projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
Makefile: missing conversion, and oslib/ dep cleanup
[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 int 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 int should_sort_io(struct thread_data *td)
194{
195 if (!td->o.verifysort_nr || !td->o.do_verify)
196 return 0;
197 if (!td_random(td))
198 return 0;
199 if (td->runstate != TD_VERIFYING)
200 return 0;
201 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
202 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
203 return 0;
204
205 return 1;
206}
207
208static int 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 1;
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 int 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
545void io_u_quiesce(struct thread_data *td)
546{
547 /*
548 * We are going to sleep, ensure that we flush anything pending as
549 * not to skew our latency numbers.
550 *
551 * Changed to only monitor 'in flight' requests here instead of the
552 * td->cur_depth, b/c td->cur_depth does not accurately represent
553 * io's that have been actually submitted to an async engine,
554 * and cur_depth is meaningless for sync engines.
555 */
556 if (td->io_u_queued || td->cur_depth) {
557 int fio_unused ret;
558
559 ret = td_io_commit(td);
560 }
561
562 while (td->io_u_in_flight) {
563 int fio_unused ret;
564
565 ret = io_u_queued_complete(td, 1);
566 }
567}
568
569static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
570{
571 enum fio_ddir odir = ddir ^ 1;
572 long usec, now;
573
574 assert(ddir_rw(ddir));
575 now = utime_since_now(&td->start);
576
577 /*
578 * if rate_next_io_time is in the past, need to catch up to rate
579 */
580 if (td->rate_next_io_time[ddir] <= now)
581 return ddir;
582
583 /*
584 * We are ahead of rate in this direction. See if we
585 * should switch.
586 */
587 if (td_rw(td) && td->o.rwmix[odir]) {
588 /*
589 * Other direction is behind rate, switch
590 */
591 if (td->rate_next_io_time[odir] <= now)
592 return odir;
593
594 /*
595 * Both directions are ahead of rate. sleep the min
596 * switch if necissary
597 */
598 if (td->rate_next_io_time[ddir] <=
599 td->rate_next_io_time[odir]) {
600 usec = td->rate_next_io_time[ddir] - now;
601 } else {
602 usec = td->rate_next_io_time[odir] - now;
603 ddir = odir;
604 }
605 } else
606 usec = td->rate_next_io_time[ddir] - now;
607
608 if (td->o.io_submit_mode == IO_MODE_INLINE)
609 io_u_quiesce(td);
610
611 usec = usec_sleep(td, usec);
612
613 return ddir;
614}
615
616/*
617 * Return the data direction for the next io_u. If the job is a
618 * mixed read/write workload, check the rwmix cycle and switch if
619 * necessary.
620 */
621static enum fio_ddir get_rw_ddir(struct thread_data *td)
622{
623 enum fio_ddir ddir;
624
625 /*
626 * see if it's time to fsync
627 */
628 if (td->o.fsync_blocks &&
629 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
630 td->io_issues[DDIR_WRITE] && should_fsync(td))
631 return DDIR_SYNC;
632
633 /*
634 * see if it's time to fdatasync
635 */
636 if (td->o.fdatasync_blocks &&
637 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
638 td->io_issues[DDIR_WRITE] && should_fsync(td))
639 return DDIR_DATASYNC;
640
641 /*
642 * see if it's time to sync_file_range
643 */
644 if (td->sync_file_range_nr &&
645 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
646 td->io_issues[DDIR_WRITE] && should_fsync(td))
647 return DDIR_SYNC_FILE_RANGE;
648
649 if (td_rw(td)) {
650 /*
651 * Check if it's time to seed a new data direction.
652 */
653 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
654 /*
655 * Put a top limit on how many bytes we do for
656 * one data direction, to avoid overflowing the
657 * ranges too much
658 */
659 ddir = get_rand_ddir(td);
660
661 if (ddir != td->rwmix_ddir)
662 set_rwmix_bytes(td);
663
664 td->rwmix_ddir = ddir;
665 }
666 ddir = td->rwmix_ddir;
667 } else if (td_read(td))
668 ddir = DDIR_READ;
669 else if (td_write(td))
670 ddir = DDIR_WRITE;
671 else
672 ddir = DDIR_TRIM;
673
674 td->rwmix_ddir = rate_ddir(td, ddir);
675 return td->rwmix_ddir;
676}
677
678static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
679{
680 enum fio_ddir ddir = get_rw_ddir(td);
681
682 if (td_trimwrite(td)) {
683 struct fio_file *f = io_u->file;
684 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
685 ddir = DDIR_TRIM;
686 else
687 ddir = DDIR_WRITE;
688 }
689
690 io_u->ddir = io_u->acct_ddir = ddir;
691
692 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
693 td->o.barrier_blocks &&
694 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
695 td->io_issues[DDIR_WRITE])
696 io_u_set(io_u, IO_U_F_BARRIER);
697}
698
699void put_file_log(struct thread_data *td, struct fio_file *f)
700{
701 unsigned int ret = put_file(td, f);
702
703 if (ret)
704 td_verror(td, ret, "file close");
705}
706
707void put_io_u(struct thread_data *td, struct io_u *io_u)
708{
709 if (td->parent)
710 td = td->parent;
711
712 td_io_u_lock(td);
713
714 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
715 put_file_log(td, io_u->file);
716
717 io_u->file = NULL;
718 io_u_set(io_u, IO_U_F_FREE);
719
720 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
721 td->cur_depth--;
722 assert(!(td->flags & TD_F_CHILD));
723 }
724 io_u_qpush(&td->io_u_freelist, io_u);
725 td_io_u_unlock(td);
726 td_io_u_free_notify(td);
727}
728
729void clear_io_u(struct thread_data *td, struct io_u *io_u)
730{
731 io_u_clear(io_u, IO_U_F_FLIGHT);
732 put_io_u(td, io_u);
733}
734
735void requeue_io_u(struct thread_data *td, struct io_u **io_u)
736{
737 struct io_u *__io_u = *io_u;
738 enum fio_ddir ddir = acct_ddir(__io_u);
739
740 dprint(FD_IO, "requeue %p\n", __io_u);
741
742 if (td->parent)
743 td = td->parent;
744
745 td_io_u_lock(td);
746
747 io_u_set(__io_u, IO_U_F_FREE);
748 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
749 td->io_issues[ddir]--;
750
751 io_u_clear(__io_u, IO_U_F_FLIGHT);
752 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
753 td->cur_depth--;
754 assert(!(td->flags & TD_F_CHILD));
755 }
756
757 io_u_rpush(&td->io_u_requeues, __io_u);
758 td_io_u_unlock(td);
759 td_io_u_free_notify(td);
760 *io_u = NULL;
761}
762
763static int fill_io_u(struct thread_data *td, struct io_u *io_u)
764{
765 unsigned int is_random;
766
767 if (td->io_ops->flags & FIO_NOIO)
768 goto out;
769
770 set_rw_ddir(td, io_u);
771
772 /*
773 * fsync() or fdatasync() or trim etc, we are done
774 */
775 if (!ddir_rw(io_u->ddir))
776 goto out;
777
778 /*
779 * See if it's time to switch to a new zone
780 */
781 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
782 struct fio_file *f = io_u->file;
783
784 td->zone_bytes = 0;
785 f->file_offset += td->o.zone_range + td->o.zone_skip;
786
787 /*
788 * Wrap from the beginning, if we exceed the file size
789 */
790 if (f->file_offset >= f->real_file_size)
791 f->file_offset = f->real_file_size - f->file_offset;
792 f->last_pos[io_u->ddir] = f->file_offset;
793 td->io_skip_bytes += td->o.zone_skip;
794 }
795
796 /*
797 * No log, let the seq/rand engine retrieve the next buflen and
798 * position.
799 */
800 if (get_next_offset(td, io_u, &is_random)) {
801 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
802 return 1;
803 }
804
805 io_u->buflen = get_next_buflen(td, io_u, is_random);
806 if (!io_u->buflen) {
807 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
808 return 1;
809 }
810
811 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
812 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
813 dprint(FD_IO, " off=%llu/%lu > %llu\n",
814 (unsigned long long) io_u->offset, io_u->buflen,
815 (unsigned long long) io_u->file->real_file_size);
816 return 1;
817 }
818
819 /*
820 * mark entry before potentially trimming io_u
821 */
822 if (td_random(td) && file_randommap(td, io_u->file))
823 mark_random_map(td, io_u);
824
825out:
826 dprint_io_u(io_u, "fill_io_u");
827 td->zone_bytes += io_u->buflen;
828 return 0;
829}
830
831static void __io_u_mark_map(unsigned int *map, unsigned int nr)
832{
833 int idx = 0;
834
835 switch (nr) {
836 default:
837 idx = 6;
838 break;
839 case 33 ... 64:
840 idx = 5;
841 break;
842 case 17 ... 32:
843 idx = 4;
844 break;
845 case 9 ... 16:
846 idx = 3;
847 break;
848 case 5 ... 8:
849 idx = 2;
850 break;
851 case 1 ... 4:
852 idx = 1;
853 case 0:
854 break;
855 }
856
857 map[idx]++;
858}
859
860void io_u_mark_submit(struct thread_data *td, unsigned int nr)
861{
862 __io_u_mark_map(td->ts.io_u_submit, nr);
863 td->ts.total_submit++;
864}
865
866void io_u_mark_complete(struct thread_data *td, unsigned int nr)
867{
868 __io_u_mark_map(td->ts.io_u_complete, nr);
869 td->ts.total_complete++;
870}
871
872void io_u_mark_depth(struct thread_data *td, unsigned int nr)
873{
874 int idx = 0;
875
876 switch (td->cur_depth) {
877 default:
878 idx = 6;
879 break;
880 case 32 ... 63:
881 idx = 5;
882 break;
883 case 16 ... 31:
884 idx = 4;
885 break;
886 case 8 ... 15:
887 idx = 3;
888 break;
889 case 4 ... 7:
890 idx = 2;
891 break;
892 case 2 ... 3:
893 idx = 1;
894 case 1:
895 break;
896 }
897
898 td->ts.io_u_map[idx] += nr;
899}
900
901static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
902{
903 int idx = 0;
904
905 assert(usec < 1000);
906
907 switch (usec) {
908 case 750 ... 999:
909 idx = 9;
910 break;
911 case 500 ... 749:
912 idx = 8;
913 break;
914 case 250 ... 499:
915 idx = 7;
916 break;
917 case 100 ... 249:
918 idx = 6;
919 break;
920 case 50 ... 99:
921 idx = 5;
922 break;
923 case 20 ... 49:
924 idx = 4;
925 break;
926 case 10 ... 19:
927 idx = 3;
928 break;
929 case 4 ... 9:
930 idx = 2;
931 break;
932 case 2 ... 3:
933 idx = 1;
934 case 0 ... 1:
935 break;
936 }
937
938 assert(idx < FIO_IO_U_LAT_U_NR);
939 td->ts.io_u_lat_u[idx]++;
940}
941
942static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
943{
944 int idx = 0;
945
946 switch (msec) {
947 default:
948 idx = 11;
949 break;
950 case 1000 ... 1999:
951 idx = 10;
952 break;
953 case 750 ... 999:
954 idx = 9;
955 break;
956 case 500 ... 749:
957 idx = 8;
958 break;
959 case 250 ... 499:
960 idx = 7;
961 break;
962 case 100 ... 249:
963 idx = 6;
964 break;
965 case 50 ... 99:
966 idx = 5;
967 break;
968 case 20 ... 49:
969 idx = 4;
970 break;
971 case 10 ... 19:
972 idx = 3;
973 break;
974 case 4 ... 9:
975 idx = 2;
976 break;
977 case 2 ... 3:
978 idx = 1;
979 case 0 ... 1:
980 break;
981 }
982
983 assert(idx < FIO_IO_U_LAT_M_NR);
984 td->ts.io_u_lat_m[idx]++;
985}
986
987static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
988{
989 if (usec < 1000)
990 io_u_mark_lat_usec(td, usec);
991 else
992 io_u_mark_lat_msec(td, usec / 1000);
993}
994
995/*
996 * Get next file to service by choosing one at random
997 */
998static struct fio_file *get_next_file_rand(struct thread_data *td,
999 enum fio_file_flags goodf,
1000 enum fio_file_flags badf)
1001{
1002 uint64_t frand_max = rand_max(&td->next_file_state);
1003 struct fio_file *f;
1004 int fno;
1005
1006 do {
1007 int opened = 0;
1008 unsigned long r;
1009
1010 r = __rand(&td->next_file_state);
1011 fno = (unsigned int) ((double) td->o.nr_files
1012 * (r / (frand_max + 1.0)));
1013
1014 f = td->files[fno];
1015 if (fio_file_done(f))
1016 continue;
1017
1018 if (!fio_file_open(f)) {
1019 int err;
1020
1021 if (td->nr_open_files >= td->o.open_files)
1022 return ERR_PTR(-EBUSY);
1023
1024 err = td_io_open_file(td, f);
1025 if (err)
1026 continue;
1027 opened = 1;
1028 }
1029
1030 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1031 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1032 return f;
1033 }
1034 if (opened)
1035 td_io_close_file(td, f);
1036 } while (1);
1037}
1038
1039/*
1040 * Get next file to service by doing round robin between all available ones
1041 */
1042static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1043 int badf)
1044{
1045 unsigned int old_next_file = td->next_file;
1046 struct fio_file *f;
1047
1048 do {
1049 int opened = 0;
1050
1051 f = td->files[td->next_file];
1052
1053 td->next_file++;
1054 if (td->next_file >= td->o.nr_files)
1055 td->next_file = 0;
1056
1057 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1058 if (fio_file_done(f)) {
1059 f = NULL;
1060 continue;
1061 }
1062
1063 if (!fio_file_open(f)) {
1064 int err;
1065
1066 if (td->nr_open_files >= td->o.open_files)
1067 return ERR_PTR(-EBUSY);
1068
1069 err = td_io_open_file(td, f);
1070 if (err) {
1071 dprint(FD_FILE, "error %d on open of %s\n",
1072 err, f->file_name);
1073 f = NULL;
1074 continue;
1075 }
1076 opened = 1;
1077 }
1078
1079 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1080 f->flags);
1081 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1082 break;
1083
1084 if (opened)
1085 td_io_close_file(td, f);
1086
1087 f = NULL;
1088 } while (td->next_file != old_next_file);
1089
1090 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1091 return f;
1092}
1093
1094static struct fio_file *__get_next_file(struct thread_data *td)
1095{
1096 struct fio_file *f;
1097
1098 assert(td->o.nr_files <= td->files_index);
1099
1100 if (td->nr_done_files >= td->o.nr_files) {
1101 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1102 " nr_files=%d\n", td->nr_open_files,
1103 td->nr_done_files,
1104 td->o.nr_files);
1105 return NULL;
1106 }
1107
1108 f = td->file_service_file;
1109 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1110 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1111 goto out;
1112 if (td->file_service_left--)
1113 goto out;
1114 }
1115
1116 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1117 td->o.file_service_type == FIO_FSERVICE_SEQ)
1118 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1119 else
1120 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1121
1122 if (IS_ERR(f))
1123 return f;
1124
1125 td->file_service_file = f;
1126 td->file_service_left = td->file_service_nr - 1;
1127out:
1128 if (f)
1129 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1130 else
1131 dprint(FD_FILE, "get_next_file: NULL\n");
1132 return f;
1133}
1134
1135static struct fio_file *get_next_file(struct thread_data *td)
1136{
1137 if (td->flags & TD_F_PROFILE_OPS) {
1138 struct prof_io_ops *ops = &td->prof_io_ops;
1139
1140 if (ops->get_next_file)
1141 return ops->get_next_file(td);
1142 }
1143
1144 return __get_next_file(td);
1145}
1146
1147static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1148{
1149 struct fio_file *f;
1150
1151 do {
1152 f = get_next_file(td);
1153 if (IS_ERR_OR_NULL(f))
1154 return PTR_ERR(f);
1155
1156 io_u->file = f;
1157 get_file(f);
1158
1159 if (!fill_io_u(td, io_u))
1160 break;
1161
1162 put_file_log(td, f);
1163 td_io_close_file(td, f);
1164 io_u->file = NULL;
1165 fio_file_set_done(f);
1166 td->nr_done_files++;
1167 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1168 td->nr_done_files, td->o.nr_files);
1169 } while (1);
1170
1171 return 0;
1172}
1173
1174static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1175 unsigned long tusec, unsigned long max_usec)
1176{
1177 if (!td->error)
1178 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1179 td_verror(td, ETIMEDOUT, "max latency exceeded");
1180 icd->error = ETIMEDOUT;
1181}
1182
1183static void lat_new_cycle(struct thread_data *td)
1184{
1185 fio_gettime(&td->latency_ts, NULL);
1186 td->latency_ios = ddir_rw_sum(td->io_blocks);
1187 td->latency_failed = 0;
1188}
1189
1190/*
1191 * We had an IO outside the latency target. Reduce the queue depth. If we
1192 * are at QD=1, then it's time to give up.
1193 */
1194static int __lat_target_failed(struct thread_data *td)
1195{
1196 if (td->latency_qd == 1)
1197 return 1;
1198
1199 td->latency_qd_high = td->latency_qd;
1200
1201 if (td->latency_qd == td->latency_qd_low)
1202 td->latency_qd_low--;
1203
1204 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1205
1206 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1207
1208 /*
1209 * When we ramp QD down, quiesce existing IO to prevent
1210 * a storm of ramp downs due to pending higher depth.
1211 */
1212 io_u_quiesce(td);
1213 lat_new_cycle(td);
1214 return 0;
1215}
1216
1217static int lat_target_failed(struct thread_data *td)
1218{
1219 if (td->o.latency_percentile.u.f == 100.0)
1220 return __lat_target_failed(td);
1221
1222 td->latency_failed++;
1223 return 0;
1224}
1225
1226void lat_target_init(struct thread_data *td)
1227{
1228 td->latency_end_run = 0;
1229
1230 if (td->o.latency_target) {
1231 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1232 fio_gettime(&td->latency_ts, NULL);
1233 td->latency_qd = 1;
1234 td->latency_qd_high = td->o.iodepth;
1235 td->latency_qd_low = 1;
1236 td->latency_ios = ddir_rw_sum(td->io_blocks);
1237 } else
1238 td->latency_qd = td->o.iodepth;
1239}
1240
1241void lat_target_reset(struct thread_data *td)
1242{
1243 if (!td->latency_end_run)
1244 lat_target_init(td);
1245}
1246
1247static void lat_target_success(struct thread_data *td)
1248{
1249 const unsigned int qd = td->latency_qd;
1250 struct thread_options *o = &td->o;
1251
1252 td->latency_qd_low = td->latency_qd;
1253
1254 /*
1255 * If we haven't failed yet, we double up to a failing value instead
1256 * of bisecting from highest possible queue depth. If we have set
1257 * a limit other than td->o.iodepth, bisect between that.
1258 */
1259 if (td->latency_qd_high != o->iodepth)
1260 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1261 else
1262 td->latency_qd *= 2;
1263
1264 if (td->latency_qd > o->iodepth)
1265 td->latency_qd = o->iodepth;
1266
1267 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1268
1269 /*
1270 * Same as last one, we are done. Let it run a latency cycle, so
1271 * we get only the results from the targeted depth.
1272 */
1273 if (td->latency_qd == qd) {
1274 if (td->latency_end_run) {
1275 dprint(FD_RATE, "We are done\n");
1276 td->done = 1;
1277 } else {
1278 dprint(FD_RATE, "Quiesce and final run\n");
1279 io_u_quiesce(td);
1280 td->latency_end_run = 1;
1281 reset_all_stats(td);
1282 reset_io_stats(td);
1283 }
1284 }
1285
1286 lat_new_cycle(td);
1287}
1288
1289/*
1290 * Check if we can bump the queue depth
1291 */
1292void lat_target_check(struct thread_data *td)
1293{
1294 uint64_t usec_window;
1295 uint64_t ios;
1296 double success_ios;
1297
1298 usec_window = utime_since_now(&td->latency_ts);
1299 if (usec_window < td->o.latency_window)
1300 return;
1301
1302 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1303 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1304 success_ios *= 100.0;
1305
1306 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1307
1308 if (success_ios >= td->o.latency_percentile.u.f)
1309 lat_target_success(td);
1310 else
1311 __lat_target_failed(td);
1312}
1313
1314/*
1315 * If latency target is enabled, we might be ramping up or down and not
1316 * using the full queue depth available.
1317 */
1318int queue_full(const struct thread_data *td)
1319{
1320 const int qempty = io_u_qempty(&td->io_u_freelist);
1321
1322 if (qempty)
1323 return 1;
1324 if (!td->o.latency_target)
1325 return 0;
1326
1327 return td->cur_depth >= td->latency_qd;
1328}
1329
1330struct io_u *__get_io_u(struct thread_data *td)
1331{
1332 struct io_u *io_u = NULL;
1333
1334 if (td->stop_io)
1335 return NULL;
1336
1337 td_io_u_lock(td);
1338
1339again:
1340 if (!io_u_rempty(&td->io_u_requeues))
1341 io_u = io_u_rpop(&td->io_u_requeues);
1342 else if (!queue_full(td)) {
1343 io_u = io_u_qpop(&td->io_u_freelist);
1344
1345 io_u->file = NULL;
1346 io_u->buflen = 0;
1347 io_u->resid = 0;
1348 io_u->end_io = NULL;
1349 }
1350
1351 if (io_u) {
1352 assert(io_u->flags & IO_U_F_FREE);
1353 io_u_clear(io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1354 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1355 IO_U_F_VER_LIST);
1356
1357 io_u->error = 0;
1358 io_u->acct_ddir = -1;
1359 td->cur_depth++;
1360 assert(!(td->flags & TD_F_CHILD));
1361 io_u_set(io_u, IO_U_F_IN_CUR_DEPTH);
1362 io_u->ipo = NULL;
1363 } else if (td_async_processing(td)) {
1364 /*
1365 * We ran out, wait for async verify threads to finish and
1366 * return one
1367 */
1368 assert(!(td->flags & TD_F_CHILD));
1369 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1370 goto again;
1371 }
1372
1373 td_io_u_unlock(td);
1374 return io_u;
1375}
1376
1377static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1378{
1379 if (!(td->flags & TD_F_TRIM_BACKLOG))
1380 return 0;
1381
1382 if (td->trim_entries) {
1383 int get_trim = 0;
1384
1385 if (td->trim_batch) {
1386 td->trim_batch--;
1387 get_trim = 1;
1388 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1389 td->last_ddir != DDIR_READ) {
1390 td->trim_batch = td->o.trim_batch;
1391 if (!td->trim_batch)
1392 td->trim_batch = td->o.trim_backlog;
1393 get_trim = 1;
1394 }
1395
1396 if (get_trim && !get_next_trim(td, io_u))
1397 return 1;
1398 }
1399
1400 return 0;
1401}
1402
1403static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1404{
1405 if (!(td->flags & TD_F_VER_BACKLOG))
1406 return 0;
1407
1408 if (td->io_hist_len) {
1409 int get_verify = 0;
1410
1411 if (td->verify_batch)
1412 get_verify = 1;
1413 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1414 td->last_ddir != DDIR_READ) {
1415 td->verify_batch = td->o.verify_batch;
1416 if (!td->verify_batch)
1417 td->verify_batch = td->o.verify_backlog;
1418 get_verify = 1;
1419 }
1420
1421 if (get_verify && !get_next_verify(td, io_u)) {
1422 td->verify_batch--;
1423 return 1;
1424 }
1425 }
1426
1427 return 0;
1428}
1429
1430/*
1431 * Fill offset and start time into the buffer content, to prevent too
1432 * easy compressible data for simple de-dupe attempts. Do this for every
1433 * 512b block in the range, since that should be the smallest block size
1434 * we can expect from a device.
1435 */
1436static void small_content_scramble(struct io_u *io_u)
1437{
1438 unsigned int i, nr_blocks = io_u->buflen / 512;
1439 uint64_t boffset;
1440 unsigned int offset;
1441 void *p, *end;
1442
1443 if (!nr_blocks)
1444 return;
1445
1446 p = io_u->xfer_buf;
1447 boffset = io_u->offset;
1448 io_u->buf_filled_len = 0;
1449
1450 for (i = 0; i < nr_blocks; i++) {
1451 /*
1452 * Fill the byte offset into a "random" start offset of
1453 * the buffer, given by the product of the usec time
1454 * and the actual offset.
1455 */
1456 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1457 offset &= ~(sizeof(uint64_t) - 1);
1458 if (offset >= 512 - sizeof(uint64_t))
1459 offset -= sizeof(uint64_t);
1460 memcpy(p + offset, &boffset, sizeof(boffset));
1461
1462 end = p + 512 - sizeof(io_u->start_time);
1463 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1464 p += 512;
1465 boffset += 512;
1466 }
1467}
1468
1469/*
1470 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1471 * etc. The returned io_u is fully ready to be prepped and submitted.
1472 */
1473struct io_u *get_io_u(struct thread_data *td)
1474{
1475 struct fio_file *f;
1476 struct io_u *io_u;
1477 int do_scramble = 0;
1478 long ret = 0;
1479
1480 io_u = __get_io_u(td);
1481 if (!io_u) {
1482 dprint(FD_IO, "__get_io_u failed\n");
1483 return NULL;
1484 }
1485
1486 if (check_get_verify(td, io_u))
1487 goto out;
1488 if (check_get_trim(td, io_u))
1489 goto out;
1490
1491 /*
1492 * from a requeue, io_u already setup
1493 */
1494 if (io_u->file)
1495 goto out;
1496
1497 /*
1498 * If using an iolog, grab next piece if any available.
1499 */
1500 if (td->flags & TD_F_READ_IOLOG) {
1501 if (read_iolog_get(td, io_u))
1502 goto err_put;
1503 } else if (set_io_u_file(td, io_u)) {
1504 ret = -EBUSY;
1505 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1506 goto err_put;
1507 }
1508
1509 f = io_u->file;
1510 if (!f) {
1511 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1512 goto err_put;
1513 }
1514
1515 assert(fio_file_open(f));
1516
1517 if (ddir_rw(io_u->ddir)) {
1518 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1519 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1520 goto err_put;
1521 }
1522
1523 f->last_start[io_u->ddir] = io_u->offset;
1524 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1525
1526 if (io_u->ddir == DDIR_WRITE) {
1527 if (td->flags & TD_F_REFILL_BUFFERS) {
1528 io_u_fill_buffer(td, io_u,
1529 td->o.min_bs[DDIR_WRITE],
1530 io_u->buflen);
1531 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1532 !(td->flags & TD_F_COMPRESS))
1533 do_scramble = 1;
1534 if (td->flags & TD_F_VER_NONE) {
1535 populate_verify_io_u(td, io_u);
1536 do_scramble = 0;
1537 }
1538 } else if (io_u->ddir == DDIR_READ) {
1539 /*
1540 * Reset the buf_filled parameters so next time if the
1541 * buffer is used for writes it is refilled.
1542 */
1543 io_u->buf_filled_len = 0;
1544 }
1545 }
1546
1547 /*
1548 * Set io data pointers.
1549 */
1550 io_u->xfer_buf = io_u->buf;
1551 io_u->xfer_buflen = io_u->buflen;
1552
1553out:
1554 assert(io_u->file);
1555 if (!td_io_prep(td, io_u)) {
1556 if (!td->o.disable_slat)
1557 fio_gettime(&io_u->start_time, NULL);
1558 if (do_scramble)
1559 small_content_scramble(io_u);
1560 return io_u;
1561 }
1562err_put:
1563 dprint(FD_IO, "get_io_u failed\n");
1564 put_io_u(td, io_u);
1565 return ERR_PTR(ret);
1566}
1567
1568static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1569{
1570 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1571
1572 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1573 return;
1574
1575 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1576 io_u->file ? " on file " : "",
1577 io_u->file ? io_u->file->file_name : "",
1578 strerror(io_u->error),
1579 io_ddir_name(io_u->ddir),
1580 io_u->offset, io_u->xfer_buflen);
1581
1582 if (td->io_ops->errdetails) {
1583 char *err = td->io_ops->errdetails(io_u);
1584
1585 log_err("fio: %s\n", err);
1586 free(err);
1587 }
1588
1589 if (!td->error)
1590 td_verror(td, io_u->error, "io_u error");
1591}
1592
1593void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1594{
1595 __io_u_log_error(td, io_u);
1596 if (td->parent)
1597 __io_u_log_error(td, io_u);
1598}
1599
1600static inline int gtod_reduce(struct thread_data *td)
1601{
1602 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1603 && td->o.disable_bw;
1604}
1605
1606static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1607 struct io_completion_data *icd,
1608 const enum fio_ddir idx, unsigned int bytes)
1609{
1610 const int no_reduce = !gtod_reduce(td);
1611 unsigned long lusec = 0;
1612
1613 if (td->parent)
1614 td = td->parent;
1615
1616 if (no_reduce)
1617 lusec = utime_since(&io_u->issue_time, &icd->time);
1618
1619 if (!td->o.disable_lat) {
1620 unsigned long tusec;
1621
1622 tusec = utime_since(&io_u->start_time, &icd->time);
1623 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1624
1625 if (td->flags & TD_F_PROFILE_OPS) {
1626 struct prof_io_ops *ops = &td->prof_io_ops;
1627
1628 if (ops->io_u_lat)
1629 icd->error = ops->io_u_lat(td, tusec);
1630 }
1631
1632 if (td->o.max_latency && tusec > td->o.max_latency)
1633 lat_fatal(td, icd, tusec, td->o.max_latency);
1634 if (td->o.latency_target && tusec > td->o.latency_target) {
1635 if (lat_target_failed(td))
1636 lat_fatal(td, icd, tusec, td->o.latency_target);
1637 }
1638 }
1639
1640 if (!td->o.disable_clat) {
1641 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1642 io_u_mark_latency(td, lusec);
1643 }
1644
1645 if (!td->o.disable_bw)
1646 add_bw_sample(td, idx, bytes, &icd->time);
1647
1648 if (no_reduce)
1649 add_iops_sample(td, idx, bytes, &icd->time);
1650
1651 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1652 uint32_t *info = io_u_block_info(td, io_u);
1653 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1654 if (io_u->ddir == DDIR_TRIM) {
1655 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1656 BLOCK_INFO_TRIMS(*info) + 1);
1657 } else if (io_u->ddir == DDIR_WRITE) {
1658 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1659 *info);
1660 }
1661 }
1662 }
1663}
1664
1665static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1666 struct io_completion_data *icd)
1667{
1668 struct io_u *io_u = *io_u_ptr;
1669 enum fio_ddir ddir = io_u->ddir;
1670 struct fio_file *f = io_u->file;
1671
1672 dprint_io_u(io_u, "io complete");
1673
1674 assert(io_u->flags & IO_U_F_FLIGHT);
1675 io_u_clear(io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1676
1677 /*
1678 * Mark IO ok to verify
1679 */
1680 if (io_u->ipo) {
1681 /*
1682 * Remove errored entry from the verification list
1683 */
1684 if (io_u->error)
1685 unlog_io_piece(td, io_u);
1686 else {
1687 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1688 write_barrier();
1689 }
1690 }
1691
1692 if (ddir_sync(ddir)) {
1693 td->last_was_sync = 1;
1694 if (f) {
1695 f->first_write = -1ULL;
1696 f->last_write = -1ULL;
1697 }
1698 return;
1699 }
1700
1701 td->last_was_sync = 0;
1702 td->last_ddir = ddir;
1703
1704 if (!io_u->error && ddir_rw(ddir)) {
1705 unsigned int bytes = io_u->buflen - io_u->resid;
1706 int ret;
1707
1708 td->io_blocks[ddir]++;
1709 td->this_io_blocks[ddir]++;
1710 td->io_bytes[ddir] += bytes;
1711
1712 if (!(io_u->flags & IO_U_F_VER_LIST))
1713 td->this_io_bytes[ddir] += bytes;
1714
1715 if (ddir == DDIR_WRITE) {
1716 if (f) {
1717 if (f->first_write == -1ULL ||
1718 io_u->offset < f->first_write)
1719 f->first_write = io_u->offset;
1720 if (f->last_write == -1ULL ||
1721 ((io_u->offset + bytes) > f->last_write))
1722 f->last_write = io_u->offset + bytes;
1723 }
1724 if (td->last_write_comp) {
1725 int idx = td->last_write_idx++;
1726
1727 td->last_write_comp[idx] = io_u->offset;
1728 if (td->last_write_idx == td->o.iodepth)
1729 td->last_write_idx = 0;
1730 }
1731 }
1732
1733 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1734 td->runstate == TD_VERIFYING))
1735 account_io_completion(td, io_u, icd, ddir, bytes);
1736
1737 icd->bytes_done[ddir] += bytes;
1738
1739 if (io_u->end_io) {
1740 ret = io_u->end_io(td, io_u_ptr);
1741 io_u = *io_u_ptr;
1742 if (ret && !icd->error)
1743 icd->error = ret;
1744 }
1745 } else if (io_u->error) {
1746 icd->error = io_u->error;
1747 io_u_log_error(td, io_u);
1748 }
1749 if (icd->error) {
1750 enum error_type_bit eb = td_error_type(ddir, icd->error);
1751
1752 if (!td_non_fatal_error(td, eb, icd->error))
1753 return;
1754
1755 /*
1756 * If there is a non_fatal error, then add to the error count
1757 * and clear all the errors.
1758 */
1759 update_error_count(td, icd->error);
1760 td_clear_error(td);
1761 icd->error = 0;
1762 if (io_u)
1763 io_u->error = 0;
1764 }
1765}
1766
1767static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1768 int nr)
1769{
1770 int ddir;
1771
1772 if (!gtod_reduce(td))
1773 fio_gettime(&icd->time, NULL);
1774
1775 icd->nr = nr;
1776
1777 icd->error = 0;
1778 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1779 icd->bytes_done[ddir] = 0;
1780}
1781
1782static void ios_completed(struct thread_data *td,
1783 struct io_completion_data *icd)
1784{
1785 struct io_u *io_u;
1786 int i;
1787
1788 for (i = 0; i < icd->nr; i++) {
1789 io_u = td->io_ops->event(td, i);
1790
1791 io_completed(td, &io_u, icd);
1792
1793 if (io_u)
1794 put_io_u(td, io_u);
1795 }
1796}
1797
1798/*
1799 * Complete a single io_u for the sync engines.
1800 */
1801int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
1802{
1803 struct io_completion_data icd;
1804 int ddir;
1805
1806 init_icd(td, &icd, 1);
1807 io_completed(td, &io_u, &icd);
1808
1809 if (io_u)
1810 put_io_u(td, io_u);
1811
1812 if (icd.error) {
1813 td_verror(td, icd.error, "io_u_sync_complete");
1814 return -1;
1815 }
1816
1817 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1818 td->bytes_done[ddir] += icd.bytes_done[ddir];
1819
1820 return 0;
1821}
1822
1823/*
1824 * Called to complete min_events number of io for the async engines.
1825 */
1826int io_u_queued_complete(struct thread_data *td, int min_evts)
1827{
1828 struct io_completion_data icd;
1829 struct timespec *tvp = NULL;
1830 int ret, ddir;
1831 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1832
1833 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1834
1835 if (!min_evts)
1836 tvp = &ts;
1837 else if (min_evts > td->cur_depth)
1838 min_evts = td->cur_depth;
1839
1840 /* No worries, td_io_getevents fixes min and max if they are
1841 * set incorrectly */
1842 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
1843 if (ret < 0) {
1844 td_verror(td, -ret, "td_io_getevents");
1845 return ret;
1846 } else if (!ret)
1847 return ret;
1848
1849 init_icd(td, &icd, ret);
1850 ios_completed(td, &icd);
1851 if (icd.error) {
1852 td_verror(td, icd.error, "io_u_queued_complete");
1853 return -1;
1854 }
1855
1856 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1857 td->bytes_done[ddir] += icd.bytes_done[ddir];
1858
1859 return 0;
1860}
1861
1862/*
1863 * Call when io_u is really queued, to update the submission latency.
1864 */
1865void io_u_queued(struct thread_data *td, struct io_u *io_u)
1866{
1867 if (!td->o.disable_slat) {
1868 unsigned long slat_time;
1869
1870 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1871
1872 if (td->parent)
1873 td = td->parent;
1874
1875 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1876 io_u->offset);
1877 }
1878}
1879
1880/*
1881 * See if we should reuse the last seed, if dedupe is enabled
1882 */
1883static struct frand_state *get_buf_state(struct thread_data *td)
1884{
1885 uint64_t frand_max;
1886 unsigned int v;
1887 unsigned long r;
1888
1889 if (!td->o.dedupe_percentage)
1890 return &td->buf_state;
1891 else if (td->o.dedupe_percentage == 100) {
1892 frand_copy(&td->buf_state_prev, &td->buf_state);
1893 return &td->buf_state;
1894 }
1895
1896 frand_max = rand_max(&td->dedupe_state);
1897 r = __rand(&td->dedupe_state);
1898 v = 1 + (int) (100.0 * (r / (frand_max + 1.0)));
1899
1900 if (v <= td->o.dedupe_percentage)
1901 return &td->buf_state_prev;
1902
1903 return &td->buf_state;
1904}
1905
1906static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1907{
1908 if (td->o.dedupe_percentage == 100)
1909 frand_copy(rs, &td->buf_state_prev);
1910 else if (rs == &td->buf_state)
1911 frand_copy(&td->buf_state_prev, rs);
1912}
1913
1914void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1915 unsigned int max_bs)
1916{
1917 struct thread_options *o = &td->o;
1918
1919 if (o->compress_percentage || o->dedupe_percentage) {
1920 unsigned int perc = td->o.compress_percentage;
1921 struct frand_state *rs;
1922 unsigned int left = max_bs;
1923 unsigned int this_write;
1924
1925 do {
1926 rs = get_buf_state(td);
1927
1928 min_write = min(min_write, left);
1929
1930 if (perc) {
1931 this_write = min_not_zero(min_write,
1932 td->o.compress_chunk);
1933
1934 fill_random_buf_percentage(rs, buf, perc,
1935 this_write, this_write,
1936 o->buffer_pattern,
1937 o->buffer_pattern_bytes);
1938 } else {
1939 fill_random_buf(rs, buf, min_write);
1940 this_write = min_write;
1941 }
1942
1943 buf += this_write;
1944 left -= this_write;
1945 save_buf_state(td, rs);
1946 } while (left);
1947 } else if (o->buffer_pattern_bytes)
1948 fill_buffer_pattern(td, buf, max_bs);
1949 else if (o->zero_buffers)
1950 memset(buf, 0, max_bs);
1951 else
1952 fill_random_buf(get_buf_state(td), buf, max_bs);
1953}
1954
1955/*
1956 * "randomly" fill the buffer contents
1957 */
1958void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1959 unsigned int min_write, unsigned int max_bs)
1960{
1961 io_u->buf_filled_len = 0;
1962 fill_io_buffer(td, io_u->buf, min_write, max_bs);
1963}
fio - Flexible IO Tester