projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Kill flist_sort from Makefiles
[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] & (1 << 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 this_blocks, mask;
54 unsigned int idx, bit;
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 = -1U;
79 else
80 mask = ((1U << this_blocks) - 1) << bit;
81
82 if (!(f->file_map[idx] & mask))
83 break;
84
85 this_blocks--;
86 } while (this_blocks);
87
88 if (!this_blocks)
89 break;
90
91 f->file_map[idx] |= mask;
92 nr_blocks -= this_blocks;
93 blocks += this_blocks;
94 block += this_blocks;
95 }
96
97 if ((blocks * min_bs) < io_u->buflen)
98 io_u->buflen = blocks * min_bs;
99}
100
101static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
102 enum fio_ddir ddir)
103{
104 unsigned long long max_blocks;
105 unsigned long long max_size;
106
107 assert(ddir_rw(ddir));
108
109 /*
110 * Hmm, should we make sure that ->io_size <= ->real_file_size?
111 */
112 max_size = f->io_size;
113 if (max_size > f->real_file_size)
114 max_size = f->real_file_size;
115
116 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
117 if (!max_blocks)
118 return 0;
119
120 return max_blocks;
121}
122
123/*
124 * Return the next free block in the map.
125 */
126static int get_next_free_block(struct thread_data *td, struct fio_file *f,
127 enum fio_ddir ddir, unsigned long long *b)
128{
129 unsigned long long min_bs = td->o.rw_min_bs, lastb;
130 int i;
131
132 lastb = last_block(td, f, ddir);
133 if (!lastb)
134 return 1;
135
136 i = f->last_free_lookup;
137 *b = (i * BLOCKS_PER_MAP);
138 while ((*b) * min_bs < f->real_file_size &&
139 (*b) * min_bs < f->io_size) {
140 if (f->file_map[i] != (unsigned int) -1) {
141 *b += ffz(f->file_map[i]);
142 if (*b > lastb)
143 break;
144 f->last_free_lookup = i;
145 return 0;
146 }
147
148 *b += BLOCKS_PER_MAP;
149 i++;
150 }
151
152 dprint(FD_IO, "failed finding a free block\n");
153 return 1;
154}
155
156static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
157 enum fio_ddir ddir, unsigned long long *b)
158{
159 unsigned long long r, lastb;
160 int loops = 5;
161
162 lastb = last_block(td, f, ddir);
163 if (!lastb)
164 return 1;
165
166 do {
167 r = os_random_long(&td->random_state);
168 dprint(FD_RANDOM, "off rand %llu\n", r);
169 *b = (lastb - 1) * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
170
171 /*
172 * if we are not maintaining a random map, we are done.
173 */
174 if (!file_randommap(td, f))
175 return 0;
176
177 /*
178 * calculate map offset and check if it's free
179 */
180 if (random_map_free(f, *b))
181 return 0;
182
183 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
184 *b);
185 } while (--loops);
186
187 /*
188 * we get here, if we didn't suceed in looking up a block. generate
189 * a random start offset into the filemap, and find the first free
190 * block from there.
191 */
192 loops = 10;
193 do {
194 f->last_free_lookup = (f->num_maps - 1) *
195 (r / (OS_RAND_MAX + 1.0));
196 if (!get_next_free_block(td, f, ddir, b))
197 return 0;
198
199 r = os_random_long(&td->random_state);
200 } while (--loops);
201
202 /*
203 * that didn't work either, try exhaustive search from the start
204 */
205 f->last_free_lookup = 0;
206 return get_next_free_block(td, f, ddir, b);
207}
208
209static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
210 enum fio_ddir ddir, unsigned long long *b)
211{
212 if (get_next_rand_offset(td, f, ddir, b)) {
213 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
214 f->file_name, f->last_pos, f->real_file_size);
215 return 1;
216 }
217
218 return 0;
219}
220
221static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
222 enum fio_ddir ddir, unsigned long long *b)
223{
224 assert(ddir_rw(ddir));
225
226 if (f->last_pos < f->real_file_size) {
227 *b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
228 return 0;
229 }
230
231 return 1;
232}
233
234static int get_next_block(struct thread_data *td, struct io_u *io_u,
235 enum fio_ddir ddir, int rw_seq, unsigned long long *b)
236{
237 struct fio_file *f = io_u->file;
238 int ret;
239
240 assert(ddir_rw(ddir));
241
242 if (rw_seq) {
243 if (td_random(td))
244 ret = get_next_rand_block(td, f, ddir, b);
245 else
246 ret = get_next_seq_block(td, f, ddir, b);
247 } else {
248 io_u->flags |= IO_U_F_BUSY_OK;
249
250 if (td->o.rw_seq == RW_SEQ_SEQ) {
251 ret = get_next_seq_block(td, f, ddir, b);
252 if (ret)
253 ret = get_next_rand_block(td, f, ddir, b);
254 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
255 if (f->last_start != -1ULL)
256 *b = (f->last_start - f->file_offset)
257 / td->o.min_bs[ddir];
258 else
259 *b = 0;
260 ret = 0;
261 } else {
262 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
263 ret = 1;
264 }
265 }
266
267 return ret;
268}
269
270/*
271 * For random io, generate a random new block and see if it's used. Repeat
272 * until we find a free one. For sequential io, just return the end of
273 * the last io issued.
274 */
275static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
276{
277 struct fio_file *f = io_u->file;
278 unsigned long long b;
279 enum fio_ddir ddir = io_u->ddir;
280 int rw_seq_hit = 0;
281
282 assert(ddir_rw(ddir));
283
284 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
285 rw_seq_hit = 1;
286 td->ddir_seq_nr = td->o.ddir_seq_nr;
287 }
288
289 if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
290 return 1;
291
292 io_u->offset = b * td->o.ba[ddir];
293 if (io_u->offset >= f->io_size) {
294 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
295 io_u->offset, f->io_size);
296 return 1;
297 }
298
299 io_u->offset += f->file_offset;
300 if (io_u->offset >= f->real_file_size) {
301 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
302 io_u->offset, f->real_file_size);
303 return 1;
304 }
305
306 return 0;
307}
308
309static int get_next_offset(struct thread_data *td, struct io_u *io_u)
310{
311 struct prof_io_ops *ops = &td->prof_io_ops;
312
313 if (ops->fill_io_u_off)
314 return ops->fill_io_u_off(td, io_u);
315
316 return __get_next_offset(td, io_u);
317}
318
319static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
320{
321 const int ddir = io_u->ddir;
322 unsigned int uninitialized_var(buflen);
323 unsigned int minbs, maxbs;
324 long r;
325
326 assert(ddir_rw(ddir));
327
328 minbs = td->o.min_bs[ddir];
329 maxbs = td->o.max_bs[ddir];
330
331 if (minbs == maxbs)
332 buflen = minbs;
333 else {
334 r = os_random_long(&td->bsrange_state);
335 if (!td->o.bssplit_nr[ddir]) {
336 buflen = 1 + (unsigned int) ((double) maxbs *
337 (r / (OS_RAND_MAX + 1.0)));
338 if (buflen < minbs)
339 buflen = minbs;
340 } else {
341 long perc = 0;
342 unsigned int i;
343
344 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
345 struct bssplit *bsp = &td->o.bssplit[ddir][i];
346
347 buflen = bsp->bs;
348 perc += bsp->perc;
349 if (r <= ((OS_RAND_MAX / 100L) * perc))
350 break;
351 }
352 }
353 if (!td->o.bs_unaligned && is_power_of_2(minbs))
354 buflen = (buflen + minbs - 1) & ~(minbs - 1);
355 }
356
357 if (io_u->offset + buflen > io_u->file->real_file_size) {
358 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
359 minbs, ddir);
360 buflen = minbs;
361 }
362
363 return buflen;
364}
365
366static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
367{
368 struct prof_io_ops *ops = &td->prof_io_ops;
369
370 if (ops->fill_io_u_size)
371 return ops->fill_io_u_size(td, io_u);
372
373 return __get_next_buflen(td, io_u);
374}
375
376static void set_rwmix_bytes(struct thread_data *td)
377{
378 unsigned int diff;
379
380 /*
381 * we do time or byte based switch. this is needed because
382 * buffered writes may issue a lot quicker than they complete,
383 * whereas reads do not.
384 */
385 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
386 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
387}
388
389static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
390{
391 unsigned int v;
392 long r;
393
394 r = os_random_long(&td->rwmix_state);
395 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
396 if (v <= td->o.rwmix[DDIR_READ])
397 return DDIR_READ;
398
399 return DDIR_WRITE;
400}
401
402static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
403{
404 enum fio_ddir odir = ddir ^ 1;
405 struct timeval t;
406 long usec;
407
408 assert(ddir_rw(ddir));
409
410 if (td->rate_pending_usleep[ddir] <= 0)
411 return ddir;
412
413 /*
414 * We have too much pending sleep in this direction. See if we
415 * should switch.
416 */
417 if (td_rw(td)) {
418 /*
419 * Other direction does not have too much pending, switch
420 */
421 if (td->rate_pending_usleep[odir] < 100000)
422 return odir;
423
424 /*
425 * Both directions have pending sleep. Sleep the minimum time
426 * and deduct from both.
427 */
428 if (td->rate_pending_usleep[ddir] <=
429 td->rate_pending_usleep[odir]) {
430 usec = td->rate_pending_usleep[ddir];
431 } else {
432 usec = td->rate_pending_usleep[odir];
433 ddir = odir;
434 }
435 } else
436 usec = td->rate_pending_usleep[ddir];
437
438 fio_gettime(&t, NULL);
439 usec_sleep(td, usec);
440 usec = utime_since_now(&t);
441
442 td->rate_pending_usleep[ddir] -= usec;
443
444 odir = ddir ^ 1;
445 if (td_rw(td) && __should_check_rate(td, odir))
446 td->rate_pending_usleep[odir] -= usec;
447
448 return ddir;
449}
450
451/*
452 * Return the data direction for the next io_u. If the job is a
453 * mixed read/write workload, check the rwmix cycle and switch if
454 * necessary.
455 */
456static enum fio_ddir get_rw_ddir(struct thread_data *td)
457{
458 enum fio_ddir ddir;
459
460 /*
461 * see if it's time to fsync
462 */
463 if (td->o.fsync_blocks &&
464 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
465 td->io_issues[DDIR_WRITE] && should_fsync(td))
466 return DDIR_SYNC;
467
468 /*
469 * see if it's time to fdatasync
470 */
471 if (td->o.fdatasync_blocks &&
472 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
473 td->io_issues[DDIR_WRITE] && should_fsync(td))
474 return DDIR_DATASYNC;
475
476 /*
477 * see if it's time to sync_file_range
478 */
479 if (td->sync_file_range_nr &&
480 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
481 td->io_issues[DDIR_WRITE] && should_fsync(td))
482 return DDIR_SYNC_FILE_RANGE;
483
484 if (td_rw(td)) {
485 /*
486 * Check if it's time to seed a new data direction.
487 */
488 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
489 /*
490 * Put a top limit on how many bytes we do for
491 * one data direction, to avoid overflowing the
492 * ranges too much
493 */
494 ddir = get_rand_ddir(td);
495
496 if (ddir != td->rwmix_ddir)
497 set_rwmix_bytes(td);
498
499 td->rwmix_ddir = ddir;
500 }
501 ddir = td->rwmix_ddir;
502 } else if (td_read(td))
503 ddir = DDIR_READ;
504 else
505 ddir = DDIR_WRITE;
506
507 td->rwmix_ddir = rate_ddir(td, ddir);
508 return td->rwmix_ddir;
509}
510
511static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
512{
513 io_u->ddir = get_rw_ddir(td);
514
515 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
516 td->o.barrier_blocks &&
517 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
518 td->io_issues[DDIR_WRITE])
519 io_u->flags |= IO_U_F_BARRIER;
520}
521
522void put_file_log(struct thread_data *td, struct fio_file *f)
523{
524 int ret = put_file(td, f);
525
526 if (ret)
527 td_verror(td, ret, "file close");
528}
529
530void put_io_u(struct thread_data *td, struct io_u *io_u)
531{
532 td_io_u_lock(td);
533
534 io_u->flags |= IO_U_F_FREE;
535 io_u->flags &= ~IO_U_F_FREE_DEF;
536
537 if (io_u->file)
538 put_file_log(td, io_u->file);
539
540 io_u->file = NULL;
541 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
542 td->cur_depth--;
543 flist_del_init(&io_u->list);
544 flist_add(&io_u->list, &td->io_u_freelist);
545 td_io_u_unlock(td);
546 td_io_u_free_notify(td);
547}
548
549void clear_io_u(struct thread_data *td, struct io_u *io_u)
550{
551 io_u->flags &= ~IO_U_F_FLIGHT;
552 put_io_u(td, io_u);
553}
554
555void requeue_io_u(struct thread_data *td, struct io_u **io_u)
556{
557 struct io_u *__io_u = *io_u;
558
559 dprint(FD_IO, "requeue %p\n", __io_u);
560
561 td_io_u_lock(td);
562
563 __io_u->flags |= IO_U_F_FREE;
564 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
565 td->io_issues[__io_u->ddir]--;
566
567 __io_u->flags &= ~IO_U_F_FLIGHT;
568 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
569 td->cur_depth--;
570 flist_del(&__io_u->list);
571 flist_add_tail(&__io_u->list, &td->io_u_requeues);
572 td_io_u_unlock(td);
573 *io_u = NULL;
574}
575
576static int fill_io_u(struct thread_data *td, struct io_u *io_u)
577{
578 if (td->io_ops->flags & FIO_NOIO)
579 goto out;
580
581 set_rw_ddir(td, io_u);
582
583 /*
584 * fsync() or fdatasync() or trim etc, we are done
585 */
586 if (!ddir_rw(io_u->ddir))
587 goto out;
588
589 /*
590 * See if it's time to switch to a new zone
591 */
592 if (td->zone_bytes >= td->o.zone_size) {
593 td->zone_bytes = 0;
594 io_u->file->last_pos += td->o.zone_skip;
595 td->io_skip_bytes += td->o.zone_skip;
596 }
597
598 /*
599 * No log, let the seq/rand engine retrieve the next buflen and
600 * position.
601 */
602 if (get_next_offset(td, io_u)) {
603 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
604 return 1;
605 }
606
607 io_u->buflen = get_next_buflen(td, io_u);
608 if (!io_u->buflen) {
609 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
610 return 1;
611 }
612
613 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
614 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
615 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
616 io_u->buflen, io_u->file->real_file_size);
617 return 1;
618 }
619
620 /*
621 * mark entry before potentially trimming io_u
622 */
623 if (td_random(td) && file_randommap(td, io_u->file))
624 mark_random_map(td, io_u);
625
626 /*
627 * If using a write iolog, store this entry.
628 */
629out:
630 dprint_io_u(io_u, "fill_io_u");
631 td->zone_bytes += io_u->buflen;
632 log_io_u(td, io_u);
633 return 0;
634}
635
636static void __io_u_mark_map(unsigned int *map, unsigned int nr)
637{
638 int idx = 0;
639
640 switch (nr) {
641 default:
642 idx = 6;
643 break;
644 case 33 ... 64:
645 idx = 5;
646 break;
647 case 17 ... 32:
648 idx = 4;
649 break;
650 case 9 ... 16:
651 idx = 3;
652 break;
653 case 5 ... 8:
654 idx = 2;
655 break;
656 case 1 ... 4:
657 idx = 1;
658 case 0:
659 break;
660 }
661
662 map[idx]++;
663}
664
665void io_u_mark_submit(struct thread_data *td, unsigned int nr)
666{
667 __io_u_mark_map(td->ts.io_u_submit, nr);
668 td->ts.total_submit++;
669}
670
671void io_u_mark_complete(struct thread_data *td, unsigned int nr)
672{
673 __io_u_mark_map(td->ts.io_u_complete, nr);
674 td->ts.total_complete++;
675}
676
677void io_u_mark_depth(struct thread_data *td, unsigned int nr)
678{
679 int idx = 0;
680
681 switch (td->cur_depth) {
682 default:
683 idx = 6;
684 break;
685 case 32 ... 63:
686 idx = 5;
687 break;
688 case 16 ... 31:
689 idx = 4;
690 break;
691 case 8 ... 15:
692 idx = 3;
693 break;
694 case 4 ... 7:
695 idx = 2;
696 break;
697 case 2 ... 3:
698 idx = 1;
699 case 1:
700 break;
701 }
702
703 td->ts.io_u_map[idx] += nr;
704}
705
706static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
707{
708 int idx = 0;
709
710 assert(usec < 1000);
711
712 switch (usec) {
713 case 750 ... 999:
714 idx = 9;
715 break;
716 case 500 ... 749:
717 idx = 8;
718 break;
719 case 250 ... 499:
720 idx = 7;
721 break;
722 case 100 ... 249:
723 idx = 6;
724 break;
725 case 50 ... 99:
726 idx = 5;
727 break;
728 case 20 ... 49:
729 idx = 4;
730 break;
731 case 10 ... 19:
732 idx = 3;
733 break;
734 case 4 ... 9:
735 idx = 2;
736 break;
737 case 2 ... 3:
738 idx = 1;
739 case 0 ... 1:
740 break;
741 }
742
743 assert(idx < FIO_IO_U_LAT_U_NR);
744 td->ts.io_u_lat_u[idx]++;
745}
746
747static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
748{
749 int idx = 0;
750
751 switch (msec) {
752 default:
753 idx = 11;
754 break;
755 case 1000 ... 1999:
756 idx = 10;
757 break;
758 case 750 ... 999:
759 idx = 9;
760 break;
761 case 500 ... 749:
762 idx = 8;
763 break;
764 case 250 ... 499:
765 idx = 7;
766 break;
767 case 100 ... 249:
768 idx = 6;
769 break;
770 case 50 ... 99:
771 idx = 5;
772 break;
773 case 20 ... 49:
774 idx = 4;
775 break;
776 case 10 ... 19:
777 idx = 3;
778 break;
779 case 4 ... 9:
780 idx = 2;
781 break;
782 case 2 ... 3:
783 idx = 1;
784 case 0 ... 1:
785 break;
786 }
787
788 assert(idx < FIO_IO_U_LAT_M_NR);
789 td->ts.io_u_lat_m[idx]++;
790}
791
792static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
793{
794 if (usec < 1000)
795 io_u_mark_lat_usec(td, usec);
796 else
797 io_u_mark_lat_msec(td, usec / 1000);
798}
799
800/*
801 * Get next file to service by choosing one at random
802 */
803static struct fio_file *get_next_file_rand(struct thread_data *td,
804 enum fio_file_flags goodf,
805 enum fio_file_flags badf)
806{
807 struct fio_file *f;
808 int fno;
809
810 do {
811 long r = os_random_long(&td->next_file_state);
812 int opened = 0;
813
814 fno = (unsigned int) ((double) td->o.nr_files
815 * (r / (OS_RAND_MAX + 1.0)));
816 f = td->files[fno];
817 if (fio_file_done(f))
818 continue;
819
820 if (!fio_file_open(f)) {
821 int err;
822
823 err = td_io_open_file(td, f);
824 if (err)
825 continue;
826 opened = 1;
827 }
828
829 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
830 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
831 return f;
832 }
833 if (opened)
834 td_io_close_file(td, f);
835 } while (1);
836}
837
838/*
839 * Get next file to service by doing round robin between all available ones
840 */
841static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
842 int badf)
843{
844 unsigned int old_next_file = td->next_file;
845 struct fio_file *f;
846
847 do {
848 int opened = 0;
849
850 f = td->files[td->next_file];
851
852 td->next_file++;
853 if (td->next_file >= td->o.nr_files)
854 td->next_file = 0;
855
856 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
857 if (fio_file_done(f)) {
858 f = NULL;
859 continue;
860 }
861
862 if (!fio_file_open(f)) {
863 int err;
864
865 err = td_io_open_file(td, f);
866 if (err) {
867 dprint(FD_FILE, "error %d on open of %s\n",
868 err, f->file_name);
869 f = NULL;
870 continue;
871 }
872 opened = 1;
873 }
874
875 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
876 f->flags);
877 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
878 break;
879
880 if (opened)
881 td_io_close_file(td, f);
882
883 f = NULL;
884 } while (td->next_file != old_next_file);
885
886 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
887 return f;
888}
889
890static struct fio_file *__get_next_file(struct thread_data *td)
891{
892 struct fio_file *f;
893
894 assert(td->o.nr_files <= td->files_index);
895
896 if (td->nr_done_files >= td->o.nr_files) {
897 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
898 " nr_files=%d\n", td->nr_open_files,
899 td->nr_done_files,
900 td->o.nr_files);
901 return NULL;
902 }
903
904 f = td->file_service_file;
905 if (f && fio_file_open(f) && !fio_file_closing(f)) {
906 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
907 goto out;
908 if (td->file_service_left--)
909 goto out;
910 }
911
912 if (td->o.file_service_type == FIO_FSERVICE_RR ||
913 td->o.file_service_type == FIO_FSERVICE_SEQ)
914 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
915 else
916 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
917
918 td->file_service_file = f;
919 td->file_service_left = td->file_service_nr - 1;
920out:
921 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
922 return f;
923}
924
925static struct fio_file *get_next_file(struct thread_data *td)
926{
927 struct prof_io_ops *ops = &td->prof_io_ops;
928
929 if (ops->get_next_file)
930 return ops->get_next_file(td);
931
932 return __get_next_file(td);
933}
934
935static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
936{
937 struct fio_file *f;
938
939 do {
940 f = get_next_file(td);
941 if (!f)
942 return 1;
943
944 io_u->file = f;
945 get_file(f);
946
947 if (!fill_io_u(td, io_u))
948 break;
949
950 put_file_log(td, f);
951 td_io_close_file(td, f);
952 io_u->file = NULL;
953 fio_file_set_done(f);
954 td->nr_done_files++;
955 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
956 td->nr_done_files, td->o.nr_files);
957 } while (1);
958
959 return 0;
960}
961
962
963struct io_u *__get_io_u(struct thread_data *td)
964{
965 struct io_u *io_u = NULL;
966
967 td_io_u_lock(td);
968
969again:
970 if (!flist_empty(&td->io_u_requeues))
971 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
972 else if (!queue_full(td)) {
973 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
974
975 io_u->buflen = 0;
976 io_u->resid = 0;
977 io_u->file = NULL;
978 io_u->end_io = NULL;
979 }
980
981 if (io_u) {
982 assert(io_u->flags & IO_U_F_FREE);
983 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
984 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
985
986 io_u->error = 0;
987 flist_del(&io_u->list);
988 flist_add(&io_u->list, &td->io_u_busylist);
989 td->cur_depth++;
990 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
991 } else if (td->o.verify_async) {
992 /*
993 * We ran out, wait for async verify threads to finish and
994 * return one
995 */
996 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
997 goto again;
998 }
999
1000 td_io_u_unlock(td);
1001 return io_u;
1002}
1003
1004static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1005{
1006 if (td->o.trim_backlog && td->trim_entries) {
1007 int get_trim = 0;
1008
1009 if (td->trim_batch) {
1010 td->trim_batch--;
1011 get_trim = 1;
1012 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1013 td->last_ddir != DDIR_READ) {
1014 td->trim_batch = td->o.trim_batch;
1015 if (!td->trim_batch)
1016 td->trim_batch = td->o.trim_backlog;
1017 get_trim = 1;
1018 }
1019
1020 if (get_trim && !get_next_trim(td, io_u))
1021 return 1;
1022 }
1023
1024 return 0;
1025}
1026
1027static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1028{
1029 if (td->o.verify_backlog && td->io_hist_len) {
1030 int get_verify = 0;
1031
1032 if (td->verify_batch) {
1033 td->verify_batch--;
1034 get_verify = 1;
1035 } else if (!(td->io_hist_len % td->o.verify_backlog) &&
1036 td->last_ddir != DDIR_READ) {
1037 td->verify_batch = td->o.verify_batch;
1038 if (!td->verify_batch)
1039 td->verify_batch = td->o.verify_backlog;
1040 get_verify = 1;
1041 }
1042
1043 if (get_verify && !get_next_verify(td, io_u))
1044 return 1;
1045 }
1046
1047 return 0;
1048}
1049
1050/*
1051 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1052 * etc. The returned io_u is fully ready to be prepped and submitted.
1053 */
1054struct io_u *get_io_u(struct thread_data *td)
1055{
1056 struct fio_file *f;
1057 struct io_u *io_u;
1058
1059 io_u = __get_io_u(td);
1060 if (!io_u) {
1061 dprint(FD_IO, "__get_io_u failed\n");
1062 return NULL;
1063 }
1064
1065 if (check_get_verify(td, io_u))
1066 goto out;
1067 if (check_get_trim(td, io_u))
1068 goto out;
1069
1070 /*
1071 * from a requeue, io_u already setup
1072 */
1073 if (io_u->file)
1074 goto out;
1075
1076 /*
1077 * If using an iolog, grab next piece if any available.
1078 */
1079 if (td->o.read_iolog_file) {
1080 if (read_iolog_get(td, io_u))
1081 goto err_put;
1082 } else if (set_io_u_file(td, io_u)) {
1083 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1084 goto err_put;
1085 }
1086
1087 f = io_u->file;
1088 assert(fio_file_open(f));
1089
1090 if (ddir_rw(io_u->ddir)) {
1091 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1092 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1093 goto err_put;
1094 }
1095
1096 f->last_start = io_u->offset;
1097 f->last_pos = io_u->offset + io_u->buflen;
1098
1099 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
1100 populate_verify_io_u(td, io_u);
1101 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
1102 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
1103 else if (io_u->ddir == DDIR_READ) {
1104 /*
1105 * Reset the buf_filled parameters so next time if the
1106 * buffer is used for writes it is refilled.
1107 */
1108 io_u->buf_filled_len = 0;
1109 }
1110 }
1111
1112 /*
1113 * Set io data pointers.
1114 */
1115 io_u->xfer_buf = io_u->buf;
1116 io_u->xfer_buflen = io_u->buflen;
1117
1118out:
1119 assert(io_u->file);
1120 if (!td_io_prep(td, io_u)) {
1121 if (!td->o.disable_slat)
1122 fio_gettime(&io_u->start_time, NULL);
1123 return io_u;
1124 }
1125err_put:
1126 dprint(FD_IO, "get_io_u failed\n");
1127 put_io_u(td, io_u);
1128 return NULL;
1129}
1130
1131void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1132{
1133 const char *msg[] = { "read", "write", "sync", "datasync",
1134 "sync_file_range", "wait", "trim" };
1135
1136
1137
1138 log_err("fio: io_u error");
1139
1140 if (io_u->file)
1141 log_err(" on file %s", io_u->file->file_name);
1142
1143 log_err(": %s\n", strerror(io_u->error));
1144
1145 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1146 io_u->offset, io_u->xfer_buflen);
1147
1148 if (!td->error)
1149 td_verror(td, io_u->error, "io_u error");
1150}
1151
1152static void io_completed(struct thread_data *td, struct io_u *io_u,
1153 struct io_completion_data *icd)
1154{
1155 /*
1156 * Older gcc's are too dumb to realize that usec is always used
1157 * initialized, silence that warning.
1158 */
1159 unsigned long uninitialized_var(usec);
1160 struct fio_file *f;
1161
1162 dprint_io_u(io_u, "io complete");
1163
1164 td_io_u_lock(td);
1165 assert(io_u->flags & IO_U_F_FLIGHT);
1166 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1167 td_io_u_unlock(td);
1168
1169 if (ddir_sync(io_u->ddir)) {
1170 td->last_was_sync = 1;
1171 f = io_u->file;
1172 if (f) {
1173 f->first_write = -1ULL;
1174 f->last_write = -1ULL;
1175 }
1176 return;
1177 }
1178
1179 td->last_was_sync = 0;
1180 td->last_ddir = io_u->ddir;
1181
1182 if (!io_u->error && ddir_rw(io_u->ddir)) {
1183 unsigned int bytes = io_u->buflen - io_u->resid;
1184 const enum fio_ddir idx = io_u->ddir;
1185 const enum fio_ddir odx = io_u->ddir ^ 1;
1186 int ret;
1187
1188 td->io_blocks[idx]++;
1189 td->io_bytes[idx] += bytes;
1190 td->this_io_bytes[idx] += bytes;
1191
1192 if (idx == DDIR_WRITE) {
1193 f = io_u->file;
1194 if (f) {
1195 if (f->first_write == -1ULL ||
1196 io_u->offset < f->first_write)
1197 f->first_write = io_u->offset;
1198 if (f->last_write == -1ULL ||
1199 ((io_u->offset + bytes) > f->last_write))
1200 f->last_write = io_u->offset + bytes;
1201 }
1202 }
1203
1204 if (ramp_time_over(td)) {
1205 unsigned long uninitialized_var(lusec);
1206
1207 if (!td->o.disable_clat || !td->o.disable_bw)
1208 lusec = utime_since(&io_u->issue_time,
1209 &icd->time);
1210 if (!td->o.disable_lat) {
1211 unsigned long tusec;
1212
1213 tusec = utime_since(&io_u->start_time,
1214 &icd->time);
1215 add_lat_sample(td, idx, tusec, bytes);
1216 }
1217 if (!td->o.disable_clat) {
1218 add_clat_sample(td, idx, lusec, bytes);
1219 io_u_mark_latency(td, lusec);
1220 }
1221 if (!td->o.disable_bw)
1222 add_bw_sample(td, idx, bytes, &icd->time);
1223 if (__should_check_rate(td, idx)) {
1224 td->rate_pending_usleep[idx] =
1225 ((td->this_io_bytes[idx] *
1226 td->rate_nsec_cycle[idx]) / 1000 -
1227 utime_since_now(&td->start));
1228 }
1229 if (__should_check_rate(td, idx ^ 1))
1230 td->rate_pending_usleep[odx] =
1231 ((td->this_io_bytes[odx] *
1232 td->rate_nsec_cycle[odx]) / 1000 -
1233 utime_since_now(&td->start));
1234 }
1235
1236 if (td_write(td) && idx == DDIR_WRITE &&
1237 td->o.do_verify &&
1238 td->o.verify != VERIFY_NONE)
1239 log_io_piece(td, io_u);
1240
1241 icd->bytes_done[idx] += bytes;
1242
1243 if (io_u->end_io) {
1244 ret = io_u->end_io(td, io_u);
1245 if (ret && !icd->error)
1246 icd->error = ret;
1247 }
1248 } else if (io_u->error) {
1249 icd->error = io_u->error;
1250 io_u_log_error(td, io_u);
1251 }
1252 if (td->o.continue_on_error && icd->error &&
1253 td_non_fatal_error(icd->error)) {
1254 /*
1255 * If there is a non_fatal error, then add to the error count
1256 * and clear all the errors.
1257 */
1258 update_error_count(td, icd->error);
1259 td_clear_error(td);
1260 icd->error = 0;
1261 io_u->error = 0;
1262 }
1263}
1264
1265static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1266 int nr)
1267{
1268 if (!td->o.disable_clat || !td->o.disable_bw)
1269 fio_gettime(&icd->time, NULL);
1270
1271 icd->nr = nr;
1272
1273 icd->error = 0;
1274 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1275}
1276
1277static void ios_completed(struct thread_data *td,
1278 struct io_completion_data *icd)
1279{
1280 struct io_u *io_u;
1281 int i;
1282
1283 for (i = 0; i < icd->nr; i++) {
1284 io_u = td->io_ops->event(td, i);
1285
1286 io_completed(td, io_u, icd);
1287
1288 if (!(io_u->flags & IO_U_F_FREE_DEF))
1289 put_io_u(td, io_u);
1290 }
1291}
1292
1293/*
1294 * Complete a single io_u for the sync engines.
1295 */
1296int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1297 unsigned long *bytes)
1298{
1299 struct io_completion_data icd;
1300
1301 init_icd(td, &icd, 1);
1302 io_completed(td, io_u, &icd);
1303
1304 if (!(io_u->flags & IO_U_F_FREE_DEF))
1305 put_io_u(td, io_u);
1306
1307 if (icd.error) {
1308 td_verror(td, icd.error, "io_u_sync_complete");
1309 return -1;
1310 }
1311
1312 if (bytes) {
1313 bytes[0] += icd.bytes_done[0];
1314 bytes[1] += icd.bytes_done[1];
1315 }
1316
1317 return 0;
1318}
1319
1320/*
1321 * Called to complete min_events number of io for the async engines.
1322 */
1323int io_u_queued_complete(struct thread_data *td, int min_evts,
1324 unsigned long *bytes)
1325{
1326 struct io_completion_data icd;
1327 struct timespec *tvp = NULL;
1328 int ret;
1329 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1330
1331 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1332
1333 if (!min_evts)
1334 tvp = &ts;
1335
1336 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1337 if (ret < 0) {
1338 td_verror(td, -ret, "td_io_getevents");
1339 return ret;
1340 } else if (!ret)
1341 return ret;
1342
1343 init_icd(td, &icd, ret);
1344 ios_completed(td, &icd);
1345 if (icd.error) {
1346 td_verror(td, icd.error, "io_u_queued_complete");
1347 return -1;
1348 }
1349
1350 if (bytes) {
1351 bytes[0] += icd.bytes_done[0];
1352 bytes[1] += icd.bytes_done[1];
1353 }
1354
1355 return 0;
1356}
1357
1358/*
1359 * Call when io_u is really queued, to update the submission latency.
1360 */
1361void io_u_queued(struct thread_data *td, struct io_u *io_u)
1362{
1363 if (!td->o.disable_slat) {
1364 unsigned long slat_time;
1365
1366 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1367 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1368 }
1369}
1370
1371/*
1372 * "randomly" fill the buffer contents
1373 */
1374void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1375 unsigned int max_bs)
1376{
1377 io_u->buf_filled_len = 0;
1378
1379 if (!td->o.zero_buffers)
1380 fill_random_buf(io_u->buf, max_bs);
1381 else
1382 memset(io_u->buf, 0, max_bs);
1383}
fio - Flexible IO Tester