projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Fix verification error with bsrange and split read/write phase
[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/axmap.h"
14#include "err.h"
15#include "lib/pow2.h"
16
17struct io_completion_data {
18 int nr; /* input */
19
20 int error; /* output */
21 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
22 struct timeval time; /* output */
23};
24
25/*
26 * The ->io_axmap contains a map of blocks we have or have not done io
27 * to yet. Used to make sure we cover the entire range in a fair fashion.
28 */
29static int random_map_free(struct fio_file *f, const uint64_t block)
30{
31 return !axmap_isset(f->io_axmap, block);
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 int nr_blocks;
42 uint64_t block;
43
44 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
45 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
46
47 if (!(io_u->flags & IO_U_F_BUSY_OK))
48 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
49
50 if ((nr_blocks * min_bs) < io_u->buflen)
51 io_u->buflen = nr_blocks * min_bs;
52}
53
54static uint64_t last_block(struct thread_data *td, struct fio_file *f,
55 enum fio_ddir ddir)
56{
57 uint64_t max_blocks;
58 uint64_t max_size;
59
60 assert(ddir_rw(ddir));
61
62 /*
63 * Hmm, should we make sure that ->io_size <= ->real_file_size?
64 */
65 max_size = f->io_size;
66 if (max_size > f->real_file_size)
67 max_size = f->real_file_size;
68
69 if (td->o.zone_range)
70 max_size = td->o.zone_range;
71
72 if (td->o.min_bs[ddir] > td->o.ba[ddir])
73 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
74
75 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
76 if (!max_blocks)
77 return 0;
78
79 return max_blocks;
80}
81
82struct rand_off {
83 struct flist_head list;
84 uint64_t off;
85};
86
87static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
88 enum fio_ddir ddir, uint64_t *b)
89{
90 uint64_t r;
91
92 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
93 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
94 uint64_t frand_max, lastb;
95
96 lastb = last_block(td, f, ddir);
97 if (!lastb)
98 return 1;
99
100 frand_max = rand_max(&td->random_state);
101 r = __rand(&td->random_state);
102
103 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
104
105 *b = lastb * (r / ((uint64_t) frand_max + 1.0));
106 } else {
107 uint64_t off = 0;
108
109 assert(fio_file_lfsr(f));
110
111 if (lfsr_next(&f->lfsr, &off))
112 return 1;
113
114 *b = off;
115 }
116
117 /*
118 * if we are not maintaining a random map, we are done.
119 */
120 if (!file_randommap(td, f))
121 goto ret;
122
123 /*
124 * calculate map offset and check if it's free
125 */
126 if (random_map_free(f, *b))
127 goto ret;
128
129 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
130 (unsigned long long) *b);
131
132 *b = axmap_next_free(f->io_axmap, *b);
133 if (*b == (uint64_t) -1ULL)
134 return 1;
135ret:
136 return 0;
137}
138
139static int __get_next_rand_offset_zipf(struct thread_data *td,
140 struct fio_file *f, enum fio_ddir ddir,
141 uint64_t *b)
142{
143 *b = zipf_next(&f->zipf);
144 return 0;
145}
146
147static int __get_next_rand_offset_pareto(struct thread_data *td,
148 struct fio_file *f, enum fio_ddir ddir,
149 uint64_t *b)
150{
151 *b = pareto_next(&f->zipf);
152 return 0;
153}
154
155static int __get_next_rand_offset_gauss(struct thread_data *td,
156 struct fio_file *f, enum fio_ddir ddir,
157 uint64_t *b)
158{
159 *b = gauss_next(&f->gauss);
160 return 0;
161}
162
163
164static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
165{
166 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
167 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
168
169 return r1->off - r2->off;
170}
171
172static int get_off_from_method(struct thread_data *td, struct fio_file *f,
173 enum fio_ddir ddir, uint64_t *b)
174{
175 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
176 return __get_next_rand_offset(td, f, ddir, b);
177 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
178 return __get_next_rand_offset_zipf(td, f, ddir, b);
179 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
180 return __get_next_rand_offset_pareto(td, f, ddir, b);
181 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
182 return __get_next_rand_offset_gauss(td, f, ddir, b);
183
184 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
185 return 1;
186}
187
188/*
189 * Sort the reads for a verify phase in batches of verifysort_nr, if
190 * specified.
191 */
192static inline int should_sort_io(struct thread_data *td)
193{
194 if (!td->o.verifysort_nr || !td->o.do_verify)
195 return 0;
196 if (!td_random(td))
197 return 0;
198 if (td->runstate != TD_VERIFYING)
199 return 0;
200 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
201 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
202 return 0;
203
204 return 1;
205}
206
207static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
208{
209 uint64_t frand_max;
210 unsigned int v;
211 unsigned long r;
212
213 if (td->o.perc_rand[ddir] == 100)
214 return 1;
215
216 frand_max = rand_max(&td->seq_rand_state[ddir]);
217 r = __rand(&td->seq_rand_state[ddir]);
218 v = 1 + (int) (100.0 * (r / (frand_max + 1.0)));
219
220 return v <= td->o.perc_rand[ddir];
221}
222
223static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
224 enum fio_ddir ddir, uint64_t *b)
225{
226 struct rand_off *r;
227 int i, ret = 1;
228
229 if (!should_sort_io(td))
230 return get_off_from_method(td, f, ddir, b);
231
232 if (!flist_empty(&td->next_rand_list)) {
233fetch:
234 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
235 flist_del(&r->list);
236 *b = r->off;
237 free(r);
238 return 0;
239 }
240
241 for (i = 0; i < td->o.verifysort_nr; i++) {
242 r = malloc(sizeof(*r));
243
244 ret = get_off_from_method(td, f, ddir, &r->off);
245 if (ret) {
246 free(r);
247 break;
248 }
249
250 flist_add(&r->list, &td->next_rand_list);
251 }
252
253 if (ret && !i)
254 return ret;
255
256 assert(!flist_empty(&td->next_rand_list));
257 flist_sort(NULL, &td->next_rand_list, flist_cmp);
258 goto fetch;
259}
260
261static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
262 enum fio_ddir ddir, uint64_t *b)
263{
264 if (!get_next_rand_offset(td, f, ddir, b))
265 return 0;
266
267 if (td->o.time_based) {
268 fio_file_reset(td, f);
269 if (!get_next_rand_offset(td, f, ddir, b))
270 return 0;
271 }
272
273 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
274 f->file_name, (unsigned long long) f->last_pos[ddir],
275 (unsigned long long) f->real_file_size);
276 return 1;
277}
278
279static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
280 enum fio_ddir ddir, uint64_t *offset)
281{
282 struct thread_options *o = &td->o;
283
284 assert(ddir_rw(ddir));
285
286 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
287 o->time_based)
288 f->last_pos[ddir] = f->last_pos[ddir] - f->io_size;
289
290 if (f->last_pos[ddir] < f->real_file_size) {
291 uint64_t pos;
292
293 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0)
294 f->last_pos[ddir] = f->real_file_size;
295
296 pos = f->last_pos[ddir] - f->file_offset;
297 if (pos && o->ddir_seq_add) {
298 pos += o->ddir_seq_add;
299
300 /*
301 * If we reach beyond the end of the file
302 * with holed IO, wrap around to the
303 * beginning again.
304 */
305 if (pos >= f->real_file_size)
306 pos = f->file_offset;
307 }
308
309 *offset = pos;
310 return 0;
311 }
312
313 return 1;
314}
315
316static int get_next_block(struct thread_data *td, struct io_u *io_u,
317 enum fio_ddir ddir, int rw_seq,
318 unsigned int *is_random)
319{
320 struct fio_file *f = io_u->file;
321 uint64_t b, offset;
322 int ret;
323
324 assert(ddir_rw(ddir));
325
326 b = offset = -1ULL;
327
328 if (rw_seq) {
329 if (td_random(td)) {
330 if (should_do_random(td, ddir)) {
331 ret = get_next_rand_block(td, f, ddir, &b);
332 *is_random = 1;
333 } else {
334 *is_random = 0;
335 io_u_set(io_u, IO_U_F_BUSY_OK);
336 ret = get_next_seq_offset(td, f, ddir, &offset);
337 if (ret)
338 ret = get_next_rand_block(td, f, ddir, &b);
339 }
340 } else {
341 *is_random = 0;
342 ret = get_next_seq_offset(td, f, ddir, &offset);
343 }
344 } else {
345 io_u_set(io_u, IO_U_F_BUSY_OK);
346 *is_random = 0;
347
348 if (td->o.rw_seq == RW_SEQ_SEQ) {
349 ret = get_next_seq_offset(td, f, ddir, &offset);
350 if (ret) {
351 ret = get_next_rand_block(td, f, ddir, &b);
352 *is_random = 0;
353 }
354 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
355 if (f->last_start[ddir] != -1ULL)
356 offset = f->last_start[ddir] - f->file_offset;
357 else
358 offset = 0;
359 ret = 0;
360 } else {
361 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
362 ret = 1;
363 }
364 }
365
366 if (!ret) {
367 if (offset != -1ULL)
368 io_u->offset = offset;
369 else if (b != -1ULL)
370 io_u->offset = b * td->o.ba[ddir];
371 else {
372 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
373 ret = 1;
374 }
375 }
376
377 return ret;
378}
379
380/*
381 * For random io, generate a random new block and see if it's used. Repeat
382 * until we find a free one. For sequential io, just return the end of
383 * the last io issued.
384 */
385static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
386 unsigned int *is_random)
387{
388 struct fio_file *f = io_u->file;
389 enum fio_ddir ddir = io_u->ddir;
390 int rw_seq_hit = 0;
391
392 assert(ddir_rw(ddir));
393
394 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
395 rw_seq_hit = 1;
396 td->ddir_seq_nr = td->o.ddir_seq_nr;
397 }
398
399 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
400 return 1;
401
402 if (io_u->offset >= f->io_size) {
403 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
404 (unsigned long long) io_u->offset,
405 (unsigned long long) f->io_size);
406 return 1;
407 }
408
409 io_u->offset += f->file_offset;
410 if (io_u->offset >= f->real_file_size) {
411 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
412 (unsigned long long) io_u->offset,
413 (unsigned long long) f->real_file_size);
414 return 1;
415 }
416
417 return 0;
418}
419
420static int get_next_offset(struct thread_data *td, struct io_u *io_u,
421 unsigned int *is_random)
422{
423 if (td->flags & TD_F_PROFILE_OPS) {
424 struct prof_io_ops *ops = &td->prof_io_ops;
425
426 if (ops->fill_io_u_off)
427 return ops->fill_io_u_off(td, io_u, is_random);
428 }
429
430 return __get_next_offset(td, io_u, is_random);
431}
432
433static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
434 unsigned int buflen)
435{
436 struct fio_file *f = io_u->file;
437
438 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
439}
440
441static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
442 unsigned int is_random)
443{
444 int ddir = io_u->ddir;
445 unsigned int buflen = 0;
446 unsigned int minbs, maxbs;
447 uint64_t frand_max;
448 unsigned long r;
449
450 assert(ddir_rw(ddir));
451
452 if (td->o.bs_is_seq_rand)
453 ddir = is_random ? DDIR_WRITE: DDIR_READ;
454
455 minbs = td->o.min_bs[ddir];
456 maxbs = td->o.max_bs[ddir];
457
458 if (minbs == maxbs)
459 return minbs;
460
461 /*
462 * If we can't satisfy the min block size from here, then fail
463 */
464 if (!io_u_fits(td, io_u, minbs))
465 return 0;
466
467 frand_max = rand_max(&td->bsrange_state);
468 do {
469 r = __rand(&td->bsrange_state);
470
471 if (!td->o.bssplit_nr[ddir]) {
472 buflen = 1 + (unsigned int) ((double) maxbs *
473 (r / (frand_max + 1.0)));
474 if (buflen < minbs)
475 buflen = minbs;
476 } else {
477 long perc = 0;
478 unsigned int i;
479
480 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
481 struct bssplit *bsp = &td->o.bssplit[ddir][i];
482
483 buflen = bsp->bs;
484 perc += bsp->perc;
485 if ((r <= ((frand_max / 100L) * perc)) &&
486 io_u_fits(td, io_u, buflen))
487 break;
488 }
489 }
490
491 if (td->o.verify != VERIFY_NONE)
492 buflen = (buflen + td->o.verify_interval - 1) &
493 ~(td->o.verify_interval - 1);
494
495 if (!td->o.bs_unaligned && is_power_of_2(minbs))
496 buflen &= ~(minbs - 1);
497
498 } while (!io_u_fits(td, io_u, buflen));
499
500 return buflen;
501}
502
503static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
504 unsigned int is_random)
505{
506 if (td->flags & TD_F_PROFILE_OPS) {
507 struct prof_io_ops *ops = &td->prof_io_ops;
508
509 if (ops->fill_io_u_size)
510 return ops->fill_io_u_size(td, io_u, is_random);
511 }
512
513 return __get_next_buflen(td, io_u, is_random);
514}
515
516static void set_rwmix_bytes(struct thread_data *td)
517{
518 unsigned int diff;
519
520 /*
521 * we do time or byte based switch. this is needed because
522 * buffered writes may issue a lot quicker than they complete,
523 * whereas reads do not.
524 */
525 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
526 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
527}
528
529static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
530{
531 uint64_t frand_max = rand_max(&td->rwmix_state);
532 unsigned int v;
533 unsigned long r;
534
535 r = __rand(&td->rwmix_state);
536 v = 1 + (int) (100.0 * (r / (frand_max + 1.0)));
537
538 if (v <= td->o.rwmix[DDIR_READ])
539 return DDIR_READ;
540
541 return DDIR_WRITE;
542}
543
544void io_u_quiesce(struct thread_data *td)
545{
546 /*
547 * We are going to sleep, ensure that we flush anything pending as
548 * not to skew our latency numbers.
549 *
550 * Changed to only monitor 'in flight' requests here instead of the
551 * td->cur_depth, b/c td->cur_depth does not accurately represent
552 * io's that have been actually submitted to an async engine,
553 * and cur_depth is meaningless for sync engines.
554 */
555 if (td->io_u_queued || td->cur_depth) {
556 int fio_unused ret;
557
558 ret = td_io_commit(td);
559 }
560
561 while (td->io_u_in_flight) {
562 int fio_unused ret;
563
564 ret = io_u_queued_complete(td, 1);
565 }
566}
567
568static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
569{
570 enum fio_ddir odir = ddir ^ 1;
571 long usec, now;
572
573 assert(ddir_rw(ddir));
574 now = utime_since_now(&td->start);
575
576 /*
577 * if rate_next_io_time is in the past, need to catch up to rate
578 */
579 if (td->rate_next_io_time[ddir] <= now)
580 return ddir;
581
582 /*
583 * We are ahead of rate in this direction. See if we
584 * should switch.
585 */
586 if (td_rw(td) && td->o.rwmix[odir]) {
587 /*
588 * Other direction is behind rate, switch
589 */
590 if (td->rate_next_io_time[odir] <= now)
591 return odir;
592
593 /*
594 * Both directions are ahead of rate. sleep the min
595 * switch if necissary
596 */
597 if (td->rate_next_io_time[ddir] <=
598 td->rate_next_io_time[odir]) {
599 usec = td->rate_next_io_time[ddir] - now;
600 } else {
601 usec = td->rate_next_io_time[odir] - now;
602 ddir = odir;
603 }
604 } else
605 usec = td->rate_next_io_time[ddir] - now;
606
607 if (td->o.io_submit_mode == IO_MODE_INLINE)
608 io_u_quiesce(td);
609
610 usec = usec_sleep(td, usec);
611
612 return ddir;
613}
614
615/*
616 * Return the data direction for the next io_u. If the job is a
617 * mixed read/write workload, check the rwmix cycle and switch if
618 * necessary.
619 */
620static enum fio_ddir get_rw_ddir(struct thread_data *td)
621{
622 enum fio_ddir ddir;
623
624 /*
625 * see if it's time to fsync
626 */
627 if (td->o.fsync_blocks &&
628 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
629 td->io_issues[DDIR_WRITE] && should_fsync(td))
630 return DDIR_SYNC;
631
632 /*
633 * see if it's time to fdatasync
634 */
635 if (td->o.fdatasync_blocks &&
636 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
637 td->io_issues[DDIR_WRITE] && should_fsync(td))
638 return DDIR_DATASYNC;
639
640 /*
641 * see if it's time to sync_file_range
642 */
643 if (td->sync_file_range_nr &&
644 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
645 td->io_issues[DDIR_WRITE] && should_fsync(td))
646 return DDIR_SYNC_FILE_RANGE;
647
648 if (td_rw(td)) {
649 /*
650 * Check if it's time to seed a new data direction.
651 */
652 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
653 /*
654 * Put a top limit on how many bytes we do for
655 * one data direction, to avoid overflowing the
656 * ranges too much
657 */
658 ddir = get_rand_ddir(td);
659
660 if (ddir != td->rwmix_ddir)
661 set_rwmix_bytes(td);
662
663 td->rwmix_ddir = ddir;
664 }
665 ddir = td->rwmix_ddir;
666 } else if (td_read(td))
667 ddir = DDIR_READ;
668 else if (td_write(td))
669 ddir = DDIR_WRITE;
670 else
671 ddir = DDIR_TRIM;
672
673 td->rwmix_ddir = rate_ddir(td, ddir);
674 return td->rwmix_ddir;
675}
676
677static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
678{
679 enum fio_ddir ddir = get_rw_ddir(td);
680
681 if (td_trimwrite(td)) {
682 struct fio_file *f = io_u->file;
683 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
684 ddir = DDIR_TRIM;
685 else
686 ddir = DDIR_WRITE;
687 }
688
689 io_u->ddir = io_u->acct_ddir = ddir;
690
691 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
692 td->o.barrier_blocks &&
693 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
694 td->io_issues[DDIR_WRITE])
695 io_u_set(io_u, IO_U_F_BARRIER);
696}
697
698void put_file_log(struct thread_data *td, struct fio_file *f)
699{
700 unsigned int ret = put_file(td, f);
701
702 if (ret)
703 td_verror(td, ret, "file close");
704}
705
706void put_io_u(struct thread_data *td, struct io_u *io_u)
707{
708 if (td->parent)
709 td = td->parent;
710
711 td_io_u_lock(td);
712
713 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
714 put_file_log(td, io_u->file);
715
716 io_u->file = NULL;
717 io_u_set(io_u, IO_U_F_FREE);
718
719 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
720 td->cur_depth--;
721 assert(!(td->flags & TD_F_CHILD));
722 }
723 io_u_qpush(&td->io_u_freelist, io_u);
724 td_io_u_unlock(td);
725 td_io_u_free_notify(td);
726}
727
728void clear_io_u(struct thread_data *td, struct io_u *io_u)
729{
730 io_u_clear(io_u, IO_U_F_FLIGHT);
731 put_io_u(td, io_u);
732}
733
734void requeue_io_u(struct thread_data *td, struct io_u **io_u)
735{
736 struct io_u *__io_u = *io_u;
737 enum fio_ddir ddir = acct_ddir(__io_u);
738
739 dprint(FD_IO, "requeue %p\n", __io_u);
740
741 if (td->parent)
742 td = td->parent;
743
744 td_io_u_lock(td);
745
746 io_u_set(__io_u, IO_U_F_FREE);
747 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
748 td->io_issues[ddir]--;
749
750 io_u_clear(__io_u, IO_U_F_FLIGHT);
751 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
752 td->cur_depth--;
753 assert(!(td->flags & TD_F_CHILD));
754 }
755
756 io_u_rpush(&td->io_u_requeues, __io_u);
757 td_io_u_unlock(td);
758 td_io_u_free_notify(td);
759 *io_u = NULL;
760}
761
762static int fill_io_u(struct thread_data *td, struct io_u *io_u)
763{
764 unsigned int is_random;
765
766 if (td->io_ops->flags & FIO_NOIO)
767 goto out;
768
769 set_rw_ddir(td, io_u);
770
771 /*
772 * fsync() or fdatasync() or trim etc, we are done
773 */
774 if (!ddir_rw(io_u->ddir))
775 goto out;
776
777 /*
778 * See if it's time to switch to a new zone
779 */
780 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
781 struct fio_file *f = io_u->file;
782
783 td->zone_bytes = 0;
784 f->file_offset += td->o.zone_range + td->o.zone_skip;
785
786 /*
787 * Wrap from the beginning, if we exceed the file size
788 */
789 if (f->file_offset >= f->real_file_size)
790 f->file_offset = f->real_file_size - f->file_offset;
791 f->last_pos[io_u->ddir] = f->file_offset;
792 td->io_skip_bytes += td->o.zone_skip;
793 }
794
795 /*
796 * No log, let the seq/rand engine retrieve the next buflen and
797 * position.
798 */
799 if (get_next_offset(td, io_u, &is_random)) {
800 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
801 return 1;
802 }
803
804 io_u->buflen = get_next_buflen(td, io_u, is_random);
805 if (!io_u->buflen) {
806 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
807 return 1;
808 }
809
810 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
811 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
812 dprint(FD_IO, " off=%llu/%lu > %llu\n",
813 (unsigned long long) io_u->offset, io_u->buflen,
814 (unsigned long long) io_u->file->real_file_size);
815 return 1;
816 }
817
818 /*
819 * mark entry before potentially trimming io_u
820 */
821 if (td_random(td) && file_randommap(td, io_u->file))
822 mark_random_map(td, io_u);
823
824out:
825 dprint_io_u(io_u, "fill_io_u");
826 td->zone_bytes += io_u->buflen;
827 return 0;
828}
829
830static void __io_u_mark_map(unsigned int *map, unsigned int nr)
831{
832 int idx = 0;
833
834 switch (nr) {
835 default:
836 idx = 6;
837 break;
838 case 33 ... 64:
839 idx = 5;
840 break;
841 case 17 ... 32:
842 idx = 4;
843 break;
844 case 9 ... 16:
845 idx = 3;
846 break;
847 case 5 ... 8:
848 idx = 2;
849 break;
850 case 1 ... 4:
851 idx = 1;
852 case 0:
853 break;
854 }
855
856 map[idx]++;
857}
858
859void io_u_mark_submit(struct thread_data *td, unsigned int nr)
860{
861 __io_u_mark_map(td->ts.io_u_submit, nr);
862 td->ts.total_submit++;
863}
864
865void io_u_mark_complete(struct thread_data *td, unsigned int nr)
866{
867 __io_u_mark_map(td->ts.io_u_complete, nr);
868 td->ts.total_complete++;
869}
870
871void io_u_mark_depth(struct thread_data *td, unsigned int nr)
872{
873 int idx = 0;
874
875 switch (td->cur_depth) {
876 default:
877 idx = 6;
878 break;
879 case 32 ... 63:
880 idx = 5;
881 break;
882 case 16 ... 31:
883 idx = 4;
884 break;
885 case 8 ... 15:
886 idx = 3;
887 break;
888 case 4 ... 7:
889 idx = 2;
890 break;
891 case 2 ... 3:
892 idx = 1;
893 case 1:
894 break;
895 }
896
897 td->ts.io_u_map[idx] += nr;
898}
899
900static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
901{
902 int idx = 0;
903
904 assert(usec < 1000);
905
906 switch (usec) {
907 case 750 ... 999:
908 idx = 9;
909 break;
910 case 500 ... 749:
911 idx = 8;
912 break;
913 case 250 ... 499:
914 idx = 7;
915 break;
916 case 100 ... 249:
917 idx = 6;
918 break;
919 case 50 ... 99:
920 idx = 5;
921 break;
922 case 20 ... 49:
923 idx = 4;
924 break;
925 case 10 ... 19:
926 idx = 3;
927 break;
928 case 4 ... 9:
929 idx = 2;
930 break;
931 case 2 ... 3:
932 idx = 1;
933 case 0 ... 1:
934 break;
935 }
936
937 assert(idx < FIO_IO_U_LAT_U_NR);
938 td->ts.io_u_lat_u[idx]++;
939}
940
941static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
942{
943 int idx = 0;
944
945 switch (msec) {
946 default:
947 idx = 11;
948 break;
949 case 1000 ... 1999:
950 idx = 10;
951 break;
952 case 750 ... 999:
953 idx = 9;
954 break;
955 case 500 ... 749:
956 idx = 8;
957 break;
958 case 250 ... 499:
959 idx = 7;
960 break;
961 case 100 ... 249:
962 idx = 6;
963 break;
964 case 50 ... 99:
965 idx = 5;
966 break;
967 case 20 ... 49:
968 idx = 4;
969 break;
970 case 10 ... 19:
971 idx = 3;
972 break;
973 case 4 ... 9:
974 idx = 2;
975 break;
976 case 2 ... 3:
977 idx = 1;
978 case 0 ... 1:
979 break;
980 }
981
982 assert(idx < FIO_IO_U_LAT_M_NR);
983 td->ts.io_u_lat_m[idx]++;
984}
985
986static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
987{
988 if (usec < 1000)
989 io_u_mark_lat_usec(td, usec);
990 else
991 io_u_mark_lat_msec(td, usec / 1000);
992}
993
994/*
995 * Get next file to service by choosing one at random
996 */
997static struct fio_file *get_next_file_rand(struct thread_data *td,
998 enum fio_file_flags goodf,
999 enum fio_file_flags badf)
1000{
1001 uint64_t frand_max = rand_max(&td->next_file_state);
1002 struct fio_file *f;
1003 int fno;
1004
1005 do {
1006 int opened = 0;
1007 unsigned long r;
1008
1009 r = __rand(&td->next_file_state);
1010 fno = (unsigned int) ((double) td->o.nr_files
1011 * (r / (frand_max + 1.0)));
1012
1013 f = td->files[fno];
1014 if (fio_file_done(f))
1015 continue;
1016
1017 if (!fio_file_open(f)) {
1018 int err;
1019
1020 if (td->nr_open_files >= td->o.open_files)
1021 return ERR_PTR(-EBUSY);
1022
1023 err = td_io_open_file(td, f);
1024 if (err)
1025 continue;
1026 opened = 1;
1027 }
1028
1029 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1030 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1031 return f;
1032 }
1033 if (opened)
1034 td_io_close_file(td, f);
1035 } while (1);
1036}
1037
1038/*
1039 * Get next file to service by doing round robin between all available ones
1040 */
1041static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1042 int badf)
1043{
1044 unsigned int old_next_file = td->next_file;
1045 struct fio_file *f;
1046
1047 do {
1048 int opened = 0;
1049
1050 f = td->files[td->next_file];
1051
1052 td->next_file++;
1053 if (td->next_file >= td->o.nr_files)
1054 td->next_file = 0;
1055
1056 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1057 if (fio_file_done(f)) {
1058 f = NULL;
1059 continue;
1060 }
1061
1062 if (!fio_file_open(f)) {
1063 int err;
1064
1065 if (td->nr_open_files >= td->o.open_files)
1066 return ERR_PTR(-EBUSY);
1067
1068 err = td_io_open_file(td, f);
1069 if (err) {
1070 dprint(FD_FILE, "error %d on open of %s\n",
1071 err, f->file_name);
1072 f = NULL;
1073 continue;
1074 }
1075 opened = 1;
1076 }
1077
1078 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1079 f->flags);
1080 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1081 break;
1082
1083 if (opened)
1084 td_io_close_file(td, f);
1085
1086 f = NULL;
1087 } while (td->next_file != old_next_file);
1088
1089 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1090 return f;
1091}
1092
1093static struct fio_file *__get_next_file(struct thread_data *td)
1094{
1095 struct fio_file *f;
1096
1097 assert(td->o.nr_files <= td->files_index);
1098
1099 if (td->nr_done_files >= td->o.nr_files) {
1100 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1101 " nr_files=%d\n", td->nr_open_files,
1102 td->nr_done_files,
1103 td->o.nr_files);
1104 return NULL;
1105 }
1106
1107 f = td->file_service_file;
1108 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1109 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1110 goto out;
1111 if (td->file_service_left--)
1112 goto out;
1113 }
1114
1115 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1116 td->o.file_service_type == FIO_FSERVICE_SEQ)
1117 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1118 else
1119 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1120
1121 if (IS_ERR(f))
1122 return f;
1123
1124 td->file_service_file = f;
1125 td->file_service_left = td->file_service_nr - 1;
1126out:
1127 if (f)
1128 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1129 else
1130 dprint(FD_FILE, "get_next_file: NULL\n");
1131 return f;
1132}
1133
1134static struct fio_file *get_next_file(struct thread_data *td)
1135{
1136 if (td->flags & TD_F_PROFILE_OPS) {
1137 struct prof_io_ops *ops = &td->prof_io_ops;
1138
1139 if (ops->get_next_file)
1140 return ops->get_next_file(td);
1141 }
1142
1143 return __get_next_file(td);
1144}
1145
1146static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1147{
1148 struct fio_file *f;
1149
1150 do {
1151 f = get_next_file(td);
1152 if (IS_ERR_OR_NULL(f))
1153 return PTR_ERR(f);
1154
1155 io_u->file = f;
1156 get_file(f);
1157
1158 if (!fill_io_u(td, io_u))
1159 break;
1160
1161 put_file_log(td, f);
1162 td_io_close_file(td, f);
1163 io_u->file = NULL;
1164 fio_file_set_done(f);
1165 td->nr_done_files++;
1166 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1167 td->nr_done_files, td->o.nr_files);
1168 } while (1);
1169
1170 return 0;
1171}
1172
1173static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1174 unsigned long tusec, unsigned long max_usec)
1175{
1176 if (!td->error)
1177 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1178 td_verror(td, ETIMEDOUT, "max latency exceeded");
1179 icd->error = ETIMEDOUT;
1180}
1181
1182static void lat_new_cycle(struct thread_data *td)
1183{
1184 fio_gettime(&td->latency_ts, NULL);
1185 td->latency_ios = ddir_rw_sum(td->io_blocks);
1186 td->latency_failed = 0;
1187}
1188
1189/*
1190 * We had an IO outside the latency target. Reduce the queue depth. If we
1191 * are at QD=1, then it's time to give up.
1192 */
1193static int __lat_target_failed(struct thread_data *td)
1194{
1195 if (td->latency_qd == 1)
1196 return 1;
1197
1198 td->latency_qd_high = td->latency_qd;
1199
1200 if (td->latency_qd == td->latency_qd_low)
1201 td->latency_qd_low--;
1202
1203 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1204
1205 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1206
1207 /*
1208 * When we ramp QD down, quiesce existing IO to prevent
1209 * a storm of ramp downs due to pending higher depth.
1210 */
1211 io_u_quiesce(td);
1212 lat_new_cycle(td);
1213 return 0;
1214}
1215
1216static int lat_target_failed(struct thread_data *td)
1217{
1218 if (td->o.latency_percentile.u.f == 100.0)
1219 return __lat_target_failed(td);
1220
1221 td->latency_failed++;
1222 return 0;
1223}
1224
1225void lat_target_init(struct thread_data *td)
1226{
1227 td->latency_end_run = 0;
1228
1229 if (td->o.latency_target) {
1230 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1231 fio_gettime(&td->latency_ts, NULL);
1232 td->latency_qd = 1;
1233 td->latency_qd_high = td->o.iodepth;
1234 td->latency_qd_low = 1;
1235 td->latency_ios = ddir_rw_sum(td->io_blocks);
1236 } else
1237 td->latency_qd = td->o.iodepth;
1238}
1239
1240void lat_target_reset(struct thread_data *td)
1241{
1242 if (!td->latency_end_run)
1243 lat_target_init(td);
1244}
1245
1246static void lat_target_success(struct thread_data *td)
1247{
1248 const unsigned int qd = td->latency_qd;
1249 struct thread_options *o = &td->o;
1250
1251 td->latency_qd_low = td->latency_qd;
1252
1253 /*
1254 * If we haven't failed yet, we double up to a failing value instead
1255 * of bisecting from highest possible queue depth. If we have set
1256 * a limit other than td->o.iodepth, bisect between that.
1257 */
1258 if (td->latency_qd_high != o->iodepth)
1259 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1260 else
1261 td->latency_qd *= 2;
1262
1263 if (td->latency_qd > o->iodepth)
1264 td->latency_qd = o->iodepth;
1265
1266 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1267
1268 /*
1269 * Same as last one, we are done. Let it run a latency cycle, so
1270 * we get only the results from the targeted depth.
1271 */
1272 if (td->latency_qd == qd) {
1273 if (td->latency_end_run) {
1274 dprint(FD_RATE, "We are done\n");
1275 td->done = 1;
1276 } else {
1277 dprint(FD_RATE, "Quiesce and final run\n");
1278 io_u_quiesce(td);
1279 td->latency_end_run = 1;
1280 reset_all_stats(td);
1281 reset_io_stats(td);
1282 }
1283 }
1284
1285 lat_new_cycle(td);
1286}
1287
1288/*
1289 * Check if we can bump the queue depth
1290 */
1291void lat_target_check(struct thread_data *td)
1292{
1293 uint64_t usec_window;
1294 uint64_t ios;
1295 double success_ios;
1296
1297 usec_window = utime_since_now(&td->latency_ts);
1298 if (usec_window < td->o.latency_window)
1299 return;
1300
1301 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1302 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1303 success_ios *= 100.0;
1304
1305 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1306
1307 if (success_ios >= td->o.latency_percentile.u.f)
1308 lat_target_success(td);
1309 else
1310 __lat_target_failed(td);
1311}
1312
1313/*
1314 * If latency target is enabled, we might be ramping up or down and not
1315 * using the full queue depth available.
1316 */
1317int queue_full(const struct thread_data *td)
1318{
1319 const int qempty = io_u_qempty(&td->io_u_freelist);
1320
1321 if (qempty)
1322 return 1;
1323 if (!td->o.latency_target)
1324 return 0;
1325
1326 return td->cur_depth >= td->latency_qd;
1327}
1328
1329struct io_u *__get_io_u(struct thread_data *td)
1330{
1331 struct io_u *io_u = NULL;
1332
1333 if (td->stop_io)
1334 return NULL;
1335
1336 td_io_u_lock(td);
1337
1338again:
1339 if (!io_u_rempty(&td->io_u_requeues))
1340 io_u = io_u_rpop(&td->io_u_requeues);
1341 else if (!queue_full(td)) {
1342 io_u = io_u_qpop(&td->io_u_freelist);
1343
1344 io_u->file = NULL;
1345 io_u->buflen = 0;
1346 io_u->resid = 0;
1347 io_u->end_io = NULL;
1348 }
1349
1350 if (io_u) {
1351 assert(io_u->flags & IO_U_F_FREE);
1352 io_u_clear(io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1353 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1354 IO_U_F_VER_LIST);
1355
1356 io_u->error = 0;
1357 io_u->acct_ddir = -1;
1358 td->cur_depth++;
1359 assert(!(td->flags & TD_F_CHILD));
1360 io_u_set(io_u, IO_U_F_IN_CUR_DEPTH);
1361 io_u->ipo = NULL;
1362 } else if (td_async_processing(td)) {
1363 /*
1364 * We ran out, wait for async verify threads to finish and
1365 * return one
1366 */
1367 assert(!(td->flags & TD_F_CHILD));
1368 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1369 goto again;
1370 }
1371
1372 td_io_u_unlock(td);
1373 return io_u;
1374}
1375
1376static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1377{
1378 if (!(td->flags & TD_F_TRIM_BACKLOG))
1379 return 0;
1380
1381 if (td->trim_entries) {
1382 int get_trim = 0;
1383
1384 if (td->trim_batch) {
1385 td->trim_batch--;
1386 get_trim = 1;
1387 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1388 td->last_ddir != DDIR_READ) {
1389 td->trim_batch = td->o.trim_batch;
1390 if (!td->trim_batch)
1391 td->trim_batch = td->o.trim_backlog;
1392 get_trim = 1;
1393 }
1394
1395 if (get_trim && !get_next_trim(td, io_u))
1396 return 1;
1397 }
1398
1399 return 0;
1400}
1401
1402static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1403{
1404 if (!(td->flags & TD_F_VER_BACKLOG))
1405 return 0;
1406
1407 if (td->io_hist_len) {
1408 int get_verify = 0;
1409
1410 if (td->verify_batch)
1411 get_verify = 1;
1412 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1413 td->last_ddir != DDIR_READ) {
1414 td->verify_batch = td->o.verify_batch;
1415 if (!td->verify_batch)
1416 td->verify_batch = td->o.verify_backlog;
1417 get_verify = 1;
1418 }
1419
1420 if (get_verify && !get_next_verify(td, io_u)) {
1421 td->verify_batch--;
1422 return 1;
1423 }
1424 }
1425
1426 return 0;
1427}
1428
1429/*
1430 * Fill offset and start time into the buffer content, to prevent too
1431 * easy compressible data for simple de-dupe attempts. Do this for every
1432 * 512b block in the range, since that should be the smallest block size
1433 * we can expect from a device.
1434 */
1435static void small_content_scramble(struct io_u *io_u)
1436{
1437 unsigned int i, nr_blocks = io_u->buflen / 512;
1438 uint64_t boffset;
1439 unsigned int offset;
1440 void *p, *end;
1441
1442 if (!nr_blocks)
1443 return;
1444
1445 p = io_u->xfer_buf;
1446 boffset = io_u->offset;
1447 io_u->buf_filled_len = 0;
1448
1449 for (i = 0; i < nr_blocks; i++) {
1450 /*
1451 * Fill the byte offset into a "random" start offset of
1452 * the buffer, given by the product of the usec time
1453 * and the actual offset.
1454 */
1455 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1456 offset &= ~(sizeof(uint64_t) - 1);
1457 if (offset >= 512 - sizeof(uint64_t))
1458 offset -= sizeof(uint64_t);
1459 memcpy(p + offset, &boffset, sizeof(boffset));
1460
1461 end = p + 512 - sizeof(io_u->start_time);
1462 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1463 p += 512;
1464 boffset += 512;
1465 }
1466}
1467
1468/*
1469 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1470 * etc. The returned io_u is fully ready to be prepped and submitted.
1471 */
1472struct io_u *get_io_u(struct thread_data *td)
1473{
1474 struct fio_file *f;
1475 struct io_u *io_u;
1476 int do_scramble = 0;
1477 long ret = 0;
1478
1479 io_u = __get_io_u(td);
1480 if (!io_u) {
1481 dprint(FD_IO, "__get_io_u failed\n");
1482 return NULL;
1483 }
1484
1485 if (check_get_verify(td, io_u))
1486 goto out;
1487 if (check_get_trim(td, io_u))
1488 goto out;
1489
1490 /*
1491 * from a requeue, io_u already setup
1492 */
1493 if (io_u->file)
1494 goto out;
1495
1496 /*
1497 * If using an iolog, grab next piece if any available.
1498 */
1499 if (td->flags & TD_F_READ_IOLOG) {
1500 if (read_iolog_get(td, io_u))
1501 goto err_put;
1502 } else if (set_io_u_file(td, io_u)) {
1503 ret = -EBUSY;
1504 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1505 goto err_put;
1506 }
1507
1508 f = io_u->file;
1509 if (!f) {
1510 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1511 goto err_put;
1512 }
1513
1514 assert(fio_file_open(f));
1515
1516 if (ddir_rw(io_u->ddir)) {
1517 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1518 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1519 goto err_put;
1520 }
1521
1522 f->last_start[io_u->ddir] = io_u->offset;
1523 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1524
1525 if (io_u->ddir == DDIR_WRITE) {
1526 if (td->flags & TD_F_REFILL_BUFFERS) {
1527 io_u_fill_buffer(td, io_u,
1528 td->o.min_bs[DDIR_WRITE],
1529 io_u->xfer_buflen);
1530 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1531 !(td->flags & TD_F_COMPRESS))
1532 do_scramble = 1;
1533 if (td->flags & TD_F_VER_NONE) {
1534 populate_verify_io_u(td, io_u);
1535 do_scramble = 0;
1536 }
1537 } else if (io_u->ddir == DDIR_READ) {
1538 /*
1539 * Reset the buf_filled parameters so next time if the
1540 * buffer is used for writes it is refilled.
1541 */
1542 io_u->buf_filled_len = 0;
1543 }
1544 }
1545
1546 /*
1547 * Set io data pointers.
1548 */
1549 io_u->xfer_buf = io_u->buf;
1550 io_u->xfer_buflen = io_u->buflen;
1551
1552out:
1553 assert(io_u->file);
1554 if (!td_io_prep(td, io_u)) {
1555 if (!td->o.disable_slat)
1556 fio_gettime(&io_u->start_time, NULL);
1557 if (do_scramble)
1558 small_content_scramble(io_u);
1559 return io_u;
1560 }
1561err_put:
1562 dprint(FD_IO, "get_io_u failed\n");
1563 put_io_u(td, io_u);
1564 return ERR_PTR(ret);
1565}
1566
1567static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1568{
1569 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1570
1571 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1572 return;
1573
1574 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1575 io_u->file ? " on file " : "",
1576 io_u->file ? io_u->file->file_name : "",
1577 strerror(io_u->error),
1578 io_ddir_name(io_u->ddir),
1579 io_u->offset, io_u->xfer_buflen);
1580
1581 if (td->io_ops->errdetails) {
1582 char *err = td->io_ops->errdetails(io_u);
1583
1584 log_err("fio: %s\n", err);
1585 free(err);
1586 }
1587
1588 if (!td->error)
1589 td_verror(td, io_u->error, "io_u error");
1590}
1591
1592void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1593{
1594 __io_u_log_error(td, io_u);
1595 if (td->parent)
1596 __io_u_log_error(td, io_u);
1597}
1598
1599static inline int gtod_reduce(struct thread_data *td)
1600{
1601 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1602 && td->o.disable_bw;
1603}
1604
1605static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1606 struct io_completion_data *icd,
1607 const enum fio_ddir idx, unsigned int bytes)
1608{
1609 const int no_reduce = !gtod_reduce(td);
1610 unsigned long lusec = 0;
1611
1612 if (td->parent)
1613 td = td->parent;
1614
1615 if (no_reduce)
1616 lusec = utime_since(&io_u->issue_time, &icd->time);
1617
1618 if (!td->o.disable_lat) {
1619 unsigned long tusec;
1620
1621 tusec = utime_since(&io_u->start_time, &icd->time);
1622 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1623
1624 if (td->flags & TD_F_PROFILE_OPS) {
1625 struct prof_io_ops *ops = &td->prof_io_ops;
1626
1627 if (ops->io_u_lat)
1628 icd->error = ops->io_u_lat(td, tusec);
1629 }
1630
1631 if (td->o.max_latency && tusec > td->o.max_latency)
1632 lat_fatal(td, icd, tusec, td->o.max_latency);
1633 if (td->o.latency_target && tusec > td->o.latency_target) {
1634 if (lat_target_failed(td))
1635 lat_fatal(td, icd, tusec, td->o.latency_target);
1636 }
1637 }
1638
1639 if (!td->o.disable_clat) {
1640 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1641 io_u_mark_latency(td, lusec);
1642 }
1643
1644 if (!td->o.disable_bw)
1645 add_bw_sample(td, idx, bytes, &icd->time);
1646
1647 if (no_reduce)
1648 add_iops_sample(td, idx, bytes, &icd->time);
1649
1650 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1651 uint32_t *info = io_u_block_info(td, io_u);
1652 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1653 if (io_u->ddir == DDIR_TRIM) {
1654 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1655 BLOCK_INFO_TRIMS(*info) + 1);
1656 } else if (io_u->ddir == DDIR_WRITE) {
1657 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1658 *info);
1659 }
1660 }
1661 }
1662}
1663
1664static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1665 struct io_completion_data *icd)
1666{
1667 struct io_u *io_u = *io_u_ptr;
1668 enum fio_ddir ddir = io_u->ddir;
1669 struct fio_file *f = io_u->file;
1670
1671 dprint_io_u(io_u, "io complete");
1672
1673 assert(io_u->flags & IO_U_F_FLIGHT);
1674 io_u_clear(io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1675
1676 /*
1677 * Mark IO ok to verify
1678 */
1679 if (io_u->ipo) {
1680 /*
1681 * Remove errored entry from the verification list
1682 */
1683 if (io_u->error)
1684 unlog_io_piece(td, io_u);
1685 else {
1686 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1687 write_barrier();
1688 }
1689 }
1690
1691 if (ddir_sync(ddir)) {
1692 td->last_was_sync = 1;
1693 if (f) {
1694 f->first_write = -1ULL;
1695 f->last_write = -1ULL;
1696 }
1697 return;
1698 }
1699
1700 td->last_was_sync = 0;
1701 td->last_ddir = ddir;
1702
1703 if (!io_u->error && ddir_rw(ddir)) {
1704 unsigned int bytes = io_u->buflen - io_u->resid;
1705 int ret;
1706
1707 td->io_blocks[ddir]++;
1708 td->this_io_blocks[ddir]++;
1709 td->io_bytes[ddir] += bytes;
1710
1711 if (!(io_u->flags & IO_U_F_VER_LIST))
1712 td->this_io_bytes[ddir] += bytes;
1713
1714 if (ddir == DDIR_WRITE) {
1715 if (f) {
1716 if (f->first_write == -1ULL ||
1717 io_u->offset < f->first_write)
1718 f->first_write = io_u->offset;
1719 if (f->last_write == -1ULL ||
1720 ((io_u->offset + bytes) > f->last_write))
1721 f->last_write = io_u->offset + bytes;
1722 }
1723 if (td->last_write_comp) {
1724 int idx = td->last_write_idx++;
1725
1726 td->last_write_comp[idx] = io_u->offset;
1727 if (td->last_write_idx == td->o.iodepth)
1728 td->last_write_idx = 0;
1729 }
1730 }
1731
1732 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1733 td->runstate == TD_VERIFYING))
1734 account_io_completion(td, io_u, icd, ddir, bytes);
1735
1736 icd->bytes_done[ddir] += bytes;
1737
1738 if (io_u->end_io) {
1739 ret = io_u->end_io(td, io_u_ptr);
1740 io_u = *io_u_ptr;
1741 if (ret && !icd->error)
1742 icd->error = ret;
1743 }
1744 } else if (io_u->error) {
1745 icd->error = io_u->error;
1746 io_u_log_error(td, io_u);
1747 }
1748 if (icd->error) {
1749 enum error_type_bit eb = td_error_type(ddir, icd->error);
1750
1751 if (!td_non_fatal_error(td, eb, icd->error))
1752 return;
1753
1754 /*
1755 * If there is a non_fatal error, then add to the error count
1756 * and clear all the errors.
1757 */
1758 update_error_count(td, icd->error);
1759 td_clear_error(td);
1760 icd->error = 0;
1761 if (io_u)
1762 io_u->error = 0;
1763 }
1764}
1765
1766static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1767 int nr)
1768{
1769 int ddir;
1770
1771 if (!gtod_reduce(td))
1772 fio_gettime(&icd->time, NULL);
1773
1774 icd->nr = nr;
1775
1776 icd->error = 0;
1777 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1778 icd->bytes_done[ddir] = 0;
1779}
1780
1781static void ios_completed(struct thread_data *td,
1782 struct io_completion_data *icd)
1783{
1784 struct io_u *io_u;
1785 int i;
1786
1787 for (i = 0; i < icd->nr; i++) {
1788 io_u = td->io_ops->event(td, i);
1789
1790 io_completed(td, &io_u, icd);
1791
1792 if (io_u)
1793 put_io_u(td, io_u);
1794 }
1795}
1796
1797/*
1798 * Complete a single io_u for the sync engines.
1799 */
1800int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1801{
1802 struct io_completion_data icd;
1803 int ddir;
1804
1805 init_icd(td, &icd, 1);
1806 io_completed(td, &io_u, &icd);
1807
1808 if (io_u)
1809 put_io_u(td, io_u);
1810
1811 if (icd.error) {
1812 td_verror(td, icd.error, "io_u_sync_complete");
1813 return -1;
1814 }
1815
1816 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1817 td->bytes_done[ddir] += icd.bytes_done[ddir];
1818
1819 return 0;
1820}
1821
1822/*
1823 * Called to complete min_events number of io for the async engines.
1824 */
1825int io_u_queued_complete(struct thread_data *td, int min_evts)
1826{
1827 struct io_completion_data icd;
1828 struct timespec *tvp = NULL;
1829 int ret, ddir;
1830 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1831
1832 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1833
1834 if (!min_evts)
1835 tvp = &ts;
1836 else if (min_evts > td->cur_depth)
1837 min_evts = td->cur_depth;
1838
1839 /* No worries, td_io_getevents fixes min and max if they are
1840 * set incorrectly */
1841 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
1842 if (ret < 0) {
1843 td_verror(td, -ret, "td_io_getevents");
1844 return ret;
1845 } else if (!ret)
1846 return ret;
1847
1848 init_icd(td, &icd, ret);
1849 ios_completed(td, &icd);
1850 if (icd.error) {
1851 td_verror(td, icd.error, "io_u_queued_complete");
1852 return -1;
1853 }
1854
1855 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1856 td->bytes_done[ddir] += icd.bytes_done[ddir];
1857
1858 return 0;
1859}
1860
1861/*
1862 * Call when io_u is really queued, to update the submission latency.
1863 */
1864void io_u_queued(struct thread_data *td, struct io_u *io_u)
1865{
1866 if (!td->o.disable_slat) {
1867 unsigned long slat_time;
1868
1869 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1870
1871 if (td->parent)
1872 td = td->parent;
1873
1874 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1875 io_u->offset);
1876 }
1877}
1878
1879/*
1880 * See if we should reuse the last seed, if dedupe is enabled
1881 */
1882static struct frand_state *get_buf_state(struct thread_data *td)
1883{
1884 uint64_t frand_max;
1885 unsigned int v;
1886 unsigned long r;
1887
1888 if (!td->o.dedupe_percentage)
1889 return &td->buf_state;
1890 else if (td->o.dedupe_percentage == 100) {
1891 frand_copy(&td->buf_state_prev, &td->buf_state);
1892 return &td->buf_state;
1893 }
1894
1895 frand_max = rand_max(&td->dedupe_state);
1896 r = __rand(&td->dedupe_state);
1897 v = 1 + (int) (100.0 * (r / (frand_max + 1.0)));
1898
1899 if (v <= td->o.dedupe_percentage)
1900 return &td->buf_state_prev;
1901
1902 return &td->buf_state;
1903}
1904
1905static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1906{
1907 if (td->o.dedupe_percentage == 100)
1908 frand_copy(rs, &td->buf_state_prev);
1909 else if (rs == &td->buf_state)
1910 frand_copy(&td->buf_state_prev, rs);
1911}
1912
1913void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1914 unsigned int max_bs)
1915{
1916 struct thread_options *o = &td->o;
1917
1918 if (o->compress_percentage || o->dedupe_percentage) {
1919 unsigned int perc = td->o.compress_percentage;
1920 struct frand_state *rs;
1921 unsigned int left = max_bs;
1922
1923 do {
1924 rs = get_buf_state(td);
1925
1926 min_write = min(min_write, left);
1927
1928 if (perc) {
1929 unsigned int seg = min_write;
1930
1931 seg = min(min_write, td->o.compress_chunk);
1932 if (!seg)
1933 seg = min_write;
1934
1935 fill_random_buf_percentage(rs, buf, perc, seg,
1936 min_write, o->buffer_pattern,
1937 o->buffer_pattern_bytes);
1938 } else
1939 fill_random_buf(rs, buf, min_write);
1940
1941 buf += min_write;
1942 left -= min_write;
1943 save_buf_state(td, rs);
1944 } while (left);
1945 } else if (o->buffer_pattern_bytes)
1946 fill_buffer_pattern(td, buf, max_bs);
1947 else if (o->zero_buffers)
1948 memset(buf, 0, max_bs);
1949 else
1950 fill_random_buf(get_buf_state(td), buf, max_bs);
1951}
1952
1953/*
1954 * "randomly" fill the buffer contents
1955 */
1956void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1957 unsigned int min_write, unsigned int max_bs)
1958{
1959 io_u->buf_filled_len = 0;
1960 fill_io_buffer(td, io_u->buf, min_write, max_bs);
1961}
fio - Flexible IO Tester