projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Initial suppor for sync_file_range()
[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
12struct io_completion_data {
13 int nr; /* input */
14
15 int error; /* output */
16 unsigned long bytes_done[2]; /* output */
17 struct timeval time; /* output */
18};
19
20/*
21 * The ->file_map[] contains a map of blocks we have or have not done io
22 * to yet. Used to make sure we cover the entire range in a fair fashion.
23 */
24static int random_map_free(struct fio_file *f, const unsigned long long block)
25{
26 unsigned int idx = RAND_MAP_IDX(f, block);
27 unsigned int bit = RAND_MAP_BIT(f, block);
28
29 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
30
31 return (f->file_map[idx] & (1 << bit)) == 0;
32}
33
34/*
35 * Mark a given offset as used in the map.
36 */
37static void mark_random_map(struct thread_data *td, struct io_u *io_u)
38{
39 unsigned int min_bs = td->o.rw_min_bs;
40 struct fio_file *f = io_u->file;
41 unsigned long long block;
42 unsigned int blocks, nr_blocks;
43
44 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
45 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
46 blocks = 0;
47
48 while (nr_blocks) {
49 unsigned int this_blocks, mask;
50 unsigned int idx, bit;
51
52 /*
53 * If we have a mixed random workload, we may
54 * encounter blocks we already did IO to.
55 */
56 if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) {
57 if (!blocks)
58 blocks = 1;
59 break;
60 }
61
62 idx = RAND_MAP_IDX(f, block);
63 bit = RAND_MAP_BIT(f, block);
64
65 fio_assert(td, idx < f->num_maps);
66
67 this_blocks = nr_blocks;
68 if (this_blocks + bit > BLOCKS_PER_MAP)
69 this_blocks = BLOCKS_PER_MAP - bit;
70
71 if (this_blocks == BLOCKS_PER_MAP)
72 mask = -1U;
73 else
74 mask = ((1U << this_blocks) - 1) << bit;
75
76 f->file_map[idx] |= mask;
77 nr_blocks -= this_blocks;
78 blocks += this_blocks;
79 block += this_blocks;
80 }
81
82 if ((blocks * min_bs) < io_u->buflen)
83 io_u->buflen = blocks * min_bs;
84}
85
86static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
87 enum fio_ddir ddir)
88{
89 unsigned long long max_blocks;
90 unsigned long long max_size;
91
92 /*
93 * Hmm, should we make sure that ->io_size <= ->real_file_size?
94 */
95 max_size = f->io_size;
96 if (max_size > f->real_file_size)
97 max_size = f->real_file_size;
98
99 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
100 if (!max_blocks)
101 return 0;
102
103 return max_blocks;
104}
105
106/*
107 * Return the next free block in the map.
108 */
109static int get_next_free_block(struct thread_data *td, struct fio_file *f,
110 enum fio_ddir ddir, unsigned long long *b)
111{
112 unsigned long long min_bs = td->o.rw_min_bs;
113 int i;
114
115 i = f->last_free_lookup;
116 *b = (i * BLOCKS_PER_MAP);
117 while ((*b) * min_bs < f->real_file_size &&
118 (*b) * min_bs < f->io_size) {
119 if (f->file_map[i] != (unsigned int) -1) {
120 *b += ffz(f->file_map[i]);
121 if (*b > last_block(td, f, ddir))
122 break;
123 f->last_free_lookup = i;
124 return 0;
125 }
126
127 *b += BLOCKS_PER_MAP;
128 i++;
129 }
130
131 dprint(FD_IO, "failed finding a free block\n");
132 return 1;
133}
134
135static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
136 enum fio_ddir ddir, unsigned long long *b)
137{
138 unsigned long long r;
139 int loops = 5;
140
141 do {
142 r = os_random_long(&td->random_state);
143 dprint(FD_RANDOM, "off rand %llu\n", r);
144 *b = (last_block(td, f, ddir) - 1)
145 * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
146
147 /*
148 * if we are not maintaining a random map, we are done.
149 */
150 if (!file_randommap(td, f))
151 return 0;
152
153 /*
154 * calculate map offset and check if it's free
155 */
156 if (random_map_free(f, *b))
157 return 0;
158
159 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
160 *b);
161 } while (--loops);
162
163 /*
164 * we get here, if we didn't suceed in looking up a block. generate
165 * a random start offset into the filemap, and find the first free
166 * block from there.
167 */
168 loops = 10;
169 do {
170 f->last_free_lookup = (f->num_maps - 1) *
171 (r / (OS_RAND_MAX + 1.0));
172 if (!get_next_free_block(td, f, ddir, b))
173 return 0;
174
175 r = os_random_long(&td->random_state);
176 } while (--loops);
177
178 /*
179 * that didn't work either, try exhaustive search from the start
180 */
181 f->last_free_lookup = 0;
182 return get_next_free_block(td, f, ddir, b);
183}
184
185/*
186 * For random io, generate a random new block and see if it's used. Repeat
187 * until we find a free one. For sequential io, just return the end of
188 * the last io issued.
189 */
190static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
191{
192 struct fio_file *f = io_u->file;
193 unsigned long long b;
194 enum fio_ddir ddir = io_u->ddir;
195
196 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
197 td->ddir_nr = td->o.ddir_nr;
198
199 if (get_next_rand_offset(td, f, ddir, &b)) {
200 dprint(FD_IO, "%s: getting rand offset failed\n",
201 f->file_name);
202 return 1;
203 }
204 } else {
205 if (f->last_pos >= f->real_file_size) {
206 if (!td_random(td) ||
207 get_next_rand_offset(td, f, ddir, &b)) {
208 dprint(FD_IO, "%s: pos %llu > size %llu\n",
209 f->file_name, f->last_pos,
210 f->real_file_size);
211 return 1;
212 }
213 } else
214 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
215 }
216
217 io_u->offset = b * td->o.ba[ddir];
218 if (io_u->offset >= f->io_size) {
219 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
220 io_u->offset, f->io_size);
221 return 1;
222 }
223
224 io_u->offset += f->file_offset;
225 if (io_u->offset >= f->real_file_size) {
226 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
227 io_u->offset, f->real_file_size);
228 return 1;
229 }
230
231 return 0;
232}
233
234static int get_next_offset(struct thread_data *td, struct io_u *io_u)
235{
236 struct prof_io_ops *ops = &td->prof_io_ops;
237
238 if (ops->fill_io_u_off)
239 return ops->fill_io_u_off(td, io_u);
240
241 return __get_next_offset(td, io_u);
242}
243
244static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
245{
246 const int ddir = io_u->ddir;
247 unsigned int uninitialized_var(buflen);
248 unsigned int minbs, maxbs;
249 long r;
250
251 minbs = td->o.min_bs[ddir];
252 maxbs = td->o.max_bs[ddir];
253
254 if (minbs == maxbs)
255 buflen = minbs;
256 else {
257 r = os_random_long(&td->bsrange_state);
258 if (!td->o.bssplit_nr[ddir]) {
259 buflen = 1 + (unsigned int) ((double) maxbs *
260 (r / (OS_RAND_MAX + 1.0)));
261 if (buflen < minbs)
262 buflen = minbs;
263 } else {
264 long perc = 0;
265 unsigned int i;
266
267 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
268 struct bssplit *bsp = &td->o.bssplit[ddir][i];
269
270 buflen = bsp->bs;
271 perc += bsp->perc;
272 if (r <= ((OS_RAND_MAX / 100L) * perc))
273 break;
274 }
275 }
276 if (!td->o.bs_unaligned && is_power_of_2(minbs))
277 buflen = (buflen + minbs - 1) & ~(minbs - 1);
278 }
279
280 if (io_u->offset + buflen > io_u->file->real_file_size) {
281 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
282 minbs, ddir);
283 buflen = minbs;
284 }
285
286 return buflen;
287}
288
289static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
290{
291 struct prof_io_ops *ops = &td->prof_io_ops;
292
293 if (ops->fill_io_u_size)
294 return ops->fill_io_u_size(td, io_u);
295
296 return __get_next_buflen(td, io_u);
297}
298
299static void set_rwmix_bytes(struct thread_data *td)
300{
301 unsigned int diff;
302
303 /*
304 * we do time or byte based switch. this is needed because
305 * buffered writes may issue a lot quicker than they complete,
306 * whereas reads do not.
307 */
308 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
309 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
310}
311
312static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
313{
314 unsigned int v;
315 long r;
316
317 r = os_random_long(&td->rwmix_state);
318 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
319 if (v <= td->o.rwmix[DDIR_READ])
320 return DDIR_READ;
321
322 return DDIR_WRITE;
323}
324
325static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
326{
327 enum fio_ddir odir = ddir ^ 1;
328 struct timeval t;
329 long usec;
330
331 if (td->rate_pending_usleep[ddir] <= 0)
332 return ddir;
333
334 /*
335 * We have too much pending sleep in this direction. See if we
336 * should switch.
337 */
338 if (td_rw(td)) {
339 /*
340 * Other direction does not have too much pending, switch
341 */
342 if (td->rate_pending_usleep[odir] < 100000)
343 return odir;
344
345 /*
346 * Both directions have pending sleep. Sleep the minimum time
347 * and deduct from both.
348 */
349 if (td->rate_pending_usleep[ddir] <=
350 td->rate_pending_usleep[odir]) {
351 usec = td->rate_pending_usleep[ddir];
352 } else {
353 usec = td->rate_pending_usleep[odir];
354 ddir = odir;
355 }
356 } else
357 usec = td->rate_pending_usleep[ddir];
358
359 fio_gettime(&t, NULL);
360 usec_sleep(td, usec);
361 usec = utime_since_now(&t);
362
363 td->rate_pending_usleep[ddir] -= usec;
364
365 odir = ddir ^ 1;
366 if (td_rw(td) && __should_check_rate(td, odir))
367 td->rate_pending_usleep[odir] -= usec;
368
369 return ddir;
370}
371
372/*
373 * Return the data direction for the next io_u. If the job is a
374 * mixed read/write workload, check the rwmix cycle and switch if
375 * necessary.
376 */
377static enum fio_ddir get_rw_ddir(struct thread_data *td)
378{
379 enum fio_ddir ddir;
380
381 /*
382 * see if it's time to fsync
383 */
384 if (td->o.fsync_blocks &&
385 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
386 td->io_issues[DDIR_WRITE] && should_fsync(td))
387 return DDIR_SYNC;
388
389 /*
390 * see if it's time to fdatasync
391 */
392 if (td->o.fdatasync_blocks &&
393 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
394 td->io_issues[DDIR_WRITE] && should_fsync(td))
395 return DDIR_DATASYNC;
396
397 /*
398 * see if it's time to sync_file_range
399 */
400 if (td->sync_file_range_nr &&
401 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
402 td->io_issues[DDIR_WRITE] && should_fsync(td))
403 return DDIR_SYNC_FILE_RANGE;
404
405 if (td_rw(td)) {
406 /*
407 * Check if it's time to seed a new data direction.
408 */
409 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
410 /*
411 * Put a top limit on how many bytes we do for
412 * one data direction, to avoid overflowing the
413 * ranges too much
414 */
415 ddir = get_rand_ddir(td);
416
417 if (ddir != td->rwmix_ddir)
418 set_rwmix_bytes(td);
419
420 td->rwmix_ddir = ddir;
421 }
422 ddir = td->rwmix_ddir;
423 } else if (td_read(td))
424 ddir = DDIR_READ;
425 else
426 ddir = DDIR_WRITE;
427
428 td->rwmix_ddir = rate_ddir(td, ddir);
429 return td->rwmix_ddir;
430}
431
432void put_file_log(struct thread_data *td, struct fio_file *f)
433{
434 int ret = put_file(td, f);
435
436 if (ret)
437 td_verror(td, ret, "file close");
438}
439
440void put_io_u(struct thread_data *td, struct io_u *io_u)
441{
442 td_io_u_lock(td);
443
444 io_u->flags |= IO_U_F_FREE;
445 io_u->flags &= ~IO_U_F_FREE_DEF;
446
447 if (io_u->file)
448 put_file_log(td, io_u->file);
449
450 io_u->file = NULL;
451 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
452 td->cur_depth--;
453 flist_del_init(&io_u->list);
454 flist_add(&io_u->list, &td->io_u_freelist);
455 td_io_u_unlock(td);
456 td_io_u_free_notify(td);
457}
458
459void clear_io_u(struct thread_data *td, struct io_u *io_u)
460{
461 io_u->flags &= ~IO_U_F_FLIGHT;
462 put_io_u(td, io_u);
463}
464
465void requeue_io_u(struct thread_data *td, struct io_u **io_u)
466{
467 struct io_u *__io_u = *io_u;
468
469 dprint(FD_IO, "requeue %p\n", __io_u);
470
471 td_io_u_lock(td);
472
473 __io_u->flags |= IO_U_F_FREE;
474 if ((__io_u->flags & IO_U_F_FLIGHT) && !ddir_sync(__io_u->ddir))
475 td->io_issues[__io_u->ddir]--;
476
477 __io_u->flags &= ~IO_U_F_FLIGHT;
478 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
479 td->cur_depth--;
480 flist_del(&__io_u->list);
481 flist_add_tail(&__io_u->list, &td->io_u_requeues);
482 td_io_u_unlock(td);
483 *io_u = NULL;
484}
485
486static int fill_io_u(struct thread_data *td, struct io_u *io_u)
487{
488 if (td->io_ops->flags & FIO_NOIO)
489 goto out;
490
491 io_u->ddir = get_rw_ddir(td);
492
493 /*
494 * fsync() or fdatasync(), we are done
495 */
496 if (ddir_sync(io_u->ddir))
497 goto out;
498
499 /*
500 * See if it's time to switch to a new zone
501 */
502 if (td->zone_bytes >= td->o.zone_size) {
503 td->zone_bytes = 0;
504 io_u->file->last_pos += td->o.zone_skip;
505 td->io_skip_bytes += td->o.zone_skip;
506 }
507
508 /*
509 * No log, let the seq/rand engine retrieve the next buflen and
510 * position.
511 */
512 if (get_next_offset(td, io_u)) {
513 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
514 return 1;
515 }
516
517 io_u->buflen = get_next_buflen(td, io_u);
518 if (!io_u->buflen) {
519 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
520 return 1;
521 }
522
523 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
524 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
525 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
526 io_u->buflen, io_u->file->real_file_size);
527 return 1;
528 }
529
530 /*
531 * mark entry before potentially trimming io_u
532 */
533 if (td_random(td) && file_randommap(td, io_u->file))
534 mark_random_map(td, io_u);
535
536 /*
537 * If using a write iolog, store this entry.
538 */
539out:
540 dprint_io_u(io_u, "fill_io_u");
541 td->zone_bytes += io_u->buflen;
542 log_io_u(td, io_u);
543 return 0;
544}
545
546static void __io_u_mark_map(unsigned int *map, unsigned int nr)
547{
548 int index = 0;
549
550 switch (nr) {
551 default:
552 index = 6;
553 break;
554 case 33 ... 64:
555 index = 5;
556 break;
557 case 17 ... 32:
558 index = 4;
559 break;
560 case 9 ... 16:
561 index = 3;
562 break;
563 case 5 ... 8:
564 index = 2;
565 break;
566 case 1 ... 4:
567 index = 1;
568 case 0:
569 break;
570 }
571
572 map[index]++;
573}
574
575void io_u_mark_submit(struct thread_data *td, unsigned int nr)
576{
577 __io_u_mark_map(td->ts.io_u_submit, nr);
578 td->ts.total_submit++;
579}
580
581void io_u_mark_complete(struct thread_data *td, unsigned int nr)
582{
583 __io_u_mark_map(td->ts.io_u_complete, nr);
584 td->ts.total_complete++;
585}
586
587void io_u_mark_depth(struct thread_data *td, unsigned int nr)
588{
589 int index = 0;
590
591 switch (td->cur_depth) {
592 default:
593 index = 6;
594 break;
595 case 32 ... 63:
596 index = 5;
597 break;
598 case 16 ... 31:
599 index = 4;
600 break;
601 case 8 ... 15:
602 index = 3;
603 break;
604 case 4 ... 7:
605 index = 2;
606 break;
607 case 2 ... 3:
608 index = 1;
609 case 1:
610 break;
611 }
612
613 td->ts.io_u_map[index] += nr;
614}
615
616static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
617{
618 int index = 0;
619
620 assert(usec < 1000);
621
622 switch (usec) {
623 case 750 ... 999:
624 index = 9;
625 break;
626 case 500 ... 749:
627 index = 8;
628 break;
629 case 250 ... 499:
630 index = 7;
631 break;
632 case 100 ... 249:
633 index = 6;
634 break;
635 case 50 ... 99:
636 index = 5;
637 break;
638 case 20 ... 49:
639 index = 4;
640 break;
641 case 10 ... 19:
642 index = 3;
643 break;
644 case 4 ... 9:
645 index = 2;
646 break;
647 case 2 ... 3:
648 index = 1;
649 case 0 ... 1:
650 break;
651 }
652
653 assert(index < FIO_IO_U_LAT_U_NR);
654 td->ts.io_u_lat_u[index]++;
655}
656
657static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
658{
659 int index = 0;
660
661 switch (msec) {
662 default:
663 index = 11;
664 break;
665 case 1000 ... 1999:
666 index = 10;
667 break;
668 case 750 ... 999:
669 index = 9;
670 break;
671 case 500 ... 749:
672 index = 8;
673 break;
674 case 250 ... 499:
675 index = 7;
676 break;
677 case 100 ... 249:
678 index = 6;
679 break;
680 case 50 ... 99:
681 index = 5;
682 break;
683 case 20 ... 49:
684 index = 4;
685 break;
686 case 10 ... 19:
687 index = 3;
688 break;
689 case 4 ... 9:
690 index = 2;
691 break;
692 case 2 ... 3:
693 index = 1;
694 case 0 ... 1:
695 break;
696 }
697
698 assert(index < FIO_IO_U_LAT_M_NR);
699 td->ts.io_u_lat_m[index]++;
700}
701
702static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
703{
704 if (usec < 1000)
705 io_u_mark_lat_usec(td, usec);
706 else
707 io_u_mark_lat_msec(td, usec / 1000);
708}
709
710/*
711 * Get next file to service by choosing one at random
712 */
713static struct fio_file *get_next_file_rand(struct thread_data *td,
714 enum fio_file_flags goodf,
715 enum fio_file_flags badf)
716{
717 struct fio_file *f;
718 int fno;
719
720 do {
721 long r = os_random_long(&td->next_file_state);
722 int opened = 0;
723
724 fno = (unsigned int) ((double) td->o.nr_files
725 * (r / (OS_RAND_MAX + 1.0)));
726 f = td->files[fno];
727 if (fio_file_done(f))
728 continue;
729
730 if (!fio_file_open(f)) {
731 int err;
732
733 err = td_io_open_file(td, f);
734 if (err)
735 continue;
736 opened = 1;
737 }
738
739 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
740 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
741 return f;
742 }
743 if (opened)
744 td_io_close_file(td, f);
745 } while (1);
746}
747
748/*
749 * Get next file to service by doing round robin between all available ones
750 */
751static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
752 int badf)
753{
754 unsigned int old_next_file = td->next_file;
755 struct fio_file *f;
756
757 do {
758 int opened = 0;
759
760 f = td->files[td->next_file];
761
762 td->next_file++;
763 if (td->next_file >= td->o.nr_files)
764 td->next_file = 0;
765
766 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
767 if (fio_file_done(f)) {
768 f = NULL;
769 continue;
770 }
771
772 if (!fio_file_open(f)) {
773 int err;
774
775 err = td_io_open_file(td, f);
776 if (err) {
777 dprint(FD_FILE, "error %d on open of %s\n",
778 err, f->file_name);
779 f = NULL;
780 continue;
781 }
782 opened = 1;
783 }
784
785 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
786 f->flags);
787 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
788 break;
789
790 if (opened)
791 td_io_close_file(td, f);
792
793 f = NULL;
794 } while (td->next_file != old_next_file);
795
796 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
797 return f;
798}
799
800static struct fio_file *__get_next_file(struct thread_data *td)
801{
802 struct fio_file *f;
803
804 assert(td->o.nr_files <= td->files_index);
805
806 if (td->nr_done_files >= td->o.nr_files) {
807 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
808 " nr_files=%d\n", td->nr_open_files,
809 td->nr_done_files,
810 td->o.nr_files);
811 return NULL;
812 }
813
814 f = td->file_service_file;
815 if (f && fio_file_open(f) && !fio_file_closing(f)) {
816 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
817 goto out;
818 if (td->file_service_left--)
819 goto out;
820 }
821
822 if (td->o.file_service_type == FIO_FSERVICE_RR ||
823 td->o.file_service_type == FIO_FSERVICE_SEQ)
824 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
825 else
826 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
827
828 td->file_service_file = f;
829 td->file_service_left = td->file_service_nr - 1;
830out:
831 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
832 return f;
833}
834
835static struct fio_file *get_next_file(struct thread_data *td)
836{
837 struct prof_io_ops *ops = &td->prof_io_ops;
838
839 if (ops->get_next_file)
840 return ops->get_next_file(td);
841
842 return __get_next_file(td);
843}
844
845static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
846{
847 struct fio_file *f;
848
849 do {
850 f = get_next_file(td);
851 if (!f)
852 return 1;
853
854 io_u->file = f;
855 get_file(f);
856
857 if (!fill_io_u(td, io_u))
858 break;
859
860 put_file_log(td, f);
861 td_io_close_file(td, f);
862 io_u->file = NULL;
863 fio_file_set_done(f);
864 td->nr_done_files++;
865 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
866 td->nr_done_files, td->o.nr_files);
867 } while (1);
868
869 return 0;
870}
871
872
873struct io_u *__get_io_u(struct thread_data *td)
874{
875 struct io_u *io_u = NULL;
876
877 td_io_u_lock(td);
878
879again:
880 if (!flist_empty(&td->io_u_requeues))
881 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
882 else if (!queue_full(td)) {
883 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
884
885 io_u->buflen = 0;
886 io_u->resid = 0;
887 io_u->file = NULL;
888 io_u->end_io = NULL;
889 }
890
891 /*
892 * We ran out, wait for async verify threads to finish and return one
893 */
894 if (!io_u && td->o.verify_async) {
895 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
896 goto again;
897 }
898
899 if (io_u) {
900 assert(io_u->flags & IO_U_F_FREE);
901 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
902
903 io_u->error = 0;
904 flist_del(&io_u->list);
905 flist_add(&io_u->list, &td->io_u_busylist);
906 td->cur_depth++;
907 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
908 }
909
910 td_io_u_unlock(td);
911 return io_u;
912}
913
914/*
915 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
916 * etc. The returned io_u is fully ready to be prepped and submitted.
917 */
918struct io_u *get_io_u(struct thread_data *td)
919{
920 struct fio_file *f;
921 struct io_u *io_u;
922
923 io_u = __get_io_u(td);
924 if (!io_u) {
925 dprint(FD_IO, "__get_io_u failed\n");
926 return NULL;
927 }
928
929 /*
930 * from a requeue, io_u already setup
931 */
932 if (io_u->file)
933 goto out;
934
935 /*
936 * If using an iolog, grab next piece if any available.
937 */
938 if (td->o.read_iolog_file) {
939 if (read_iolog_get(td, io_u))
940 goto err_put;
941 } else if (set_io_u_file(td, io_u)) {
942 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
943 goto err_put;
944 }
945
946 f = io_u->file;
947 assert(fio_file_open(f));
948
949 if (!ddir_sync(io_u->ddir)) {
950 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
951 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
952 goto err_put;
953 }
954
955 f->last_pos = io_u->offset + io_u->buflen;
956
957 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
958 populate_verify_io_u(td, io_u);
959 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
960 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
961 }
962
963 /*
964 * Set io data pointers.
965 */
966 io_u->xfer_buf = io_u->buf;
967 io_u->xfer_buflen = io_u->buflen;
968
969out:
970 if (!td_io_prep(td, io_u)) {
971 if (!td->o.disable_slat)
972 fio_gettime(&io_u->start_time, NULL);
973 return io_u;
974 }
975err_put:
976 dprint(FD_IO, "get_io_u failed\n");
977 put_io_u(td, io_u);
978 return NULL;
979}
980
981void io_u_log_error(struct thread_data *td, struct io_u *io_u)
982{
983 const char *msg[] = { "read", "write", "sync" };
984
985 log_err("fio: io_u error");
986
987 if (io_u->file)
988 log_err(" on file %s", io_u->file->file_name);
989
990 log_err(": %s\n", strerror(io_u->error));
991
992 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
993 io_u->offset, io_u->xfer_buflen);
994
995 if (!td->error)
996 td_verror(td, io_u->error, "io_u error");
997}
998
999static void io_completed(struct thread_data *td, struct io_u *io_u,
1000 struct io_completion_data *icd)
1001{
1002 /*
1003 * Older gcc's are too dumb to realize that usec is always used
1004 * initialized, silence that warning.
1005 */
1006 unsigned long uninitialized_var(usec);
1007 struct fio_file *f;
1008
1009 dprint_io_u(io_u, "io complete");
1010
1011 td_io_u_lock(td);
1012 assert(io_u->flags & IO_U_F_FLIGHT);
1013 io_u->flags &= ~IO_U_F_FLIGHT;
1014 td_io_u_unlock(td);
1015
1016 if (ddir_sync(io_u->ddir)) {
1017 td->last_was_sync = 1;
1018 f = io_u->file;
1019 if (f) {
1020 f->first_write = -1ULL;
1021 f->last_write = -1ULL;
1022 }
1023 return;
1024 }
1025
1026 td->last_was_sync = 0;
1027
1028 if (!io_u->error) {
1029 unsigned int bytes = io_u->buflen - io_u->resid;
1030 const enum fio_ddir idx = io_u->ddir;
1031 const enum fio_ddir odx = io_u->ddir ^ 1;
1032 int ret;
1033
1034 td->io_blocks[idx]++;
1035 td->io_bytes[idx] += bytes;
1036 td->this_io_bytes[idx] += bytes;
1037
1038 if (idx == DDIR_WRITE) {
1039 f = io_u->file;
1040 if (f) {
1041 if (f->first_write == -1ULL ||
1042 io_u->offset < f->first_write)
1043 f->first_write = io_u->offset;
1044 if (f->last_write == -1ULL ||
1045 ((io_u->offset + bytes) > f->last_write))
1046 f->last_write = io_u->offset + bytes;
1047 }
1048 }
1049
1050 if (ramp_time_over(td)) {
1051 unsigned long uninitialized_var(lusec);
1052
1053 if (!td->o.disable_clat || !td->o.disable_bw)
1054 lusec = utime_since(&io_u->issue_time,
1055 &icd->time);
1056
1057 if (!td->o.disable_clat) {
1058 add_clat_sample(td, idx, lusec, bytes);
1059 io_u_mark_latency(td, lusec);
1060 }
1061 if (!td->o.disable_bw)
1062 add_bw_sample(td, idx, bytes, &icd->time);
1063 if (__should_check_rate(td, idx)) {
1064 td->rate_pending_usleep[idx] =
1065 ((td->this_io_bytes[idx] *
1066 td->rate_nsec_cycle[idx]) / 1000 -
1067 utime_since_now(&td->start));
1068 }
1069 if (__should_check_rate(td, idx ^ 1))
1070 td->rate_pending_usleep[odx] =
1071 ((td->this_io_bytes[odx] *
1072 td->rate_nsec_cycle[odx]) / 1000 -
1073 utime_since_now(&td->start));
1074 }
1075
1076 if (td_write(td) && idx == DDIR_WRITE &&
1077 td->o.do_verify &&
1078 td->o.verify != VERIFY_NONE)
1079 log_io_piece(td, io_u);
1080
1081 icd->bytes_done[idx] += bytes;
1082
1083 if (io_u->end_io) {
1084 ret = io_u->end_io(td, io_u);
1085 if (ret && !icd->error)
1086 icd->error = ret;
1087 }
1088 } else {
1089 icd->error = io_u->error;
1090 io_u_log_error(td, io_u);
1091 }
1092 if (td->o.continue_on_error && icd->error &&
1093 td_non_fatal_error(icd->error)) {
1094 /*
1095 * If there is a non_fatal error, then add to the error count
1096 * and clear all the errors.
1097 */
1098 update_error_count(td, icd->error);
1099 td_clear_error(td);
1100 icd->error = 0;
1101 io_u->error = 0;
1102 }
1103}
1104
1105static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1106 int nr)
1107{
1108 if (!td->o.disable_clat || !td->o.disable_bw)
1109 fio_gettime(&icd->time, NULL);
1110
1111 icd->nr = nr;
1112
1113 icd->error = 0;
1114 icd->bytes_done[0] = icd->bytes_done[1] = 0;
1115}
1116
1117static void ios_completed(struct thread_data *td,
1118 struct io_completion_data *icd)
1119{
1120 struct io_u *io_u;
1121 int i;
1122
1123 for (i = 0; i < icd->nr; i++) {
1124 io_u = td->io_ops->event(td, i);
1125
1126 io_completed(td, io_u, icd);
1127
1128 if (!(io_u->flags & IO_U_F_FREE_DEF))
1129 put_io_u(td, io_u);
1130 }
1131}
1132
1133/*
1134 * Complete a single io_u for the sync engines.
1135 */
1136int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1137 unsigned long *bytes)
1138{
1139 struct io_completion_data icd;
1140
1141 init_icd(td, &icd, 1);
1142 io_completed(td, io_u, &icd);
1143
1144 if (!(io_u->flags & IO_U_F_FREE_DEF))
1145 put_io_u(td, io_u);
1146
1147 if (icd.error) {
1148 td_verror(td, icd.error, "io_u_sync_complete");
1149 return -1;
1150 }
1151
1152 if (bytes) {
1153 bytes[0] += icd.bytes_done[0];
1154 bytes[1] += icd.bytes_done[1];
1155 }
1156
1157 return 0;
1158}
1159
1160/*
1161 * Called to complete min_events number of io for the async engines.
1162 */
1163int io_u_queued_complete(struct thread_data *td, int min_evts,
1164 unsigned long *bytes)
1165{
1166 struct io_completion_data icd;
1167 struct timespec *tvp = NULL;
1168 int ret;
1169 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1170
1171 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1172
1173 if (!min_evts)
1174 tvp = &ts;
1175
1176 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1177 if (ret < 0) {
1178 td_verror(td, -ret, "td_io_getevents");
1179 return ret;
1180 } else if (!ret)
1181 return ret;
1182
1183 init_icd(td, &icd, ret);
1184 ios_completed(td, &icd);
1185 if (icd.error) {
1186 td_verror(td, icd.error, "io_u_queued_complete");
1187 return -1;
1188 }
1189
1190 if (bytes) {
1191 bytes[0] += icd.bytes_done[0];
1192 bytes[1] += icd.bytes_done[1];
1193 }
1194
1195 return 0;
1196}
1197
1198/*
1199 * Call when io_u is really queued, to update the submission latency.
1200 */
1201void io_u_queued(struct thread_data *td, struct io_u *io_u)
1202{
1203 if (!td->o.disable_slat) {
1204 unsigned long slat_time;
1205
1206 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1207 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
1208 }
1209}
1210
1211/*
1212 * "randomly" fill the buffer contents
1213 */
1214void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1215 unsigned int max_bs)
1216{
1217 long *ptr = io_u->buf;
1218
1219 if (!td->o.zero_buffers) {
1220 while ((void *) ptr - io_u->buf < max_bs) {
1221 *ptr = rand() * GOLDEN_RATIO_PRIME;
1222 ptr++;
1223 }
1224 } else
1225 memset(ptr, 0, max_bs);
1226}
fio - Flexible IO Tester