projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Kill now unused bitmap defines from legacy code
[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#include "lib/bitmap.h"
14
15struct io_completion_data {
16 int nr; /* input */
17
18 int error; /* output */
19 unsigned long bytes_done[DDIR_RWDIR_CNT]; /* output */
20 struct timeval time; /* output */
21};
22
23/*
24 * The ->io_bitmap contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
26 */
27static int random_map_free(struct fio_file *f, const unsigned long long block)
28{
29 return !bitmap_isset(f->io_bitmap, block);
30}
31
32/*
33 * Mark a given offset as used in the map.
34 */
35static void mark_random_map(struct thread_data *td, struct io_u *io_u)
36{
37 unsigned int min_bs = td->o.rw_min_bs;
38 struct fio_file *f = io_u->file;
39 unsigned long long block;
40 unsigned int nr_blocks;
41
42 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
43 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
44
45 nr_blocks = bitmap_set_nr(f->io_bitmap, block, nr_blocks);
46
47 if ((nr_blocks * min_bs) < io_u->buflen)
48 io_u->buflen = nr_blocks * min_bs;
49}
50
51static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
52 enum fio_ddir ddir)
53{
54 unsigned long long max_blocks;
55 unsigned long long max_size;
56
57 assert(ddir_rw(ddir));
58
59 /*
60 * Hmm, should we make sure that ->io_size <= ->real_file_size?
61 */
62 max_size = f->io_size;
63 if (max_size > f->real_file_size)
64 max_size = f->real_file_size;
65
66 if (td->o.zone_range)
67 max_size = td->o.zone_range;
68
69 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
70 if (!max_blocks)
71 return 0;
72
73 return max_blocks;
74}
75
76static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
77 enum fio_ddir ddir, unsigned long long *b)
78{
79 unsigned long long rmax, r, lastb;
80
81 lastb = last_block(td, f, ddir);
82 if (!lastb)
83 return 1;
84
85 rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
86
87 if (td->o.use_os_rand) {
88 rmax = OS_RAND_MAX;
89 r = os_random_long(&td->random_state);
90 } else {
91 rmax = FRAND_MAX;
92 r = __rand(&td->__random_state);
93 }
94
95 *b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0));
96
97 dprint(FD_RANDOM, "off rand %llu\n", r);
98
99 /*
100 * if we are not maintaining a random map, we are done.
101 */
102 if (!file_randommap(td, f))
103 goto ret;
104
105 /*
106 * calculate map offset and check if it's free
107 */
108 if (random_map_free(f, *b))
109 goto ret;
110
111 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", *b);
112
113 *b = bitmap_next_free(f->io_bitmap, *b);
114 if (*b == (uint64_t) -1ULL)
115 return 1;
116ret:
117 return 0;
118}
119
120static int __get_next_rand_offset_zipf(struct thread_data *td,
121 struct fio_file *f, enum fio_ddir ddir,
122 unsigned long long *b)
123{
124 *b = zipf_next(&f->zipf);
125 return 0;
126}
127
128static int __get_next_rand_offset_pareto(struct thread_data *td,
129 struct fio_file *f, enum fio_ddir ddir,
130 unsigned long long *b)
131{
132 *b = pareto_next(&f->zipf);
133 return 0;
134}
135
136static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
137 enum fio_ddir ddir, unsigned long long *b)
138{
139 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
140 return __get_next_rand_offset(td, f, ddir, b);
141 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
142 return __get_next_rand_offset_zipf(td, f, ddir, b);
143 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
144 return __get_next_rand_offset_pareto(td, f, ddir, b);
145
146 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
147 return 1;
148}
149
150static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
151 enum fio_ddir ddir, unsigned long long *b)
152{
153 if (!get_next_rand_offset(td, f, ddir, b))
154 return 0;
155
156 if (td->o.time_based) {
157 fio_file_reset(f);
158 if (!get_next_rand_offset(td, f, ddir, b))
159 return 0;
160 }
161
162 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
163 f->file_name, f->last_pos, f->real_file_size);
164 return 1;
165}
166
167static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
168 enum fio_ddir ddir, unsigned long long *offset)
169{
170 assert(ddir_rw(ddir));
171
172 if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
173 f->last_pos = f->last_pos - f->io_size;
174
175 if (f->last_pos < f->real_file_size) {
176 unsigned long long pos;
177
178 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
179 f->last_pos = f->real_file_size;
180
181 pos = f->last_pos - f->file_offset;
182 if (pos)
183 pos += td->o.ddir_seq_add;
184
185 *offset = pos;
186 return 0;
187 }
188
189 return 1;
190}
191
192static int get_next_block(struct thread_data *td, struct io_u *io_u,
193 enum fio_ddir ddir, int rw_seq)
194{
195 struct fio_file *f = io_u->file;
196 unsigned long long b, offset;
197 int ret;
198
199 assert(ddir_rw(ddir));
200
201 b = offset = -1ULL;
202
203 if (rw_seq) {
204 if (td_random(td))
205 ret = get_next_rand_block(td, f, ddir, &b);
206 else
207 ret = get_next_seq_offset(td, f, ddir, &offset);
208 } else {
209 io_u->flags |= IO_U_F_BUSY_OK;
210
211 if (td->o.rw_seq == RW_SEQ_SEQ) {
212 ret = get_next_seq_offset(td, f, ddir, &offset);
213 if (ret)
214 ret = get_next_rand_block(td, f, ddir, &b);
215 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
216 if (f->last_start != -1ULL)
217 offset = f->last_start - f->file_offset;
218 else
219 offset = 0;
220 ret = 0;
221 } else {
222 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
223 ret = 1;
224 }
225 }
226
227 if (!ret) {
228 if (offset != -1ULL)
229 io_u->offset = offset;
230 else if (b != -1ULL)
231 io_u->offset = b * td->o.ba[ddir];
232 else {
233 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n",
234 offset, b);
235 ret = 1;
236 }
237 }
238
239 return ret;
240}
241
242/*
243 * For random io, generate a random new block and see if it's used. Repeat
244 * until we find a free one. For sequential io, just return the end of
245 * the last io issued.
246 */
247static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
248{
249 struct fio_file *f = io_u->file;
250 enum fio_ddir ddir = io_u->ddir;
251 int rw_seq_hit = 0;
252
253 assert(ddir_rw(ddir));
254
255 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
256 rw_seq_hit = 1;
257 td->ddir_seq_nr = td->o.ddir_seq_nr;
258 }
259
260 if (get_next_block(td, io_u, ddir, rw_seq_hit))
261 return 1;
262
263 if (io_u->offset >= f->io_size) {
264 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
265 io_u->offset, f->io_size);
266 return 1;
267 }
268
269 io_u->offset += f->file_offset;
270 if (io_u->offset >= f->real_file_size) {
271 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
272 io_u->offset, f->real_file_size);
273 return 1;
274 }
275
276 return 0;
277}
278
279static int get_next_offset(struct thread_data *td, struct io_u *io_u)
280{
281 struct prof_io_ops *ops = &td->prof_io_ops;
282
283 if (ops->fill_io_u_off)
284 return ops->fill_io_u_off(td, io_u);
285
286 return __get_next_offset(td, io_u);
287}
288
289static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
290 unsigned int buflen)
291{
292 struct fio_file *f = io_u->file;
293
294 return io_u->offset + buflen <= f->io_size + get_start_offset(td);
295}
296
297static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
298{
299 const int ddir = io_u->ddir;
300 unsigned int buflen = 0;
301 unsigned int minbs, maxbs;
302 unsigned long r, rand_max;
303
304 assert(ddir_rw(ddir));
305
306 minbs = td->o.min_bs[ddir];
307 maxbs = td->o.max_bs[ddir];
308
309 if (minbs == maxbs)
310 return minbs;
311
312 /*
313 * If we can't satisfy the min block size from here, then fail
314 */
315 if (!io_u_fits(td, io_u, minbs))
316 return 0;
317
318 if (td->o.use_os_rand)
319 rand_max = OS_RAND_MAX;
320 else
321 rand_max = FRAND_MAX;
322
323 do {
324 if (td->o.use_os_rand)
325 r = os_random_long(&td->bsrange_state);
326 else
327 r = __rand(&td->__bsrange_state);
328
329 if (!td->o.bssplit_nr[ddir]) {
330 buflen = 1 + (unsigned int) ((double) maxbs *
331 (r / (rand_max + 1.0)));
332 if (buflen < minbs)
333 buflen = minbs;
334 } else {
335 long perc = 0;
336 unsigned int i;
337
338 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
339 struct bssplit *bsp = &td->o.bssplit[ddir][i];
340
341 buflen = bsp->bs;
342 perc += bsp->perc;
343 if ((r <= ((rand_max / 100L) * perc)) &&
344 io_u_fits(td, io_u, buflen))
345 break;
346 }
347 }
348
349 if (!td->o.bs_unaligned && is_power_of_2(minbs))
350 buflen = (buflen + minbs - 1) & ~(minbs - 1);
351
352 } while (!io_u_fits(td, io_u, buflen));
353
354 return buflen;
355}
356
357static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
358{
359 struct prof_io_ops *ops = &td->prof_io_ops;
360
361 if (ops->fill_io_u_size)
362 return ops->fill_io_u_size(td, io_u);
363
364 return __get_next_buflen(td, io_u);
365}
366
367static void set_rwmix_bytes(struct thread_data *td)
368{
369 unsigned int diff;
370
371 /*
372 * we do time or byte based switch. this is needed because
373 * buffered writes may issue a lot quicker than they complete,
374 * whereas reads do not.
375 */
376 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
377 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
378}
379
380static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
381{
382 unsigned int v;
383 unsigned long r;
384
385 if (td->o.use_os_rand) {
386 r = os_random_long(&td->rwmix_state);
387 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
388 } else {
389 r = __rand(&td->__rwmix_state);
390 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
391 }
392
393 if (v <= td->o.rwmix[DDIR_READ])
394 return DDIR_READ;
395
396 return DDIR_WRITE;
397}
398
399static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
400{
401 enum fio_ddir odir = ddir ^ 1;
402 struct timeval t;
403 long usec;
404
405 assert(ddir_rw(ddir));
406
407 if (td->rate_pending_usleep[ddir] <= 0)
408 return ddir;
409
410 /*
411 * We have too much pending sleep in this direction. See if we
412 * should switch.
413 */
414 if (td_rw(td)) {
415 /*
416 * Other direction does not have too much pending, switch
417 */
418 if (td->rate_pending_usleep[odir] < 100000)
419 return odir;
420
421 /*
422 * Both directions have pending sleep. Sleep the minimum time
423 * and deduct from both.
424 */
425 if (td->rate_pending_usleep[ddir] <=
426 td->rate_pending_usleep[odir]) {
427 usec = td->rate_pending_usleep[ddir];
428 } else {
429 usec = td->rate_pending_usleep[odir];
430 ddir = odir;
431 }
432 } else
433 usec = td->rate_pending_usleep[ddir];
434
435 /*
436 * We are going to sleep, ensure that we flush anything pending as
437 * not to skew our latency numbers.
438 *
439 * Changed to only monitor 'in flight' requests here instead of the
440 * td->cur_depth, b/c td->cur_depth does not accurately represent
441 * io's that have been actually submitted to an async engine,
442 * and cur_depth is meaningless for sync engines.
443 */
444 if (td->io_u_in_flight) {
445 int fio_unused ret;
446
447 ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
448 }
449
450 fio_gettime(&t, NULL);
451 usec_sleep(td, usec);
452 usec = utime_since_now(&t);
453
454 td->rate_pending_usleep[ddir] -= usec;
455
456 odir = ddir ^ 1;
457 if (td_rw(td) && __should_check_rate(td, odir))
458 td->rate_pending_usleep[odir] -= usec;
459
460 if (ddir_trim(ddir))
461 return ddir;
462 return ddir;
463}
464
465/*
466 * Return the data direction for the next io_u. If the job is a
467 * mixed read/write workload, check the rwmix cycle and switch if
468 * necessary.
469 */
470static enum fio_ddir get_rw_ddir(struct thread_data *td)
471{
472 enum fio_ddir ddir;
473
474 /*
475 * see if it's time to fsync
476 */
477 if (td->o.fsync_blocks &&
478 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
479 td->io_issues[DDIR_WRITE] && should_fsync(td))
480 return DDIR_SYNC;
481
482 /*
483 * see if it's time to fdatasync
484 */
485 if (td->o.fdatasync_blocks &&
486 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
487 td->io_issues[DDIR_WRITE] && should_fsync(td))
488 return DDIR_DATASYNC;
489
490 /*
491 * see if it's time to sync_file_range
492 */
493 if (td->sync_file_range_nr &&
494 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
495 td->io_issues[DDIR_WRITE] && should_fsync(td))
496 return DDIR_SYNC_FILE_RANGE;
497
498 if (td_rw(td)) {
499 /*
500 * Check if it's time to seed a new data direction.
501 */
502 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
503 /*
504 * Put a top limit on how many bytes we do for
505 * one data direction, to avoid overflowing the
506 * ranges too much
507 */
508 ddir = get_rand_ddir(td);
509
510 if (ddir != td->rwmix_ddir)
511 set_rwmix_bytes(td);
512
513 td->rwmix_ddir = ddir;
514 }
515 ddir = td->rwmix_ddir;
516 } else if (td_read(td))
517 ddir = DDIR_READ;
518 else if (td_write(td))
519 ddir = DDIR_WRITE;
520 else
521 ddir = DDIR_TRIM;
522
523 td->rwmix_ddir = rate_ddir(td, ddir);
524 return td->rwmix_ddir;
525}
526
527static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
528{
529 io_u->ddir = get_rw_ddir(td);
530
531 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
532 td->o.barrier_blocks &&
533 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
534 td->io_issues[DDIR_WRITE])
535 io_u->flags |= IO_U_F_BARRIER;
536}
537
538void put_file_log(struct thread_data *td, struct fio_file *f)
539{
540 int ret = put_file(td, f);
541
542 if (ret)
543 td_verror(td, ret, "file close");
544}
545
546void put_io_u(struct thread_data *td, struct io_u *io_u)
547{
548 td_io_u_lock(td);
549
550 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
551 put_file_log(td, io_u->file);
552 io_u->file = NULL;
553 io_u->flags &= ~IO_U_F_FREE_DEF;
554 io_u->flags |= IO_U_F_FREE;
555
556 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
557 td->cur_depth--;
558 flist_del_init(&io_u->list);
559 flist_add(&io_u->list, &td->io_u_freelist);
560 td_io_u_unlock(td);
561 td_io_u_free_notify(td);
562}
563
564void clear_io_u(struct thread_data *td, struct io_u *io_u)
565{
566 io_u->flags &= ~IO_U_F_FLIGHT;
567 put_io_u(td, io_u);
568}
569
570void requeue_io_u(struct thread_data *td, struct io_u **io_u)
571{
572 struct io_u *__io_u = *io_u;
573
574 dprint(FD_IO, "requeue %p\n", __io_u);
575
576 td_io_u_lock(td);
577
578 __io_u->flags |= IO_U_F_FREE;
579 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
580 td->io_issues[__io_u->ddir]--;
581
582 __io_u->flags &= ~IO_U_F_FLIGHT;
583 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
584 td->cur_depth--;
585 flist_del(&__io_u->list);
586 flist_add_tail(&__io_u->list, &td->io_u_requeues);
587 td_io_u_unlock(td);
588 *io_u = NULL;
589}
590
591static int fill_io_u(struct thread_data *td, struct io_u *io_u)
592{
593 if (td->io_ops->flags & FIO_NOIO)
594 goto out;
595
596 set_rw_ddir(td, io_u);
597
598 /*
599 * fsync() or fdatasync() or trim etc, we are done
600 */
601 if (!ddir_rw(io_u->ddir))
602 goto out;
603
604 /*
605 * See if it's time to switch to a new zone
606 */
607 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
608 td->zone_bytes = 0;
609 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
610 io_u->file->last_pos = io_u->file->file_offset;
611 td->io_skip_bytes += td->o.zone_skip;
612 }
613
614 /*
615 * No log, let the seq/rand engine retrieve the next buflen and
616 * position.
617 */
618 if (get_next_offset(td, io_u)) {
619 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
620 return 1;
621 }
622
623 io_u->buflen = get_next_buflen(td, io_u);
624 if (!io_u->buflen) {
625 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
626 return 1;
627 }
628
629 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
630 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
631 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
632 io_u->buflen, io_u->file->real_file_size);
633 return 1;
634 }
635
636 /*
637 * mark entry before potentially trimming io_u
638 */
639 if (td_random(td) && file_randommap(td, io_u->file))
640 mark_random_map(td, io_u);
641
642 /*
643 * If using a write iolog, store this entry.
644 */
645out:
646 dprint_io_u(io_u, "fill_io_u");
647 td->zone_bytes += io_u->buflen;
648 log_io_u(td, io_u);
649 return 0;
650}
651
652static void __io_u_mark_map(unsigned int *map, unsigned int nr)
653{
654 int idx = 0;
655
656 switch (nr) {
657 default:
658 idx = 6;
659 break;
660 case 33 ... 64:
661 idx = 5;
662 break;
663 case 17 ... 32:
664 idx = 4;
665 break;
666 case 9 ... 16:
667 idx = 3;
668 break;
669 case 5 ... 8:
670 idx = 2;
671 break;
672 case 1 ... 4:
673 idx = 1;
674 case 0:
675 break;
676 }
677
678 map[idx]++;
679}
680
681void io_u_mark_submit(struct thread_data *td, unsigned int nr)
682{
683 __io_u_mark_map(td->ts.io_u_submit, nr);
684 td->ts.total_submit++;
685}
686
687void io_u_mark_complete(struct thread_data *td, unsigned int nr)
688{
689 __io_u_mark_map(td->ts.io_u_complete, nr);
690 td->ts.total_complete++;
691}
692
693void io_u_mark_depth(struct thread_data *td, unsigned int nr)
694{
695 int idx = 0;
696
697 switch (td->cur_depth) {
698 default:
699 idx = 6;
700 break;
701 case 32 ... 63:
702 idx = 5;
703 break;
704 case 16 ... 31:
705 idx = 4;
706 break;
707 case 8 ... 15:
708 idx = 3;
709 break;
710 case 4 ... 7:
711 idx = 2;
712 break;
713 case 2 ... 3:
714 idx = 1;
715 case 1:
716 break;
717 }
718
719 td->ts.io_u_map[idx] += nr;
720}
721
722static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
723{
724 int idx = 0;
725
726 assert(usec < 1000);
727
728 switch (usec) {
729 case 750 ... 999:
730 idx = 9;
731 break;
732 case 500 ... 749:
733 idx = 8;
734 break;
735 case 250 ... 499:
736 idx = 7;
737 break;
738 case 100 ... 249:
739 idx = 6;
740 break;
741 case 50 ... 99:
742 idx = 5;
743 break;
744 case 20 ... 49:
745 idx = 4;
746 break;
747 case 10 ... 19:
748 idx = 3;
749 break;
750 case 4 ... 9:
751 idx = 2;
752 break;
753 case 2 ... 3:
754 idx = 1;
755 case 0 ... 1:
756 break;
757 }
758
759 assert(idx < FIO_IO_U_LAT_U_NR);
760 td->ts.io_u_lat_u[idx]++;
761}
762
763static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
764{
765 int idx = 0;
766
767 switch (msec) {
768 default:
769 idx = 11;
770 break;
771 case 1000 ... 1999:
772 idx = 10;
773 break;
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_M_NR);
805 td->ts.io_u_lat_m[idx]++;
806}
807
808static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
809{
810 if (usec < 1000)
811 io_u_mark_lat_usec(td, usec);
812 else
813 io_u_mark_lat_msec(td, usec / 1000);
814}
815
816/*
817 * Get next file to service by choosing one at random
818 */
819static struct fio_file *get_next_file_rand(struct thread_data *td,
820 enum fio_file_flags goodf,
821 enum fio_file_flags badf)
822{
823 struct fio_file *f;
824 int fno;
825
826 do {
827 int opened = 0;
828 unsigned long r;
829
830 if (td->o.use_os_rand) {
831 r = os_random_long(&td->next_file_state);
832 fno = (unsigned int) ((double) td->o.nr_files
833 * (r / (OS_RAND_MAX + 1.0)));
834 } else {
835 r = __rand(&td->__next_file_state);
836 fno = (unsigned int) ((double) td->o.nr_files
837 * (r / (FRAND_MAX + 1.0)));
838 }
839
840 f = td->files[fno];
841 if (fio_file_done(f))
842 continue;
843
844 if (!fio_file_open(f)) {
845 int err;
846
847 err = td_io_open_file(td, f);
848 if (err)
849 continue;
850 opened = 1;
851 }
852
853 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
854 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
855 return f;
856 }
857 if (opened)
858 td_io_close_file(td, f);
859 } while (1);
860}
861
862/*
863 * Get next file to service by doing round robin between all available ones
864 */
865static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
866 int badf)
867{
868 unsigned int old_next_file = td->next_file;
869 struct fio_file *f;
870
871 do {
872 int opened = 0;
873
874 f = td->files[td->next_file];
875
876 td->next_file++;
877 if (td->next_file >= td->o.nr_files)
878 td->next_file = 0;
879
880 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
881 if (fio_file_done(f)) {
882 f = NULL;
883 continue;
884 }
885
886 if (!fio_file_open(f)) {
887 int err;
888
889 err = td_io_open_file(td, f);
890 if (err) {
891 dprint(FD_FILE, "error %d on open of %s\n",
892 err, f->file_name);
893 f = NULL;
894 continue;
895 }
896 opened = 1;
897 }
898
899 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
900 f->flags);
901 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
902 break;
903
904 if (opened)
905 td_io_close_file(td, f);
906
907 f = NULL;
908 } while (td->next_file != old_next_file);
909
910 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
911 return f;
912}
913
914static struct fio_file *__get_next_file(struct thread_data *td)
915{
916 struct fio_file *f;
917
918 assert(td->o.nr_files <= td->files_index);
919
920 if (td->nr_done_files >= td->o.nr_files) {
921 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
922 " nr_files=%d\n", td->nr_open_files,
923 td->nr_done_files,
924 td->o.nr_files);
925 return NULL;
926 }
927
928 f = td->file_service_file;
929 if (f && fio_file_open(f) && !fio_file_closing(f)) {
930 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
931 goto out;
932 if (td->file_service_left--)
933 goto out;
934 }
935
936 if (td->o.file_service_type == FIO_FSERVICE_RR ||
937 td->o.file_service_type == FIO_FSERVICE_SEQ)
938 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
939 else
940 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
941
942 td->file_service_file = f;
943 td->file_service_left = td->file_service_nr - 1;
944out:
945 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
946 return f;
947}
948
949static struct fio_file *get_next_file(struct thread_data *td)
950{
951 struct prof_io_ops *ops = &td->prof_io_ops;
952
953 if (ops->get_next_file)
954 return ops->get_next_file(td);
955
956 return __get_next_file(td);
957}
958
959static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
960{
961 struct fio_file *f;
962
963 do {
964 f = get_next_file(td);
965 if (!f)
966 return 1;
967
968 io_u->file = f;
969 get_file(f);
970
971 if (!fill_io_u(td, io_u))
972 break;
973
974 put_file_log(td, f);
975 td_io_close_file(td, f);
976 io_u->file = NULL;
977 fio_file_set_done(f);
978 td->nr_done_files++;
979 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
980 td->nr_done_files, td->o.nr_files);
981 } while (1);
982
983 return 0;
984}
985
986
987struct io_u *__get_io_u(struct thread_data *td)
988{
989 struct io_u *io_u = NULL;
990
991 td_io_u_lock(td);
992
993again:
994 if (!flist_empty(&td->io_u_requeues))
995 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
996 else if (!queue_full(td)) {
997 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
998
999 io_u->buflen = 0;
1000 io_u->resid = 0;
1001 io_u->file = NULL;
1002 io_u->end_io = NULL;
1003 }
1004
1005 if (io_u) {
1006 assert(io_u->flags & IO_U_F_FREE);
1007 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1008 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
1009 io_u->flags &= ~IO_U_F_VER_LIST;
1010
1011 io_u->error = 0;
1012 flist_del(&io_u->list);
1013 flist_add_tail(&io_u->list, &td->io_u_busylist);
1014 td->cur_depth++;
1015 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1016 } else if (td->o.verify_async) {
1017 /*
1018 * We ran out, wait for async verify threads to finish and
1019 * return one
1020 */
1021 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1022 goto again;
1023 }
1024
1025 td_io_u_unlock(td);
1026 return io_u;
1027}
1028
1029static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1030{
1031 if (td->o.trim_backlog && td->trim_entries) {
1032 int get_trim = 0;
1033
1034 if (td->trim_batch) {
1035 td->trim_batch--;
1036 get_trim = 1;
1037 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1038 td->last_ddir != DDIR_READ) {
1039 td->trim_batch = td->o.trim_batch;
1040 if (!td->trim_batch)
1041 td->trim_batch = td->o.trim_backlog;
1042 get_trim = 1;
1043 }
1044
1045 if (get_trim && !get_next_trim(td, io_u))
1046 return 1;
1047 }
1048
1049 return 0;
1050}
1051
1052static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1053{
1054 if (td->o.verify_backlog && td->io_hist_len) {
1055 int get_verify = 0;
1056
1057 if (td->verify_batch)
1058 get_verify = 1;
1059 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1060 td->last_ddir != DDIR_READ) {
1061 td->verify_batch = td->o.verify_batch;
1062 if (!td->verify_batch)
1063 td->verify_batch = td->o.verify_backlog;
1064 get_verify = 1;
1065 }
1066
1067 if (get_verify && !get_next_verify(td, io_u)) {
1068 td->verify_batch--;
1069 return 1;
1070 }
1071 }
1072
1073 return 0;
1074}
1075
1076/*
1077 * Fill offset and start time into the buffer content, to prevent too
1078 * easy compressible data for simple de-dupe attempts. Do this for every
1079 * 512b block in the range, since that should be the smallest block size
1080 * we can expect from a device.
1081 */
1082static void small_content_scramble(struct io_u *io_u)
1083{
1084 unsigned int i, nr_blocks = io_u->buflen / 512;
1085 unsigned long long boffset;
1086 unsigned int offset;
1087 void *p, *end;
1088
1089 if (!nr_blocks)
1090 return;
1091
1092 p = io_u->xfer_buf;
1093 boffset = io_u->offset;
1094 io_u->buf_filled_len = 0;
1095
1096 for (i = 0; i < nr_blocks; i++) {
1097 /*
1098 * Fill the byte offset into a "random" start offset of
1099 * the buffer, given by the product of the usec time
1100 * and the actual offset.
1101 */
1102 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1103 offset &= ~(sizeof(unsigned long long) - 1);
1104 if (offset >= 512 - sizeof(unsigned long long))
1105 offset -= sizeof(unsigned long long);
1106 memcpy(p + offset, &boffset, sizeof(boffset));
1107
1108 end = p + 512 - sizeof(io_u->start_time);
1109 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1110 p += 512;
1111 boffset += 512;
1112 }
1113}
1114
1115/*
1116 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1117 * etc. The returned io_u is fully ready to be prepped and submitted.
1118 */
1119struct io_u *get_io_u(struct thread_data *td)
1120{
1121 struct fio_file *f;
1122 struct io_u *io_u;
1123 int do_scramble = 0;
1124
1125 io_u = __get_io_u(td);
1126 if (!io_u) {
1127 dprint(FD_IO, "__get_io_u failed\n");
1128 return NULL;
1129 }
1130
1131 if (check_get_verify(td, io_u))
1132 goto out;
1133 if (check_get_trim(td, io_u))
1134 goto out;
1135
1136 /*
1137 * from a requeue, io_u already setup
1138 */
1139 if (io_u->file)
1140 goto out;
1141
1142 /*
1143 * If using an iolog, grab next piece if any available.
1144 */
1145 if (td->o.read_iolog_file) {
1146 if (read_iolog_get(td, io_u))
1147 goto err_put;
1148 } else if (set_io_u_file(td, io_u)) {
1149 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1150 goto err_put;
1151 }
1152
1153 f = io_u->file;
1154 assert(fio_file_open(f));
1155
1156 if (ddir_rw(io_u->ddir)) {
1157 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1158 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1159 goto err_put;
1160 }
1161
1162 f->last_start = io_u->offset;
1163 f->last_pos = io_u->offset + io_u->buflen;
1164
1165 if (io_u->ddir == DDIR_WRITE) {
1166 if (td->o.refill_buffers) {
1167 io_u_fill_buffer(td, io_u,
1168 io_u->xfer_buflen, io_u->xfer_buflen);
1169 } else if (td->o.scramble_buffers)
1170 do_scramble = 1;
1171 if (td->o.verify != VERIFY_NONE) {
1172 populate_verify_io_u(td, io_u);
1173 do_scramble = 0;
1174 }
1175 } else if (io_u->ddir == DDIR_READ) {
1176 /*
1177 * Reset the buf_filled parameters so next time if the
1178 * buffer is used for writes it is refilled.
1179 */
1180 io_u->buf_filled_len = 0;
1181 }
1182 }
1183
1184 /*
1185 * Set io data pointers.
1186 */
1187 io_u->xfer_buf = io_u->buf;
1188 io_u->xfer_buflen = io_u->buflen;
1189
1190out:
1191 assert(io_u->file);
1192 if (!td_io_prep(td, io_u)) {
1193 if (!td->o.disable_slat)
1194 fio_gettime(&io_u->start_time, NULL);
1195 if (do_scramble)
1196 small_content_scramble(io_u);
1197 return io_u;
1198 }
1199err_put:
1200 dprint(FD_IO, "get_io_u failed\n");
1201 put_io_u(td, io_u);
1202 return NULL;
1203}
1204
1205void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1206{
1207 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1208 const char *msg[] = { "read", "write", "sync", "datasync",
1209 "sync_file_range", "wait", "trim" };
1210
1211 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1212 return;
1213
1214 log_err("fio: io_u error");
1215
1216 if (io_u->file)
1217 log_err(" on file %s", io_u->file->file_name);
1218
1219 log_err(": %s\n", strerror(io_u->error));
1220
1221 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1222 io_u->offset, io_u->xfer_buflen);
1223
1224 if (!td->error)
1225 td_verror(td, io_u->error, "io_u error");
1226}
1227
1228static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1229 struct io_completion_data *icd,
1230 const enum fio_ddir idx, unsigned int bytes)
1231{
1232 unsigned long lusec = 0;
1233
1234 if (!td->o.disable_clat || !td->o.disable_bw)
1235 lusec = utime_since(&io_u->issue_time, &icd->time);
1236
1237 if (!td->o.disable_lat) {
1238 unsigned long tusec;
1239
1240 tusec = utime_since(&io_u->start_time, &icd->time);
1241 add_lat_sample(td, idx, tusec, bytes);
1242
1243 if (td->o.max_latency && tusec > td->o.max_latency) {
1244 if (!td->error)
1245 log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
1246 td_verror(td, ETIMEDOUT, "max latency exceeded");
1247 icd->error = ETIMEDOUT;
1248 }
1249 }
1250
1251 if (!td->o.disable_clat) {
1252 add_clat_sample(td, idx, lusec, bytes);
1253 io_u_mark_latency(td, lusec);
1254 }
1255
1256 if (!td->o.disable_bw)
1257 add_bw_sample(td, idx, bytes, &icd->time);
1258
1259 add_iops_sample(td, idx, &icd->time);
1260}
1261
1262static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1263{
1264 unsigned long long secs, remainder, bps, bytes;
1265 bytes = td->this_io_bytes[ddir];
1266 bps = td->rate_bps[ddir];
1267 secs = bytes / bps;
1268 remainder = bytes % bps;
1269 return remainder * 1000000 / bps + secs * 1000000;
1270}
1271
1272static void io_completed(struct thread_data *td, struct io_u *io_u,
1273 struct io_completion_data *icd)
1274{
1275 struct fio_file *f;
1276
1277 dprint_io_u(io_u, "io complete");
1278
1279 td_io_u_lock(td);
1280 assert(io_u->flags & IO_U_F_FLIGHT);
1281 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1282 td_io_u_unlock(td);
1283
1284 if (ddir_sync(io_u->ddir)) {
1285 td->last_was_sync = 1;
1286 f = io_u->file;
1287 if (f) {
1288 f->first_write = -1ULL;
1289 f->last_write = -1ULL;
1290 }
1291 return;
1292 }
1293
1294 td->last_was_sync = 0;
1295 td->last_ddir = io_u->ddir;
1296
1297 if (!io_u->error && ddir_rw(io_u->ddir)) {
1298 unsigned int bytes = io_u->buflen - io_u->resid;
1299 const enum fio_ddir idx = io_u->ddir;
1300 const enum fio_ddir odx = io_u->ddir ^ 1;
1301 int ret;
1302
1303 td->io_blocks[idx]++;
1304 td->this_io_blocks[idx]++;
1305 td->io_bytes[idx] += bytes;
1306
1307 if (!(io_u->flags & IO_U_F_VER_LIST))
1308 td->this_io_bytes[idx] += bytes;
1309
1310 if (idx == DDIR_WRITE) {
1311 f = io_u->file;
1312 if (f) {
1313 if (f->first_write == -1ULL ||
1314 io_u->offset < f->first_write)
1315 f->first_write = io_u->offset;
1316 if (f->last_write == -1ULL ||
1317 ((io_u->offset + bytes) > f->last_write))
1318 f->last_write = io_u->offset + bytes;
1319 }
1320 }
1321
1322 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1323 td->runstate == TD_VERIFYING)) {
1324 account_io_completion(td, io_u, icd, idx, bytes);
1325
1326 if (__should_check_rate(td, idx)) {
1327 td->rate_pending_usleep[idx] =
1328 (usec_for_io(td, idx) -
1329 utime_since_now(&td->start));
1330 }
1331 if (idx != DDIR_TRIM && __should_check_rate(td, odx))
1332 td->rate_pending_usleep[odx] =
1333 (usec_for_io(td, odx) -
1334 utime_since_now(&td->start));
1335 }
1336
1337 if (td_write(td) && idx == DDIR_WRITE &&
1338 td->o.do_verify &&
1339 td->o.verify != VERIFY_NONE)
1340 log_io_piece(td, io_u);
1341
1342 icd->bytes_done[idx] += bytes;
1343
1344 if (io_u->end_io) {
1345 ret = io_u->end_io(td, io_u);
1346 if (ret && !icd->error)
1347 icd->error = ret;
1348 }
1349 } else if (io_u->error) {
1350 icd->error = io_u->error;
1351 io_u_log_error(td, io_u);
1352 }
1353 if (icd->error) {
1354 enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1355 if (!td_non_fatal_error(td, eb, icd->error))
1356 return;
1357 /*
1358 * If there is a non_fatal error, then add to the error count
1359 * and clear all the errors.
1360 */
1361 update_error_count(td, icd->error);
1362 td_clear_error(td);
1363 icd->error = 0;
1364 io_u->error = 0;
1365 }
1366}
1367
1368static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1369 int nr)
1370{
1371 int ddir;
1372 if (!td->o.disable_clat || !td->o.disable_bw)
1373 fio_gettime(&icd->time, NULL);
1374
1375 icd->nr = nr;
1376
1377 icd->error = 0;
1378 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1379 icd->bytes_done[ddir] = 0;
1380}
1381
1382static void ios_completed(struct thread_data *td,
1383 struct io_completion_data *icd)
1384{
1385 struct io_u *io_u;
1386 int i;
1387
1388 for (i = 0; i < icd->nr; i++) {
1389 io_u = td->io_ops->event(td, i);
1390
1391 io_completed(td, io_u, icd);
1392
1393 if (!(io_u->flags & IO_U_F_FREE_DEF))
1394 put_io_u(td, io_u);
1395 }
1396}
1397
1398/*
1399 * Complete a single io_u for the sync engines.
1400 */
1401int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1402 unsigned long *bytes)
1403{
1404 struct io_completion_data icd;
1405
1406 init_icd(td, &icd, 1);
1407 io_completed(td, io_u, &icd);
1408
1409 if (!(io_u->flags & IO_U_F_FREE_DEF))
1410 put_io_u(td, io_u);
1411
1412 if (icd.error) {
1413 td_verror(td, icd.error, "io_u_sync_complete");
1414 return -1;
1415 }
1416
1417 if (bytes) {
1418 int ddir;
1419
1420 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1421 bytes[ddir] += icd.bytes_done[ddir];
1422 }
1423
1424 return 0;
1425}
1426
1427/*
1428 * Called to complete min_events number of io for the async engines.
1429 */
1430int io_u_queued_complete(struct thread_data *td, int min_evts,
1431 unsigned long *bytes)
1432{
1433 struct io_completion_data icd;
1434 struct timespec *tvp = NULL;
1435 int ret;
1436 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1437
1438 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1439
1440 if (!min_evts)
1441 tvp = &ts;
1442
1443 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1444 if (ret < 0) {
1445 td_verror(td, -ret, "td_io_getevents");
1446 return ret;
1447 } else if (!ret)
1448 return ret;
1449
1450 init_icd(td, &icd, ret);
1451 ios_completed(td, &icd);
1452 if (icd.error) {
1453 td_verror(td, icd.error, "io_u_queued_complete");
1454 return -1;
1455 }
1456
1457 if (bytes) {
1458 int ddir;
1459
1460 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1461 bytes[ddir] += icd.bytes_done[ddir];
1462 }
1463
1464 return 0;
1465}
1466
1467/*
1468 * Call when io_u is really queued, to update the submission latency.
1469 */
1470void io_u_queued(struct thread_data *td, struct io_u *io_u)
1471{
1472 if (!td->o.disable_slat) {
1473 unsigned long slat_time;
1474
1475 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1476 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1477 }
1478}
1479
1480/*
1481 * "randomly" fill the buffer contents
1482 */
1483void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1484 unsigned int min_write, unsigned int max_bs)
1485{
1486 io_u->buf_filled_len = 0;
1487
1488 if (!td->o.zero_buffers) {
1489 unsigned int perc = td->o.compress_percentage;
1490
1491 if (perc) {
1492 unsigned int seg = min_write;
1493
1494 seg = min(min_write, td->o.compress_chunk);
1495 fill_random_buf_percentage(&td->buf_state, io_u->buf,
1496 perc, seg, max_bs);
1497 } else
1498 fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1499 } else
1500 memset(io_u->buf, 0, max_bs);
1501}
fio - Flexible IO Tester