projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
t/verify-state: one more printf type fix
[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#include "minmax.h"
17
18struct io_completion_data {
19 int nr; /* input */
20
21 int error; /* output */
22 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
23 struct timeval time; /* output */
24};
25
26/*
27 * The ->io_axmap contains a map of blocks we have or have not done io
28 * to yet. Used to make sure we cover the entire range in a fair fashion.
29 */
30static bool random_map_free(struct fio_file *f, const uint64_t block)
31{
32 return !axmap_isset(f->io_axmap, block);
33}
34
35/*
36 * Mark a given offset as used in the map.
37 */
38static void mark_random_map(struct thread_data *td, struct io_u *io_u)
39{
40 unsigned int min_bs = td->o.rw_min_bs;
41 struct fio_file *f = io_u->file;
42 unsigned int nr_blocks;
43 uint64_t block;
44
45 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
46 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
47
48 if (!(io_u->flags & IO_U_F_BUSY_OK))
49 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
50
51 if ((nr_blocks * min_bs) < io_u->buflen)
52 io_u->buflen = nr_blocks * min_bs;
53}
54
55static uint64_t last_block(struct thread_data *td, struct fio_file *f,
56 enum fio_ddir ddir)
57{
58 uint64_t max_blocks;
59 uint64_t max_size;
60
61 assert(ddir_rw(ddir));
62
63 /*
64 * Hmm, should we make sure that ->io_size <= ->real_file_size?
65 */
66 max_size = f->io_size;
67 if (max_size > f->real_file_size)
68 max_size = f->real_file_size;
69
70 if (td->o.zone_range)
71 max_size = td->o.zone_range;
72
73 if (td->o.min_bs[ddir] > td->o.ba[ddir])
74 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
75
76 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
77 if (!max_blocks)
78 return 0;
79
80 return max_blocks;
81}
82
83struct rand_off {
84 struct flist_head list;
85 uint64_t off;
86};
87
88static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
89 enum fio_ddir ddir, uint64_t *b)
90{
91 uint64_t r;
92
93 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
94 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
95 uint64_t frand_max, lastb;
96
97 lastb = last_block(td, f, ddir);
98 if (!lastb)
99 return 1;
100
101 frand_max = rand_max(&td->random_state);
102 r = __rand(&td->random_state);
103
104 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
105
106 *b = lastb * (r / ((uint64_t) frand_max + 1.0));
107 } else {
108 uint64_t off = 0;
109
110 assert(fio_file_lfsr(f));
111
112 if (lfsr_next(&f->lfsr, &off))
113 return 1;
114
115 *b = off;
116 }
117
118 /*
119 * if we are not maintaining a random map, we are done.
120 */
121 if (!file_randommap(td, f))
122 goto ret;
123
124 /*
125 * calculate map offset and check if it's free
126 */
127 if (random_map_free(f, *b))
128 goto ret;
129
130 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
131 (unsigned long long) *b);
132
133 *b = axmap_next_free(f->io_axmap, *b);
134 if (*b == (uint64_t) -1ULL)
135 return 1;
136ret:
137 return 0;
138}
139
140static int __get_next_rand_offset_zipf(struct thread_data *td,
141 struct fio_file *f, enum fio_ddir ddir,
142 uint64_t *b)
143{
144 *b = zipf_next(&f->zipf);
145 return 0;
146}
147
148static int __get_next_rand_offset_pareto(struct thread_data *td,
149 struct fio_file *f, enum fio_ddir ddir,
150 uint64_t *b)
151{
152 *b = pareto_next(&f->zipf);
153 return 0;
154}
155
156static int __get_next_rand_offset_gauss(struct thread_data *td,
157 struct fio_file *f, enum fio_ddir ddir,
158 uint64_t *b)
159{
160 *b = gauss_next(&f->gauss);
161 return 0;
162}
163
164
165static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
166{
167 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
168 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
169
170 return r1->off - r2->off;
171}
172
173static int get_off_from_method(struct thread_data *td, struct fio_file *f,
174 enum fio_ddir ddir, uint64_t *b)
175{
176 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
177 return __get_next_rand_offset(td, f, ddir, b);
178 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
179 return __get_next_rand_offset_zipf(td, f, ddir, b);
180 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
181 return __get_next_rand_offset_pareto(td, f, ddir, b);
182 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
183 return __get_next_rand_offset_gauss(td, f, ddir, b);
184
185 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
186 return 1;
187}
188
189/*
190 * Sort the reads for a verify phase in batches of verifysort_nr, if
191 * specified.
192 */
193static inline bool should_sort_io(struct thread_data *td)
194{
195 if (!td->o.verifysort_nr || !td->o.do_verify)
196 return false;
197 if (!td_random(td))
198 return false;
199 if (td->runstate != TD_VERIFYING)
200 return false;
201 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
202 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
203 return false;
204
205 return true;
206}
207
208static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
209{
210 uint64_t frand_max;
211 unsigned int v;
212 unsigned long r;
213
214 if (td->o.perc_rand[ddir] == 100)
215 return true;
216
217 frand_max = rand_max(&td->seq_rand_state[ddir]);
218 r = __rand(&td->seq_rand_state[ddir]);
219 v = 1 + (int) (100.0 * (r / (frand_max + 1.0)));
220
221 return v <= td->o.perc_rand[ddir];
222}
223
224static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
225 enum fio_ddir ddir, uint64_t *b)
226{
227 struct rand_off *r;
228 int i, ret = 1;
229
230 if (!should_sort_io(td))
231 return get_off_from_method(td, f, ddir, b);
232
233 if (!flist_empty(&td->next_rand_list)) {
234fetch:
235 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
236 flist_del(&r->list);
237 *b = r->off;
238 free(r);
239 return 0;
240 }
241
242 for (i = 0; i < td->o.verifysort_nr; i++) {
243 r = malloc(sizeof(*r));
244
245 ret = get_off_from_method(td, f, ddir, &r->off);
246 if (ret) {
247 free(r);
248 break;
249 }
250
251 flist_add(&r->list, &td->next_rand_list);
252 }
253
254 if (ret && !i)
255 return ret;
256
257 assert(!flist_empty(&td->next_rand_list));
258 flist_sort(NULL, &td->next_rand_list, flist_cmp);
259 goto fetch;
260}
261
262static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
263 enum fio_ddir ddir, uint64_t *b)
264{
265 if (!get_next_rand_offset(td, f, ddir, b))
266 return 0;
267
268 if (td->o.time_based) {
269 fio_file_reset(td, f);
270 if (!get_next_rand_offset(td, f, ddir, b))
271 return 0;
272 }
273
274 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
275 f->file_name, (unsigned long long) f->last_pos[ddir],
276 (unsigned long long) f->real_file_size);
277 return 1;
278}
279
280static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
281 enum fio_ddir ddir, uint64_t *offset)
282{
283 struct thread_options *o = &td->o;
284
285 assert(ddir_rw(ddir));
286
287 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
288 o->time_based)
289 f->last_pos[ddir] = f->last_pos[ddir] - f->io_size;
290
291 if (f->last_pos[ddir] < f->real_file_size) {
292 uint64_t pos;
293
294 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0)
295 f->last_pos[ddir] = f->real_file_size;
296
297 pos = f->last_pos[ddir] - f->file_offset;
298 if (pos && o->ddir_seq_add) {
299 pos += o->ddir_seq_add;
300
301 /*
302 * If we reach beyond the end of the file
303 * with holed IO, wrap around to the
304 * beginning again.
305 */
306 if (pos >= f->real_file_size)
307 pos = f->file_offset;
308 }
309
310 *offset = pos;
311 return 0;
312 }
313
314 return 1;
315}
316
317static int get_next_block(struct thread_data *td, struct io_u *io_u,
318 enum fio_ddir ddir, int rw_seq,
319 unsigned int *is_random)
320{
321 struct fio_file *f = io_u->file;
322 uint64_t b, offset;
323 int ret;
324
325 assert(ddir_rw(ddir));
326
327 b = offset = -1ULL;
328
329 if (rw_seq) {
330 if (td_random(td)) {
331 if (should_do_random(td, ddir)) {
332 ret = get_next_rand_block(td, f, ddir, &b);
333 *is_random = 1;
334 } else {
335 *is_random = 0;
336 io_u_set(io_u, IO_U_F_BUSY_OK);
337 ret = get_next_seq_offset(td, f, ddir, &offset);
338 if (ret)
339 ret = get_next_rand_block(td, f, ddir, &b);
340 }
341 } else {
342 *is_random = 0;
343 ret = get_next_seq_offset(td, f, ddir, &offset);
344 }
345 } else {
346 io_u_set(io_u, IO_U_F_BUSY_OK);
347 *is_random = 0;
348
349 if (td->o.rw_seq == RW_SEQ_SEQ) {
350 ret = get_next_seq_offset(td, f, ddir, &offset);
351 if (ret) {
352 ret = get_next_rand_block(td, f, ddir, &b);
353 *is_random = 0;
354 }
355 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
356 if (f->last_start[ddir] != -1ULL)
357 offset = f->last_start[ddir] - f->file_offset;
358 else
359 offset = 0;
360 ret = 0;
361 } else {
362 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
363 ret = 1;
364 }
365 }
366
367 if (!ret) {
368 if (offset != -1ULL)
369 io_u->offset = offset;
370 else if (b != -1ULL)
371 io_u->offset = b * td->o.ba[ddir];
372 else {
373 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
374 ret = 1;
375 }
376 }
377
378 return ret;
379}
380
381/*
382 * For random io, generate a random new block and see if it's used. Repeat
383 * until we find a free one. For sequential io, just return the end of
384 * the last io issued.
385 */
386static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
387 unsigned int *is_random)
388{
389 struct fio_file *f = io_u->file;
390 enum fio_ddir ddir = io_u->ddir;
391 int rw_seq_hit = 0;
392
393 assert(ddir_rw(ddir));
394
395 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
396 rw_seq_hit = 1;
397 td->ddir_seq_nr = td->o.ddir_seq_nr;
398 }
399
400 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
401 return 1;
402
403 if (io_u->offset >= f->io_size) {
404 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
405 (unsigned long long) io_u->offset,
406 (unsigned long long) f->io_size);
407 return 1;
408 }
409
410 io_u->offset += f->file_offset;
411 if (io_u->offset >= f->real_file_size) {
412 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
413 (unsigned long long) io_u->offset,
414 (unsigned long long) f->real_file_size);
415 return 1;
416 }
417
418 return 0;
419}
420
421static int get_next_offset(struct thread_data *td, struct io_u *io_u,
422 unsigned int *is_random)
423{
424 if (td->flags & TD_F_PROFILE_OPS) {
425 struct prof_io_ops *ops = &td->prof_io_ops;
426
427 if (ops->fill_io_u_off)
428 return ops->fill_io_u_off(td, io_u, is_random);
429 }
430
431 return __get_next_offset(td, io_u, is_random);
432}
433
434static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
435 unsigned int buflen)
436{
437 struct fio_file *f = io_u->file;
438
439 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
440}
441
442static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
443 unsigned int is_random)
444{
445 int ddir = io_u->ddir;
446 unsigned int buflen = 0;
447 unsigned int minbs, maxbs;
448 uint64_t frand_max;
449 unsigned long r;
450
451 assert(ddir_rw(ddir));
452
453 if (td->o.bs_is_seq_rand)
454 ddir = is_random ? DDIR_WRITE: DDIR_READ;
455
456 minbs = td->o.min_bs[ddir];
457 maxbs = td->o.max_bs[ddir];
458
459 if (minbs == maxbs)
460 return minbs;
461
462 /*
463 * If we can't satisfy the min block size from here, then fail
464 */
465 if (!io_u_fits(td, io_u, minbs))
466 return 0;
467
468 frand_max = rand_max(&td->bsrange_state);
469 do {
470 r = __rand(&td->bsrange_state);
471
472 if (!td->o.bssplit_nr[ddir]) {
473 buflen = 1 + (unsigned int) ((double) maxbs *
474 (r / (frand_max + 1.0)));
475 if (buflen < minbs)
476 buflen = minbs;
477 } else {
478 long perc = 0;
479 unsigned int i;
480
481 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
482 struct bssplit *bsp = &td->o.bssplit[ddir][i];
483
484 buflen = bsp->bs;
485 perc += bsp->perc;
486 if ((r <= ((frand_max / 100L) * perc)) &&
487 io_u_fits(td, io_u, buflen))
488 break;
489 }
490 }
491
492 if (td->o.verify != VERIFY_NONE)
493 buflen = (buflen + td->o.verify_interval - 1) &
494 ~(td->o.verify_interval - 1);
495
496 if (!td->o.bs_unaligned && is_power_of_2(minbs))
497 buflen &= ~(minbs - 1);
498
499 } while (!io_u_fits(td, io_u, buflen));
500
501 return buflen;
502}
503
504static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
505 unsigned int is_random)
506{
507 if (td->flags & TD_F_PROFILE_OPS) {
508 struct prof_io_ops *ops = &td->prof_io_ops;
509
510 if (ops->fill_io_u_size)
511 return ops->fill_io_u_size(td, io_u, is_random);
512 }
513
514 return __get_next_buflen(td, io_u, is_random);
515}
516
517static void set_rwmix_bytes(struct thread_data *td)
518{
519 unsigned int diff;
520
521 /*
522 * we do time or byte based switch. this is needed because
523 * buffered writes may issue a lot quicker than they complete,
524 * whereas reads do not.
525 */
526 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
527 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
528}
529
530static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
531{
532 uint64_t frand_max = rand_max(&td->rwmix_state);
533 unsigned int v;
534 unsigned long r;
535
536 r = __rand(&td->rwmix_state);
537 v = 1 + (int) (100.0 * (r / (frand_max + 1.0)));
538
539 if (v <= td->o.rwmix[DDIR_READ])
540 return DDIR_READ;
541
542 return DDIR_WRITE;
543}
544
545int io_u_quiesce(struct thread_data *td)
546{
547 int completed = 0;
548
549 /*
550 * We are going to sleep, ensure that we flush anything pending as
551 * not to skew our latency numbers.
552 *
553 * Changed to only monitor 'in flight' requests here instead of the
554 * td->cur_depth, b/c td->cur_depth does not accurately represent
555 * io's that have been actually submitted to an async engine,
556 * and cur_depth is meaningless for sync engines.
557 */
558 if (td->io_u_queued || td->cur_depth) {
559 int fio_unused ret;
560
561 ret = td_io_commit(td);
562 }
563
564 while (td->io_u_in_flight) {
565 int fio_unused ret;
566
567 ret = io_u_queued_complete(td, 1);
568 if (ret > 0)
569 completed += ret;
570 }
571
572 return completed;
573}
574
575static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
576{
577 enum fio_ddir odir = ddir ^ 1;
578 long usec, now;
579
580 assert(ddir_rw(ddir));
581 now = utime_since_now(&td->start);
582
583 /*
584 * if rate_next_io_time is in the past, need to catch up to rate
585 */
586 if (td->rate_next_io_time[ddir] <= now)
587 return ddir;
588
589 /*
590 * We are ahead of rate in this direction. See if we
591 * should switch.
592 */
593 if (td_rw(td) && td->o.rwmix[odir]) {
594 /*
595 * Other direction is behind rate, switch
596 */
597 if (td->rate_next_io_time[odir] <= now)
598 return odir;
599
600 /*
601 * Both directions are ahead of rate. sleep the min
602 * switch if necissary
603 */
604 if (td->rate_next_io_time[ddir] <=
605 td->rate_next_io_time[odir]) {
606 usec = td->rate_next_io_time[ddir] - now;
607 } else {
608 usec = td->rate_next_io_time[odir] - now;
609 ddir = odir;
610 }
611 } else
612 usec = td->rate_next_io_time[ddir] - now;
613
614 if (td->o.io_submit_mode == IO_MODE_INLINE)
615 io_u_quiesce(td);
616
617 usec = usec_sleep(td, usec);
618
619 return ddir;
620}
621
622/*
623 * Return the data direction for the next io_u. If the job is a
624 * mixed read/write workload, check the rwmix cycle and switch if
625 * necessary.
626 */
627static enum fio_ddir get_rw_ddir(struct thread_data *td)
628{
629 enum fio_ddir ddir;
630
631 /*
632 * see if it's time to fsync
633 */
634 if (td->o.fsync_blocks &&
635 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
636 td->io_issues[DDIR_WRITE] && should_fsync(td))
637 return DDIR_SYNC;
638
639 /*
640 * see if it's time to fdatasync
641 */
642 if (td->o.fdatasync_blocks &&
643 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
644 td->io_issues[DDIR_WRITE] && should_fsync(td))
645 return DDIR_DATASYNC;
646
647 /*
648 * see if it's time to sync_file_range
649 */
650 if (td->sync_file_range_nr &&
651 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
652 td->io_issues[DDIR_WRITE] && should_fsync(td))
653 return DDIR_SYNC_FILE_RANGE;
654
655 if (td_rw(td)) {
656 /*
657 * Check if it's time to seed a new data direction.
658 */
659 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
660 /*
661 * Put a top limit on how many bytes we do for
662 * one data direction, to avoid overflowing the
663 * ranges too much
664 */
665 ddir = get_rand_ddir(td);
666
667 if (ddir != td->rwmix_ddir)
668 set_rwmix_bytes(td);
669
670 td->rwmix_ddir = ddir;
671 }
672 ddir = td->rwmix_ddir;
673 } else if (td_read(td))
674 ddir = DDIR_READ;
675 else if (td_write(td))
676 ddir = DDIR_WRITE;
677 else
678 ddir = DDIR_TRIM;
679
680 td->rwmix_ddir = rate_ddir(td, ddir);
681 return td->rwmix_ddir;
682}
683
684static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
685{
686 enum fio_ddir ddir = get_rw_ddir(td);
687
688 if (td_trimwrite(td)) {
689 struct fio_file *f = io_u->file;
690 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
691 ddir = DDIR_TRIM;
692 else
693 ddir = DDIR_WRITE;
694 }
695
696 io_u->ddir = io_u->acct_ddir = ddir;
697
698 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
699 td->o.barrier_blocks &&
700 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
701 td->io_issues[DDIR_WRITE])
702 io_u_set(io_u, IO_U_F_BARRIER);
703}
704
705void put_file_log(struct thread_data *td, struct fio_file *f)
706{
707 unsigned int ret = put_file(td, f);
708
709 if (ret)
710 td_verror(td, ret, "file close");
711}
712
713void put_io_u(struct thread_data *td, struct io_u *io_u)
714{
715 if (td->parent)
716 td = td->parent;
717
718 td_io_u_lock(td);
719
720 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
721 put_file_log(td, io_u->file);
722
723 io_u->file = NULL;
724 io_u_set(io_u, IO_U_F_FREE);
725
726 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
727 td->cur_depth--;
728 assert(!(td->flags & TD_F_CHILD));
729 }
730 io_u_qpush(&td->io_u_freelist, io_u);
731 td_io_u_unlock(td);
732 td_io_u_free_notify(td);
733}
734
735void clear_io_u(struct thread_data *td, struct io_u *io_u)
736{
737 io_u_clear(io_u, IO_U_F_FLIGHT);
738 put_io_u(td, io_u);
739}
740
741void requeue_io_u(struct thread_data *td, struct io_u **io_u)
742{
743 struct io_u *__io_u = *io_u;
744 enum fio_ddir ddir = acct_ddir(__io_u);
745
746 dprint(FD_IO, "requeue %p\n", __io_u);
747
748 if (td->parent)
749 td = td->parent;
750
751 td_io_u_lock(td);
752
753 io_u_set(__io_u, IO_U_F_FREE);
754 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
755 td->io_issues[ddir]--;
756
757 io_u_clear(__io_u, IO_U_F_FLIGHT);
758 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
759 td->cur_depth--;
760 assert(!(td->flags & TD_F_CHILD));
761 }
762
763 io_u_rpush(&td->io_u_requeues, __io_u);
764 td_io_u_unlock(td);
765 td_io_u_free_notify(td);
766 *io_u = NULL;
767}
768
769static int fill_io_u(struct thread_data *td, struct io_u *io_u)
770{
771 unsigned int is_random;
772
773 if (td->io_ops->flags & FIO_NOIO)
774 goto out;
775
776 set_rw_ddir(td, io_u);
777
778 /*
779 * fsync() or fdatasync() or trim etc, we are done
780 */
781 if (!ddir_rw(io_u->ddir))
782 goto out;
783
784 /*
785 * See if it's time to switch to a new zone
786 */
787 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
788 struct fio_file *f = io_u->file;
789
790 td->zone_bytes = 0;
791 f->file_offset += td->o.zone_range + td->o.zone_skip;
792
793 /*
794 * Wrap from the beginning, if we exceed the file size
795 */
796 if (f->file_offset >= f->real_file_size)
797 f->file_offset = f->real_file_size - f->file_offset;
798 f->last_pos[io_u->ddir] = f->file_offset;
799 td->io_skip_bytes += td->o.zone_skip;
800 }
801
802 /*
803 * No log, let the seq/rand engine retrieve the next buflen and
804 * position.
805 */
806 if (get_next_offset(td, io_u, &is_random)) {
807 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
808 return 1;
809 }
810
811 io_u->buflen = get_next_buflen(td, io_u, is_random);
812 if (!io_u->buflen) {
813 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
814 return 1;
815 }
816
817 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
818 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
819 dprint(FD_IO, " off=%llu/%lu > %llu\n",
820 (unsigned long long) io_u->offset, io_u->buflen,
821 (unsigned long long) io_u->file->real_file_size);
822 return 1;
823 }
824
825 /*
826 * mark entry before potentially trimming io_u
827 */
828 if (td_random(td) && file_randommap(td, io_u->file))
829 mark_random_map(td, io_u);
830
831out:
832 dprint_io_u(io_u, "fill_io_u");
833 td->zone_bytes += io_u->buflen;
834 return 0;
835}
836
837static void __io_u_mark_map(unsigned int *map, unsigned int nr)
838{
839 int idx = 0;
840
841 switch (nr) {
842 default:
843 idx = 6;
844 break;
845 case 33 ... 64:
846 idx = 5;
847 break;
848 case 17 ... 32:
849 idx = 4;
850 break;
851 case 9 ... 16:
852 idx = 3;
853 break;
854 case 5 ... 8:
855 idx = 2;
856 break;
857 case 1 ... 4:
858 idx = 1;
859 case 0:
860 break;
861 }
862
863 map[idx]++;
864}
865
866void io_u_mark_submit(struct thread_data *td, unsigned int nr)
867{
868 __io_u_mark_map(td->ts.io_u_submit, nr);
869 td->ts.total_submit++;
870}
871
872void io_u_mark_complete(struct thread_data *td, unsigned int nr)
873{
874 __io_u_mark_map(td->ts.io_u_complete, nr);
875 td->ts.total_complete++;
876}
877
878void io_u_mark_depth(struct thread_data *td, unsigned int nr)
879{
880 int idx = 0;
881
882 switch (td->cur_depth) {
883 default:
884 idx = 6;
885 break;
886 case 32 ... 63:
887 idx = 5;
888 break;
889 case 16 ... 31:
890 idx = 4;
891 break;
892 case 8 ... 15:
893 idx = 3;
894 break;
895 case 4 ... 7:
896 idx = 2;
897 break;
898 case 2 ... 3:
899 idx = 1;
900 case 1:
901 break;
902 }
903
904 td->ts.io_u_map[idx] += nr;
905}
906
907static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
908{
909 int idx = 0;
910
911 assert(usec < 1000);
912
913 switch (usec) {
914 case 750 ... 999:
915 idx = 9;
916 break;
917 case 500 ... 749:
918 idx = 8;
919 break;
920 case 250 ... 499:
921 idx = 7;
922 break;
923 case 100 ... 249:
924 idx = 6;
925 break;
926 case 50 ... 99:
927 idx = 5;
928 break;
929 case 20 ... 49:
930 idx = 4;
931 break;
932 case 10 ... 19:
933 idx = 3;
934 break;
935 case 4 ... 9:
936 idx = 2;
937 break;
938 case 2 ... 3:
939 idx = 1;
940 case 0 ... 1:
941 break;
942 }
943
944 assert(idx < FIO_IO_U_LAT_U_NR);
945 td->ts.io_u_lat_u[idx]++;
946}
947
948static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
949{
950 int idx = 0;
951
952 switch (msec) {
953 default:
954 idx = 11;
955 break;
956 case 1000 ... 1999:
957 idx = 10;
958 break;
959 case 750 ... 999:
960 idx = 9;
961 break;
962 case 500 ... 749:
963 idx = 8;
964 break;
965 case 250 ... 499:
966 idx = 7;
967 break;
968 case 100 ... 249:
969 idx = 6;
970 break;
971 case 50 ... 99:
972 idx = 5;
973 break;
974 case 20 ... 49:
975 idx = 4;
976 break;
977 case 10 ... 19:
978 idx = 3;
979 break;
980 case 4 ... 9:
981 idx = 2;
982 break;
983 case 2 ... 3:
984 idx = 1;
985 case 0 ... 1:
986 break;
987 }
988
989 assert(idx < FIO_IO_U_LAT_M_NR);
990 td->ts.io_u_lat_m[idx]++;
991}
992
993static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
994{
995 if (usec < 1000)
996 io_u_mark_lat_usec(td, usec);
997 else
998 io_u_mark_lat_msec(td, usec / 1000);
999}
1000
1001/*
1002 * Get next file to service by choosing one at random
1003 */
1004static struct fio_file *get_next_file_rand(struct thread_data *td,
1005 enum fio_file_flags goodf,
1006 enum fio_file_flags badf)
1007{
1008 uint64_t frand_max = rand_max(&td->next_file_state);
1009 struct fio_file *f;
1010 int fno;
1011
1012 do {
1013 int opened = 0;
1014 unsigned long r;
1015
1016 r = __rand(&td->next_file_state);
1017 fno = (unsigned int) ((double) td->o.nr_files
1018 * (r / (frand_max + 1.0)));
1019
1020 f = td->files[fno];
1021 if (fio_file_done(f))
1022 continue;
1023
1024 if (!fio_file_open(f)) {
1025 int err;
1026
1027 if (td->nr_open_files >= td->o.open_files)
1028 return ERR_PTR(-EBUSY);
1029
1030 err = td_io_open_file(td, f);
1031 if (err)
1032 continue;
1033 opened = 1;
1034 }
1035
1036 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1037 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1038 return f;
1039 }
1040 if (opened)
1041 td_io_close_file(td, f);
1042 } while (1);
1043}
1044
1045/*
1046 * Get next file to service by doing round robin between all available ones
1047 */
1048static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1049 int badf)
1050{
1051 unsigned int old_next_file = td->next_file;
1052 struct fio_file *f;
1053
1054 do {
1055 int opened = 0;
1056
1057 f = td->files[td->next_file];
1058
1059 td->next_file++;
1060 if (td->next_file >= td->o.nr_files)
1061 td->next_file = 0;
1062
1063 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1064 if (fio_file_done(f)) {
1065 f = NULL;
1066 continue;
1067 }
1068
1069 if (!fio_file_open(f)) {
1070 int err;
1071
1072 if (td->nr_open_files >= td->o.open_files)
1073 return ERR_PTR(-EBUSY);
1074
1075 err = td_io_open_file(td, f);
1076 if (err) {
1077 dprint(FD_FILE, "error %d on open of %s\n",
1078 err, f->file_name);
1079 f = NULL;
1080 continue;
1081 }
1082 opened = 1;
1083 }
1084
1085 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1086 f->flags);
1087 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1088 break;
1089
1090 if (opened)
1091 td_io_close_file(td, f);
1092
1093 f = NULL;
1094 } while (td->next_file != old_next_file);
1095
1096 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1097 return f;
1098}
1099
1100static struct fio_file *__get_next_file(struct thread_data *td)
1101{
1102 struct fio_file *f;
1103
1104 assert(td->o.nr_files <= td->files_index);
1105
1106 if (td->nr_done_files >= td->o.nr_files) {
1107 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1108 " nr_files=%d\n", td->nr_open_files,
1109 td->nr_done_files,
1110 td->o.nr_files);
1111 return NULL;
1112 }
1113
1114 f = td->file_service_file;
1115 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1116 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1117 goto out;
1118 if (td->file_service_left--)
1119 goto out;
1120 }
1121
1122 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1123 td->o.file_service_type == FIO_FSERVICE_SEQ)
1124 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1125 else
1126 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1127
1128 if (IS_ERR(f))
1129 return f;
1130
1131 td->file_service_file = f;
1132 td->file_service_left = td->file_service_nr - 1;
1133out:
1134 if (f)
1135 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1136 else
1137 dprint(FD_FILE, "get_next_file: NULL\n");
1138 return f;
1139}
1140
1141static struct fio_file *get_next_file(struct thread_data *td)
1142{
1143 if (td->flags & TD_F_PROFILE_OPS) {
1144 struct prof_io_ops *ops = &td->prof_io_ops;
1145
1146 if (ops->get_next_file)
1147 return ops->get_next_file(td);
1148 }
1149
1150 return __get_next_file(td);
1151}
1152
1153static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1154{
1155 struct fio_file *f;
1156
1157 do {
1158 f = get_next_file(td);
1159 if (IS_ERR_OR_NULL(f))
1160 return PTR_ERR(f);
1161
1162 io_u->file = f;
1163 get_file(f);
1164
1165 if (!fill_io_u(td, io_u))
1166 break;
1167
1168 put_file_log(td, f);
1169 td_io_close_file(td, f);
1170 io_u->file = NULL;
1171 fio_file_set_done(f);
1172 td->nr_done_files++;
1173 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1174 td->nr_done_files, td->o.nr_files);
1175 } while (1);
1176
1177 return 0;
1178}
1179
1180static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1181 unsigned long tusec, unsigned long max_usec)
1182{
1183 if (!td->error)
1184 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1185 td_verror(td, ETIMEDOUT, "max latency exceeded");
1186 icd->error = ETIMEDOUT;
1187}
1188
1189static void lat_new_cycle(struct thread_data *td)
1190{
1191 fio_gettime(&td->latency_ts, NULL);
1192 td->latency_ios = ddir_rw_sum(td->io_blocks);
1193 td->latency_failed = 0;
1194}
1195
1196/*
1197 * We had an IO outside the latency target. Reduce the queue depth. If we
1198 * are at QD=1, then it's time to give up.
1199 */
1200static bool __lat_target_failed(struct thread_data *td)
1201{
1202 if (td->latency_qd == 1)
1203 return true;
1204
1205 td->latency_qd_high = td->latency_qd;
1206
1207 if (td->latency_qd == td->latency_qd_low)
1208 td->latency_qd_low--;
1209
1210 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1211
1212 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1213
1214 /*
1215 * When we ramp QD down, quiesce existing IO to prevent
1216 * a storm of ramp downs due to pending higher depth.
1217 */
1218 io_u_quiesce(td);
1219 lat_new_cycle(td);
1220 return false;
1221}
1222
1223static bool lat_target_failed(struct thread_data *td)
1224{
1225 if (td->o.latency_percentile.u.f == 100.0)
1226 return __lat_target_failed(td);
1227
1228 td->latency_failed++;
1229 return false;
1230}
1231
1232void lat_target_init(struct thread_data *td)
1233{
1234 td->latency_end_run = 0;
1235
1236 if (td->o.latency_target) {
1237 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1238 fio_gettime(&td->latency_ts, NULL);
1239 td->latency_qd = 1;
1240 td->latency_qd_high = td->o.iodepth;
1241 td->latency_qd_low = 1;
1242 td->latency_ios = ddir_rw_sum(td->io_blocks);
1243 } else
1244 td->latency_qd = td->o.iodepth;
1245}
1246
1247void lat_target_reset(struct thread_data *td)
1248{
1249 if (!td->latency_end_run)
1250 lat_target_init(td);
1251}
1252
1253static void lat_target_success(struct thread_data *td)
1254{
1255 const unsigned int qd = td->latency_qd;
1256 struct thread_options *o = &td->o;
1257
1258 td->latency_qd_low = td->latency_qd;
1259
1260 /*
1261 * If we haven't failed yet, we double up to a failing value instead
1262 * of bisecting from highest possible queue depth. If we have set
1263 * a limit other than td->o.iodepth, bisect between that.
1264 */
1265 if (td->latency_qd_high != o->iodepth)
1266 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1267 else
1268 td->latency_qd *= 2;
1269
1270 if (td->latency_qd > o->iodepth)
1271 td->latency_qd = o->iodepth;
1272
1273 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1274
1275 /*
1276 * Same as last one, we are done. Let it run a latency cycle, so
1277 * we get only the results from the targeted depth.
1278 */
1279 if (td->latency_qd == qd) {
1280 if (td->latency_end_run) {
1281 dprint(FD_RATE, "We are done\n");
1282 td->done = 1;
1283 } else {
1284 dprint(FD_RATE, "Quiesce and final run\n");
1285 io_u_quiesce(td);
1286 td->latency_end_run = 1;
1287 reset_all_stats(td);
1288 reset_io_stats(td);
1289 }
1290 }
1291
1292 lat_new_cycle(td);
1293}
1294
1295/*
1296 * Check if we can bump the queue depth
1297 */
1298void lat_target_check(struct thread_data *td)
1299{
1300 uint64_t usec_window;
1301 uint64_t ios;
1302 double success_ios;
1303
1304 usec_window = utime_since_now(&td->latency_ts);
1305 if (usec_window < td->o.latency_window)
1306 return;
1307
1308 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1309 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1310 success_ios *= 100.0;
1311
1312 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1313
1314 if (success_ios >= td->o.latency_percentile.u.f)
1315 lat_target_success(td);
1316 else
1317 __lat_target_failed(td);
1318}
1319
1320/*
1321 * If latency target is enabled, we might be ramping up or down and not
1322 * using the full queue depth available.
1323 */
1324bool queue_full(const struct thread_data *td)
1325{
1326 const int qempty = io_u_qempty(&td->io_u_freelist);
1327
1328 if (qempty)
1329 return true;
1330 if (!td->o.latency_target)
1331 return false;
1332
1333 return td->cur_depth >= td->latency_qd;
1334}
1335
1336struct io_u *__get_io_u(struct thread_data *td)
1337{
1338 struct io_u *io_u = NULL;
1339
1340 if (td->stop_io)
1341 return NULL;
1342
1343 td_io_u_lock(td);
1344
1345again:
1346 if (!io_u_rempty(&td->io_u_requeues))
1347 io_u = io_u_rpop(&td->io_u_requeues);
1348 else if (!queue_full(td)) {
1349 io_u = io_u_qpop(&td->io_u_freelist);
1350
1351 io_u->file = NULL;
1352 io_u->buflen = 0;
1353 io_u->resid = 0;
1354 io_u->end_io = NULL;
1355 }
1356
1357 if (io_u) {
1358 assert(io_u->flags & IO_U_F_FREE);
1359 io_u_clear(io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1360 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1361 IO_U_F_VER_LIST);
1362
1363 io_u->error = 0;
1364 io_u->acct_ddir = -1;
1365 td->cur_depth++;
1366 assert(!(td->flags & TD_F_CHILD));
1367 io_u_set(io_u, IO_U_F_IN_CUR_DEPTH);
1368 io_u->ipo = NULL;
1369 } else if (td_async_processing(td)) {
1370 /*
1371 * We ran out, wait for async verify threads to finish and
1372 * return one
1373 */
1374 assert(!(td->flags & TD_F_CHILD));
1375 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1376 goto again;
1377 }
1378
1379 td_io_u_unlock(td);
1380 return io_u;
1381}
1382
1383static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1384{
1385 if (!(td->flags & TD_F_TRIM_BACKLOG))
1386 return false;
1387
1388 if (td->trim_entries) {
1389 int get_trim = 0;
1390
1391 if (td->trim_batch) {
1392 td->trim_batch--;
1393 get_trim = 1;
1394 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1395 td->last_ddir != DDIR_READ) {
1396 td->trim_batch = td->o.trim_batch;
1397 if (!td->trim_batch)
1398 td->trim_batch = td->o.trim_backlog;
1399 get_trim = 1;
1400 }
1401
1402 if (get_trim && !get_next_trim(td, io_u))
1403 return true;
1404 }
1405
1406 return false;
1407}
1408
1409static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1410{
1411 if (!(td->flags & TD_F_VER_BACKLOG))
1412 return false;
1413
1414 if (td->io_hist_len) {
1415 int get_verify = 0;
1416
1417 if (td->verify_batch)
1418 get_verify = 1;
1419 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1420 td->last_ddir != DDIR_READ) {
1421 td->verify_batch = td->o.verify_batch;
1422 if (!td->verify_batch)
1423 td->verify_batch = td->o.verify_backlog;
1424 get_verify = 1;
1425 }
1426
1427 if (get_verify && !get_next_verify(td, io_u)) {
1428 td->verify_batch--;
1429 return true;
1430 }
1431 }
1432
1433 return false;
1434}
1435
1436/*
1437 * Fill offset and start time into the buffer content, to prevent too
1438 * easy compressible data for simple de-dupe attempts. Do this for every
1439 * 512b block in the range, since that should be the smallest block size
1440 * we can expect from a device.
1441 */
1442static void small_content_scramble(struct io_u *io_u)
1443{
1444 unsigned int i, nr_blocks = io_u->buflen / 512;
1445 uint64_t boffset;
1446 unsigned int offset;
1447 void *p, *end;
1448
1449 if (!nr_blocks)
1450 return;
1451
1452 p = io_u->xfer_buf;
1453 boffset = io_u->offset;
1454 io_u->buf_filled_len = 0;
1455
1456 for (i = 0; i < nr_blocks; i++) {
1457 /*
1458 * Fill the byte offset into a "random" start offset of
1459 * the buffer, given by the product of the usec time
1460 * and the actual offset.
1461 */
1462 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1463 offset &= ~(sizeof(uint64_t) - 1);
1464 if (offset >= 512 - sizeof(uint64_t))
1465 offset -= sizeof(uint64_t);
1466 memcpy(p + offset, &boffset, sizeof(boffset));
1467
1468 end = p + 512 - sizeof(io_u->start_time);
1469 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1470 p += 512;
1471 boffset += 512;
1472 }
1473}
1474
1475/*
1476 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1477 * etc. The returned io_u is fully ready to be prepped and submitted.
1478 */
1479struct io_u *get_io_u(struct thread_data *td)
1480{
1481 struct fio_file *f;
1482 struct io_u *io_u;
1483 int do_scramble = 0;
1484 long ret = 0;
1485
1486 io_u = __get_io_u(td);
1487 if (!io_u) {
1488 dprint(FD_IO, "__get_io_u failed\n");
1489 return NULL;
1490 }
1491
1492 if (check_get_verify(td, io_u))
1493 goto out;
1494 if (check_get_trim(td, io_u))
1495 goto out;
1496
1497 /*
1498 * from a requeue, io_u already setup
1499 */
1500 if (io_u->file)
1501 goto out;
1502
1503 /*
1504 * If using an iolog, grab next piece if any available.
1505 */
1506 if (td->flags & TD_F_READ_IOLOG) {
1507 if (read_iolog_get(td, io_u))
1508 goto err_put;
1509 } else if (set_io_u_file(td, io_u)) {
1510 ret = -EBUSY;
1511 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1512 goto err_put;
1513 }
1514
1515 f = io_u->file;
1516 if (!f) {
1517 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1518 goto err_put;
1519 }
1520
1521 assert(fio_file_open(f));
1522
1523 if (ddir_rw(io_u->ddir)) {
1524 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1525 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1526 goto err_put;
1527 }
1528
1529 f->last_start[io_u->ddir] = io_u->offset;
1530 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1531
1532 if (io_u->ddir == DDIR_WRITE) {
1533 if (td->flags & TD_F_REFILL_BUFFERS) {
1534 io_u_fill_buffer(td, io_u,
1535 td->o.min_bs[DDIR_WRITE],
1536 io_u->buflen);
1537 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1538 !(td->flags & TD_F_COMPRESS))
1539 do_scramble = 1;
1540 if (td->flags & TD_F_VER_NONE) {
1541 populate_verify_io_u(td, io_u);
1542 do_scramble = 0;
1543 }
1544 } else if (io_u->ddir == DDIR_READ) {
1545 /*
1546 * Reset the buf_filled parameters so next time if the
1547 * buffer is used for writes it is refilled.
1548 */
1549 io_u->buf_filled_len = 0;
1550 }
1551 }
1552
1553 /*
1554 * Set io data pointers.
1555 */
1556 io_u->xfer_buf = io_u->buf;
1557 io_u->xfer_buflen = io_u->buflen;
1558
1559out:
1560 assert(io_u->file);
1561 if (!td_io_prep(td, io_u)) {
1562 if (!td->o.disable_lat)
1563 fio_gettime(&io_u->start_time, NULL);
1564 if (do_scramble)
1565 small_content_scramble(io_u);
1566 return io_u;
1567 }
1568err_put:
1569 dprint(FD_IO, "get_io_u failed\n");
1570 put_io_u(td, io_u);
1571 return ERR_PTR(ret);
1572}
1573
1574static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1575{
1576 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1577
1578 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1579 return;
1580
1581 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1582 io_u->file ? " on file " : "",
1583 io_u->file ? io_u->file->file_name : "",
1584 strerror(io_u->error),
1585 io_ddir_name(io_u->ddir),
1586 io_u->offset, io_u->xfer_buflen);
1587
1588 if (td->io_ops->errdetails) {
1589 char *err = td->io_ops->errdetails(io_u);
1590
1591 log_err("fio: %s\n", err);
1592 free(err);
1593 }
1594
1595 if (!td->error)
1596 td_verror(td, io_u->error, "io_u error");
1597}
1598
1599void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1600{
1601 __io_u_log_error(td, io_u);
1602 if (td->parent)
1603 __io_u_log_error(td, io_u);
1604}
1605
1606static inline bool gtod_reduce(struct thread_data *td)
1607{
1608 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1609 || td->o.gtod_reduce;
1610}
1611
1612static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1613 struct io_completion_data *icd,
1614 const enum fio_ddir idx, unsigned int bytes)
1615{
1616 const int no_reduce = !gtod_reduce(td);
1617 unsigned long lusec = 0;
1618
1619 if (td->parent)
1620 td = td->parent;
1621
1622 if (no_reduce)
1623 lusec = utime_since(&io_u->issue_time, &icd->time);
1624
1625 if (!td->o.disable_lat) {
1626 unsigned long tusec;
1627
1628 tusec = utime_since(&io_u->start_time, &icd->time);
1629 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1630
1631 if (td->flags & TD_F_PROFILE_OPS) {
1632 struct prof_io_ops *ops = &td->prof_io_ops;
1633
1634 if (ops->io_u_lat)
1635 icd->error = ops->io_u_lat(td, tusec);
1636 }
1637
1638 if (td->o.max_latency && tusec > td->o.max_latency)
1639 lat_fatal(td, icd, tusec, td->o.max_latency);
1640 if (td->o.latency_target && tusec > td->o.latency_target) {
1641 if (lat_target_failed(td))
1642 lat_fatal(td, icd, tusec, td->o.latency_target);
1643 }
1644 }
1645
1646 if (!td->o.disable_clat) {
1647 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1648 io_u_mark_latency(td, lusec);
1649 }
1650
1651 if (!td->o.disable_bw)
1652 add_bw_sample(td, idx, bytes, &icd->time);
1653
1654 if (no_reduce)
1655 add_iops_sample(td, idx, bytes, &icd->time);
1656
1657 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1658 uint32_t *info = io_u_block_info(td, io_u);
1659 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1660 if (io_u->ddir == DDIR_TRIM) {
1661 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1662 BLOCK_INFO_TRIMS(*info) + 1);
1663 } else if (io_u->ddir == DDIR_WRITE) {
1664 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1665 *info);
1666 }
1667 }
1668 }
1669}
1670
1671static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1672 struct io_completion_data *icd)
1673{
1674 struct io_u *io_u = *io_u_ptr;
1675 enum fio_ddir ddir = io_u->ddir;
1676 struct fio_file *f = io_u->file;
1677
1678 dprint_io_u(io_u, "io complete");
1679
1680 assert(io_u->flags & IO_U_F_FLIGHT);
1681 io_u_clear(io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1682
1683 /*
1684 * Mark IO ok to verify
1685 */
1686 if (io_u->ipo) {
1687 /*
1688 * Remove errored entry from the verification list
1689 */
1690 if (io_u->error)
1691 unlog_io_piece(td, io_u);
1692 else {
1693 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1694 write_barrier();
1695 }
1696 }
1697
1698 if (ddir_sync(ddir)) {
1699 td->last_was_sync = 1;
1700 if (f) {
1701 f->first_write = -1ULL;
1702 f->last_write = -1ULL;
1703 }
1704 return;
1705 }
1706
1707 td->last_was_sync = 0;
1708 td->last_ddir = ddir;
1709
1710 if (!io_u->error && ddir_rw(ddir)) {
1711 unsigned int bytes = io_u->buflen - io_u->resid;
1712 int ret;
1713
1714 td->io_blocks[ddir]++;
1715 td->this_io_blocks[ddir]++;
1716 td->io_bytes[ddir] += bytes;
1717
1718 if (!(io_u->flags & IO_U_F_VER_LIST))
1719 td->this_io_bytes[ddir] += bytes;
1720
1721 if (ddir == DDIR_WRITE) {
1722 if (f) {
1723 if (f->first_write == -1ULL ||
1724 io_u->offset < f->first_write)
1725 f->first_write = io_u->offset;
1726 if (f->last_write == -1ULL ||
1727 ((io_u->offset + bytes) > f->last_write))
1728 f->last_write = io_u->offset + bytes;
1729 }
1730 if (td->last_write_comp) {
1731 int idx = td->last_write_idx++;
1732
1733 td->last_write_comp[idx] = io_u->offset;
1734 if (td->last_write_idx == td->o.iodepth)
1735 td->last_write_idx = 0;
1736 }
1737 }
1738
1739 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1740 td->runstate == TD_VERIFYING))
1741 account_io_completion(td, io_u, icd, ddir, bytes);
1742
1743 icd->bytes_done[ddir] += bytes;
1744
1745 if (io_u->end_io) {
1746 ret = io_u->end_io(td, io_u_ptr);
1747 io_u = *io_u_ptr;
1748 if (ret && !icd->error)
1749 icd->error = ret;
1750 }
1751 } else if (io_u->error) {
1752 icd->error = io_u->error;
1753 io_u_log_error(td, io_u);
1754 }
1755 if (icd->error) {
1756 enum error_type_bit eb = td_error_type(ddir, icd->error);
1757
1758 if (!td_non_fatal_error(td, eb, icd->error))
1759 return;
1760
1761 /*
1762 * If there is a non_fatal error, then add to the error count
1763 * and clear all the errors.
1764 */
1765 update_error_count(td, icd->error);
1766 td_clear_error(td);
1767 icd->error = 0;
1768 if (io_u)
1769 io_u->error = 0;
1770 }
1771}
1772
1773static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1774 int nr)
1775{
1776 int ddir;
1777
1778 if (!gtod_reduce(td))
1779 fio_gettime(&icd->time, NULL);
1780
1781 icd->nr = nr;
1782
1783 icd->error = 0;
1784 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1785 icd->bytes_done[ddir] = 0;
1786}
1787
1788static void ios_completed(struct thread_data *td,
1789 struct io_completion_data *icd)
1790{
1791 struct io_u *io_u;
1792 int i;
1793
1794 for (i = 0; i < icd->nr; i++) {
1795 io_u = td->io_ops->event(td, i);
1796
1797 io_completed(td, &io_u, icd);
1798
1799 if (io_u)
1800 put_io_u(td, io_u);
1801 }
1802}
1803
1804/*
1805 * Complete a single io_u for the sync engines.
1806 */
1807int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1808{
1809 struct io_completion_data icd;
1810 int ddir;
1811
1812 init_icd(td, &icd, 1);
1813 io_completed(td, &io_u, &icd);
1814
1815 if (io_u)
1816 put_io_u(td, io_u);
1817
1818 if (icd.error) {
1819 td_verror(td, icd.error, "io_u_sync_complete");
1820 return -1;
1821 }
1822
1823 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1824 td->bytes_done[ddir] += icd.bytes_done[ddir];
1825
1826 return 0;
1827}
1828
1829/*
1830 * Called to complete min_events number of io for the async engines.
1831 */
1832int io_u_queued_complete(struct thread_data *td, int min_evts)
1833{
1834 struct io_completion_data icd;
1835 struct timespec *tvp = NULL;
1836 int ret, ddir;
1837 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1838
1839 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1840
1841 if (!min_evts)
1842 tvp = &ts;
1843 else if (min_evts > td->cur_depth)
1844 min_evts = td->cur_depth;
1845
1846 /* No worries, td_io_getevents fixes min and max if they are
1847 * set incorrectly */
1848 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
1849 if (ret < 0) {
1850 td_verror(td, -ret, "td_io_getevents");
1851 return ret;
1852 } else if (!ret)
1853 return ret;
1854
1855 init_icd(td, &icd, ret);
1856 ios_completed(td, &icd);
1857 if (icd.error) {
1858 td_verror(td, icd.error, "io_u_queued_complete");
1859 return -1;
1860 }
1861
1862 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1863 td->bytes_done[ddir] += icd.bytes_done[ddir];
1864
1865 return ret;
1866}
1867
1868/*
1869 * Call when io_u is really queued, to update the submission latency.
1870 */
1871void io_u_queued(struct thread_data *td, struct io_u *io_u)
1872{
1873 if (!td->o.disable_slat) {
1874 unsigned long slat_time;
1875
1876 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1877
1878 if (td->parent)
1879 td = td->parent;
1880
1881 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1882 io_u->offset);
1883 }
1884}
1885
1886/*
1887 * See if we should reuse the last seed, if dedupe is enabled
1888 */
1889static struct frand_state *get_buf_state(struct thread_data *td)
1890{
1891 uint64_t frand_max;
1892 unsigned int v;
1893 unsigned long r;
1894
1895 if (!td->o.dedupe_percentage)
1896 return &td->buf_state;
1897 else if (td->o.dedupe_percentage == 100) {
1898 frand_copy(&td->buf_state_prev, &td->buf_state);
1899 return &td->buf_state;
1900 }
1901
1902 frand_max = rand_max(&td->dedupe_state);
1903 r = __rand(&td->dedupe_state);
1904 v = 1 + (int) (100.0 * (r / (frand_max + 1.0)));
1905
1906 if (v <= td->o.dedupe_percentage)
1907 return &td->buf_state_prev;
1908
1909 return &td->buf_state;
1910}
1911
1912static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1913{
1914 if (td->o.dedupe_percentage == 100)
1915 frand_copy(rs, &td->buf_state_prev);
1916 else if (rs == &td->buf_state)
1917 frand_copy(&td->buf_state_prev, rs);
1918}
1919
1920void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1921 unsigned int max_bs)
1922{
1923 struct thread_options *o = &td->o;
1924
1925 if (o->compress_percentage || o->dedupe_percentage) {
1926 unsigned int perc = td->o.compress_percentage;
1927 struct frand_state *rs;
1928 unsigned int left = max_bs;
1929 unsigned int this_write;
1930
1931 do {
1932 rs = get_buf_state(td);
1933
1934 min_write = min(min_write, left);
1935
1936 if (perc) {
1937 this_write = min_not_zero(min_write,
1938 td->o.compress_chunk);
1939
1940 fill_random_buf_percentage(rs, buf, perc,
1941 this_write, this_write,
1942 o->buffer_pattern,
1943 o->buffer_pattern_bytes);
1944 } else {
1945 fill_random_buf(rs, buf, min_write);
1946 this_write = min_write;
1947 }
1948
1949 buf += this_write;
1950 left -= this_write;
1951 save_buf_state(td, rs);
1952 } while (left);
1953 } else if (o->buffer_pattern_bytes)
1954 fill_buffer_pattern(td, buf, max_bs);
1955 else if (o->zero_buffers)
1956 memset(buf, 0, max_bs);
1957 else
1958 fill_random_buf(get_buf_state(td), buf, max_bs);
1959}
1960
1961/*
1962 * "randomly" fill the buffer contents
1963 */
1964void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1965 unsigned int min_write, unsigned int max_bs)
1966{
1967 io_u->buf_filled_len = 0;
1968 fill_io_buffer(td, io_u->buf, min_write, max_bs);
1969}
fio - Flexible IO Tester