projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
iolog: account for wasted time in ipo stall
[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
16struct io_completion_data {
17 int nr; /* input */
18
19 int error; /* output */
20 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
21 struct timeval time; /* output */
22};
23
24/*
25 * The ->io_axmap contains a map of blocks we have or have not done io
26 * to yet. Used to make sure we cover the entire range in a fair fashion.
27 */
28static int random_map_free(struct fio_file *f, const uint64_t block)
29{
30 return !axmap_isset(f->io_axmap, block);
31}
32
33/*
34 * Mark a given offset as used in the map.
35 */
36static void mark_random_map(struct thread_data *td, struct io_u *io_u)
37{
38 unsigned int min_bs = td->o.rw_min_bs;
39 struct fio_file *f = io_u->file;
40 unsigned int nr_blocks;
41 uint64_t block;
42
43 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
44 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
45
46 if (!(io_u->flags & IO_U_F_BUSY_OK))
47 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
48
49 if ((nr_blocks * min_bs) < io_u->buflen)
50 io_u->buflen = nr_blocks * min_bs;
51}
52
53static uint64_t last_block(struct thread_data *td, struct fio_file *f,
54 enum fio_ddir ddir)
55{
56 uint64_t max_blocks;
57 uint64_t max_size;
58
59 assert(ddir_rw(ddir));
60
61 /*
62 * Hmm, should we make sure that ->io_size <= ->real_file_size?
63 */
64 max_size = f->io_size;
65 if (max_size > f->real_file_size)
66 max_size = f->real_file_size;
67
68 if (td->o.zone_range)
69 max_size = td->o.zone_range;
70
71 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
72 if (!max_blocks)
73 return 0;
74
75 return max_blocks;
76}
77
78struct rand_off {
79 struct flist_head list;
80 uint64_t off;
81};
82
83static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
84 enum fio_ddir ddir, uint64_t *b)
85{
86 uint64_t r, lastb;
87
88 lastb = last_block(td, f, ddir);
89 if (!lastb)
90 return 1;
91
92 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
93 r = __rand(&td->random_state);
94
95 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
96
97 *b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0));
98 } else {
99 uint64_t off = 0;
100
101 if (lfsr_next(&f->lfsr, &off, lastb))
102 return 1;
103
104 *b = off;
105 }
106
107 /*
108 * if we are not maintaining a random map, we are done.
109 */
110 if (!file_randommap(td, f))
111 goto ret;
112
113 /*
114 * calculate map offset and check if it's free
115 */
116 if (random_map_free(f, *b))
117 goto ret;
118
119 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
120 (unsigned long long) *b);
121
122 *b = axmap_next_free(f->io_axmap, *b);
123 if (*b == (uint64_t) -1ULL)
124 return 1;
125ret:
126 return 0;
127}
128
129static int __get_next_rand_offset_zipf(struct thread_data *td,
130 struct fio_file *f, enum fio_ddir ddir,
131 uint64_t *b)
132{
133 *b = zipf_next(&f->zipf);
134 return 0;
135}
136
137static int __get_next_rand_offset_pareto(struct thread_data *td,
138 struct fio_file *f, enum fio_ddir ddir,
139 uint64_t *b)
140{
141 *b = pareto_next(&f->zipf);
142 return 0;
143}
144
145static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
146{
147 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
148 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
149
150 return r1->off - r2->off;
151}
152
153static int get_off_from_method(struct thread_data *td, struct fio_file *f,
154 enum fio_ddir ddir, uint64_t *b)
155{
156 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
157 return __get_next_rand_offset(td, f, ddir, b);
158 else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
159 return __get_next_rand_offset_zipf(td, f, ddir, b);
160 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
161 return __get_next_rand_offset_pareto(td, f, ddir, b);
162
163 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
164 return 1;
165}
166
167/*
168 * Sort the reads for a verify phase in batches of verifysort_nr, if
169 * specified.
170 */
171static inline int should_sort_io(struct thread_data *td)
172{
173 if (!td->o.verifysort_nr || !td->o.do_verify)
174 return 0;
175 if (!td_random(td))
176 return 0;
177 if (td->runstate != TD_VERIFYING)
178 return 0;
179 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
180 return 0;
181
182 return 1;
183}
184
185static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
186{
187 unsigned int v;
188 unsigned long r;
189
190 if (td->o.perc_rand[ddir] == 100)
191 return 1;
192
193 r = __rand(&td->seq_rand_state[ddir]);
194 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
195
196 return v <= td->o.perc_rand[ddir];
197}
198
199static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
200 enum fio_ddir ddir, uint64_t *b)
201{
202 struct rand_off *r;
203 int i, ret = 1;
204
205 if (!should_sort_io(td))
206 return get_off_from_method(td, f, ddir, b);
207
208 if (!flist_empty(&td->next_rand_list)) {
209fetch:
210 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
211 flist_del(&r->list);
212 *b = r->off;
213 free(r);
214 return 0;
215 }
216
217 for (i = 0; i < td->o.verifysort_nr; i++) {
218 r = malloc(sizeof(*r));
219
220 ret = get_off_from_method(td, f, ddir, &r->off);
221 if (ret) {
222 free(r);
223 break;
224 }
225
226 flist_add(&r->list, &td->next_rand_list);
227 }
228
229 if (ret && !i)
230 return ret;
231
232 assert(!flist_empty(&td->next_rand_list));
233 flist_sort(NULL, &td->next_rand_list, flist_cmp);
234 goto fetch;
235}
236
237static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
238 enum fio_ddir ddir, uint64_t *b)
239{
240 if (!get_next_rand_offset(td, f, ddir, b))
241 return 0;
242
243 if (td->o.time_based) {
244 fio_file_reset(td, f);
245 if (!get_next_rand_offset(td, f, ddir, b))
246 return 0;
247 }
248
249 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
250 f->file_name, (unsigned long long) f->last_pos,
251 (unsigned long long) f->real_file_size);
252 return 1;
253}
254
255static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
256 enum fio_ddir ddir, uint64_t *offset)
257{
258 struct thread_options *o = &td->o;
259
260 assert(ddir_rw(ddir));
261
262 if (f->last_pos >= f->io_size + get_start_offset(td, f) &&
263 o->time_based)
264 f->last_pos = f->last_pos - f->io_size;
265
266 if (f->last_pos < f->real_file_size) {
267 uint64_t pos;
268
269 if (f->last_pos == f->file_offset && o->ddir_seq_add < 0)
270 f->last_pos = f->real_file_size;
271
272 pos = f->last_pos - f->file_offset;
273 if (pos && o->ddir_seq_add) {
274 pos += o->ddir_seq_add;
275
276 /*
277 * If we reach beyond the end of the file
278 * with holed IO, wrap around to the
279 * beginning again.
280 */
281 if (pos >= f->real_file_size)
282 pos = f->file_offset;
283 }
284
285 *offset = pos;
286 return 0;
287 }
288
289 return 1;
290}
291
292static int get_next_block(struct thread_data *td, struct io_u *io_u,
293 enum fio_ddir ddir, int rw_seq,
294 unsigned int *is_random)
295{
296 struct fio_file *f = io_u->file;
297 uint64_t b, offset;
298 int ret;
299
300 assert(ddir_rw(ddir));
301
302 b = offset = -1ULL;
303
304 if (rw_seq) {
305 if (td_random(td)) {
306 if (should_do_random(td, ddir)) {
307 ret = get_next_rand_block(td, f, ddir, &b);
308 *is_random = 1;
309 } else {
310 *is_random = 0;
311 io_u->flags |= IO_U_F_BUSY_OK;
312 ret = get_next_seq_offset(td, f, ddir, &offset);
313 if (ret)
314 ret = get_next_rand_block(td, f, ddir, &b);
315 }
316 } else {
317 *is_random = 0;
318 ret = get_next_seq_offset(td, f, ddir, &offset);
319 }
320 } else {
321 io_u->flags |= IO_U_F_BUSY_OK;
322 *is_random = 0;
323
324 if (td->o.rw_seq == RW_SEQ_SEQ) {
325 ret = get_next_seq_offset(td, f, ddir, &offset);
326 if (ret) {
327 ret = get_next_rand_block(td, f, ddir, &b);
328 *is_random = 0;
329 }
330 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
331 if (f->last_start != -1ULL)
332 offset = f->last_start - f->file_offset;
333 else
334 offset = 0;
335 ret = 0;
336 } else {
337 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
338 ret = 1;
339 }
340 }
341
342 if (!ret) {
343 if (offset != -1ULL)
344 io_u->offset = offset;
345 else if (b != -1ULL)
346 io_u->offset = b * td->o.ba[ddir];
347 else {
348 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
349 ret = 1;
350 }
351 }
352
353 return ret;
354}
355
356/*
357 * For random io, generate a random new block and see if it's used. Repeat
358 * until we find a free one. For sequential io, just return the end of
359 * the last io issued.
360 */
361static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
362 unsigned int *is_random)
363{
364 struct fio_file *f = io_u->file;
365 enum fio_ddir ddir = io_u->ddir;
366 int rw_seq_hit = 0;
367
368 assert(ddir_rw(ddir));
369
370 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
371 rw_seq_hit = 1;
372 td->ddir_seq_nr = td->o.ddir_seq_nr;
373 }
374
375 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
376 return 1;
377
378 if (io_u->offset >= f->io_size) {
379 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
380 (unsigned long long) io_u->offset,
381 (unsigned long long) f->io_size);
382 return 1;
383 }
384
385 io_u->offset += f->file_offset;
386 if (io_u->offset >= f->real_file_size) {
387 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
388 (unsigned long long) io_u->offset,
389 (unsigned long long) f->real_file_size);
390 return 1;
391 }
392
393 return 0;
394}
395
396static int get_next_offset(struct thread_data *td, struct io_u *io_u,
397 unsigned int *is_random)
398{
399 if (td->flags & TD_F_PROFILE_OPS) {
400 struct prof_io_ops *ops = &td->prof_io_ops;
401
402 if (ops->fill_io_u_off)
403 return ops->fill_io_u_off(td, io_u, is_random);
404 }
405
406 return __get_next_offset(td, io_u, is_random);
407}
408
409static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
410 unsigned int buflen)
411{
412 struct fio_file *f = io_u->file;
413
414 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
415}
416
417static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
418 unsigned int is_random)
419{
420 int ddir = io_u->ddir;
421 unsigned int buflen = 0;
422 unsigned int minbs, maxbs;
423 unsigned long r;
424
425 assert(ddir_rw(ddir));
426
427 if (td->o.bs_is_seq_rand)
428 ddir = is_random ? DDIR_WRITE: DDIR_READ;
429
430 minbs = td->o.min_bs[ddir];
431 maxbs = td->o.max_bs[ddir];
432
433 if (minbs == maxbs)
434 return minbs;
435
436 /*
437 * If we can't satisfy the min block size from here, then fail
438 */
439 if (!io_u_fits(td, io_u, minbs))
440 return 0;
441
442 do {
443 r = __rand(&td->bsrange_state);
444
445 if (!td->o.bssplit_nr[ddir]) {
446 buflen = 1 + (unsigned int) ((double) maxbs *
447 (r / (FRAND_MAX + 1.0)));
448 if (buflen < minbs)
449 buflen = minbs;
450 } else {
451 long perc = 0;
452 unsigned int i;
453
454 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
455 struct bssplit *bsp = &td->o.bssplit[ddir][i];
456
457 buflen = bsp->bs;
458 perc += bsp->perc;
459 if ((r <= ((FRAND_MAX / 100L) * perc)) &&
460 io_u_fits(td, io_u, buflen))
461 break;
462 }
463 }
464
465 if (td->o.do_verify && td->o.verify != VERIFY_NONE)
466 buflen = (buflen + td->o.verify_interval - 1) &
467 ~(td->o.verify_interval - 1);
468
469 if (!td->o.bs_unaligned && is_power_of_2(minbs))
470 buflen = (buflen + minbs - 1) & ~(minbs - 1);
471
472 } while (!io_u_fits(td, io_u, buflen));
473
474 return buflen;
475}
476
477static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
478 unsigned int is_random)
479{
480 if (td->flags & TD_F_PROFILE_OPS) {
481 struct prof_io_ops *ops = &td->prof_io_ops;
482
483 if (ops->fill_io_u_size)
484 return ops->fill_io_u_size(td, io_u, is_random);
485 }
486
487 return __get_next_buflen(td, io_u, is_random);
488}
489
490static void set_rwmix_bytes(struct thread_data *td)
491{
492 unsigned int diff;
493
494 /*
495 * we do time or byte based switch. this is needed because
496 * buffered writes may issue a lot quicker than they complete,
497 * whereas reads do not.
498 */
499 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
500 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
501}
502
503static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
504{
505 unsigned int v;
506 unsigned long r;
507
508 r = __rand(&td->rwmix_state);
509 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
510
511 if (v <= td->o.rwmix[DDIR_READ])
512 return DDIR_READ;
513
514 return DDIR_WRITE;
515}
516
517void io_u_quiesce(struct thread_data *td)
518{
519 /*
520 * We are going to sleep, ensure that we flush anything pending as
521 * not to skew our latency numbers.
522 *
523 * Changed to only monitor 'in flight' requests here instead of the
524 * td->cur_depth, b/c td->cur_depth does not accurately represent
525 * io's that have been actually submitted to an async engine,
526 * and cur_depth is meaningless for sync engines.
527 */
528 while (td->io_u_in_flight) {
529 int fio_unused ret;
530
531 ret = io_u_queued_complete(td, 1, NULL);
532 }
533}
534
535static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
536{
537 enum fio_ddir odir = ddir ^ 1;
538 struct timeval t;
539 long usec;
540
541 assert(ddir_rw(ddir));
542
543 if (td->rate_pending_usleep[ddir] <= 0)
544 return ddir;
545
546 /*
547 * We have too much pending sleep in this direction. See if we
548 * should switch.
549 */
550 if (td_rw(td) && td->o.rwmix[odir]) {
551 /*
552 * Other direction does not have too much pending, switch
553 */
554 if (td->rate_pending_usleep[odir] < 100000)
555 return odir;
556
557 /*
558 * Both directions have pending sleep. Sleep the minimum time
559 * and deduct from both.
560 */
561 if (td->rate_pending_usleep[ddir] <=
562 td->rate_pending_usleep[odir]) {
563 usec = td->rate_pending_usleep[ddir];
564 } else {
565 usec = td->rate_pending_usleep[odir];
566 ddir = odir;
567 }
568 } else
569 usec = td->rate_pending_usleep[ddir];
570
571 io_u_quiesce(td);
572
573 fio_gettime(&t, NULL);
574 usec_sleep(td, usec);
575 usec = utime_since_now(&t);
576
577 td->rate_pending_usleep[ddir] -= usec;
578
579 odir = ddir ^ 1;
580 if (td_rw(td) && __should_check_rate(td, odir))
581 td->rate_pending_usleep[odir] -= usec;
582
583 if (ddir == DDIR_TRIM)
584 return DDIR_TRIM;
585
586 return ddir;
587}
588
589/*
590 * Return the data direction for the next io_u. If the job is a
591 * mixed read/write workload, check the rwmix cycle and switch if
592 * necessary.
593 */
594static enum fio_ddir get_rw_ddir(struct thread_data *td)
595{
596 enum fio_ddir ddir;
597
598 /*
599 * see if it's time to fsync
600 */
601 if (td->o.fsync_blocks &&
602 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
603 td->io_issues[DDIR_WRITE] && should_fsync(td))
604 return DDIR_SYNC;
605
606 /*
607 * see if it's time to fdatasync
608 */
609 if (td->o.fdatasync_blocks &&
610 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
611 td->io_issues[DDIR_WRITE] && should_fsync(td))
612 return DDIR_DATASYNC;
613
614 /*
615 * see if it's time to sync_file_range
616 */
617 if (td->sync_file_range_nr &&
618 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
619 td->io_issues[DDIR_WRITE] && should_fsync(td))
620 return DDIR_SYNC_FILE_RANGE;
621
622 if (td_rw(td)) {
623 /*
624 * Check if it's time to seed a new data direction.
625 */
626 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
627 /*
628 * Put a top limit on how many bytes we do for
629 * one data direction, to avoid overflowing the
630 * ranges too much
631 */
632 ddir = get_rand_ddir(td);
633
634 if (ddir != td->rwmix_ddir)
635 set_rwmix_bytes(td);
636
637 td->rwmix_ddir = ddir;
638 }
639 ddir = td->rwmix_ddir;
640 } else if (td_read(td))
641 ddir = DDIR_READ;
642 else if (td_write(td))
643 ddir = DDIR_WRITE;
644 else
645 ddir = DDIR_TRIM;
646
647 td->rwmix_ddir = rate_ddir(td, ddir);
648 return td->rwmix_ddir;
649}
650
651static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
652{
653 io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
654
655 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
656 td->o.barrier_blocks &&
657 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
658 td->io_issues[DDIR_WRITE])
659 io_u->flags |= IO_U_F_BARRIER;
660}
661
662void put_file_log(struct thread_data *td, struct fio_file *f)
663{
664 unsigned int ret = put_file(td, f);
665
666 if (ret)
667 td_verror(td, ret, "file close");
668}
669
670void put_io_u(struct thread_data *td, struct io_u *io_u)
671{
672 td_io_u_lock(td);
673
674 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
675 put_file_log(td, io_u->file);
676
677 io_u->file = NULL;
678 io_u->flags |= IO_U_F_FREE;
679
680 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
681 td->cur_depth--;
682 io_u_qpush(&td->io_u_freelist, io_u);
683 td_io_u_unlock(td);
684 td_io_u_free_notify(td);
685}
686
687void clear_io_u(struct thread_data *td, struct io_u *io_u)
688{
689 io_u->flags &= ~IO_U_F_FLIGHT;
690 put_io_u(td, io_u);
691}
692
693void requeue_io_u(struct thread_data *td, struct io_u **io_u)
694{
695 struct io_u *__io_u = *io_u;
696 enum fio_ddir ddir = acct_ddir(__io_u);
697
698 dprint(FD_IO, "requeue %p\n", __io_u);
699
700 td_io_u_lock(td);
701
702 __io_u->flags |= IO_U_F_FREE;
703 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
704 td->io_issues[ddir]--;
705
706 __io_u->flags &= ~IO_U_F_FLIGHT;
707 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
708 td->cur_depth--;
709
710 io_u_rpush(&td->io_u_requeues, __io_u);
711 td_io_u_unlock(td);
712 *io_u = NULL;
713}
714
715static int fill_io_u(struct thread_data *td, struct io_u *io_u)
716{
717 unsigned int is_random;
718
719 if (td->io_ops->flags & FIO_NOIO)
720 goto out;
721
722 set_rw_ddir(td, io_u);
723
724 /*
725 * fsync() or fdatasync() or trim etc, we are done
726 */
727 if (!ddir_rw(io_u->ddir))
728 goto out;
729
730 /*
731 * See if it's time to switch to a new zone
732 */
733 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
734 struct fio_file *f = io_u->file;
735
736 td->zone_bytes = 0;
737 f->file_offset += td->o.zone_range + td->o.zone_skip;
738
739 /*
740 * Wrap from the beginning, if we exceed the file size
741 */
742 if (f->file_offset >= f->real_file_size)
743 f->file_offset = f->real_file_size - f->file_offset;
744 f->last_pos = f->file_offset;
745 td->io_skip_bytes += td->o.zone_skip;
746 }
747
748 /*
749 * No log, let the seq/rand engine retrieve the next buflen and
750 * position.
751 */
752 if (get_next_offset(td, io_u, &is_random)) {
753 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
754 return 1;
755 }
756
757 io_u->buflen = get_next_buflen(td, io_u, is_random);
758 if (!io_u->buflen) {
759 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
760 return 1;
761 }
762
763 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
764 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
765 dprint(FD_IO, " off=%llu/%lu > %llu\n",
766 (unsigned long long) io_u->offset, io_u->buflen,
767 (unsigned long long) io_u->file->real_file_size);
768 return 1;
769 }
770
771 /*
772 * mark entry before potentially trimming io_u
773 */
774 if (td_random(td) && file_randommap(td, io_u->file))
775 mark_random_map(td, io_u);
776
777out:
778 dprint_io_u(io_u, "fill_io_u");
779 td->zone_bytes += io_u->buflen;
780 return 0;
781}
782
783static void __io_u_mark_map(unsigned int *map, unsigned int nr)
784{
785 int idx = 0;
786
787 switch (nr) {
788 default:
789 idx = 6;
790 break;
791 case 33 ... 64:
792 idx = 5;
793 break;
794 case 17 ... 32:
795 idx = 4;
796 break;
797 case 9 ... 16:
798 idx = 3;
799 break;
800 case 5 ... 8:
801 idx = 2;
802 break;
803 case 1 ... 4:
804 idx = 1;
805 case 0:
806 break;
807 }
808
809 map[idx]++;
810}
811
812void io_u_mark_submit(struct thread_data *td, unsigned int nr)
813{
814 __io_u_mark_map(td->ts.io_u_submit, nr);
815 td->ts.total_submit++;
816}
817
818void io_u_mark_complete(struct thread_data *td, unsigned int nr)
819{
820 __io_u_mark_map(td->ts.io_u_complete, nr);
821 td->ts.total_complete++;
822}
823
824void io_u_mark_depth(struct thread_data *td, unsigned int nr)
825{
826 int idx = 0;
827
828 switch (td->cur_depth) {
829 default:
830 idx = 6;
831 break;
832 case 32 ... 63:
833 idx = 5;
834 break;
835 case 16 ... 31:
836 idx = 4;
837 break;
838 case 8 ... 15:
839 idx = 3;
840 break;
841 case 4 ... 7:
842 idx = 2;
843 break;
844 case 2 ... 3:
845 idx = 1;
846 case 1:
847 break;
848 }
849
850 td->ts.io_u_map[idx] += nr;
851}
852
853static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
854{
855 int idx = 0;
856
857 assert(usec < 1000);
858
859 switch (usec) {
860 case 750 ... 999:
861 idx = 9;
862 break;
863 case 500 ... 749:
864 idx = 8;
865 break;
866 case 250 ... 499:
867 idx = 7;
868 break;
869 case 100 ... 249:
870 idx = 6;
871 break;
872 case 50 ... 99:
873 idx = 5;
874 break;
875 case 20 ... 49:
876 idx = 4;
877 break;
878 case 10 ... 19:
879 idx = 3;
880 break;
881 case 4 ... 9:
882 idx = 2;
883 break;
884 case 2 ... 3:
885 idx = 1;
886 case 0 ... 1:
887 break;
888 }
889
890 assert(idx < FIO_IO_U_LAT_U_NR);
891 td->ts.io_u_lat_u[idx]++;
892}
893
894static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
895{
896 int idx = 0;
897
898 switch (msec) {
899 default:
900 idx = 11;
901 break;
902 case 1000 ... 1999:
903 idx = 10;
904 break;
905 case 750 ... 999:
906 idx = 9;
907 break;
908 case 500 ... 749:
909 idx = 8;
910 break;
911 case 250 ... 499:
912 idx = 7;
913 break;
914 case 100 ... 249:
915 idx = 6;
916 break;
917 case 50 ... 99:
918 idx = 5;
919 break;
920 case 20 ... 49:
921 idx = 4;
922 break;
923 case 10 ... 19:
924 idx = 3;
925 break;
926 case 4 ... 9:
927 idx = 2;
928 break;
929 case 2 ... 3:
930 idx = 1;
931 case 0 ... 1:
932 break;
933 }
934
935 assert(idx < FIO_IO_U_LAT_M_NR);
936 td->ts.io_u_lat_m[idx]++;
937}
938
939static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
940{
941 if (usec < 1000)
942 io_u_mark_lat_usec(td, usec);
943 else
944 io_u_mark_lat_msec(td, usec / 1000);
945}
946
947/*
948 * Get next file to service by choosing one at random
949 */
950static struct fio_file *get_next_file_rand(struct thread_data *td,
951 enum fio_file_flags goodf,
952 enum fio_file_flags badf)
953{
954 struct fio_file *f;
955 int fno;
956
957 do {
958 int opened = 0;
959 unsigned long r;
960
961 r = __rand(&td->next_file_state);
962 fno = (unsigned int) ((double) td->o.nr_files
963 * (r / (FRAND_MAX + 1.0)));
964
965 f = td->files[fno];
966 if (fio_file_done(f))
967 continue;
968
969 if (!fio_file_open(f)) {
970 int err;
971
972 if (td->nr_open_files >= td->o.open_files)
973 return ERR_PTR(-EBUSY);
974
975 err = td_io_open_file(td, f);
976 if (err)
977 continue;
978 opened = 1;
979 }
980
981 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
982 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
983 return f;
984 }
985 if (opened)
986 td_io_close_file(td, f);
987 } while (1);
988}
989
990/*
991 * Get next file to service by doing round robin between all available ones
992 */
993static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
994 int badf)
995{
996 unsigned int old_next_file = td->next_file;
997 struct fio_file *f;
998
999 do {
1000 int opened = 0;
1001
1002 f = td->files[td->next_file];
1003
1004 td->next_file++;
1005 if (td->next_file >= td->o.nr_files)
1006 td->next_file = 0;
1007
1008 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1009 if (fio_file_done(f)) {
1010 f = NULL;
1011 continue;
1012 }
1013
1014 if (!fio_file_open(f)) {
1015 int err;
1016
1017 if (td->nr_open_files >= td->o.open_files)
1018 return ERR_PTR(-EBUSY);
1019
1020 err = td_io_open_file(td, f);
1021 if (err) {
1022 dprint(FD_FILE, "error %d on open of %s\n",
1023 err, f->file_name);
1024 f = NULL;
1025 continue;
1026 }
1027 opened = 1;
1028 }
1029
1030 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1031 f->flags);
1032 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1033 break;
1034
1035 if (opened)
1036 td_io_close_file(td, f);
1037
1038 f = NULL;
1039 } while (td->next_file != old_next_file);
1040
1041 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1042 return f;
1043}
1044
1045static struct fio_file *__get_next_file(struct thread_data *td)
1046{
1047 struct fio_file *f;
1048
1049 assert(td->o.nr_files <= td->files_index);
1050
1051 if (td->nr_done_files >= td->o.nr_files) {
1052 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1053 " nr_files=%d\n", td->nr_open_files,
1054 td->nr_done_files,
1055 td->o.nr_files);
1056 return NULL;
1057 }
1058
1059 f = td->file_service_file;
1060 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1061 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1062 goto out;
1063 if (td->file_service_left--)
1064 goto out;
1065 }
1066
1067 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1068 td->o.file_service_type == FIO_FSERVICE_SEQ)
1069 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1070 else
1071 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1072
1073 if (IS_ERR(f))
1074 return f;
1075
1076 td->file_service_file = f;
1077 td->file_service_left = td->file_service_nr - 1;
1078out:
1079 if (f)
1080 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1081 else
1082 dprint(FD_FILE, "get_next_file: NULL\n");
1083 return f;
1084}
1085
1086static struct fio_file *get_next_file(struct thread_data *td)
1087{
1088 if (td->flags & TD_F_PROFILE_OPS) {
1089 struct prof_io_ops *ops = &td->prof_io_ops;
1090
1091 if (ops->get_next_file)
1092 return ops->get_next_file(td);
1093 }
1094
1095 return __get_next_file(td);
1096}
1097
1098static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1099{
1100 struct fio_file *f;
1101
1102 do {
1103 f = get_next_file(td);
1104 if (IS_ERR_OR_NULL(f))
1105 return PTR_ERR(f);
1106
1107 io_u->file = f;
1108 get_file(f);
1109
1110 if (!fill_io_u(td, io_u))
1111 break;
1112
1113 put_file_log(td, f);
1114 td_io_close_file(td, f);
1115 io_u->file = NULL;
1116 fio_file_set_done(f);
1117 td->nr_done_files++;
1118 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1119 td->nr_done_files, td->o.nr_files);
1120 } while (1);
1121
1122 return 0;
1123}
1124
1125static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1126 unsigned long tusec, unsigned long max_usec)
1127{
1128 if (!td->error)
1129 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1130 td_verror(td, ETIMEDOUT, "max latency exceeded");
1131 icd->error = ETIMEDOUT;
1132}
1133
1134static void lat_new_cycle(struct thread_data *td)
1135{
1136 fio_gettime(&td->latency_ts, NULL);
1137 td->latency_ios = ddir_rw_sum(td->io_blocks);
1138 td->latency_failed = 0;
1139}
1140
1141/*
1142 * We had an IO outside the latency target. Reduce the queue depth. If we
1143 * are at QD=1, then it's time to give up.
1144 */
1145static int __lat_target_failed(struct thread_data *td)
1146{
1147 if (td->latency_qd == 1)
1148 return 1;
1149
1150 td->latency_qd_high = td->latency_qd;
1151
1152 if (td->latency_qd == td->latency_qd_low)
1153 td->latency_qd_low--;
1154
1155 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1156
1157 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1158
1159 /*
1160 * When we ramp QD down, quiesce existing IO to prevent
1161 * a storm of ramp downs due to pending higher depth.
1162 */
1163 io_u_quiesce(td);
1164 lat_new_cycle(td);
1165 return 0;
1166}
1167
1168static int lat_target_failed(struct thread_data *td)
1169{
1170 if (td->o.latency_percentile.u.f == 100.0)
1171 return __lat_target_failed(td);
1172
1173 td->latency_failed++;
1174 return 0;
1175}
1176
1177void lat_target_init(struct thread_data *td)
1178{
1179 td->latency_end_run = 0;
1180
1181 if (td->o.latency_target) {
1182 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1183 fio_gettime(&td->latency_ts, NULL);
1184 td->latency_qd = 1;
1185 td->latency_qd_high = td->o.iodepth;
1186 td->latency_qd_low = 1;
1187 td->latency_ios = ddir_rw_sum(td->io_blocks);
1188 } else
1189 td->latency_qd = td->o.iodepth;
1190}
1191
1192void lat_target_reset(struct thread_data *td)
1193{
1194 if (!td->latency_end_run)
1195 lat_target_init(td);
1196}
1197
1198static void lat_target_success(struct thread_data *td)
1199{
1200 const unsigned int qd = td->latency_qd;
1201 struct thread_options *o = &td->o;
1202
1203 td->latency_qd_low = td->latency_qd;
1204
1205 /*
1206 * If we haven't failed yet, we double up to a failing value instead
1207 * of bisecting from highest possible queue depth. If we have set
1208 * a limit other than td->o.iodepth, bisect between that.
1209 */
1210 if (td->latency_qd_high != o->iodepth)
1211 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1212 else
1213 td->latency_qd *= 2;
1214
1215 if (td->latency_qd > o->iodepth)
1216 td->latency_qd = o->iodepth;
1217
1218 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1219
1220 /*
1221 * Same as last one, we are done. Let it run a latency cycle, so
1222 * we get only the results from the targeted depth.
1223 */
1224 if (td->latency_qd == qd) {
1225 if (td->latency_end_run) {
1226 dprint(FD_RATE, "We are done\n");
1227 td->done = 1;
1228 } else {
1229 dprint(FD_RATE, "Quiesce and final run\n");
1230 io_u_quiesce(td);
1231 td->latency_end_run = 1;
1232 reset_all_stats(td);
1233 reset_io_stats(td);
1234 }
1235 }
1236
1237 lat_new_cycle(td);
1238}
1239
1240/*
1241 * Check if we can bump the queue depth
1242 */
1243void lat_target_check(struct thread_data *td)
1244{
1245 uint64_t usec_window;
1246 uint64_t ios;
1247 double success_ios;
1248
1249 usec_window = utime_since_now(&td->latency_ts);
1250 if (usec_window < td->o.latency_window)
1251 return;
1252
1253 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1254 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1255 success_ios *= 100.0;
1256
1257 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1258
1259 if (success_ios >= td->o.latency_percentile.u.f)
1260 lat_target_success(td);
1261 else
1262 __lat_target_failed(td);
1263}
1264
1265/*
1266 * If latency target is enabled, we might be ramping up or down and not
1267 * using the full queue depth available.
1268 */
1269int queue_full(const struct thread_data *td)
1270{
1271 const int qempty = io_u_qempty(&td->io_u_freelist);
1272
1273 if (qempty)
1274 return 1;
1275 if (!td->o.latency_target)
1276 return 0;
1277
1278 return td->cur_depth >= td->latency_qd;
1279}
1280
1281struct io_u *__get_io_u(struct thread_data *td)
1282{
1283 struct io_u *io_u = NULL;
1284
1285 if (td->stop_io)
1286 return NULL;
1287
1288 td_io_u_lock(td);
1289
1290again:
1291 if (!io_u_rempty(&td->io_u_requeues))
1292 io_u = io_u_rpop(&td->io_u_requeues);
1293 else if (!queue_full(td)) {
1294 io_u = io_u_qpop(&td->io_u_freelist);
1295
1296 io_u->file = NULL;
1297 io_u->buflen = 0;
1298 io_u->resid = 0;
1299 io_u->end_io = NULL;
1300 }
1301
1302 if (io_u) {
1303 assert(io_u->flags & IO_U_F_FREE);
1304 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1305 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1306 IO_U_F_VER_LIST);
1307
1308 io_u->error = 0;
1309 io_u->acct_ddir = -1;
1310 td->cur_depth++;
1311 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1312 io_u->ipo = NULL;
1313 } else if (td->o.verify_async) {
1314 /*
1315 * We ran out, wait for async verify threads to finish and
1316 * return one
1317 */
1318 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1319 goto again;
1320 }
1321
1322 td_io_u_unlock(td);
1323 return io_u;
1324}
1325
1326static int check_get_trim(struct thread_data *td, struct io_u *io_u)
1327{
1328 if (!(td->flags & TD_F_TRIM_BACKLOG))
1329 return 0;
1330
1331 if (td->trim_entries) {
1332 int get_trim = 0;
1333
1334 if (td->trim_batch) {
1335 td->trim_batch--;
1336 get_trim = 1;
1337 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1338 td->last_ddir != DDIR_READ) {
1339 td->trim_batch = td->o.trim_batch;
1340 if (!td->trim_batch)
1341 td->trim_batch = td->o.trim_backlog;
1342 get_trim = 1;
1343 }
1344
1345 if (get_trim && !get_next_trim(td, io_u))
1346 return 1;
1347 }
1348
1349 return 0;
1350}
1351
1352static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1353{
1354 if (!(td->flags & TD_F_VER_BACKLOG))
1355 return 0;
1356
1357 if (td->io_hist_len) {
1358 int get_verify = 0;
1359
1360 if (td->verify_batch)
1361 get_verify = 1;
1362 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1363 td->last_ddir != DDIR_READ) {
1364 td->verify_batch = td->o.verify_batch;
1365 if (!td->verify_batch)
1366 td->verify_batch = td->o.verify_backlog;
1367 get_verify = 1;
1368 }
1369
1370 if (get_verify && !get_next_verify(td, io_u)) {
1371 td->verify_batch--;
1372 return 1;
1373 }
1374 }
1375
1376 return 0;
1377}
1378
1379/*
1380 * Fill offset and start time into the buffer content, to prevent too
1381 * easy compressible data for simple de-dupe attempts. Do this for every
1382 * 512b block in the range, since that should be the smallest block size
1383 * we can expect from a device.
1384 */
1385static void small_content_scramble(struct io_u *io_u)
1386{
1387 unsigned int i, nr_blocks = io_u->buflen / 512;
1388 uint64_t boffset;
1389 unsigned int offset;
1390 void *p, *end;
1391
1392 if (!nr_blocks)
1393 return;
1394
1395 p = io_u->xfer_buf;
1396 boffset = io_u->offset;
1397 io_u->buf_filled_len = 0;
1398
1399 for (i = 0; i < nr_blocks; i++) {
1400 /*
1401 * Fill the byte offset into a "random" start offset of
1402 * the buffer, given by the product of the usec time
1403 * and the actual offset.
1404 */
1405 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1406 offset &= ~(sizeof(uint64_t) - 1);
1407 if (offset >= 512 - sizeof(uint64_t))
1408 offset -= sizeof(uint64_t);
1409 memcpy(p + offset, &boffset, sizeof(boffset));
1410
1411 end = p + 512 - sizeof(io_u->start_time);
1412 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1413 p += 512;
1414 boffset += 512;
1415 }
1416}
1417
1418/*
1419 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1420 * etc. The returned io_u is fully ready to be prepped and submitted.
1421 */
1422struct io_u *get_io_u(struct thread_data *td)
1423{
1424 struct fio_file *f;
1425 struct io_u *io_u;
1426 int do_scramble = 0;
1427 long ret = 0;
1428
1429 io_u = __get_io_u(td);
1430 if (!io_u) {
1431 dprint(FD_IO, "__get_io_u failed\n");
1432 return NULL;
1433 }
1434
1435 if (check_get_verify(td, io_u))
1436 goto out;
1437 if (check_get_trim(td, io_u))
1438 goto out;
1439
1440 /*
1441 * from a requeue, io_u already setup
1442 */
1443 if (io_u->file)
1444 goto out;
1445
1446 /*
1447 * If using an iolog, grab next piece if any available.
1448 */
1449 if (td->flags & TD_F_READ_IOLOG) {
1450 if (read_iolog_get(td, io_u))
1451 goto err_put;
1452 } else if (set_io_u_file(td, io_u)) {
1453 ret = -EBUSY;
1454 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1455 goto err_put;
1456 }
1457
1458 f = io_u->file;
1459 if (!f) {
1460 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1461 goto err_put;
1462 }
1463
1464 assert(fio_file_open(f));
1465
1466 if (ddir_rw(io_u->ddir)) {
1467 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
1468 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1469 goto err_put;
1470 }
1471
1472 f->last_start = io_u->offset;
1473 f->last_pos = io_u->offset + io_u->buflen;
1474
1475 if (io_u->ddir == DDIR_WRITE) {
1476 if (td->flags & TD_F_REFILL_BUFFERS) {
1477 io_u_fill_buffer(td, io_u,
1478 td->o.min_bs[DDIR_WRITE],
1479 io_u->xfer_buflen);
1480 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1481 !(td->flags & TD_F_COMPRESS))
1482 do_scramble = 1;
1483 if (td->flags & TD_F_VER_NONE) {
1484 populate_verify_io_u(td, io_u);
1485 do_scramble = 0;
1486 }
1487 } else if (io_u->ddir == DDIR_READ) {
1488 /*
1489 * Reset the buf_filled parameters so next time if the
1490 * buffer is used for writes it is refilled.
1491 */
1492 io_u->buf_filled_len = 0;
1493 }
1494 }
1495
1496 /*
1497 * Set io data pointers.
1498 */
1499 io_u->xfer_buf = io_u->buf;
1500 io_u->xfer_buflen = io_u->buflen;
1501
1502out:
1503 assert(io_u->file);
1504 if (!td_io_prep(td, io_u)) {
1505 if (!td->o.disable_slat)
1506 fio_gettime(&io_u->start_time, NULL);
1507 if (do_scramble)
1508 small_content_scramble(io_u);
1509 return io_u;
1510 }
1511err_put:
1512 dprint(FD_IO, "get_io_u failed\n");
1513 put_io_u(td, io_u);
1514 return ERR_PTR(ret);
1515}
1516
1517void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1518{
1519 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1520
1521 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1522 return;
1523
1524 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1525 io_u->file ? " on file " : "",
1526 io_u->file ? io_u->file->file_name : "",
1527 strerror(io_u->error),
1528 io_ddir_name(io_u->ddir),
1529 io_u->offset, io_u->xfer_buflen);
1530
1531 if (!td->error)
1532 td_verror(td, io_u->error, "io_u error");
1533}
1534
1535static inline int gtod_reduce(struct thread_data *td)
1536{
1537 return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
1538 && td->o.disable_bw;
1539}
1540
1541static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1542 struct io_completion_data *icd,
1543 const enum fio_ddir idx, unsigned int bytes)
1544{
1545 unsigned long lusec = 0;
1546
1547 if (!gtod_reduce(td))
1548 lusec = utime_since(&io_u->issue_time, &icd->time);
1549
1550 if (!td->o.disable_lat) {
1551 unsigned long tusec;
1552
1553 tusec = utime_since(&io_u->start_time, &icd->time);
1554 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
1555
1556 if (td->flags & TD_F_PROFILE_OPS) {
1557 struct prof_io_ops *ops = &td->prof_io_ops;
1558
1559 if (ops->io_u_lat)
1560 icd->error = ops->io_u_lat(td, tusec);
1561 }
1562
1563 if (td->o.max_latency && tusec > td->o.max_latency)
1564 lat_fatal(td, icd, tusec, td->o.max_latency);
1565 if (td->o.latency_target && tusec > td->o.latency_target) {
1566 if (lat_target_failed(td))
1567 lat_fatal(td, icd, tusec, td->o.latency_target);
1568 }
1569 }
1570
1571 if (!td->o.disable_clat) {
1572 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1573 io_u_mark_latency(td, lusec);
1574 }
1575
1576 if (!td->o.disable_bw)
1577 add_bw_sample(td, idx, bytes, &icd->time);
1578
1579 if (!gtod_reduce(td))
1580 add_iops_sample(td, idx, bytes, &icd->time);
1581}
1582
1583static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1584{
1585 uint64_t secs, remainder, bps, bytes;
1586
1587 bytes = td->this_io_bytes[ddir];
1588 bps = td->rate_bps[ddir];
1589 secs = bytes / bps;
1590 remainder = bytes % bps;
1591 return remainder * 1000000 / bps + secs * 1000000;
1592}
1593
1594static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1595 struct io_completion_data *icd)
1596{
1597 struct io_u *io_u = *io_u_ptr;
1598 enum fio_ddir ddir = io_u->ddir;
1599 struct fio_file *f = io_u->file;
1600
1601 dprint_io_u(io_u, "io complete");
1602
1603 td_io_u_lock(td);
1604 assert(io_u->flags & IO_U_F_FLIGHT);
1605 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1606
1607 /*
1608 * Mark IO ok to verify
1609 */
1610 if (io_u->ipo) {
1611 /*
1612 * Remove errored entry from the verification list
1613 */
1614 if (io_u->error)
1615 unlog_io_piece(td, io_u);
1616 else {
1617 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1618 write_barrier();
1619 }
1620 }
1621
1622 td_io_u_unlock(td);
1623
1624 if (ddir_sync(ddir)) {
1625 td->last_was_sync = 1;
1626 if (f) {
1627 f->first_write = -1ULL;
1628 f->last_write = -1ULL;
1629 }
1630 return;
1631 }
1632
1633 td->last_was_sync = 0;
1634 td->last_ddir = ddir;
1635
1636 if (!io_u->error && ddir_rw(ddir)) {
1637 unsigned int bytes = io_u->buflen - io_u->resid;
1638 const enum fio_ddir oddir = ddir ^ 1;
1639 int ret;
1640
1641 td->io_blocks[ddir]++;
1642 td->this_io_blocks[ddir]++;
1643 td->io_bytes[ddir] += bytes;
1644
1645 if (!(io_u->flags & IO_U_F_VER_LIST))
1646 td->this_io_bytes[ddir] += bytes;
1647
1648 if (ddir == DDIR_WRITE) {
1649 if (f) {
1650 if (f->first_write == -1ULL ||
1651 io_u->offset < f->first_write)
1652 f->first_write = io_u->offset;
1653 if (f->last_write == -1ULL ||
1654 ((io_u->offset + bytes) > f->last_write))
1655 f->last_write = io_u->offset + bytes;
1656 }
1657 if (td->last_write_comp) {
1658 int idx = td->last_write_idx++;
1659
1660 td->last_write_comp[idx] = io_u->offset;
1661 if (td->last_write_idx == td->o.iodepth)
1662 td->last_write_idx = 0;
1663 }
1664 }
1665
1666 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1667 td->runstate == TD_VERIFYING)) {
1668 account_io_completion(td, io_u, icd, ddir, bytes);
1669
1670 if (__should_check_rate(td, ddir)) {
1671 td->rate_pending_usleep[ddir] =
1672 (usec_for_io(td, ddir) -
1673 utime_since_now(&td->start));
1674 }
1675 if (ddir != DDIR_TRIM &&
1676 __should_check_rate(td, oddir)) {
1677 td->rate_pending_usleep[oddir] =
1678 (usec_for_io(td, oddir) -
1679 utime_since_now(&td->start));
1680 }
1681 }
1682
1683 icd->bytes_done[ddir] += bytes;
1684
1685 if (io_u->end_io) {
1686 ret = io_u->end_io(td, io_u_ptr);
1687 io_u = *io_u_ptr;
1688 if (ret && !icd->error)
1689 icd->error = ret;
1690 }
1691 } else if (io_u->error) {
1692 icd->error = io_u->error;
1693 io_u_log_error(td, io_u);
1694 }
1695 if (icd->error) {
1696 enum error_type_bit eb = td_error_type(ddir, icd->error);
1697
1698 if (!td_non_fatal_error(td, eb, icd->error))
1699 return;
1700
1701 /*
1702 * If there is a non_fatal error, then add to the error count
1703 * and clear all the errors.
1704 */
1705 update_error_count(td, icd->error);
1706 td_clear_error(td);
1707 icd->error = 0;
1708 if (io_u)
1709 io_u->error = 0;
1710 }
1711}
1712
1713static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1714 int nr)
1715{
1716 int ddir;
1717
1718 if (!gtod_reduce(td))
1719 fio_gettime(&icd->time, NULL);
1720
1721 icd->nr = nr;
1722
1723 icd->error = 0;
1724 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1725 icd->bytes_done[ddir] = 0;
1726}
1727
1728static void ios_completed(struct thread_data *td,
1729 struct io_completion_data *icd)
1730{
1731 struct io_u *io_u;
1732 int i;
1733
1734 for (i = 0; i < icd->nr; i++) {
1735 io_u = td->io_ops->event(td, i);
1736
1737 io_completed(td, &io_u, icd);
1738
1739 if (io_u)
1740 put_io_u(td, io_u);
1741 }
1742}
1743
1744/*
1745 * Complete a single io_u for the sync engines.
1746 */
1747int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1748 uint64_t *bytes)
1749{
1750 struct io_completion_data icd;
1751
1752 init_icd(td, &icd, 1);
1753 io_completed(td, &io_u, &icd);
1754
1755 if (io_u)
1756 put_io_u(td, io_u);
1757
1758 if (icd.error) {
1759 td_verror(td, icd.error, "io_u_sync_complete");
1760 return -1;
1761 }
1762
1763 if (bytes) {
1764 int ddir;
1765
1766 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1767 bytes[ddir] += icd.bytes_done[ddir];
1768 }
1769
1770 return 0;
1771}
1772
1773/*
1774 * Called to complete min_events number of io for the async engines.
1775 */
1776int io_u_queued_complete(struct thread_data *td, int min_evts,
1777 uint64_t *bytes)
1778{
1779 struct io_completion_data icd;
1780 struct timespec *tvp = NULL;
1781 int ret;
1782 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
1783
1784 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
1785
1786 if (!min_evts)
1787 tvp = &ts;
1788 else if (min_evts > td->cur_depth)
1789 min_evts = td->cur_depth;
1790
1791 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
1792 if (ret < 0) {
1793 td_verror(td, -ret, "td_io_getevents");
1794 return ret;
1795 } else if (!ret)
1796 return ret;
1797
1798 init_icd(td, &icd, ret);
1799 ios_completed(td, &icd);
1800 if (icd.error) {
1801 td_verror(td, icd.error, "io_u_queued_complete");
1802 return -1;
1803 }
1804
1805 if (bytes) {
1806 int ddir;
1807
1808 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1809 bytes[ddir] += icd.bytes_done[ddir];
1810 }
1811
1812 return 0;
1813}
1814
1815/*
1816 * Call when io_u is really queued, to update the submission latency.
1817 */
1818void io_u_queued(struct thread_data *td, struct io_u *io_u)
1819{
1820 if (!td->o.disable_slat) {
1821 unsigned long slat_time;
1822
1823 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
1824 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
1825 io_u->offset);
1826 }
1827}
1828
1829/*
1830 * See if we should reuse the last seed, if dedupe is enabled
1831 */
1832static struct frand_state *get_buf_state(struct thread_data *td)
1833{
1834 unsigned int v;
1835 unsigned long r;
1836
1837 if (!td->o.dedupe_percentage)
1838 return &td->buf_state;
1839 else if (td->o.dedupe_percentage == 100)
1840 return &td->buf_state_prev;
1841
1842 r = __rand(&td->dedupe_state);
1843 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
1844
1845 if (v <= td->o.dedupe_percentage)
1846 return &td->buf_state_prev;
1847
1848 return &td->buf_state;
1849}
1850
1851static void save_buf_state(struct thread_data *td, struct frand_state *rs)
1852{
1853 if (rs == &td->buf_state)
1854 frand_copy(&td->buf_state_prev, rs);
1855}
1856
1857void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
1858 unsigned int max_bs)
1859{
1860 if (td->o.buffer_pattern_bytes)
1861 fill_buffer_pattern(td, buf, max_bs);
1862 else if (!td->o.zero_buffers) {
1863 unsigned int perc = td->o.compress_percentage;
1864 struct frand_state *rs;
1865 unsigned int left = max_bs;
1866
1867 do {
1868 rs = get_buf_state(td);
1869
1870 min_write = min(min_write, left);
1871
1872 if (perc) {
1873 unsigned int seg = min_write;
1874
1875 seg = min(min_write, td->o.compress_chunk);
1876 if (!seg)
1877 seg = min_write;
1878
1879 fill_random_buf_percentage(rs, buf, perc, seg,
1880 min_write);
1881 } else
1882 fill_random_buf(rs, buf, min_write);
1883
1884 buf += min_write;
1885 left -= min_write;
1886 save_buf_state(td, rs);
1887 } while (left);
1888 } else
1889 memset(buf, 0, max_bs);
1890}
1891
1892/*
1893 * "randomly" fill the buffer contents
1894 */
1895void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1896 unsigned int min_write, unsigned int max_bs)
1897{
1898 io_u->buf_filled_len = 0;
1899 fill_io_buffer(td, io_u->buf, min_write, max_bs);
1900}
fio - Flexible IO Tester