
1107
1108 if (get_trim && !get_next_trim(td, io_u))
1109 return 1;
1110 }
1111
1112 return 0;
1113}
1114
1115static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1116{
1117 if (td->o.verify_backlog && td->io_hist_len) {
1118 int get_verify = 0;
1119
1120 if (td->verify_batch)
1121 get_verify = 1;
1122 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1123 td->last_ddir != DDIR_READ) {
1124 td->verify_batch = td->o.verify_batch;
1125 if (!td->verify_batch)
1126 td->verify_batch = td->o.verify_backlog;
1127 get_verify = 1;
1128 }
1129
1130 if (get_verify && !get_next_verify(td, io_u)) {
1131 td->verify_batch--;
1132 return 1;
1133 }
1134 }
1135
1136 return 0;
1137}
1138
1139/*
1140 * Fill offset and start time into the buffer content, to prevent too
1141 * easy compressible data for simple de-dupe attempts. Do this for every
1142 * 512b block in the range, since that should be the smallest block size
1143 * we can expect from a device.
1144 */
1145static void small_content_scramble(struct io_u *io_u)
1146{
1147 unsigned int i, nr_blocks = io_u->buflen / 512;
1148 unsigned long long boffset;
1149 unsigned int offset;
1150 void *p, *end;
1151
1152 if (!nr_blocks)
1153 return;
1154
1155 p = io_u->xfer_buf;
1156 boffset = io_u->offset;
1157 io_u->buf_filled_len = 0;
1158
1159 for (i = 0; i < nr_blocks; i++) {
1160 /*
1161 * Fill the byte offset into a "random" start offset of
1162 * the buffer, given by the product of the usec time
1163 * and the actual offset.
1164 */
1165 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1166 offset &= ~(sizeof(unsigned long long) - 1);
1167 if (offset >= 512 - sizeof(unsigned long long))
1168 offset -= sizeof(unsigned long long);
1169 memcpy(p + offset, &boffset, sizeof(boffset));
1170
1171 end = p + 512 - sizeof(io_u->start_time);
1172 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1173 p += 512;
1174 boffset += 512;
1175 }
1176}
1177
1178/*
1179 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1180 * etc. The returned io_u is fully ready to be prepped and submitted.
1181 */
1182struct io_u *get_io_u(struct thread_data *td)
1183{
1184 struct fio_file *f;
1185 struct io_u *io_u;
1186 int do_scramble = 0;
1187
1188 io_u = __get_io_u(td);
1189 if (!io_u) {
1190 dprint(FD_IO, "__get_io_u failed\n");
1191 return NULL;
1192 }
1193
1194 if (check_get_verify(td, io_u))
1195 goto out;
1196 if (check_get_trim(td, io_u))
1197 goto out;
1198
1199 /*
1200 * from a requeue, io_u already setup
1201 */
1202 if (io_u->file)
1203 goto out;
1204
1205 /*
1206 * If using an iolog, grab next piece if any available.
1207 */
1208 if (td->o.read_iolog_file) {
1209 if (read_iolog_get(td, io_u))
1210 goto err_put;
1211 } else if (set_io_u_file(td, io_u)) {
1212 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1213 goto err_put;
1214 }
1215
1216 f = io_u->file;
1217 assert(fio_file_open(f));
1218
1219 if (ddir_rw(io_u->ddir)) {
1220 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1221 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1222 goto err_put;
1223 }
1224
1225 f->last_start = io_u->offset;
1226 f->last_pos = io_u->offset + io_u->buflen;
1227
1228 if (io_u->ddir == DDIR_WRITE) {
1229 if (td->o.refill_buffers) {
1230 io_u_fill_buffer(td, io_u,
1231 io_u->xfer_buflen, io_u->xfer_buflen);
1232 } else if (td->o.scramble_buffers)
1233 do_scramble = 1;
1234 if (td->o.verify != VERIFY_NONE) {
1235 populate_verify_io_u(td, io_u);
1236 do_scramble = 0;
1237 }
1238 } else if (io_u->ddir == DDIR_READ) {
1239 /*
1240 * Reset the buf_filled parameters so next time if the
1241 * buffer is used for writes it is refilled.
1242 */
1243 io_u->buf_filled_len = 0;
1244 }
1245 }
1246
1247 /*
1248 * Set io data pointers.
1249 */
1250 io_u->xfer_buf = io_u->buf;
1251 io_u->xfer_buflen = io_u->buflen;
1252
1253out:
1254 assert(io_u->file);
1255 if (!td_io_prep(td, io_u)) {
1256 if (!td->o.disable_slat)
1257 fio_gettime(&io_u->start_time, NULL);
1258 if (do_scramble)
1259 small_content_scramble(io_u);
1260 return io_u;
1261 }
1262err_put:
1263 dprint(FD_IO, "get_io_u failed\n");
1264 put_io_u(td, io_u);
1265 return NULL;
1266}
1267
1268void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1269{
1270 const char *msg[] = { "read", "write", "sync", "datasync",
1271 "sync_file_range", "wait", "trim" };
1272
1273
1274
1275 log_err("fio: io_u error");
1276
1277 if (io_u->file)
1278 log_err(" on file %s", io_u->file->file_name);
1279
1280 log_err(": %s\n", strerror(io_u->error));
1281
1282 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1283 io_u->offset, io_u->xfer_buflen);
1284
1285 if (!td->error)
1286 td_verror(td, io_u->error, "io_u error");
1287}
1288
1289static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1290 struct io_completion_data *icd,
1291 const enum fio_ddir idx, unsigned int bytes)
1292{
1293 unsigned long uninitialized_var(lusec);
1294
1295 if (!td->o.disable_clat || !td->o.disable_bw)
1296 lusec = utime_since(&io_u->issue_time, &icd->time);
1297
1298 if (!td->o.disable_lat) {
1299 unsigned long tusec;
1300
1301 tusec = utime_since(&io_u->start_time, &icd->time);
1302 add_lat_sample(td, idx, tusec, bytes);
1303 }
1304
1305 if (!td->o.disable_clat) {
1306 add_clat_sample(td, idx, lusec, bytes);
1307 io_u_mark_latency(td, lusec);
1308 }
1309
1310 if (!td->o.disable_bw)
1311 add_bw_sample(td, idx, bytes, &icd->time);
1312
1313 add_iops_sample(td, idx, &icd->time);
1314}
1315
1316static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1317{
1318 unsigned long long secs, remainder, bps, bytes;
1319 bytes = td->this_io_bytes[ddir];
1320 bps = td->rate_bps[ddir];
1321 secs = bytes / bps;
1322 remainder = bytes % bps;
1323 return remainder * 1000000 / bps + secs * 1000000;
1324}
1325
1326static void io_completed(struct thread_data *td, struct io_u *io_u,
1327 struct io_completion_data *icd)
1328{
1329 /*
1330 * Older gcc's are too dumb to realize that usec is always used
1331 * initialized, silence that warning.
1332 */
1333 unsigned long uninitialized_var(usec);
1334 struct fio_file *f;
1335
1336 dprint_io_u(io_u, "io complete");
1337
1338 td_io_u_lock(td);
1339 assert(io_u->flags & IO_U_F_FLIGHT);
1340 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1341 td_io_u_unlock(td);
1342
1343 if (ddir_sync(io_u->ddir)) {
1344 td->last_was_sync = 1;
1345 f = io_u->file;
1346 if (f) {
1347 f->first_write = -1ULL;
1348 f->last_write = -1ULL;
1349 }
1350 return;
1351 }
1352
1353 td->last_was_sync = 0;
1354 td->last_ddir = io_u->ddir;
1355
1356 if (!io_u->error && ddir_rw(io_u->ddir)) {
1357 unsigned int bytes = io_u->buflen - io_u->resid;
1358 const enum fio_ddir idx = io_u->ddir;
1359 const enum fio_ddir odx = io_u->ddir ^ 1;
1360 int ret;
1361
1362 td->io_blocks[idx]++;
1363 td->this_io_blocks[idx]++;
1364 td->io_bytes[idx] += bytes;
1365 td->this_io_bytes[idx] += bytes;
1366
1367 if (idx == DDIR_WRITE) {
1368 f = io_u->file;
1369 if (f) {
1370 if (f->first_write == -1ULL ||
1371 io_u->offset < f->first_write)
1372 f->first_write = io_u->offset;
1373 if (f->last_write == -1ULL ||
1374 ((io_u->offset + bytes) > f->last_write))
1375 f->last_write = io_u->offset + bytes;
1376 }
1377 }
1378
1379 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1380 td->runstate == TD_VERIFYING)) {
1381 account_io_completion(td, io_u, icd, idx, bytes);
1382
1383 if (__should_check_rate(td, idx)) {
1384 td->rate_pending_usleep[idx] =
1385 (usec_for_io(td, idx) -
1386 utime_since_now(&td->start));
1387 }
1388 if (__should_check_rate(td, odx))
1389 td->rate_pending_usleep[odx] =
1390 (usec_for_io(td, odx) -
1391 utime_since_now(&td->start));
1392 }
1393
1394 if (td_write(td) && idx == DDIR_WRITE &&
1395 td->o.do_verify &&
1396 td->o.verify != VERIFY_NONE)
1397 log_io_piece(td, io_u);
1398
1399 icd->bytes_done[idx] += bytes;
1400
1401 if (io_u->end_io) {
1402 ret = io_u->end_io(td, io_u);
1403 if (ret && !icd->error)
1404 icd->error = ret;
1405 }
1406 } else if (io_u->error) {
1407 icd->error = io_u->error;
1408 io_u_log_error(td, io_u);
1409 }
1410 if (icd->error && td_non_fatal_error(icd->error) &&
1411 (td->o.continue_on_error & td_error_type(io_u->ddir, icd->error))) {
1412 /*
1413 * If there is a non_fatal error, then add to the error count
1414 * and clear all the errors.
1415 */
1416 update_error_count(td, icd->error);
1417 td_clear_error(td);
1418 icd->error = 0;
1419 io_u->error = 0;
1420 }
1421}
1422
1423static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1424 int nr)
1425{
1426 if (!td->o.disable_clat || !td->o.disable_bw)
1427 fio_gettime(&icd->time, NULL);
1428
1429 icd->nr = nr;
1430
1431 icd->error = 0;
1432 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1433}
1434
1435static void ios_completed(struct thread_data *td,
1436 struct io_completion_data *icd)
1437{
1438 struct io_u *io_u;
1439 int i;
1440
1441 for (i = 0; i < icd->nr; i++) {
1442 io_u = td->io_ops->event(td, i);
1443
1444 io_completed(td, io_u, icd);
1445
1446 if (!(io_u->flags & IO_U_F_FREE_DEF))
1447 put_io_u(td, io_u);
1448 }
1449}
1450
1451/*
1452 * Complete a single io_u for the sync engines.
1453 */
1454int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1455 unsigned long *bytes)
1456{
1457 struct io_completion_data icd;
1458
1459 init_icd(td, &icd, 1);
1460 io_completed(td, io_u, &icd);
1461
1462 if (!(io_u->flags & IO_U_F_FREE_DEF))
1463 put_io_u(td, io_u);
1464
1465 if (icd.error) {
1466 td_verror(td, icd.error, "io_u_sync_complete");
1467 return -1;
1468 }
1469
1470 if (bytes) {
1471 bytes[0] += icd.bytes_done[0];
1472 bytes[1] += icd.bytes_done[1];
1473 }
1474
1475 return 0;
1476}
1477
1478/*
1479 * Called to complete min_events number of io for the async engines.
1480 */
1481int io_u_queued_complete(struct thread_data *td, int min_evts,
1482 unsigned long *bytes)
1483{
1484 struct io_completion_data icd;
1485 struct timespec *tvp = NULL;
1486 int ret;
1487 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1488
1489 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1490
1491 if (!min_evts)
1492 tvp = &ts;
1493
1494 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1495 if (ret < 0) {
1496 td_verror(td, -ret, "td_io_getevents");
1497 return ret;
1498 } else if (!ret)
1499 return ret;
1500
1501 init_icd(td, &icd, ret);
1502 ios_completed(td, &icd);
1503 if (icd.error) {
1504 td_verror(td, icd.error, "io_u_queued_complete");
1505 return -1;
1506 }
1507
1508 if (bytes) {
1509 bytes[0] += icd.bytes_done[0];
1510 bytes[1] += icd.bytes_done[1];
1511 }
1512
1513 return 0;
1514}
1515
1516/*
1517 * Call when io_u is really queued, to update the submission latency.
1518 */
1519void io_u_queued(struct thread_data *td, struct io_u *io_u)
1520{
1521 if (!td->o.disable_slat) {
1522 unsigned long slat_time;
1523
1524 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1525 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1526 }
1527}
1528
1529/*
1530 * "randomly" fill the buffer contents
1531 */
1532void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1533 unsigned int min_write, unsigned int max_bs)
1534{
1535 io_u->buf_filled_len = 0;
1536
1537 if (!td->o.zero_buffers) {
1538 unsigned int perc = td->o.compress_percentage;
1539
1540 if (perc) {
1541 unsigned int seg = min_write;
1542
1543 seg = min(min_write, td->o.compress_chunk);
1544 fill_random_buf_percentage(&td->buf_state, io_u->buf,
1545 perc, seg, max_bs);
1546 } else
1547 fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1548 } else
1549 memset(io_u->buf, 0, max_bs);
1550}
fio - Flexible IO Tester