projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
io_ddir: io_ddir_name array should be static
[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 struct thread_options *o = &td->o;
290 uint64_t io_size = f->io_size + (f->io_size % o->min_bs[ddir]);
291
292 if (io_size > f->last_pos[ddir])
293 f->last_pos[ddir] = 0;
294 else
295 f->last_pos[ddir] = f->last_pos[ddir] - io_size;
296 }
297
298 if (f->last_pos[ddir] < f->real_file_size) {
299 uint64_t pos;
300
301 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0)
302 f->last_pos[ddir] = f->real_file_size;
303
304 pos = f->last_pos[ddir] - f->file_offset;
305 if (pos && o->ddir_seq_add) {
306 pos += o->ddir_seq_add;
307
308 /*
309 * If we reach beyond the end of the file
310 * with holed IO, wrap around to the
311 * beginning again.
312 */
313 if (pos >= f->real_file_size)
314 pos = f->file_offset;
315 }
316
317 *offset = pos;
318 return 0;
319 }
320
321 return 1;
322}
323
324static int get_next_block(struct thread_data *td, struct io_u *io_u,
325 enum fio_ddir ddir, int rw_seq,
326 unsigned int *is_random)
327{
328 struct fio_file *f = io_u->file;
329 uint64_t b, offset;
330 int ret;
331
332 assert(ddir_rw(ddir));
333
334 b = offset = -1ULL;
335
336 if (rw_seq) {
337 if (td_random(td)) {
338 if (should_do_random(td, ddir)) {
339 ret = get_next_rand_block(td, f, ddir, &b);
340 *is_random = 1;
341 } else {
342 *is_random = 0;
343 io_u_set(io_u, IO_U_F_BUSY_OK);
344 ret = get_next_seq_offset(td, f, ddir, &offset);
345 if (ret)
346 ret = get_next_rand_block(td, f, ddir, &b);
347 }
348 } else {
349 *is_random = 0;
350 ret = get_next_seq_offset(td, f, ddir, &offset);
351 }
352 } else {
353 io_u_set(io_u, IO_U_F_BUSY_OK);
354 *is_random = 0;
355
356 if (td->o.rw_seq == RW_SEQ_SEQ) {
357 ret = get_next_seq_offset(td, f, ddir, &offset);
358 if (ret) {
359 ret = get_next_rand_block(td, f, ddir, &b);
360 *is_random = 0;
361 }
362 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
363 if (f->last_start[ddir] != -1ULL)
364 offset = f->last_start[ddir] - f->file_offset;
365 else
366 offset = 0;
367 ret = 0;
368 } else {
369 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
370 ret = 1;
371 }
372 }
373
374 if (!ret) {
375 if (offset != -1ULL)
376 io_u->offset = offset;
377 else if (b != -1ULL)
378 io_u->offset = b * td->o.ba[ddir];
379 else {
380 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
381 ret = 1;
382 }
383 }
384
385 return ret;
386}
387
388/*
389 * For random io, generate a random new block and see if it's used. Repeat
390 * until we find a free one. For sequential io, just return the end of
391 * the last io issued.
392 */
393static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
394 unsigned int *is_random)
395{
396 struct fio_file *f = io_u->file;
397 enum fio_ddir ddir = io_u->ddir;
398 int rw_seq_hit = 0;
399
400 assert(ddir_rw(ddir));
401
402 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
403 rw_seq_hit = 1;
404 td->ddir_seq_nr = td->o.ddir_seq_nr;
405 }
406
407 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
408 return 1;
409
410 if (io_u->offset >= f->io_size) {
411 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
412 (unsigned long long) io_u->offset,
413 (unsigned long long) f->io_size);
414 return 1;
415 }
416
417 io_u->offset += f->file_offset;
418 if (io_u->offset >= f->real_file_size) {
419 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
420 (unsigned long long) io_u->offset,
421 (unsigned long long) f->real_file_size);
422 return 1;
423 }
424
425 return 0;
426}
427
428static int get_next_offset(struct thread_data *td, struct io_u *io_u,
429 unsigned int *is_random)
430{
431 if (td->flags & TD_F_PROFILE_OPS) {
432 struct prof_io_ops *ops = &td->prof_io_ops;
433
434 if (ops->fill_io_u_off)
435 return ops->fill_io_u_off(td, io_u, is_random);
436 }
437
438 return __get_next_offset(td, io_u, is_random);
439}
440
441static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
442 unsigned int buflen)
443{
444 struct fio_file *f = io_u->file;
445
446 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
447}
448
449static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
450 unsigned int is_random)
451{
452 int ddir = io_u->ddir;
453 unsigned int buflen = 0;
454 unsigned int minbs, maxbs;
455 uint64_t frand_max;
456 unsigned long r;
457
458 assert(ddir_rw(ddir));
459
460 if (td->o.bs_is_seq_rand)
461 ddir = is_random ? DDIR_WRITE: DDIR_READ;
462
463 minbs = td->o.min_bs[ddir];
464 maxbs = td->o.max_bs[ddir];
465
466 if (minbs == maxbs)
467 return minbs;
468
469 /*
470 * If we can't satisfy the min block size from here, then fail
471 */
472 if (!io_u_fits(td, io_u, minbs))
473 return 0;
474
475 frand_max = rand_max(&td->bsrange_state);
476 do {
477 r = __rand(&td->bsrange_state);
478
479 if (!td->o.bssplit_nr[ddir]) {
480 buflen = 1 + (unsigned int) ((double) maxbs *
481 (r / (frand_max + 1.0)));
482 if (buflen < minbs)
483 buflen = minbs;
484 } else {
485 long perc = 0;
486 unsigned int i;
487
488 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
489 struct bssplit *bsp = &td->o.bssplit[ddir][i];
490
491 buflen = bsp->bs;
492 perc += bsp->perc;
493 if ((r <= ((frand_max / 100L) * perc)) &&
494 io_u_fits(td, io_u, buflen))
495 break;
496 }
497 }
498
499 if (td->o.verify != VERIFY_NONE)
500 buflen = (buflen + td->o.verify_interval - 1) &
501 ~(td->o.verify_interval - 1);
502
503 if (!td->o.bs_unaligned && is_power_of_2(minbs))
504 buflen &= ~(minbs - 1);
505
506 } while (!io_u_fits(td, io_u, buflen));
507
508 return buflen;
509}
510
511static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
512 unsigned int is_random)
513{
514 if (td->flags & TD_F_PROFILE_OPS) {
515 struct prof_io_ops *ops = &td->prof_io_ops;
516
517 if (ops->fill_io_u_size)
518 return ops->fill_io_u_size(td, io_u, is_random);
519 }
520
521 return __get_next_buflen(td, io_u, is_random);
522}
523
524static void set_rwmix_bytes(struct thread_data *td)
525{
526 unsigned int diff;
527
528 /*
529 * we do time or byte based switch. this is needed because
530 * buffered writes may issue a lot quicker than they complete,
531 * whereas reads do not.
532 */
533 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
534 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
535}
536
537static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
538{
539 uint64_t frand_max = rand_max(&td->rwmix_state);
540 unsigned int v;
541 unsigned long r;
542
543 r = __rand(&td->rwmix_state);
544 v = 1 + (int) (100.0 * (r / (frand_max + 1.0)));
545
546 if (v <= td->o.rwmix[DDIR_READ])
547 return DDIR_READ;
548
549 return DDIR_WRITE;
550}
551
552int io_u_quiesce(struct thread_data *td)
553{
554 int completed = 0;
555
556 /*
557 * We are going to sleep, ensure that we flush anything pending as
558 * not to skew our latency numbers.
559 *
560 * Changed to only monitor 'in flight' requests here instead of the
561 * td->cur_depth, b/c td->cur_depth does not accurately represent
562 * io's that have been actually submitted to an async engine,
563 * and cur_depth is meaningless for sync engines.
564 */
565 if (td->io_u_queued || td->cur_depth) {
566 int fio_unused ret;
567
568 ret = td_io_commit(td);
569 }
570
571 while (td->io_u_in_flight) {
572 int fio_unused ret;
573
574 ret = io_u_queued_complete(td, 1);
575 if (ret > 0)
576 completed += ret;
577 }
578
579 return completed;
580}
581
582static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
583{
584 enum fio_ddir odir = ddir ^ 1;
585 long usec, now;
586
587 assert(ddir_rw(ddir));
588 now = utime_since_now(&td->start);
589
590 /*
591 * if rate_next_io_time is in the past, need to catch up to rate
592 */
593 if (td->rate_next_io_time[ddir] <= now)
594 return ddir;
595
596 /*
597 * We are ahead of rate in this direction. See if we
598 * should switch.
599 */
600 if (td_rw(td) && td->o.rwmix[odir]) {
601 /*
602 * Other direction is behind rate, switch
603 */
604 if (td->rate_next_io_time[odir] <= now)
605 return odir;
606
607 /*
608 * Both directions are ahead of rate. sleep the min
609 * switch if necissary
610 */
611 if (td->rate_next_io_time[ddir] <=
612 td->rate_next_io_time[odir]) {
613 usec = td->rate_next_io_time[ddir] - now;
614 } else {
615 usec = td->rate_next_io_time[odir] - now;
616 ddir = odir;
617 }
618 } else
619 usec = td->rate_next_io_time[ddir] - now;
620
621 if (td->o.io_submit_mode == IO_MODE_INLINE)
622 io_u_quiesce(td);
623
624 usec = usec_sleep(td, usec);
625
626 return ddir;
627}
628
629/*
630 * Return the data direction for the next io_u. If the job is a
631 * mixed read/write workload, check the rwmix cycle and switch if
632 * necessary.
633 */
634static enum fio_ddir get_rw_ddir(struct thread_data *td)
635{
636 enum fio_ddir ddir;
637
638 /*
639 * see if it's time to fsync
640 */
641 if (td->o.fsync_blocks &&
642 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
643 td->io_issues[DDIR_WRITE] && should_fsync(td))
644 return DDIR_SYNC;
645
646 /*
647 * see if it's time to fdatasync
648 */
649 if (td->o.fdatasync_blocks &&
650 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
651 td->io_issues[DDIR_WRITE] && should_fsync(td))
652 return DDIR_DATASYNC;
653
654 /*
655 * see if it's time to sync_file_range
656 */
657 if (td->sync_file_range_nr &&
658 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
659 td->io_issues[DDIR_WRITE] && should_fsync(td))
660 return DDIR_SYNC_FILE_RANGE;
661
662 if (td_rw(td)) {
663 /*
664 * Check if it's time to seed a new data direction.
665 */
666 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
667 /*
668 * Put a top limit on how many bytes we do for
669 * one data direction, to avoid overflowing the
670 * ranges too much
671 */
672 ddir = get_rand_ddir(td);
673
674 if (ddir != td->rwmix_ddir)
675 set_rwmix_bytes(td);
676
677 td->rwmix_ddir = ddir;
678 }
679 ddir = td->rwmix_ddir;
680 } else if (td_read(td))
681 ddir = DDIR_READ;
682 else if (td_write(td))
683 ddir = DDIR_WRITE;
684 else
685 ddir = DDIR_TRIM;
686
687 td->rwmix_ddir = rate_ddir(td, ddir);
688 return td->rwmix_ddir;
689}
690
691static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
692{
693 enum fio_ddir ddir = get_rw_ddir(td);
694
695 if (td_trimwrite(td)) {
696 struct fio_file *f = io_u->file;
697 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
698 ddir = DDIR_TRIM;
699 else
700 ddir = DDIR_WRITE;
701 }
702
703 io_u->ddir = io_u->acct_ddir = ddir;
704
705 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
706 td->o.barrier_blocks &&
707 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
708 td->io_issues[DDIR_WRITE])
709 io_u_set(io_u, IO_U_F_BARRIER);
710}
711
712void put_file_log(struct thread_data *td, struct fio_file *f)
713{
714 unsigned int ret = put_file(td, f);
715
716 if (ret)
717 td_verror(td, ret, "file close");
718}
719
720void put_io_u(struct thread_data *td, struct io_u *io_u)
721{
722 if (td->parent)
723 td = td->parent;
724
725 td_io_u_lock(td);
726
727 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
728 put_file_log(td, io_u->file);
729
730 io_u->file = NULL;
731 io_u_set(io_u, IO_U_F_FREE);
732
733 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
734 td->cur_depth--;
735 assert(!(td->flags & TD_F_CHILD));
736 }
737 io_u_qpush(&td->io_u_freelist, io_u);
738 td_io_u_unlock(td);
739 td_io_u_free_notify(td);
740}
741
742void clear_io_u(struct thread_data *td, struct io_u *io_u)
743{
744 io_u_clear(io_u, IO_U_F_FLIGHT);
745 put_io_u(td, io_u);
746}
747
748void requeue_io_u(struct thread_data *td, struct io_u **io_u)
749{
750 struct io_u *__io_u = *io_u;
751 enum fio_ddir ddir = acct_ddir(__io_u);
752
753 dprint(FD_IO, "requeue %p\n", __io_u);
754
755 if (td->parent)
756 td = td->parent;
757
758 td_io_u_lock(td);
759
760 io_u_set(__io_u, IO_U_F_FREE);
761 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
762 td->io_issues[ddir]--;
763
764 io_u_clear(__io_u, IO_U_F_FLIGHT);
765 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
766 td->cur_depth--;
767 assert(!(td->flags & TD_F_CHILD));
768 }
769
770 io_u_rpush(&td->io_u_requeues, __io_u);
771 td_io_u_unlock(td);
772 td_io_u_free_notify(td);
773 *io_u = NULL;
774}
775
776static int fill_io_u(struct thread_data *td, struct io_u *io_u)
777{
778 unsigned int is_random;
779
780 if (td->io_ops->flags & FIO_NOIO)
781 goto out;
782
783 set_rw_ddir(td, io_u);
784
785 /*
786 * fsync() or fdatasync() or trim etc, we are done
787 */
788 if (!ddir_rw(io_u->ddir))
789 goto out;
790
791 /*
792 * See if it's time to switch to a new zone
793 */
794 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
795 struct fio_file *f = io_u->file;
796
797 td->zone_bytes = 0;
798 f->file_offset += td->o.zone_range + td->o.zone_skip;
799
800 /*
801 * Wrap from the beginning, if we exceed the file size
802 */
803 if (f->file_offset >= f->real_file_size)
804 f->file_offset = f->real_file_size - f->file_offset;
805 f->last_pos[io_u->ddir] = f->file_offset;
806 td->io_skip_bytes += td->o.zone_skip;
807 }
808
809 /*
810 * No log, let the seq/rand engine retrieve the next buflen and
811 * position.
812 */
813 if (get_next_offset(td, io_u, &is_random)) {
814 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
815 return 1;
816 }
817
818 io_u->buflen = get_next_buflen(td, io_u, is_random);
819 if (!io_u->buflen) {
820 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
821 return 1;
822 }
823
824 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
825 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
826 dprint(FD_IO, " off=%llu/%lu > %llu\n",
827 (unsigned long long) io_u->offset, io_u->buflen,
828 (unsigned long long) io_u->file->real_file_size);
829 return 1;
830 }
831
832 /*
833 * mark entry before potentially trimming io_u
834 */
835 if (td_random(td) && file_randommap(td, io_u->file))
836 mark_random_map(td, io_u);
837
838out:
839 dprint_io_u(io_u, "fill_io_u");
840 td->zone_bytes += io_u->buflen;
841 return 0;
842}
843
844static void __io_u_mark_map(unsigned int *map, unsigned int nr)
845{
846 int idx = 0;
847
848 switch (nr) {
849 default:
850 idx = 6;
851 break;
852 case 33 ... 64:
853 idx = 5;
854 break;
855 case 17 ... 32:
856 idx = 4;
857 break;
858 case 9 ... 16:
859 idx = 3;
860 break;
861 case 5 ... 8:
862 idx = 2;
863 break;
864 case 1 ... 4:
865 idx = 1;
866 case 0:
867 break;
868 }
869
870 map[idx]++;
871}
872
873void io_u_mark_submit(struct thread_data *td, unsigned int nr)
874{
875 __io_u_mark_map(td->ts.io_u_submit, nr);
876 td->ts.total_submit++;
877}
878
879void io_u_mark_complete(struct thread_data *td, unsigned int nr)
880{
881 __io_u_mark_map(td->ts.io_u_complete, nr);
882 td->ts.total_complete++;
883}
884
885void io_u_mark_depth(struct thread_data *td, unsigned int nr)
886{
887 int idx = 0;
888
889 switch (td->cur_depth) {
890 default:
891 idx = 6;
892 break;
893 case 32 ... 63:
894 idx = 5;
895 break;
896 case 16 ... 31:
897 idx = 4;
898 break;
899 case 8 ... 15:
900 idx = 3;
901 break;
902 case 4 ... 7:
903 idx = 2;
904 break;
905 case 2 ... 3:
906 idx = 1;
907 case 1:
908 break;
909 }
910
911 td->ts.io_u_map[idx] += nr;
912}
913
914static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
915{
916 int idx = 0;
917
918 assert(usec < 1000);
919
920 switch (usec) {
921 case 750 ... 999:
922 idx = 9;
923 break;
924 case 500 ... 749:
925 idx = 8;
926 break;
927 case 250 ... 499:
928 idx = 7;
929 break;
930 case 100 ... 249:
931 idx = 6;
932 break;
933 case 50 ... 99:
934 idx = 5;
935 break;
936 case 20 ... 49:
937 idx = 4;
938 break;
939 case 10 ... 19:
940 idx = 3;
941 break;
942 case 4 ... 9:
943 idx = 2;
944 break;
945 case 2 ... 3:
946 idx = 1;
947 case 0 ... 1:
948 break;
949 }
950
951 assert(idx < FIO_IO_U_LAT_U_NR);
952 td->ts.io_u_lat_u[idx]++;
953}
954
955static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
956{
957 int idx = 0;
958
959 switch (msec) {
960 default:
961 idx = 11;
962 break;
963 case 1000 ... 1999:
964 idx = 10;
965 break;
966 case 750 ... 999:
967 idx = 9;
968 break;
969 case 500 ... 749:
970 idx = 8;
971 break;
972 case 250 ... 499:
973 idx = 7;
974 break;
975 case 100 ... 249:
976 idx = 6;
977 break;
978 case 50 ... 99:
979 idx = 5;
980 break;
981 case 20 ... 49:
982 idx = 4;
983 break;
984 case 10 ... 19:
985 idx = 3;
986 break;
987 case 4 ... 9:
988 idx = 2;
989 break;
990 case 2 ... 3:
991 idx = 1;
992 case 0 ... 1:
993 break;
994 }
995
996 assert(idx < FIO_IO_U_LAT_M_NR);
997 td->ts.io_u_lat_m[idx]++;
998}
999
1000static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
1001{
1002 if (usec < 1000)
1003 io_u_mark_lat_usec(td, usec);
1004 else
1005 io_u_mark_lat_msec(td, usec / 1000);
1006}
1007
1008/*
1009 * Get next file to service by choosing one at random
1010 */
1011static struct fio_file *get_next_file_rand(struct thread_data *td,
1012 enum fio_file_flags goodf,
1013 enum fio_file_flags badf)
1014{
1015 uint64_t frand_max = rand_max(&td->next_file_state);
1016 struct fio_file *f;
1017 int fno;
1018
1019 do {
1020 int opened = 0;
1021 unsigned long r;
1022
1023 r = __rand(&td->next_file_state);
1024 fno = (unsigned int) ((double) td->o.nr_files
1025 * (r / (frand_max + 1.0)));
1026
1027 f = td->files[fno];
1028 if (fio_file_done(f))
1029 continue;
1030
1031 if (!fio_file_open(f)) {
1032 int err;
1033
1034 if (td->nr_open_files >= td->o.open_files)
1035 return ERR_PTR(-EBUSY);
1036
1037 err = td_io_open_file(td, f);
1038 if (err)
1039 continue;
1040 opened = 1;
1041 }
1042
1043 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1044 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1045 return f;
1046 }
1047 if (opened)
1048 td_io_close_file(td, f);
1049 } while (1);
1050}
1051
1052/*
1053 * Get next file to service by doing round robin between all available ones
1054 */
1055static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1056 int badf)
1057{
1058 unsigned int old_next_file = td->next_file;
1059 struct fio_file *f;
1060
1061 do {
1062 int opened = 0;
1063
1064 f = td->files[td->next_file];
1065
1066 td->next_file++;
1067 if (td->next_file >= td->o.nr_files)
1068 td->next_file = 0;
1069
1070 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1071 if (fio_file_done(f)) {
1072 f = NULL;
1073 continue;
1074 }
1075
1076 if (!fio_file_open(f)) {
1077 int err;
1078
1079 if (td->nr_open_files >= td->o.open_files)
1080 return ERR_PTR(-EBUSY);
1081
1082 err = td_io_open_file(td, f);
1083 if (err) {
1084 dprint(FD_FILE, "error %d on open of %s\n",
1085 err, f->file_name);
1086 f = NULL;
1087 continue;
1088 }
1089 opened = 1;
1090 }
1091
1092 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1093 f->flags);
1094 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1095 break;
1096
1097 if (opened)
1098 td_io_close_file(td, f);
1099
1100 f = NULL;
1101 } while (td->next_file != old_next_file);
1102
1103 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1104 return f;
1105}
1106
1107static struct fio_file *__get_next_file(struct thread_data *td)
1108{
1109 struct fio_file *f;
1110
1111 assert(td->o.nr_files <= td->files_index);
1112
1113 if (td->nr_done_files >= td->o.nr_files) {
1114 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1115 " nr_files=%d\n", td->nr_open_files,
1116 td->nr_done_files,
1117 td->o.nr_files);
1118 return NULL;
1119 }
1120
1121 f = td->file_service_file;
1122 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1123 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1124 goto out;
1125 if (td->file_service_left--)
1126 goto out;
1127 }
1128
1129 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1130 td->o.file_service_type == FIO_FSERVICE_SEQ)
1131 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1132 else
1133 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1134
1135 if (IS_ERR(f))
1136 return f;
1137
1138 td->file_service_file = f;
1139 td->file_service_left = td->file_service_nr - 1;
1140out:
1141 if (f)
1142 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1143 else
1144 dprint(FD_FILE, "get_next_file: NULL\n");
1145 return f;
1146}
1147
1148static struct fio_file *get_next_file(struct thread_data *td)
1149{
1150 if (td->flags & TD_F_PROFILE_OPS) {
1151 struct prof_io_ops *ops = &td->prof_io_ops;
1152
1153 if (ops->get_next_file)
1154 return ops->get_next_file(td);
1155 }
1156
1157 return __get_next_file(td);
1158}
1159
1160static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1161{
1162 struct fio_file *f;
1163
1164 do {
1165 f = get_next_file(td);
1166 if (IS_ERR_OR_NULL(f))
1167 return PTR_ERR(f);
1168
1169 io_u->file = f;
1170 get_file(f);
1171
1172 if (!fill_io_u(td, io_u))
1173 break;
1174
1175 put_file_log(td, f);
1176 td_io_close_file(td, f);
1177 io_u->file = NULL;
1178 fio_file_set_done(f);
1179 td->nr_done_files++;
1180 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1181 td->nr_done_files, td->o.nr_files);
1182 } while (1);
1183
1184 return 0;
1185}
1186
1187static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1188 unsigned long tusec, unsigned long max_usec)
1189{
1190 if (!td->error)
1191 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1192 td_verror(td, ETIMEDOUT, "max latency exceeded");
1193 icd->error = ETIMEDOUT;
1194}
1195
1196static void lat_new_cycle(struct thread_data *td)
1197{
1198 fio_gettime(&td->latency_ts, NULL);
1199 td->latency_ios = ddir_rw_sum(td->io_blocks);
1200 td->latency_failed = 0;
1201}
1202
1203/*
1204 * We had an IO outside the latency target. Reduce the queue depth. If we
1205 * are at QD=1, then it's time to give up.
1206 */
1207static bool __lat_target_failed(struct thread_data *td)
1208{
1209 if (td->latency_qd == 1)
1210 return true;
1211
1212 td->latency_qd_high = td->latency_qd;
1213
1214 if (td->latency_qd == td->latency_qd_low)
1215 td->latency_qd_low--;
1216
1217 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1218
1219 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1220
1221 /*
1222 * When we ramp QD down, quiesce existing IO to prevent
1223 * a storm of ramp downs due to pending higher depth.
1224 */
1225 io_u_quiesce(td);
1226 lat_new_cycle(td);
1227 return false;
1228}
1229
1230static bool lat_target_failed(struct thread_data *td)
1231{
1232 if (td->o.latency_percentile.u.f == 100.0)
1233 return __lat_target_failed(td);
1234
1235 td->latency_failed++;
1236 return false;
1237}
1238
1239void lat_target_init(struct thread_data *td)
1240{
1241 td->latency_end_run = 0;
1242
1243 if (td->o.latency_target) {
1244 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1245 fio_gettime(&td->latency_ts, NULL);
1246 td->latency_qd = 1;
1247 td->latency_qd_high = td->o.iodepth;
1248 td->latency_qd_low = 1;
1249 td->latency_ios = ddir_rw_sum(td->io_blocks);
1250 } else
1251 td->latency_qd = td->o.iodepth;
1252}
1253
1254void lat_target_reset(struct thread_data *td)
1255{
1256 if (!td->latency_end_run)
1257 lat_target_init(td);
1258}
1259
1260static void lat_target_success(struct thread_data *td)
1261{
1262 const unsigned int qd = td->latency_qd;
1263 struct thread_options *o = &td->o;
1264
1265 td->latency_qd_low = td->latency_qd;
1266
1267 /*
1268 * If we haven't failed yet, we double up to a failing value instead
1269 * of bisecting from highest possible queue depth. If we have set
1270 * a limit other than td->o.iodepth, bisect between that.
1271 */
1272 if (td->latency_qd_high != o->iodepth)
1273 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1274 else
1275 td->latency_qd *= 2;
1276
1277 if (td->latency_qd > o->iodepth)
1278 td->latency_qd = o->iodepth;
1279
1280 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1281
1282 /*
1283 * Same as last one, we are done. Let it run a latency cycle, so
1284 * we get only the results from the targeted depth.
1285 */
1286 if (td->latency_qd == qd) {
1287 if (td->latency_end_run) {
1288 dprint(FD_RATE, "We are done\n");
1289 td->done = 1;
1290 } else {
1291 dprint(FD_RATE, "Quiesce and final run\n");
1292 io_u_quiesce(td);
1293 td->latency_end_run = 1;
1294 reset_all_stats(td);
1295 reset_io_stats(td);
1296 }
1297 }
1298
1299 lat_new_cycle(td);
1300}
1301
1302/*
1303 * Check if we can bump the queue depth
1304 */
1305void lat_target_check(struct thread_data *td)
1306{
1307 uint64_t usec_window;
1308 uint64_t ios;
1309 double success_ios;
1310
1311 usec_window = utime_since_now(&td->latency_ts);
1312 if (usec_window < td->o.latency_window)
1313 return;
1314
1315 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1316 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1317 success_ios *= 100.0;
1318
1319 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1320
1321 if (success_ios >= td->o.latency_percentile.u.f)
1322 lat_target_success(td);
1323 else
1324 __lat_target_failed(td);
1325}
1326
1327/*
1328 * If latency target is enabled, we might be ramping up or down and not
1329 * using the full queue depth available.
1330 */
1331bool queue_full(const struct thread_data *td)
1332{
1333 const int qempty = io_u_qempty(&td->io_u_freelist);
1334
1335 if (qempty)
1336 return true;
1337 if (!td->o.latency_target)
1338 return false;
1339
1340 return td->cur_depth >= td->latency_qd;
1341}
1342
1343struct io_u *__get_io_u(struct thread_data *td)
1344{
1345 struct io_u *io_u = NULL;
1346
1347 if (td->stop_io)
1348 return NULL;
1349
1350 td_io_u_lock(td);
1351
1352again:
1353 if (!io_u_rempty(&td->io_u_requeues))
1354 io_u = io_u_rpop(&td->io_u_requeues);
1355 else if (!queue_full(td)) {
1356 io_u = io_u_qpop(&td->io_u_freelist);
1357
1358 io_u->file = NULL;
1359 io_u->buflen = 0;
1360 io_u->resid = 0;
1361 io_u->end_io = NULL;
1362 }
1363
1364 if (io_u) {
1365 assert(io_u->flags & IO_U_F_FREE);
1366 io_u_clear(io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1367 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1368 IO_U_F_VER_LIST);
1369
1370 io_u->error = 0;
1371 io_u->acct_ddir = -1;
1372 td->cur_depth++;
1373 assert(!(td->flags & TD_F_CHILD));
1374 io_u_set(io_u, IO_U_F_IN_CUR_DEPTH);
1375 io_u->ipo = NULL;
1376 } else if (td_async_processing(td)) {
1377 /*
1378 * We ran out, wait for async verify threads to finish and
1379 * return one
1380 */
1381 assert(!(td->flags & TD_F_CHILD));
1382 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1383 goto again;
1384 }
1385
1386 td_io_u_unlock(td);
1387 return io_u;
1388}
1389
1390static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1391{
1392 if (!(td->flags & TD_F_TRIM_BACKLOG))
1393 return false;
1394
1395 if (td->trim_entries) {
1396 int get_trim = 0;
1397
1398 if (td->trim_batch) {
1399 td->trim_batch--;
1400 get_trim = 1;
1401 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1402 td->last_ddir != DDIR_READ) {
1403 td->trim_batch = td->o.trim_batch;
1404 if (!td->trim_batch)
1405 td->trim_batch = td->o.trim_backlog;
1406 get_trim = 1;
1407 }
1408
1409 if (get_trim && !get_next_trim(td, io_u))
1410 return true;
1411 }
1412
1413 return false;
1414}
1415
1416static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1417{
1418 if (!(td->flags & TD_F_VER_BACKLOG))
1419 return false;
1420
1421 if (td->io_hist_len) {
1422 int get_verify = 0;
1423
1424 if (td->verify_batch)
1425 get_verify = 1;
1426 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1427 td->last_ddir != DDIR_READ) {
1428 td->verify_batch = td->o.verify_batch;
1429 if (!td->verify_batch)
1430 td->verify_batch = td->o.verify_backlog;
1431 get_verify = 1;
1432 }
1433
1434 if (get_verify && !get_next_verify(td, io_u)) {
1435 td->verify_batch--;
1436 return true;
1437 }
1438 }
1439
1440 return false;
1441}
1442
1443/*
1444 * Fill offset and start time into the buffer content, to prevent too
1445 * easy compressible data for simple de-dupe attempts. Do this for every
1446 * 512b block in the range, since that should be the smallest block size
1447 * we can expect from a device.
1448 */
1449static void small_content_scramble(struct io_u *io_u)
1450{
1451 unsigned int i, nr_blocks = io_u->buflen / 512;
1452 uint64_t boffset;
1453 unsigned int offset;
1454 void *p, *end;
1455
1456 if (!nr_blocks)
1457 return;
1458
1459 p = io_u->xfer_buf;
1460 boffset = io_u->offset;
1461 io_u->buf_filled_len = 0;
1462
1463 for (i = 0; i < nr_blocks; i++) {
1464 /*
1465 * Fill the byte offset into a "random" start offset of
1466 * the buffer, given by the product of the usec time
1467 * and the actual offset.
1468 */
1469 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1470 offset &= ~(sizeof(uint64_t) - 1);
1471 if (offset >= 512 - sizeof(uint64_t))
1472 offset -= sizeof(uint64_t);
1473 memcpy(p + offset, &boffset, sizeof(boffset));
1474
1475 end = p + 512 - sizeof(io_u->start_time);
1476 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1477 p += 512;
1478 boffset += 512;
1479 }
1480}
1481
1482/*
1483 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1484 * etc. The returned io_u is fully ready to be prepped and submitted.
1485 */
1486struct io_u *get_io_u(struct thread_data *td)
1487{
1488 struct fio_file *f;
1489 struct io_u *io_u;
1490 int do_scramble = 0;
1491 long ret = 0;
1492
1493 io_u = __get_io_u(td);
1494 if (!io_u) {
1495 dprint(FD_IO, "__get_io_u failed\n");
1496 return NULL;
1497 }
1498
1499 if (check_get_verify(td, io_u))
1500 goto out;
1501 if (check_get_trim(td, io_u))
1502 goto out;
1503
1504 /*
1505 * from a requeue, io_u already setup
1506 */
1507 if (io_u->file)
1508 goto out;
1509
1510 /*
1511 * If using an iolog, grab next piece if any available.
1512 */
1513 if (td->flags & TD_F_READ_IOLOG) {
1514 if (read_iolog_get(td, io_u))
1515 goto err_put;
1516 } else if (set_io_u_file(td, io_u)) {
1517 ret = -EBUSY;
1518 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1519 goto err_put;
1520 }
1521
1522 f = io_u->file;
1523 if (!f) {
1524 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1525 goto err_put;
1526 }
1527
1528 assert(fio_file_open(f));
1529
1530 if (ddir_rw(io_u->ddir)) {
1531 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1532 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1533 goto err_put;
1534 }
1535
1536 f->last_start[io_u->ddir] = io_u->offset;
1537 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1538
1539 if (io_u->ddir == DDIR_WRITE) {
1540 if (td->flags & TD_F_REFILL_BUFFERS) {
1541 io_u_fill_buffer(td, io_u,
1542 td->o.min_bs[DDIR_WRITE],
1543 io_u->buflen);
1544 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1545 !(td->flags & TD_F_COMPRESS))
1546 do_scramble = 1;
1547 if (td->flags & TD_F_VER_NONE) {
1548 populate_verify_io_u(td, io_u);
1549 do_scramble = 0;
1550 }
1551 } else if (io_u->ddir == DDIR_READ) {
1552 /*
1553 * Reset the buf_filled parameters so next time if the
1554 * buffer is used for writes it is refilled.
1555 */
1556 io_u->buf_filled_len = 0;
1557 }
1558 }
1559
1560 /*
1561 * Set io data pointers.
1562 */
1563 io_u->xfer_buf = io_u->buf;
1564 io_u->xfer_buflen = io_u->buflen;
1565
1566out:
1567 assert(io_u->file);
1568 if (!td_io_prep(td, io_u)) {
1569 if (!td->o.disable_lat)
1570 fio_gettime(&io_u->start_time, NULL);
1571 if (do_scramble)
1572 small_content_scramble(io_u);
1573 return io_u;
1574 }
1575err_put:
1576 dprint(FD_IO, "get_io_u failed\n");
1577 put_io_u(td, io_u);
1578 return ERR_PTR(ret);
1579}
1580
1581static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1582{
1583 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1584
1585 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1586 return;
1587
1588 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1589 io_u->file ? " on file " : "",
1590 io_u->file ? io_u->file->file_name : "",
1591 strerror(io_u->error),
1592 io_ddir_name(io_u->ddir),
1593 io_u->offset, io_u->xfer_buflen);
1594
1595 if (td->io_ops->errdetails) {
1596 char *err = td->io_ops->errdetails(io_u);
1597
1598 log_err("fio: %s\n", err);
1599 free(err);
1600 }
1601
1602 if (!td->error)
1603 td_verror(td, io_u->error, "io_u error");
1604}
1605
1606void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1607{
1608 __io_u_log_error(td, io_u);
1609 if (td->parent)
1610 __io_u_log_error(td, io_u);
1611}
1612
1613static inline bool gtod_reduce(struct thread_data *td)
1614{
1615 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1616 || td->o.gtod_reduce;
1617}
1618
1619static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1620 struct io_completion_data *icd,
1621 const enum fio_ddir idx, unsigned int bytes)
1622{
1623 const int no_reduce = !gtod_reduce(td);
1624 unsigned long lusec = 0;
1625
1626 if (td->parent)
1627 td = td->parent;
1628
1629 if (no_reduce)
1630 lusec = utime_since(&io_u->issue_time, &icd->time);
1631
1632 if (!td->o.disable_lat) {
1633 unsigned long tusec;
1634
1635 tusec = utime_since(&io_u->start_time, &icd->time);
1636 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1637
1638 if (td->flags & TD_F_PROFILE_OPS) {
1639 struct prof_io_ops *ops = &td->prof_io_ops;
1640
1641 if (ops->io_u_lat)
1642 icd->error = ops->io_u_lat(td, tusec);
1643 }
1644
1645 if (td->o.max_latency && tusec > td->o.max_latency)
1646 lat_fatal(td, icd, tusec, td->o.max_latency);
1647 if (td->o.latency_target && tusec > td->o.latency_target) {
1648 if (lat_target_failed(td))
1649 lat_fatal(td, icd, tusec, td->o.latency_target);
1650 }
1651 }
1652
1653 if (!td->o.disable_clat) {
1654 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1655 io_u_mark_latency(td, lusec);
1656 }
1657
1658 if (!td->o.disable_bw)
1659 add_bw_sample(td, idx, bytes, &icd->time);
1660
1661 if (no_reduce)
1662 add_iops_sample(td, idx, bytes, &icd->time);
1663
1664 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1665 uint32_t *info = io_u_block_info(td, io_u);
1666 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1667 if (io_u->ddir == DDIR_TRIM) {
1668 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1669 BLOCK_INFO_TRIMS(*info) + 1);
1670 } else if (io_u->ddir == DDIR_WRITE) {
1671 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1672 *info);
1673 }
1674 }
1675 }
1676}
1677
1678static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1679 struct io_completion_data *icd)
1680{
1681 struct io_u *io_u = *io_u_ptr;
1682 enum fio_ddir ddir = io_u->ddir;
1683 struct fio_file *f = io_u->file;
1684
1685 dprint_io_u(io_u, "io complete");
1686
1687 assert(io_u->flags & IO_U_F_FLIGHT);
1688 io_u_clear(io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1689
1690 /*
1691 * Mark IO ok to verify
1692 */
1693 if (io_u->ipo) {
1694 /*
1695 * Remove errored entry from the verification list
1696 */
1697 if (io_u->error)
1698 unlog_io_piece(td, io_u);
1699 else {
1700 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1701 write_barrier();
1702 }
1703 }
1704
1705 if (ddir_sync(ddir)) {
1706 td->last_was_sync = 1;
1707 if (f) {
1708 f->first_write = -1ULL;
1709 f->last_write = -1ULL;
1710 }
1711 return;
1712 }
1713
1714 td->last_was_sync = 0;
1715 td->last_ddir = ddir;
1716
1717 if (!io_u->error && ddir_rw(ddir)) {
1718 unsigned int bytes = io_u->buflen - io_u->resid;
1719 int ret;
1720
1721 td->io_blocks[ddir]++;
1722 td->this_io_blocks[ddir]++;
1723 td->io_bytes[ddir] += bytes;
1724
1725 if (!(io_u->flags & IO_U_F_VER_LIST))
1726 td->this_io_bytes[ddir] += bytes;
1727
1728 if (ddir == DDIR_WRITE) {
1729 if (f) {
1730 if (f->first_write == -1ULL ||
1731 io_u->offset < f->first_write)
1732 f->first_write = io_u->offset;
1733 if (f->last_write == -1ULL ||
1734 ((io_u->offset + bytes) > f->last_write))
1735 f->last_write = io_u->offset + bytes;
1736 }
1737 if (td->last_write_comp) {
1738 int idx = td->last_write_idx++;
1739
1740 td->last_write_comp[idx] = io_u->offset;
1741 if (td->last_write_idx == td->o.iodepth)
1742 td->last_write_idx = 0;
1743 }
1744 }
1745
1746 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1747 td->runstate == TD_VERIFYING))
1748 account_io_completion(td, io_u, icd, ddir, bytes);
1749
1750 icd->bytes_done[ddir] += bytes;
1751
1752 if (io_u->end_io) {
1753 ret = io_u->end_io(td, io_u_ptr);
1754 io_u = *io_u_ptr;
1755 if (ret && !icd->error)
1756 icd->error = ret;
1757 }
1758 } else if (io_u->error) {
1759 icd->error = io_u->error;
1760 io_u_log_error(td, io_u);
1761 }
1762 if (icd->error) {
1763 enum error_type_bit eb = td_error_type(ddir, icd->error);
1764
1765 if (!td_non_fatal_error(td, eb, icd->error))
1766 return;
1767
1768 /*
1769 * If there is a non_fatal error, then add to the error count
1770 * and clear all the errors.
1771 */
1772 update_error_count(td, icd->error);
1773 td_clear_error(td);
1774 icd->error = 0;
1775 if (io_u)
1776 io_u->error = 0;
1777 }
1778}
1779
1780static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1781 int nr)
1782{
1783 int ddir;
1784
1785 if (!gtod_reduce(td))
1786 fio_gettime(&icd->time, NULL);
1787
1788 icd->nr = nr;
1789
1790 icd->error = 0;
1791 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1792 icd->bytes_done[ddir] = 0;
1793}
1794
1795static void ios_completed(struct thread_data *td,
1796 struct io_completion_data *icd)
1797{
1798 struct io_u *io_u;
1799 int i;
1800
1801 for (i = 0; i < icd->nr; i++) {
1802 io_u = td->io_ops->event(td, i);
1803
1804 io_completed(td, &io_u, icd);
1805
1806 if (io_u)
1807 put_io_u(td, io_u);
1808 }
1809}
1810
1811/*
1812 * Complete a single io_u for the sync engines.
1813 */
1814int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1815{
1816 struct io_completion_data icd;
1817 int ddir;
1818
1819 init_icd(td, &icd, 1);
1820 io_completed(td, &io_u, &icd);
1821
1822 if (io_u)
1823 put_io_u(td, io_u);
1824
1825 if (icd.error) {
1826 td_verror(td, icd.error, "io_u_sync_complete");
1827 return -1;
1828 }
1829
1830 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1831 td->bytes_done[ddir] += icd.bytes_done[ddir];
1832
1833 return 0;
1834}
1835
1836/*
1837 * Called to complete min_events number of io for the async engines.
1838 */
1839int io_u_queued_complete(struct thread_data *td, int min_evts)
1840{
1841 struct io_completion_data icd;
1842 struct timespec *tvp = NULL;
1843 int ret, ddir;
1844 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1845
1846 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1847
1848 if (!min_evts)
1849 tvp = &ts;
1850 else if (min_evts > td->cur_depth)
1851 min_evts = td->cur_depth;
1852
1853 /* No worries, td_io_getevents fixes min and max if they are
1854 * set incorrectly */
1855 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
1856 if (ret < 0) {
1857 td_verror(td, -ret, "td_io_getevents");
1858 return ret;
1859 } else if (!ret)
1860 return ret;
1861
1862 init_icd(td, &icd, ret);
1863 ios_completed(td, &icd);
1864 if (icd.error) {
1865 td_verror(td, icd.error, "io_u_queued_complete");
1866 return -1;
1867 }
1868
1869 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1870 td->bytes_done[ddir] += icd.bytes_done[ddir];
1871
1872 return ret;
1873}
1874
1875/*
1876 * Call when io_u is really queued, to update the submission latency.
1877 */
1878void io_u_queued(struct thread_data *td, struct io_u *io_u)
1879{
1880 if (!td->o.disable_slat) {
1881 unsigned long slat_time;
1882
1883 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1884
1885 if (td->parent)
1886 td = td->parent;
1887
1888 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1889 io_u->offset);
1890 }
1891}
1892
1893/*
1894 * See if we should reuse the last seed, if dedupe is enabled
1895 */
1896static struct frand_state *get_buf_state(struct thread_data *td)
1897{
1898 uint64_t frand_max;
1899 unsigned int v;
1900 unsigned long r;
1901
1902 if (!td->o.dedupe_percentage)
1903 return &td->buf_state;
1904 else if (td->o.dedupe_percentage == 100) {
1905 frand_copy(&td->buf_state_prev, &td->buf_state);
1906 return &td->buf_state;
1907 }
1908
1909 frand_max = rand_max(&td->dedupe_state);
1910 r = __rand(&td->dedupe_state);
1911 v = 1 + (int) (100.0 * (r / (frand_max + 1.0)));
1912
1913 if (v <= td->o.dedupe_percentage)
1914 return &td->buf_state_prev;
1915
1916 return &td->buf_state;
1917}
1918
1919static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1920{
1921 if (td->o.dedupe_percentage == 100)
1922 frand_copy(rs, &td->buf_state_prev);
1923 else if (rs == &td->buf_state)
1924 frand_copy(&td->buf_state_prev, rs);
1925}
1926
1927void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1928 unsigned int max_bs)
1929{
1930 struct thread_options *o = &td->o;
1931
1932 if (o->compress_percentage || o->dedupe_percentage) {
1933 unsigned int perc = td->o.compress_percentage;
1934 struct frand_state *rs;
1935 unsigned int left = max_bs;
1936 unsigned int this_write;
1937
1938 do {
1939 rs = get_buf_state(td);
1940
1941 min_write = min(min_write, left);
1942
1943 if (perc) {
1944 this_write = min_not_zero(min_write,
1945 td->o.compress_chunk);
1946
1947 fill_random_buf_percentage(rs, buf, perc,
1948 this_write, this_write,
1949 o->buffer_pattern,
1950 o->buffer_pattern_bytes);
1951 } else {
1952 fill_random_buf(rs, buf, min_write);
1953 this_write = min_write;
1954 }
1955
1956 buf += this_write;
1957 left -= this_write;
1958 save_buf_state(td, rs);
1959 } while (left);
1960 } else if (o->buffer_pattern_bytes)
1961 fill_buffer_pattern(td, buf, max_bs);
1962 else if (o->zero_buffers)
1963 memset(buf, 0, max_bs);
1964 else
1965 fill_random_buf(get_buf_state(td), buf, max_bs);
1966}
1967
1968/*
1969 * "randomly" fill the buffer contents
1970 */
1971void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1972 unsigned int min_write, unsigned int max_bs)
1973{
1974 io_u->buf_filled_len = 0;
1975 fill_io_buffer(td, io_u->buf, min_write, max_bs);
1976}
fio - Flexible IO Tester