projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
time_based: Avoid restarting main I/O loop
[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 return 0;
242
243 if (td->o.time_based) {
244 fio_file_reset(f);
245 if (!get_next_rand_offset(td, f, ddir, b))
246 return 0;
247 }
248
249 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
250 f->file_name, f->last_pos, f->real_file_size);
251 return 1;
252}
253
254static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
255 enum fio_ddir ddir, unsigned long long *b)
256{
257 assert(ddir_rw(ddir));
258
259 if (f->last_pos >= f->io_size && td->o.time_based)
260 f->last_pos = f->last_pos - f->io_size;
261
262 if (f->last_pos < f->real_file_size) {
263 unsigned long long pos;
264
265 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
266 f->last_pos = f->real_file_size;
267
268 pos = f->last_pos - f->file_offset;
269 if (pos)
270 pos += td->o.ddir_seq_add;
271
272 *b = pos / td->o.min_bs[ddir];
273 return 0;
274 }
275
276 return 1;
277}
278
279static int get_next_block(struct thread_data *td, struct io_u *io_u,
280 enum fio_ddir ddir, int rw_seq, unsigned long long *b)
281{
282 struct fio_file *f = io_u->file;
283 int ret;
284
285 assert(ddir_rw(ddir));
286
287 if (rw_seq) {
288 if (td_random(td))
289 ret = get_next_rand_block(td, f, ddir, b);
290 else
291 ret = get_next_seq_block(td, f, ddir, b);
292 } else {
293 io_u->flags |= IO_U_F_BUSY_OK;
294
295 if (td->o.rw_seq == RW_SEQ_SEQ) {
296 ret = get_next_seq_block(td, f, ddir, b);
297 if (ret)
298 ret = get_next_rand_block(td, f, ddir, b);
299 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
300 if (f->last_start != -1ULL)
301 *b = (f->last_start - f->file_offset)
302 / td->o.min_bs[ddir];
303 else
304 *b = 0;
305 ret = 0;
306 } else {
307 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
308 ret = 1;
309 }
310 }
311
312 return ret;
313}
314
315/*
316 * For random io, generate a random new block and see if it's used. Repeat
317 * until we find a free one. For sequential io, just return the end of
318 * the last io issued.
319 */
320static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
321{
322 struct fio_file *f = io_u->file;
323 unsigned long long b;
324 enum fio_ddir ddir = io_u->ddir;
325 int rw_seq_hit = 0;
326
327 assert(ddir_rw(ddir));
328
329 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
330 rw_seq_hit = 1;
331 td->ddir_seq_nr = td->o.ddir_seq_nr;
332 }
333
334 if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
335 return 1;
336
337 io_u->offset = b * td->o.ba[ddir];
338 if (io_u->offset >= f->io_size) {
339 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
340 io_u->offset, f->io_size);
341 return 1;
342 }
343
344 io_u->offset += f->file_offset;
345 if (io_u->offset >= f->real_file_size) {
346 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
347 io_u->offset, f->real_file_size);
348 return 1;
349 }
350
351 return 0;
352}
353
354static int get_next_offset(struct thread_data *td, struct io_u *io_u)
355{
356 struct prof_io_ops *ops = &td->prof_io_ops;
357
358 if (ops->fill_io_u_off)
359 return ops->fill_io_u_off(td, io_u);
360
361 return __get_next_offset(td, io_u);
362}
363
364static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
365 unsigned int buflen)
366{
367 struct fio_file *f = io_u->file;
368
369 return io_u->offset + buflen <= f->io_size + td->o.start_offset;
370}
371
372static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
373{
374 const int ddir = io_u->ddir;
375 unsigned int uninitialized_var(buflen);
376 unsigned int minbs, maxbs;
377 unsigned long r, rand_max;
378
379 assert(ddir_rw(ddir));
380
381 minbs = td->o.min_bs[ddir];
382 maxbs = td->o.max_bs[ddir];
383
384 if (minbs == maxbs)
385 return minbs;
386
387 /*
388 * If we can't satisfy the min block size from here, then fail
389 */
390 if (!io_u_fits(td, io_u, minbs))
391 return 0;
392
393 if (td->o.use_os_rand)
394 rand_max = OS_RAND_MAX;
395 else
396 rand_max = FRAND_MAX;
397
398 do {
399 if (td->o.use_os_rand)
400 r = os_random_long(&td->bsrange_state);
401 else
402 r = __rand(&td->__bsrange_state);
403
404 if (!td->o.bssplit_nr[ddir]) {
405 buflen = 1 + (unsigned int) ((double) maxbs *
406 (r / (rand_max + 1.0)));
407 if (buflen < minbs)
408 buflen = minbs;
409 } else {
410 long perc = 0;
411 unsigned int i;
412
413 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
414 struct bssplit *bsp = &td->o.bssplit[ddir][i];
415
416 buflen = bsp->bs;
417 perc += bsp->perc;
418 if ((r <= ((rand_max / 100L) * perc)) &&
419 io_u_fits(td, io_u, buflen))
420 break;
421 }
422 }
423
424 if (!td->o.bs_unaligned && is_power_of_2(minbs))
425 buflen = (buflen + minbs - 1) & ~(minbs - 1);
426
427 } while (!io_u_fits(td, io_u, buflen));
428
429 return buflen;
430}
431
432static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
433{
434 struct prof_io_ops *ops = &td->prof_io_ops;
435
436 if (ops->fill_io_u_size)
437 return ops->fill_io_u_size(td, io_u);
438
439 return __get_next_buflen(td, io_u);
440}
441
442static void set_rwmix_bytes(struct thread_data *td)
443{
444 unsigned int diff;
445
446 /*
447 * we do time or byte based switch. this is needed because
448 * buffered writes may issue a lot quicker than they complete,
449 * whereas reads do not.
450 */
451 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
452 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
453}
454
455static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
456{
457 unsigned int v;
458 unsigned long r;
459
460 if (td->o.use_os_rand) {
461 r = os_random_long(&td->rwmix_state);
462 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
463 } else {
464 r = __rand(&td->__rwmix_state);
465 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
466 }
467
468 if (v <= td->o.rwmix[DDIR_READ])
469 return DDIR_READ;
470
471 return DDIR_WRITE;
472}
473
474static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
475{
476 enum fio_ddir odir = ddir ^ 1;
477 struct timeval t;
478 long usec;
479
480 assert(ddir_rw(ddir));
481
482 if (td->rate_pending_usleep[ddir] <= 0)
483 return ddir;
484
485 /*
486 * We have too much pending sleep in this direction. See if we
487 * should switch.
488 */
489 if (td_rw(td)) {
490 /*
491 * Other direction does not have too much pending, switch
492 */
493 if (td->rate_pending_usleep[odir] < 100000)
494 return odir;
495
496 /*
497 * Both directions have pending sleep. Sleep the minimum time
498 * and deduct from both.
499 */
500 if (td->rate_pending_usleep[ddir] <=
501 td->rate_pending_usleep[odir]) {
502 usec = td->rate_pending_usleep[ddir];
503 } else {
504 usec = td->rate_pending_usleep[odir];
505 ddir = odir;
506 }
507 } else
508 usec = td->rate_pending_usleep[ddir];
509
510 /*
511 * We are going to sleep, ensure that we flush anything pending as
512 * not to skew our latency numbers.
513 *
514 * Changed to only monitor 'in flight' requests here instead of the
515 * td->cur_depth, b/c td->cur_depth does not accurately represent
516 * io's that have been actually submitted to an async engine,
517 * and cur_depth is meaningless for sync engines.
518 */
519 if (td->io_u_in_flight) {
520 int fio_unused ret;
521
522 ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
523 }
524
525 fio_gettime(&t, NULL);
526 usec_sleep(td, usec);
527 usec = utime_since_now(&t);
528
529 td->rate_pending_usleep[ddir] -= usec;
530
531 odir = ddir ^ 1;
532 if (td_rw(td) && __should_check_rate(td, odir))
533 td->rate_pending_usleep[odir] -= usec;
534
535 return ddir;
536}
537
538/*
539 * Return the data direction for the next io_u. If the job is a
540 * mixed read/write workload, check the rwmix cycle and switch if
541 * necessary.
542 */
543static enum fio_ddir get_rw_ddir(struct thread_data *td)
544{
545 enum fio_ddir ddir;
546
547 /*
548 * see if it's time to fsync
549 */
550 if (td->o.fsync_blocks &&
551 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
552 td->io_issues[DDIR_WRITE] && should_fsync(td))
553 return DDIR_SYNC;
554
555 /*
556 * see if it's time to fdatasync
557 */
558 if (td->o.fdatasync_blocks &&
559 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
560 td->io_issues[DDIR_WRITE] && should_fsync(td))
561 return DDIR_DATASYNC;
562
563 /*
564 * see if it's time to sync_file_range
565 */
566 if (td->sync_file_range_nr &&
567 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
568 td->io_issues[DDIR_WRITE] && should_fsync(td))
569 return DDIR_SYNC_FILE_RANGE;
570
571 if (td_rw(td)) {
572 /*
573 * Check if it's time to seed a new data direction.
574 */
575 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
576 /*
577 * Put a top limit on how many bytes we do for
578 * one data direction, to avoid overflowing the
579 * ranges too much
580 */
581 ddir = get_rand_ddir(td);
582
583 if (ddir != td->rwmix_ddir)
584 set_rwmix_bytes(td);
585
586 td->rwmix_ddir = ddir;
587 }
588 ddir = td->rwmix_ddir;
589 } else if (td_read(td))
590 ddir = DDIR_READ;
591 else
592 ddir = DDIR_WRITE;
593
594 td->rwmix_ddir = rate_ddir(td, ddir);
595 return td->rwmix_ddir;
596}
597
598static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
599{
600 io_u->ddir = get_rw_ddir(td);
601
602 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
603 td->o.barrier_blocks &&
604 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
605 td->io_issues[DDIR_WRITE])
606 io_u->flags |= IO_U_F_BARRIER;
607}
608
609void put_file_log(struct thread_data *td, struct fio_file *f)
610{
611 int ret = put_file(td, f);
612
613 if (ret)
614 td_verror(td, ret, "file close");
615}
616
617void put_io_u(struct thread_data *td, struct io_u *io_u)
618{
619 td_io_u_lock(td);
620
621 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
622 put_file_log(td, io_u->file);
623 io_u->file = NULL;
624 io_u->flags &= ~IO_U_F_FREE_DEF;
625 io_u->flags |= IO_U_F_FREE;
626
627 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
628 td->cur_depth--;
629 flist_del_init(&io_u->list);
630 flist_add(&io_u->list, &td->io_u_freelist);
631 td_io_u_unlock(td);
632 td_io_u_free_notify(td);
633}
634
635void clear_io_u(struct thread_data *td, struct io_u *io_u)
636{
637 io_u->flags &= ~IO_U_F_FLIGHT;
638 put_io_u(td, io_u);
639}
640
641void requeue_io_u(struct thread_data *td, struct io_u **io_u)
642{
643 struct io_u *__io_u = *io_u;
644
645 dprint(FD_IO, "requeue %p\n", __io_u);
646
647 td_io_u_lock(td);
648
649 __io_u->flags |= IO_U_F_FREE;
650 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
651 td->io_issues[__io_u->ddir]--;
652
653 __io_u->flags &= ~IO_U_F_FLIGHT;
654 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
655 td->cur_depth--;
656 flist_del(&__io_u->list);
657 flist_add_tail(&__io_u->list, &td->io_u_requeues);
658 td_io_u_unlock(td);
659 *io_u = NULL;
660}
661
662static int fill_io_u(struct thread_data *td, struct io_u *io_u)
663{
664 if (td->io_ops->flags & FIO_NOIO)
665 goto out;
666
667 set_rw_ddir(td, io_u);
668
669 /*
670 * fsync() or fdatasync() or trim etc, we are done
671 */
672 if (!ddir_rw(io_u->ddir))
673 goto out;
674
675 /*
676 * See if it's time to switch to a new zone
677 */
678 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
679 td->zone_bytes = 0;
680 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
681 io_u->file->last_pos = io_u->file->file_offset;
682 td->io_skip_bytes += td->o.zone_skip;
683 }
684
685 /*
686 * No log, let the seq/rand engine retrieve the next buflen and
687 * position.
688 */
689 if (get_next_offset(td, io_u)) {
690 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
691 return 1;
692 }
693
694 io_u->buflen = get_next_buflen(td, io_u);
695 if (!io_u->buflen) {
696 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
697 return 1;
698 }
699
700 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
701 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
702 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
703 io_u->buflen, io_u->file->real_file_size);
704 return 1;
705 }
706
707 /*
708 * mark entry before potentially trimming io_u
709 */
710 if (td_random(td) && file_randommap(td, io_u->file))
711 mark_random_map(td, io_u);
712
713 /*
714 * If using a write iolog, store this entry.
715 */
716out:
717 dprint_io_u(io_u, "fill_io_u");
718 td->zone_bytes += io_u->buflen;
719 log_io_u(td, io_u);
720 return 0;
721}
722
723static void __io_u_mark_map(unsigned int *map, unsigned int nr)
724{
725 int idx = 0;
726
727 switch (nr) {
728 default:
729 idx = 6;
730 break;
731 case 33 ... 64:
732 idx = 5;
733 break;
734 case 17 ... 32:
735 idx = 4;
736 break;
737 case 9 ... 16:
738 idx = 3;
739 break;
740 case 5 ... 8:
741 idx = 2;
742 break;
743 case 1 ... 4:
744 idx = 1;
745 case 0:
746 break;
747 }
748
749 map[idx]++;
750}
751
752void io_u_mark_submit(struct thread_data *td, unsigned int nr)
753{
754 __io_u_mark_map(td->ts.io_u_submit, nr);
755 td->ts.total_submit++;
756}
757
758void io_u_mark_complete(struct thread_data *td, unsigned int nr)
759{
760 __io_u_mark_map(td->ts.io_u_complete, nr);
761 td->ts.total_complete++;
762}
763
764void io_u_mark_depth(struct thread_data *td, unsigned int nr)
765{
766 int idx = 0;
767
768 switch (td->cur_depth) {
769 default:
770 idx = 6;
771 break;
772 case 32 ... 63:
773 idx = 5;
774 break;
775 case 16 ... 31:
776 idx = 4;
777 break;
778 case 8 ... 15:
779 idx = 3;
780 break;
781 case 4 ... 7:
782 idx = 2;
783 break;
784 case 2 ... 3:
785 idx = 1;
786 case 1:
787 break;
788 }
789
790 td->ts.io_u_map[idx] += nr;
791}
792
793static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
794{
795 int idx = 0;
796
797 assert(usec < 1000);
798
799 switch (usec) {
800 case 750 ... 999:
801 idx = 9;
802 break;
803 case 500 ... 749:
804 idx = 8;
805 break;
806 case 250 ... 499:
807 idx = 7;
808 break;
809 case 100 ... 249:
810 idx = 6;
811 break;
812 case 50 ... 99:
813 idx = 5;
814 break;
815 case 20 ... 49:
816 idx = 4;
817 break;
818 case 10 ... 19:
819 idx = 3;
820 break;
821 case 4 ... 9:
822 idx = 2;
823 break;
824 case 2 ... 3:
825 idx = 1;
826 case 0 ... 1:
827 break;
828 }
829
830 assert(idx < FIO_IO_U_LAT_U_NR);
831 td->ts.io_u_lat_u[idx]++;
832}
833
834static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
835{
836 int idx = 0;
837
838 switch (msec) {
839 default:
840 idx = 11;
841 break;
842 case 1000 ... 1999:
843 idx = 10;
844 break;
845 case 750 ... 999:
846 idx = 9;
847 break;
848 case 500 ... 749:
849 idx = 8;
850 break;
851 case 250 ... 499:
852 idx = 7;
853 break;
854 case 100 ... 249:
855 idx = 6;
856 break;
857 case 50 ... 99:
858 idx = 5;
859 break;
860 case 20 ... 49:
861 idx = 4;
862 break;
863 case 10 ... 19:
864 idx = 3;
865 break;
866 case 4 ... 9:
867 idx = 2;
868 break;
869 case 2 ... 3:
870 idx = 1;
871 case 0 ... 1:
872 break;
873 }
874
875 assert(idx < FIO_IO_U_LAT_M_NR);
876 td->ts.io_u_lat_m[idx]++;
877}
878
879static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
880{
881 if (usec < 1000)
882 io_u_mark_lat_usec(td, usec);
883 else
884 io_u_mark_lat_msec(td, usec / 1000);
885}
886
887/*
888 * Get next file to service by choosing one at random
889 */
890static struct fio_file *get_next_file_rand(struct thread_data *td,
891 enum fio_file_flags goodf,
892 enum fio_file_flags badf)
893{
894 struct fio_file *f;
895 int fno;
896
897 do {
898 int opened = 0;
899 unsigned long r;
900
901 if (td->o.use_os_rand) {
902 r = os_random_long(&td->next_file_state);
903 fno = (unsigned int) ((double) td->o.nr_files
904 * (r / (OS_RAND_MAX + 1.0)));
905 } else {
906 r = __rand(&td->__next_file_state);
907 fno = (unsigned int) ((double) td->o.nr_files
908 * (r / (FRAND_MAX + 1.0)));
909 }
910
911 f = td->files[fno];
912 if (fio_file_done(f))
913 continue;
914
915 if (!fio_file_open(f)) {
916 int err;
917
918 err = td_io_open_file(td, f);
919 if (err)
920 continue;
921 opened = 1;
922 }
923
924 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
925 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
926 return f;
927 }
928 if (opened)
929 td_io_close_file(td, f);
930 } while (1);
931}
932
933/*
934 * Get next file to service by doing round robin between all available ones
935 */
936static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
937 int badf)
938{
939 unsigned int old_next_file = td->next_file;
940 struct fio_file *f;
941
942 do {
943 int opened = 0;
944
945 f = td->files[td->next_file];
946
947 td->next_file++;
948 if (td->next_file >= td->o.nr_files)
949 td->next_file = 0;
950
951 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
952 if (fio_file_done(f)) {
953 f = NULL;
954 continue;
955 }
956
957 if (!fio_file_open(f)) {
958 int err;
959
960 err = td_io_open_file(td, f);
961 if (err) {
962 dprint(FD_FILE, "error %d on open of %s\n",
963 err, f->file_name);
964 f = NULL;
965 continue;
966 }
967 opened = 1;
968 }
969
970 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
971 f->flags);
972 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
973 break;
974
975 if (opened)
976 td_io_close_file(td, f);
977
978 f = NULL;
979 } while (td->next_file != old_next_file);
980
981 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
982 return f;
983}
984
985static struct fio_file *__get_next_file(struct thread_data *td)
986{
987 struct fio_file *f;
988
989 assert(td->o.nr_files <= td->files_index);
990
991 if (td->nr_done_files >= td->o.nr_files) {
992 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
993 " nr_files=%d\n", td->nr_open_files,
994 td->nr_done_files,
995 td->o.nr_files);
996 return NULL;
997 }
998
999 f = td->file_service_file;
1000 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1001 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1002 goto out;
1003 if (td->file_service_left--)
1004 goto out;
1005 }
1006
1007 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1008 td->o.file_service_type == FIO_FSERVICE_SEQ)
1009 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1010 else
1011 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1012
1013 td->file_service_file = f;
1014 td->file_service_left = td->file_service_nr - 1;
1015out:
1016 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1017 return f;
1018}
1019
1020static struct fio_file *get_next_file(struct thread_data *td)
1021{
1022 struct prof_io_ops *ops = &td->prof_io_ops;
1023
1024 if (ops->get_next_file)
1025 return ops->get_next_file(td);
1026
1027 return __get_next_file(td);
1028}
1029
1030static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1031{
1032 struct fio_file *f;
1033
1034 do {
1035 f = get_next_file(td);
1036 if (!f)
1037 return 1;
1038
1039 io_u->file = f;
1040 get_file(f);
1041
1042 if (!fill_io_u(td, io_u))
1043 break;
1044
1045 put_file_log(td, f);
1046 td_io_close_file(td, f);
1047 io_u->file = NULL;
1048 fio_file_set_done(f);
1049 td->nr_done_files++;
1050 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1051 td->nr_done_files, td->o.nr_files);
1052 } while (1);
1053
1054 return 0;
1055}
1056
1057
1058struct io_u *__get_io_u(struct thread_data *td)
1059{
1060 struct io_u *io_u = NULL;
1061
1062 td_io_u_lock(td);
1063
1064again:
1065 if (!flist_empty(&td->io_u_requeues))
1066 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
1067 else if (!queue_full(td)) {
1068 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
1069
1070 io_u->buflen = 0;
1071 io_u->resid = 0;
1072 io_u->file = NULL;
1073 io_u->end_io = NULL;
1074 }
1075
1076 if (io_u) {
1077 assert(io_u->flags & IO_U_F_FREE);
1078 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1079 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1080 io_u->flags &= ~IO_U_F_VER_LIST;
1081
1082 io_u->error = 0;
1083 flist_del(&io_u->list);
1084 flist_add(&io_u->list, &td->io_u_busylist);
1085 td->cur_depth++;
1086 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1087 } else if (td->o.verify_async) {
1088 /*
1089 * We ran out, wait for async verify threads to finish and
1090 * return one
1091 */
1092 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1093 goto again;
1094 }
1095
1096 td_io_u_unlock(td);
1097 return io_u;
1098}
1099
1100static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1101{
1102 if (td->o.trim_backlog && td->trim_entries) {
1103 int get_trim = 0;
1104
1105 if (td->trim_batch) {
1106 td->trim_batch--;
1107 get_trim = 1;
1108 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1109 td->last_ddir != DDIR_READ) {
1110 td->trim_batch = td->o.trim_batch;
1111 if (!td->trim_batch)
1112 td->trim_batch = td->o.trim_backlog;
1113 get_trim = 1;
1114 }
1115
1116 if (get_trim && !get_next_trim(td, io_u))
1117 return 1;
1118 }
1119
1120 return 0;
1121}
1122
1123static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1124{
1125 if (td->o.verify_backlog && td->io_hist_len) {
1126 int get_verify = 0;
1127
1128 if (td->verify_batch)
1129 get_verify = 1;
1130 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1131 td->last_ddir != DDIR_READ) {
1132 td->verify_batch = td->o.verify_batch;
1133 if (!td->verify_batch)
1134 td->verify_batch = td->o.verify_backlog;
1135 get_verify = 1;
1136 }
1137
1138 if (get_verify && !get_next_verify(td, io_u)) {
1139 td->verify_batch--;
1140 return 1;
1141 }
1142 }
1143
1144 return 0;
1145}
1146
1147/*
1148 * Fill offset and start time into the buffer content, to prevent too
1149 * easy compressible data for simple de-dupe attempts. Do this for every
1150 * 512b block in the range, since that should be the smallest block size
1151 * we can expect from a device.
1152 */
1153static void small_content_scramble(struct io_u *io_u)
1154{
1155 unsigned int i, nr_blocks = io_u->buflen / 512;
1156 unsigned long long boffset;
1157 unsigned int offset;
1158 void *p, *end;
1159
1160 if (!nr_blocks)
1161 return;
1162
1163 p = io_u->xfer_buf;
1164 boffset = io_u->offset;
1165 io_u->buf_filled_len = 0;
1166
1167 for (i = 0; i < nr_blocks; i++) {
1168 /*
1169 * Fill the byte offset into a "random" start offset of
1170 * the buffer, given by the product of the usec time
1171 * and the actual offset.
1172 */
1173 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1174 offset &= ~(sizeof(unsigned long long) - 1);
1175 if (offset >= 512 - sizeof(unsigned long long))
1176 offset -= sizeof(unsigned long long);
1177 memcpy(p + offset, &boffset, sizeof(boffset));
1178
1179 end = p + 512 - sizeof(io_u->start_time);
1180 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1181 p += 512;
1182 boffset += 512;
1183 }
1184}
1185
1186/*
1187 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1188 * etc. The returned io_u is fully ready to be prepped and submitted.
1189 */
1190struct io_u *get_io_u(struct thread_data *td)
1191{
1192 struct fio_file *f;
1193 struct io_u *io_u;
1194 int do_scramble = 0;
1195
1196 io_u = __get_io_u(td);
1197 if (!io_u) {
1198 dprint(FD_IO, "__get_io_u failed\n");
1199 return NULL;
1200 }
1201
1202 if (check_get_verify(td, io_u))
1203 goto out;
1204 if (check_get_trim(td, io_u))
1205 goto out;
1206
1207 /*
1208 * from a requeue, io_u already setup
1209 */
1210 if (io_u->file)
1211 goto out;
1212
1213 /*
1214 * If using an iolog, grab next piece if any available.
1215 */
1216 if (td->o.read_iolog_file) {
1217 if (read_iolog_get(td, io_u))
1218 goto err_put;
1219 } else if (set_io_u_file(td, io_u)) {
1220 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1221 goto err_put;
1222 }
1223
1224 f = io_u->file;
1225 assert(fio_file_open(f));
1226
1227 if (ddir_rw(io_u->ddir)) {
1228 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1229 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1230 goto err_put;
1231 }
1232
1233 f->last_start = io_u->offset;
1234 f->last_pos = io_u->offset + io_u->buflen;
1235
1236 if (io_u->ddir == DDIR_WRITE) {
1237 if (td->o.refill_buffers) {
1238 io_u_fill_buffer(td, io_u,
1239 io_u->xfer_buflen, io_u->xfer_buflen);
1240 } else if (td->o.scramble_buffers)
1241 do_scramble = 1;
1242 if (td->o.verify != VERIFY_NONE) {
1243 populate_verify_io_u(td, io_u);
1244 do_scramble = 0;
1245 }
1246 } else if (io_u->ddir == DDIR_READ) {
1247 /*
1248 * Reset the buf_filled parameters so next time if the
1249 * buffer is used for writes it is refilled.
1250 */
1251 io_u->buf_filled_len = 0;
1252 }
1253 }
1254
1255 /*
1256 * Set io data pointers.
1257 */
1258 io_u->xfer_buf = io_u->buf;
1259 io_u->xfer_buflen = io_u->buflen;
1260
1261out:
1262 assert(io_u->file);
1263 if (!td_io_prep(td, io_u)) {
1264 if (!td->o.disable_slat)
1265 fio_gettime(&io_u->start_time, NULL);
1266 if (do_scramble)
1267 small_content_scramble(io_u);
1268 return io_u;
1269 }
1270err_put:
1271 dprint(FD_IO, "get_io_u failed\n");
1272 put_io_u(td, io_u);
1273 return NULL;
1274}
1275
1276void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1277{
1278 const char *msg[] = { "read", "write", "sync", "datasync",
1279 "sync_file_range", "wait", "trim" };
1280
1281
1282
1283 log_err("fio: io_u error");
1284
1285 if (io_u->file)
1286 log_err(" on file %s", io_u->file->file_name);
1287
1288 log_err(": %s\n", strerror(io_u->error));
1289
1290 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1291 io_u->offset, io_u->xfer_buflen);
1292
1293 if (!td->error)
1294 td_verror(td, io_u->error, "io_u error");
1295}
1296
1297static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1298 struct io_completion_data *icd,
1299 const enum fio_ddir idx, unsigned int bytes)
1300{
1301 unsigned long uninitialized_var(lusec);
1302
1303 if (!td->o.disable_clat || !td->o.disable_bw)
1304 lusec = utime_since(&io_u->issue_time, &icd->time);
1305
1306 if (!td->o.disable_lat) {
1307 unsigned long tusec;
1308
1309 tusec = utime_since(&io_u->start_time, &icd->time);
1310 add_lat_sample(td, idx, tusec, bytes);
1311 }
1312
1313 if (!td->o.disable_clat) {
1314 add_clat_sample(td, idx, lusec, bytes);
1315 io_u_mark_latency(td, lusec);
1316 }
1317
1318 if (!td->o.disable_bw)
1319 add_bw_sample(td, idx, bytes, &icd->time);
1320
1321 add_iops_sample(td, idx, &icd->time);
1322}
1323
1324static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1325{
1326 unsigned long long secs, remainder, bps, bytes;
1327 bytes = td->this_io_bytes[ddir];
1328 bps = td->rate_bps[ddir];
1329 secs = bytes / bps;
1330 remainder = bytes % bps;
1331 return remainder * 1000000 / bps + secs * 1000000;
1332}
1333
1334static void io_completed(struct thread_data *td, struct io_u *io_u,
1335 struct io_completion_data *icd)
1336{
1337 /*
1338 * Older gcc's are too dumb to realize that usec is always used
1339 * initialized, silence that warning.
1340 */
1341 unsigned long uninitialized_var(usec);
1342 struct fio_file *f;
1343
1344 dprint_io_u(io_u, "io complete");
1345
1346 td_io_u_lock(td);
1347 assert(io_u->flags & IO_U_F_FLIGHT);
1348 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1349 td_io_u_unlock(td);
1350
1351 if (ddir_sync(io_u->ddir)) {
1352 td->last_was_sync = 1;
1353 f = io_u->file;
1354 if (f) {
1355 f->first_write = -1ULL;
1356 f->last_write = -1ULL;
1357 }
1358 return;
1359 }
1360
1361 td->last_was_sync = 0;
1362 td->last_ddir = io_u->ddir;
1363
1364 if (!io_u->error && ddir_rw(io_u->ddir)) {
1365 unsigned int bytes = io_u->buflen - io_u->resid;
1366 const enum fio_ddir idx = io_u->ddir;
1367 const enum fio_ddir odx = io_u->ddir ^ 1;
1368 int ret;
1369
1370 td->io_blocks[idx]++;
1371 td->this_io_blocks[idx]++;
1372 td->io_bytes[idx] += bytes;
1373 td->this_io_bytes[idx] += bytes;
1374
1375 if (idx == DDIR_WRITE) {
1376 f = io_u->file;
1377 if (f) {
1378 if (f->first_write == -1ULL ||
1379 io_u->offset < f->first_write)
1380 f->first_write = io_u->offset;
1381 if (f->last_write == -1ULL ||
1382 ((io_u->offset + bytes) > f->last_write))
1383 f->last_write = io_u->offset + bytes;
1384 }
1385 }
1386
1387 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1388 td->runstate == TD_VERIFYING)) {
1389 account_io_completion(td, io_u, icd, idx, bytes);
1390
1391 if (__should_check_rate(td, idx)) {
1392 td->rate_pending_usleep[idx] =
1393 (usec_for_io(td, idx) -
1394 utime_since_now(&td->start));
1395 }
1396 if (__should_check_rate(td, odx))
1397 td->rate_pending_usleep[odx] =
1398 (usec_for_io(td, odx) -
1399 utime_since_now(&td->start));
1400 }
1401
1402 if (td_write(td) && idx == DDIR_WRITE &&
1403 td->o.do_verify &&
1404 td->o.verify != VERIFY_NONE)
1405 log_io_piece(td, io_u);
1406
1407 icd->bytes_done[idx] += bytes;
1408
1409 if (io_u->end_io) {
1410 ret = io_u->end_io(td, io_u);
1411 if (ret && !icd->error)
1412 icd->error = ret;
1413 }
1414 } else if (io_u->error) {
1415 icd->error = io_u->error;
1416 io_u_log_error(td, io_u);
1417 }
1418 if (icd->error && td_non_fatal_error(icd->error) &&
1419 (td->o.continue_on_error & td_error_type(io_u->ddir, icd->error))) {
1420 /*
1421 * If there is a non_fatal error, then add to the error count
1422 * and clear all the errors.
1423 */
1424 update_error_count(td, icd->error);
1425 td_clear_error(td);
1426 icd->error = 0;
1427 io_u->error = 0;
1428 }
1429}
1430
1431static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1432 int nr)
1433{
1434 if (!td->o.disable_clat || !td->o.disable_bw)
1435 fio_gettime(&icd->time, NULL);
1436
1437 icd->nr = nr;
1438
1439 icd->error = 0;
1440 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1441}
1442
1443static void ios_completed(struct thread_data *td,
1444 struct io_completion_data *icd)
1445{
1446 struct io_u *io_u;
1447 int i;
1448
1449 for (i = 0; i < icd->nr; i++) {
1450 io_u = td->io_ops->event(td, i);
1451
1452 io_completed(td, io_u, icd);
1453
1454 if (!(io_u->flags & IO_U_F_FREE_DEF))
1455 put_io_u(td, io_u);
1456 }
1457}
1458
1459/*
1460 * Complete a single io_u for the sync engines.
1461 */
1462int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1463 unsigned long *bytes)
1464{
1465 struct io_completion_data icd;
1466
1467 init_icd(td, &icd, 1);
1468 io_completed(td, io_u, &icd);
1469
1470 if (!(io_u->flags & IO_U_F_FREE_DEF))
1471 put_io_u(td, io_u);
1472
1473 if (icd.error) {
1474 td_verror(td, icd.error, "io_u_sync_complete");
1475 return -1;
1476 }
1477
1478 if (bytes) {
1479 bytes[0] += icd.bytes_done[0];
1480 bytes[1] += icd.bytes_done[1];
1481 }
1482
1483 return 0;
1484}
1485
1486/*
1487 * Called to complete min_events number of io for the async engines.
1488 */
1489int io_u_queued_complete(struct thread_data *td, int min_evts,
1490 unsigned long *bytes)
1491{
1492 struct io_completion_data icd;
1493 struct timespec *tvp = NULL;
1494 int ret;
1495 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1496
1497 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1498
1499 if (!min_evts)
1500 tvp = &ts;
1501
1502 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1503 if (ret < 0) {
1504 td_verror(td, -ret, "td_io_getevents");
1505 return ret;
1506 } else if (!ret)
1507 return ret;
1508
1509 init_icd(td, &icd, ret);
1510 ios_completed(td, &icd);
1511 if (icd.error) {
1512 td_verror(td, icd.error, "io_u_queued_complete");
1513 return -1;
1514 }
1515
1516 if (bytes) {
1517 bytes[0] += icd.bytes_done[0];
1518 bytes[1] += icd.bytes_done[1];
1519 }
1520
1521 return 0;
1522}
1523
1524/*
1525 * Call when io_u is really queued, to update the submission latency.
1526 */
1527void io_u_queued(struct thread_data *td, struct io_u *io_u)
1528{
1529 if (!td->o.disable_slat) {
1530 unsigned long slat_time;
1531
1532 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1533 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1534 }
1535}
1536
1537/*
1538 * "randomly" fill the buffer contents
1539 */
1540void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1541 unsigned int min_write, unsigned int max_bs)
1542{
1543 io_u->buf_filled_len = 0;
1544
1545 if (!td->o.zero_buffers) {
1546 unsigned int perc = td->o.compress_percentage;
1547
1548 if (perc) {
1549 unsigned int seg = min_write;
1550
1551 seg = min(min_write, td->o.compress_chunk);
1552 fill_random_buf_percentage(&td->buf_state, io_u->buf,
1553 perc, seg, max_bs);
1554 } else
1555 fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1556 } else
1557 memset(io_u->buf, 0, max_bs);
1558}
fio - Flexible IO Tester