projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
testing: add test script for zonemode=strided
[fio.git] / io_u.c
... / ...
CommitLineData
1#include <unistd.h>
2#include <string.h>
3#include <assert.h>
4
5#include "fio.h"
6#include "verify.h"
7#include "trim.h"
8#include "lib/rand.h"
9#include "lib/axmap.h"
10#include "err.h"
11#include "lib/pow2.h"
12#include "minmax.h"
13#include "zbd.h"
14
15struct io_completion_data {
16 int nr; /* input */
17
18 int error; /* output */
19 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
20 struct timespec time; /* output */
21};
22
23/*
24 * The ->io_axmap contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
26 */
27static bool random_map_free(struct fio_file *f, const uint64_t block)
28{
29 return !axmap_isset(f->io_axmap, block);
30}
31
32/*
33 * Mark a given offset as used in the map.
34 */
35static uint64_t mark_random_map(struct thread_data *td, struct io_u *io_u,
36 uint64_t offset, uint64_t buflen)
37{
38 unsigned long long min_bs = td->o.min_bs[io_u->ddir];
39 struct fio_file *f = io_u->file;
40 unsigned long long nr_blocks;
41 uint64_t block;
42
43 block = (offset - f->file_offset) / (uint64_t) min_bs;
44 nr_blocks = (buflen + min_bs - 1) / min_bs;
45 assert(nr_blocks > 0);
46
47 if (!(io_u->flags & IO_U_F_BUSY_OK)) {
48 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
49 assert(nr_blocks > 0);
50 }
51
52 if ((nr_blocks * min_bs) < buflen)
53 buflen = nr_blocks * min_bs;
54
55 return buflen;
56}
57
58static uint64_t last_block(struct thread_data *td, struct fio_file *f,
59 enum fio_ddir ddir)
60{
61 uint64_t max_blocks;
62 uint64_t max_size;
63
64 assert(ddir_rw(ddir));
65
66 /*
67 * Hmm, should we make sure that ->io_size <= ->real_file_size?
68 * -> not for now since there is code assuming it could go either.
69 */
70 max_size = f->io_size;
71 if (max_size > f->real_file_size)
72 max_size = f->real_file_size;
73
74 if (td->o.zone_mode == ZONE_MODE_STRIDED && td->o.zone_range)
75 max_size = td->o.zone_range;
76
77 if (td->o.min_bs[ddir] > td->o.ba[ddir])
78 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
79
80 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
81 if (!max_blocks)
82 return 0;
83
84 return max_blocks;
85}
86
87static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
88 enum fio_ddir ddir, uint64_t *b,
89 uint64_t lastb)
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
96 r = __rand(&td->random_state);
97
98 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
99
100 *b = lastb * (r / (rand_max(&td->random_state) + 1.0));
101 } else {
102 uint64_t off = 0;
103
104 assert(fio_file_lfsr(f));
105
106 if (lfsr_next(&f->lfsr, &off))
107 return 1;
108
109 *b = off;
110 }
111
112 /*
113 * if we are not maintaining a random map, we are done.
114 */
115 if (!file_randommap(td, f))
116 goto ret;
117
118 /*
119 * calculate map offset and check if it's free
120 */
121 if (random_map_free(f, *b))
122 goto ret;
123
124 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
125 (unsigned long long) *b);
126
127 *b = axmap_next_free(f->io_axmap, *b);
128 if (*b == (uint64_t) -1ULL)
129 return 1;
130ret:
131 return 0;
132}
133
134static int __get_next_rand_offset_zipf(struct thread_data *td,
135 struct fio_file *f, enum fio_ddir ddir,
136 uint64_t *b)
137{
138 *b = zipf_next(&f->zipf);
139 return 0;
140}
141
142static int __get_next_rand_offset_pareto(struct thread_data *td,
143 struct fio_file *f, enum fio_ddir ddir,
144 uint64_t *b)
145{
146 *b = pareto_next(&f->zipf);
147 return 0;
148}
149
150static int __get_next_rand_offset_gauss(struct thread_data *td,
151 struct fio_file *f, enum fio_ddir ddir,
152 uint64_t *b)
153{
154 *b = gauss_next(&f->gauss);
155 return 0;
156}
157
158static int __get_next_rand_offset_zoned_abs(struct thread_data *td,
159 struct fio_file *f,
160 enum fio_ddir ddir, uint64_t *b)
161{
162 struct zone_split_index *zsi;
163 uint64_t lastb, send, stotal;
164 unsigned int v;
165
166 lastb = last_block(td, f, ddir);
167 if (!lastb)
168 return 1;
169
170 if (!td->o.zone_split_nr[ddir]) {
171bail:
172 return __get_next_rand_offset(td, f, ddir, b, lastb);
173 }
174
175 /*
176 * Generate a value, v, between 1 and 100, both inclusive
177 */
178 v = rand_between(&td->zone_state, 1, 100);
179
180 /*
181 * Find our generated table. 'send' is the end block of this zone,
182 * 'stotal' is our start offset.
183 */
184 zsi = &td->zone_state_index[ddir][v - 1];
185 stotal = zsi->size_prev / td->o.ba[ddir];
186 send = zsi->size / td->o.ba[ddir];
187
188 /*
189 * Should never happen
190 */
191 if (send == -1U) {
192 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
193 log_err("fio: bug in zoned generation\n");
194 goto bail;
195 } else if (send > lastb) {
196 /*
197 * This happens if the user specifies ranges that exceed
198 * the file/device size. We can't handle that gracefully,
199 * so error and exit.
200 */
201 log_err("fio: zoned_abs sizes exceed file size\n");
202 return 1;
203 }
204
205 /*
206 * Generate index from 0..send-stotal
207 */
208 if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1)
209 return 1;
210
211 *b += stotal;
212 return 0;
213}
214
215static int __get_next_rand_offset_zoned(struct thread_data *td,
216 struct fio_file *f, enum fio_ddir ddir,
217 uint64_t *b)
218{
219 unsigned int v, send, stotal;
220 uint64_t offset, lastb;
221 struct zone_split_index *zsi;
222
223 lastb = last_block(td, f, ddir);
224 if (!lastb)
225 return 1;
226
227 if (!td->o.zone_split_nr[ddir]) {
228bail:
229 return __get_next_rand_offset(td, f, ddir, b, lastb);
230 }
231
232 /*
233 * Generate a value, v, between 1 and 100, both inclusive
234 */
235 v = rand_between(&td->zone_state, 1, 100);
236
237 zsi = &td->zone_state_index[ddir][v - 1];
238 stotal = zsi->size_perc_prev;
239 send = zsi->size_perc;
240
241 /*
242 * Should never happen
243 */
244 if (send == -1U) {
245 if (!fio_did_warn(FIO_WARN_ZONED_BUG))
246 log_err("fio: bug in zoned generation\n");
247 goto bail;
248 }
249
250 /*
251 * 'send' is some percentage below or equal to 100 that
252 * marks the end of the current IO range. 'stotal' marks
253 * the start, in percent.
254 */
255 if (stotal)
256 offset = stotal * lastb / 100ULL;
257 else
258 offset = 0;
259
260 lastb = lastb * (send - stotal) / 100ULL;
261
262 /*
263 * Generate index from 0..send-of-lastb
264 */
265 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
266 return 1;
267
268 /*
269 * Add our start offset, if any
270 */
271 if (offset)
272 *b += offset;
273
274 return 0;
275}
276
277static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
278 enum fio_ddir ddir, uint64_t *b)
279{
280 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
281 uint64_t lastb;
282
283 lastb = last_block(td, f, ddir);
284 if (!lastb)
285 return 1;
286
287 return __get_next_rand_offset(td, f, ddir, b, lastb);
288 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
289 return __get_next_rand_offset_zipf(td, f, ddir, b);
290 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
291 return __get_next_rand_offset_pareto(td, f, ddir, b);
292 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
293 return __get_next_rand_offset_gauss(td, f, ddir, b);
294 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
295 return __get_next_rand_offset_zoned(td, f, ddir, b);
296 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
297 return __get_next_rand_offset_zoned_abs(td, f, ddir, b);
298
299 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
300 return 1;
301}
302
303static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
304{
305 unsigned int v;
306
307 if (td->o.perc_rand[ddir] == 100)
308 return true;
309
310 v = rand_between(&td->seq_rand_state[ddir], 1, 100);
311
312 return v <= td->o.perc_rand[ddir];
313}
314
315static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f)
316{
317 struct thread_options *o = &td->o;
318
319 if (o->invalidate_cache && !o->odirect) {
320 int fio_unused ret;
321
322 ret = file_invalidate_cache(td, f);
323 }
324}
325
326static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
327 enum fio_ddir ddir, uint64_t *b)
328{
329 if (!get_next_rand_offset(td, f, ddir, b))
330 return 0;
331
332 if (td->o.time_based ||
333 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
334 fio_file_reset(td, f);
335 loop_cache_invalidate(td, f);
336 if (!get_next_rand_offset(td, f, ddir, b))
337 return 0;
338 }
339
340 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
341 f->file_name, (unsigned long long) f->last_pos[ddir],
342 (unsigned long long) f->real_file_size);
343 return 1;
344}
345
346static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
347 enum fio_ddir ddir, uint64_t *offset)
348{
349 struct thread_options *o = &td->o;
350
351 assert(ddir_rw(ddir));
352
353 /*
354 * If we reach the end for a time based run, reset us back to 0
355 * and invalidate the cache, if we need to.
356 */
357 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
358 o->time_based) {
359 f->last_pos[ddir] = f->file_offset;
360 loop_cache_invalidate(td, f);
361 }
362
363 if (f->last_pos[ddir] < f->real_file_size) {
364 uint64_t pos;
365
366 /*
367 * Only rewind if we already hit the end
368 */
369 if (f->last_pos[ddir] == f->file_offset &&
370 f->file_offset && o->ddir_seq_add < 0) {
371 if (f->real_file_size > f->io_size)
372 f->last_pos[ddir] = f->io_size;
373 else
374 f->last_pos[ddir] = f->real_file_size;
375 }
376
377 pos = f->last_pos[ddir] - f->file_offset;
378 if (pos && o->ddir_seq_add) {
379 pos += o->ddir_seq_add;
380
381 /*
382 * If we reach beyond the end of the file
383 * with holed IO, wrap around to the
384 * beginning again. If we're doing backwards IO,
385 * wrap to the end.
386 */
387 if (pos >= f->real_file_size) {
388 if (o->ddir_seq_add > 0)
389 pos = f->file_offset;
390 else {
391 if (f->real_file_size > f->io_size)
392 pos = f->io_size;
393 else
394 pos = f->real_file_size;
395
396 pos += o->ddir_seq_add;
397 }
398 }
399 }
400
401 *offset = pos;
402 return 0;
403 }
404
405 return 1;
406}
407
408static int get_next_block(struct thread_data *td, struct io_u *io_u,
409 enum fio_ddir ddir, int rw_seq,
410 bool *is_random)
411{
412 struct fio_file *f = io_u->file;
413 uint64_t b, offset;
414 int ret;
415
416 assert(ddir_rw(ddir));
417
418 b = offset = -1ULL;
419
420 if (rw_seq) {
421 if (td_random(td)) {
422 if (should_do_random(td, ddir)) {
423 ret = get_next_rand_block(td, f, ddir, &b);
424 *is_random = true;
425 } else {
426 *is_random = false;
427 io_u_set(td, io_u, IO_U_F_BUSY_OK);
428 ret = get_next_seq_offset(td, f, ddir, &offset);
429 if (ret)
430 ret = get_next_rand_block(td, f, ddir, &b);
431 }
432 } else {
433 *is_random = false;
434 ret = get_next_seq_offset(td, f, ddir, &offset);
435 }
436 } else {
437 io_u_set(td, io_u, IO_U_F_BUSY_OK);
438 *is_random = false;
439
440 if (td->o.rw_seq == RW_SEQ_SEQ) {
441 ret = get_next_seq_offset(td, f, ddir, &offset);
442 if (ret) {
443 ret = get_next_rand_block(td, f, ddir, &b);
444 *is_random = false;
445 }
446 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
447 if (f->last_start[ddir] != -1ULL)
448 offset = f->last_start[ddir] - f->file_offset;
449 else
450 offset = 0;
451 ret = 0;
452 } else {
453 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
454 ret = 1;
455 }
456 }
457
458 if (!ret) {
459 if (offset != -1ULL)
460 io_u->offset = offset;
461 else if (b != -1ULL)
462 io_u->offset = b * td->o.ba[ddir];
463 else {
464 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
465 ret = 1;
466 }
467 }
468
469 return ret;
470}
471
472/*
473 * For random io, generate a random new block and see if it's used. Repeat
474 * until we find a free one. For sequential io, just return the end of
475 * the last io issued.
476 */
477static int get_next_offset(struct thread_data *td, struct io_u *io_u,
478 bool *is_random)
479{
480 struct fio_file *f = io_u->file;
481 enum fio_ddir ddir = io_u->ddir;
482 int rw_seq_hit = 0;
483
484 assert(ddir_rw(ddir));
485
486 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
487 rw_seq_hit = 1;
488 td->ddir_seq_nr = td->o.ddir_seq_nr;
489 }
490
491 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
492 return 1;
493
494 if (io_u->offset >= f->io_size) {
495 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
496 (unsigned long long) io_u->offset,
497 (unsigned long long) f->io_size);
498 return 1;
499 }
500
501 io_u->offset += f->file_offset;
502 if (io_u->offset >= f->real_file_size) {
503 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
504 (unsigned long long) io_u->offset,
505 (unsigned long long) f->real_file_size);
506 return 1;
507 }
508
509 return 0;
510}
511
512static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
513 unsigned long long buflen)
514{
515 struct fio_file *f = io_u->file;
516
517 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
518}
519
520static unsigned long long get_next_buflen(struct thread_data *td, struct io_u *io_u,
521 bool is_random)
522{
523 int ddir = io_u->ddir;
524 unsigned long long buflen = 0;
525 unsigned long long minbs, maxbs;
526 uint64_t frand_max, r;
527 bool power_2;
528
529 assert(ddir_rw(ddir));
530
531 if (td->o.bs_is_seq_rand)
532 ddir = is_random ? DDIR_WRITE : DDIR_READ;
533
534 minbs = td->o.min_bs[ddir];
535 maxbs = td->o.max_bs[ddir];
536
537 if (minbs == maxbs)
538 return minbs;
539
540 /*
541 * If we can't satisfy the min block size from here, then fail
542 */
543 if (!io_u_fits(td, io_u, minbs))
544 return 0;
545
546 frand_max = rand_max(&td->bsrange_state[ddir]);
547 do {
548 r = __rand(&td->bsrange_state[ddir]);
549
550 if (!td->o.bssplit_nr[ddir]) {
551 buflen = minbs + (unsigned long long) ((double) maxbs *
552 (r / (frand_max + 1.0)));
553 } else {
554 long long perc = 0;
555 unsigned int i;
556
557 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
558 struct bssplit *bsp = &td->o.bssplit[ddir][i];
559
560 if (!bsp->perc)
561 continue;
562 buflen = bsp->bs;
563 perc += bsp->perc;
564 if ((r / perc <= frand_max / 100ULL) &&
565 io_u_fits(td, io_u, buflen))
566 break;
567 }
568 }
569
570 power_2 = is_power_of_2(minbs);
571 if (!td->o.bs_unaligned && power_2)
572 buflen &= ~(minbs - 1);
573 else if (!td->o.bs_unaligned && !power_2)
574 buflen -= buflen % minbs;
575 if (buflen > maxbs)
576 buflen = maxbs;
577 } while (!io_u_fits(td, io_u, buflen));
578
579 return buflen;
580}
581
582static void set_rwmix_bytes(struct thread_data *td)
583{
584 unsigned int diff;
585
586 /*
587 * we do time or byte based switch. this is needed because
588 * buffered writes may issue a lot quicker than they complete,
589 * whereas reads do not.
590 */
591 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
592 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
593}
594
595static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
596{
597 unsigned int v;
598
599 v = rand_between(&td->rwmix_state, 1, 100);
600
601 if (v <= td->o.rwmix[DDIR_READ])
602 return DDIR_READ;
603
604 return DDIR_WRITE;
605}
606
607int io_u_quiesce(struct thread_data *td)
608{
609 int ret = 0, completed = 0;
610
611 /*
612 * We are going to sleep, ensure that we flush anything pending as
613 * not to skew our latency numbers.
614 *
615 * Changed to only monitor 'in flight' requests here instead of the
616 * td->cur_depth, b/c td->cur_depth does not accurately represent
617 * io's that have been actually submitted to an async engine,
618 * and cur_depth is meaningless for sync engines.
619 */
620 if (td->io_u_queued || td->cur_depth)
621 td_io_commit(td);
622
623 while (td->io_u_in_flight) {
624 ret = io_u_queued_complete(td, 1);
625 if (ret > 0)
626 completed += ret;
627 else if (ret < 0)
628 break;
629 }
630
631 if (td->flags & TD_F_REGROW_LOGS)
632 regrow_logs(td);
633
634 if (completed)
635 return completed;
636
637 return ret;
638}
639
640static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
641{
642 enum fio_ddir odir = ddir ^ 1;
643 uint64_t usec;
644 uint64_t now;
645
646 assert(ddir_rw(ddir));
647 now = utime_since_now(&td->start);
648
649 /*
650 * if rate_next_io_time is in the past, need to catch up to rate
651 */
652 if (td->rate_next_io_time[ddir] <= now)
653 return ddir;
654
655 /*
656 * We are ahead of rate in this direction. See if we
657 * should switch.
658 */
659 if (td_rw(td) && td->o.rwmix[odir]) {
660 /*
661 * Other direction is behind rate, switch
662 */
663 if (td->rate_next_io_time[odir] <= now)
664 return odir;
665
666 /*
667 * Both directions are ahead of rate. sleep the min,
668 * switch if necessary
669 */
670 if (td->rate_next_io_time[ddir] <=
671 td->rate_next_io_time[odir]) {
672 usec = td->rate_next_io_time[ddir] - now;
673 } else {
674 usec = td->rate_next_io_time[odir] - now;
675 ddir = odir;
676 }
677 } else
678 usec = td->rate_next_io_time[ddir] - now;
679
680 if (td->o.io_submit_mode == IO_MODE_INLINE)
681 io_u_quiesce(td);
682
683 usec_sleep(td, usec);
684 return ddir;
685}
686
687/*
688 * Return the data direction for the next io_u. If the job is a
689 * mixed read/write workload, check the rwmix cycle and switch if
690 * necessary.
691 */
692static enum fio_ddir get_rw_ddir(struct thread_data *td)
693{
694 enum fio_ddir ddir;
695
696 /*
697 * See if it's time to fsync/fdatasync/sync_file_range first,
698 * and if not then move on to check regular I/Os.
699 */
700 if (should_fsync(td)) {
701 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
702 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
703 return DDIR_SYNC;
704
705 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
706 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
707 return DDIR_DATASYNC;
708
709 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
710 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
711 return DDIR_SYNC_FILE_RANGE;
712 }
713
714 if (td_rw(td)) {
715 /*
716 * Check if it's time to seed a new data direction.
717 */
718 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
719 /*
720 * Put a top limit on how many bytes we do for
721 * one data direction, to avoid overflowing the
722 * ranges too much
723 */
724 ddir = get_rand_ddir(td);
725
726 if (ddir != td->rwmix_ddir)
727 set_rwmix_bytes(td);
728
729 td->rwmix_ddir = ddir;
730 }
731 ddir = td->rwmix_ddir;
732 } else if (td_read(td))
733 ddir = DDIR_READ;
734 else if (td_write(td))
735 ddir = DDIR_WRITE;
736 else if (td_trim(td))
737 ddir = DDIR_TRIM;
738 else
739 ddir = DDIR_INVAL;
740
741 td->rwmix_ddir = rate_ddir(td, ddir);
742 return td->rwmix_ddir;
743}
744
745static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
746{
747 enum fio_ddir ddir = get_rw_ddir(td);
748
749 if (td_trimwrite(td)) {
750 struct fio_file *f = io_u->file;
751 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
752 ddir = DDIR_TRIM;
753 else
754 ddir = DDIR_WRITE;
755 }
756
757 io_u->ddir = io_u->acct_ddir = ddir;
758
759 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
760 td->o.barrier_blocks &&
761 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
762 td->io_issues[DDIR_WRITE])
763 io_u_set(td, io_u, IO_U_F_BARRIER);
764}
765
766void put_file_log(struct thread_data *td, struct fio_file *f)
767{
768 unsigned int ret = put_file(td, f);
769
770 if (ret)
771 td_verror(td, ret, "file close");
772}
773
774void put_io_u(struct thread_data *td, struct io_u *io_u)
775{
776 const bool needs_lock = td_async_processing(td);
777
778 zbd_put_io_u(io_u);
779
780 if (td->parent)
781 td = td->parent;
782
783 if (needs_lock)
784 __td_io_u_lock(td);
785
786 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
787 put_file_log(td, io_u->file);
788
789 io_u->file = NULL;
790 io_u_set(td, io_u, IO_U_F_FREE);
791
792 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
793 td->cur_depth--;
794 assert(!(td->flags & TD_F_CHILD));
795 }
796 io_u_qpush(&td->io_u_freelist, io_u);
797 td_io_u_free_notify(td);
798
799 if (needs_lock)
800 __td_io_u_unlock(td);
801}
802
803void clear_io_u(struct thread_data *td, struct io_u *io_u)
804{
805 io_u_clear(td, io_u, IO_U_F_FLIGHT);
806 put_io_u(td, io_u);
807}
808
809void requeue_io_u(struct thread_data *td, struct io_u **io_u)
810{
811 const bool needs_lock = td_async_processing(td);
812 struct io_u *__io_u = *io_u;
813 enum fio_ddir ddir = acct_ddir(__io_u);
814
815 dprint(FD_IO, "requeue %p\n", __io_u);
816
817 if (td->parent)
818 td = td->parent;
819
820 if (needs_lock)
821 __td_io_u_lock(td);
822
823 io_u_set(td, __io_u, IO_U_F_FREE);
824 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
825 td->io_issues[ddir]--;
826
827 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
828 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
829 td->cur_depth--;
830 assert(!(td->flags & TD_F_CHILD));
831 }
832
833 io_u_rpush(&td->io_u_requeues, __io_u);
834 td_io_u_free_notify(td);
835
836 if (needs_lock)
837 __td_io_u_unlock(td);
838
839 *io_u = NULL;
840}
841
842static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u)
843{
844 struct fio_file *f = io_u->file;
845
846 assert(td->o.zone_mode == ZONE_MODE_STRIDED);
847 assert(td->o.zone_size);
848 assert(td->o.zone_range);
849
850 /*
851 * See if it's time to switch to a new zone
852 */
853 if (td->zone_bytes >= td->o.zone_size &&
854 fio_option_is_set(&td->o, zone_skip)) {
855 td->zone_bytes = 0;
856 f->file_offset += td->o.zone_range + td->o.zone_skip;
857
858 /*
859 * Wrap from the beginning, if we exceed the file size
860 */
861 if (f->file_offset >= f->real_file_size)
862 f->file_offset = get_start_offset(td, f);
863
864 f->last_pos[io_u->ddir] = f->file_offset;
865 td->io_skip_bytes += td->o.zone_skip;
866 }
867
868 /*
869 * If zone_size > zone_range, then maintain the same zone until
870 * zone_bytes >= zone_size.
871 */
872 if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
873 dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
874 f->file_offset, f->last_pos[io_u->ddir]);
875 f->last_pos[io_u->ddir] = f->file_offset;
876 }
877
878 /*
879 * For random: if 'norandommap' is not set and zone_size > zone_range,
880 * map needs to be reset as it's done with zone_range everytime.
881 */
882 if ((td->zone_bytes % td->o.zone_range) == 0)
883 fio_file_reset(td, f);
884}
885
886static int fill_io_u(struct thread_data *td, struct io_u *io_u)
887{
888 bool is_random;
889 uint64_t offset;
890 enum io_u_action ret;
891
892 if (td_ioengine_flagged(td, FIO_NOIO))
893 goto out;
894
895 set_rw_ddir(td, io_u);
896
897 /*
898 * fsync() or fdatasync() or trim etc, we are done
899 */
900 if (!ddir_rw(io_u->ddir))
901 goto out;
902
903 if (td->o.zone_mode == ZONE_MODE_STRIDED)
904 setup_strided_zone_mode(td, io_u);
905 else if (td->o.zone_mode == ZONE_MODE_ZBD)
906 setup_zbd_zone_mode(td, io_u);
907
908 /*
909 * No log, let the seq/rand engine retrieve the next buflen and
910 * position.
911 */
912 if (get_next_offset(td, io_u, &is_random)) {
913 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
914 return 1;
915 }
916
917 io_u->buflen = get_next_buflen(td, io_u, is_random);
918 if (!io_u->buflen) {
919 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
920 return 1;
921 }
922
923 offset = io_u->offset;
924 if (td->o.zone_mode == ZONE_MODE_ZBD) {
925 ret = zbd_adjust_block(td, io_u);
926 if (ret == io_u_eof)
927 return 1;
928 }
929
930 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
931 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n",
932 io_u,
933 (unsigned long long) io_u->offset, io_u->buflen,
934 (unsigned long long) io_u->file->real_file_size);
935 return 1;
936 }
937
938 /*
939 * mark entry before potentially trimming io_u
940 */
941 if (td_random(td) && file_randommap(td, io_u->file))
942 io_u->buflen = mark_random_map(td, io_u, offset, io_u->buflen);
943
944out:
945 dprint_io_u(io_u, "fill");
946 td->zone_bytes += io_u->buflen;
947 return 0;
948}
949
950static void __io_u_mark_map(uint64_t *map, unsigned int nr)
951{
952 int idx = 0;
953
954 switch (nr) {
955 default:
956 idx = 6;
957 break;
958 case 33 ... 64:
959 idx = 5;
960 break;
961 case 17 ... 32:
962 idx = 4;
963 break;
964 case 9 ... 16:
965 idx = 3;
966 break;
967 case 5 ... 8:
968 idx = 2;
969 break;
970 case 1 ... 4:
971 idx = 1;
972 case 0:
973 break;
974 }
975
976 map[idx]++;
977}
978
979void io_u_mark_submit(struct thread_data *td, unsigned int nr)
980{
981 __io_u_mark_map(td->ts.io_u_submit, nr);
982 td->ts.total_submit++;
983}
984
985void io_u_mark_complete(struct thread_data *td, unsigned int nr)
986{
987 __io_u_mark_map(td->ts.io_u_complete, nr);
988 td->ts.total_complete++;
989}
990
991void io_u_mark_depth(struct thread_data *td, unsigned int nr)
992{
993 int idx = 0;
994
995 switch (td->cur_depth) {
996 default:
997 idx = 6;
998 break;
999 case 32 ... 63:
1000 idx = 5;
1001 break;
1002 case 16 ... 31:
1003 idx = 4;
1004 break;
1005 case 8 ... 15:
1006 idx = 3;
1007 break;
1008 case 4 ... 7:
1009 idx = 2;
1010 break;
1011 case 2 ... 3:
1012 idx = 1;
1013 case 1:
1014 break;
1015 }
1016
1017 td->ts.io_u_map[idx] += nr;
1018}
1019
1020static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
1021{
1022 int idx = 0;
1023
1024 assert(nsec < 1000);
1025
1026 switch (nsec) {
1027 case 750 ... 999:
1028 idx = 9;
1029 break;
1030 case 500 ... 749:
1031 idx = 8;
1032 break;
1033 case 250 ... 499:
1034 idx = 7;
1035 break;
1036 case 100 ... 249:
1037 idx = 6;
1038 break;
1039 case 50 ... 99:
1040 idx = 5;
1041 break;
1042 case 20 ... 49:
1043 idx = 4;
1044 break;
1045 case 10 ... 19:
1046 idx = 3;
1047 break;
1048 case 4 ... 9:
1049 idx = 2;
1050 break;
1051 case 2 ... 3:
1052 idx = 1;
1053 case 0 ... 1:
1054 break;
1055 }
1056
1057 assert(idx < FIO_IO_U_LAT_N_NR);
1058 td->ts.io_u_lat_n[idx]++;
1059}
1060
1061static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1062{
1063 int idx = 0;
1064
1065 assert(usec < 1000 && usec >= 1);
1066
1067 switch (usec) {
1068 case 750 ... 999:
1069 idx = 9;
1070 break;
1071 case 500 ... 749:
1072 idx = 8;
1073 break;
1074 case 250 ... 499:
1075 idx = 7;
1076 break;
1077 case 100 ... 249:
1078 idx = 6;
1079 break;
1080 case 50 ... 99:
1081 idx = 5;
1082 break;
1083 case 20 ... 49:
1084 idx = 4;
1085 break;
1086 case 10 ... 19:
1087 idx = 3;
1088 break;
1089 case 4 ... 9:
1090 idx = 2;
1091 break;
1092 case 2 ... 3:
1093 idx = 1;
1094 case 0 ... 1:
1095 break;
1096 }
1097
1098 assert(idx < FIO_IO_U_LAT_U_NR);
1099 td->ts.io_u_lat_u[idx]++;
1100}
1101
1102static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
1103{
1104 int idx = 0;
1105
1106 assert(msec >= 1);
1107
1108 switch (msec) {
1109 default:
1110 idx = 11;
1111 break;
1112 case 1000 ... 1999:
1113 idx = 10;
1114 break;
1115 case 750 ... 999:
1116 idx = 9;
1117 break;
1118 case 500 ... 749:
1119 idx = 8;
1120 break;
1121 case 250 ... 499:
1122 idx = 7;
1123 break;
1124 case 100 ... 249:
1125 idx = 6;
1126 break;
1127 case 50 ... 99:
1128 idx = 5;
1129 break;
1130 case 20 ... 49:
1131 idx = 4;
1132 break;
1133 case 10 ... 19:
1134 idx = 3;
1135 break;
1136 case 4 ... 9:
1137 idx = 2;
1138 break;
1139 case 2 ... 3:
1140 idx = 1;
1141 case 0 ... 1:
1142 break;
1143 }
1144
1145 assert(idx < FIO_IO_U_LAT_M_NR);
1146 td->ts.io_u_lat_m[idx]++;
1147}
1148
1149static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
1150{
1151 if (nsec < 1000)
1152 io_u_mark_lat_nsec(td, nsec);
1153 else if (nsec < 1000000)
1154 io_u_mark_lat_usec(td, nsec / 1000);
1155 else
1156 io_u_mark_lat_msec(td, nsec / 1000000);
1157}
1158
1159static unsigned int __get_next_fileno_rand(struct thread_data *td)
1160{
1161 unsigned long fileno;
1162
1163 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1164 uint64_t frand_max = rand_max(&td->next_file_state);
1165 unsigned long r;
1166
1167 r = __rand(&td->next_file_state);
1168 return (unsigned int) ((double) td->o.nr_files
1169 * (r / (frand_max + 1.0)));
1170 }
1171
1172 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1173 fileno = zipf_next(&td->next_file_zipf);
1174 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1175 fileno = pareto_next(&td->next_file_zipf);
1176 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1177 fileno = gauss_next(&td->next_file_gauss);
1178 else {
1179 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1180 assert(0);
1181 return 0;
1182 }
1183
1184 return fileno >> FIO_FSERVICE_SHIFT;
1185}
1186
1187/*
1188 * Get next file to service by choosing one at random
1189 */
1190static struct fio_file *get_next_file_rand(struct thread_data *td,
1191 enum fio_file_flags goodf,
1192 enum fio_file_flags badf)
1193{
1194 struct fio_file *f;
1195 int fno;
1196
1197 do {
1198 int opened = 0;
1199
1200 fno = __get_next_fileno_rand(td);
1201
1202 f = td->files[fno];
1203 if (fio_file_done(f))
1204 continue;
1205
1206 if (!fio_file_open(f)) {
1207 int err;
1208
1209 if (td->nr_open_files >= td->o.open_files)
1210 return ERR_PTR(-EBUSY);
1211
1212 err = td_io_open_file(td, f);
1213 if (err)
1214 continue;
1215 opened = 1;
1216 }
1217
1218 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1219 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
1220 return f;
1221 }
1222 if (opened)
1223 td_io_close_file(td, f);
1224 } while (1);
1225}
1226
1227/*
1228 * Get next file to service by doing round robin between all available ones
1229 */
1230static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1231 int badf)
1232{
1233 unsigned int old_next_file = td->next_file;
1234 struct fio_file *f;
1235
1236 do {
1237 int opened = 0;
1238
1239 f = td->files[td->next_file];
1240
1241 td->next_file++;
1242 if (td->next_file >= td->o.nr_files)
1243 td->next_file = 0;
1244
1245 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
1246 if (fio_file_done(f)) {
1247 f = NULL;
1248 continue;
1249 }
1250
1251 if (!fio_file_open(f)) {
1252 int err;
1253
1254 if (td->nr_open_files >= td->o.open_files)
1255 return ERR_PTR(-EBUSY);
1256
1257 err = td_io_open_file(td, f);
1258 if (err) {
1259 dprint(FD_FILE, "error %d on open of %s\n",
1260 err, f->file_name);
1261 f = NULL;
1262 continue;
1263 }
1264 opened = 1;
1265 }
1266
1267 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1268 f->flags);
1269 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
1270 break;
1271
1272 if (opened)
1273 td_io_close_file(td, f);
1274
1275 f = NULL;
1276 } while (td->next_file != old_next_file);
1277
1278 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
1279 return f;
1280}
1281
1282static struct fio_file *__get_next_file(struct thread_data *td)
1283{
1284 struct fio_file *f;
1285
1286 assert(td->o.nr_files <= td->files_index);
1287
1288 if (td->nr_done_files >= td->o.nr_files) {
1289 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1290 " nr_files=%d\n", td->nr_open_files,
1291 td->nr_done_files,
1292 td->o.nr_files);
1293 return NULL;
1294 }
1295
1296 f = td->file_service_file;
1297 if (f && fio_file_open(f) && !fio_file_closing(f)) {
1298 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1299 goto out;
1300 if (td->file_service_left--)
1301 goto out;
1302 }
1303
1304 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1305 td->o.file_service_type == FIO_FSERVICE_SEQ)
1306 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
1307 else
1308 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1309
1310 if (IS_ERR(f))
1311 return f;
1312
1313 td->file_service_file = f;
1314 td->file_service_left = td->file_service_nr - 1;
1315out:
1316 if (f)
1317 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1318 else
1319 dprint(FD_FILE, "get_next_file: NULL\n");
1320 return f;
1321}
1322
1323static struct fio_file *get_next_file(struct thread_data *td)
1324{
1325 return __get_next_file(td);
1326}
1327
1328static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
1329{
1330 struct fio_file *f;
1331
1332 do {
1333 f = get_next_file(td);
1334 if (IS_ERR_OR_NULL(f))
1335 return PTR_ERR(f);
1336
1337 io_u->file = f;
1338 get_file(f);
1339
1340 if (!fill_io_u(td, io_u))
1341 break;
1342
1343 zbd_put_io_u(io_u);
1344
1345 put_file_log(td, f);
1346 td_io_close_file(td, f);
1347 io_u->file = NULL;
1348 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1349 fio_file_reset(td, f);
1350 else {
1351 fio_file_set_done(f);
1352 td->nr_done_files++;
1353 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1354 td->nr_done_files, td->o.nr_files);
1355 }
1356 } while (1);
1357
1358 return 0;
1359}
1360
1361static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1362 unsigned long long tnsec, unsigned long long max_nsec)
1363{
1364 if (!td->error)
1365 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
1366 td_verror(td, ETIMEDOUT, "max latency exceeded");
1367 icd->error = ETIMEDOUT;
1368}
1369
1370static void lat_new_cycle(struct thread_data *td)
1371{
1372 fio_gettime(&td->latency_ts, NULL);
1373 td->latency_ios = ddir_rw_sum(td->io_blocks);
1374 td->latency_failed = 0;
1375}
1376
1377/*
1378 * We had an IO outside the latency target. Reduce the queue depth. If we
1379 * are at QD=1, then it's time to give up.
1380 */
1381static bool __lat_target_failed(struct thread_data *td)
1382{
1383 if (td->latency_qd == 1)
1384 return true;
1385
1386 td->latency_qd_high = td->latency_qd;
1387
1388 if (td->latency_qd == td->latency_qd_low)
1389 td->latency_qd_low--;
1390
1391 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1392
1393 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1394
1395 /*
1396 * When we ramp QD down, quiesce existing IO to prevent
1397 * a storm of ramp downs due to pending higher depth.
1398 */
1399 io_u_quiesce(td);
1400 lat_new_cycle(td);
1401 return false;
1402}
1403
1404static bool lat_target_failed(struct thread_data *td)
1405{
1406 if (td->o.latency_percentile.u.f == 100.0)
1407 return __lat_target_failed(td);
1408
1409 td->latency_failed++;
1410 return false;
1411}
1412
1413void lat_target_init(struct thread_data *td)
1414{
1415 td->latency_end_run = 0;
1416
1417 if (td->o.latency_target) {
1418 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1419 fio_gettime(&td->latency_ts, NULL);
1420 td->latency_qd = 1;
1421 td->latency_qd_high = td->o.iodepth;
1422 td->latency_qd_low = 1;
1423 td->latency_ios = ddir_rw_sum(td->io_blocks);
1424 } else
1425 td->latency_qd = td->o.iodepth;
1426}
1427
1428void lat_target_reset(struct thread_data *td)
1429{
1430 if (!td->latency_end_run)
1431 lat_target_init(td);
1432}
1433
1434static void lat_target_success(struct thread_data *td)
1435{
1436 const unsigned int qd = td->latency_qd;
1437 struct thread_options *o = &td->o;
1438
1439 td->latency_qd_low = td->latency_qd;
1440
1441 /*
1442 * If we haven't failed yet, we double up to a failing value instead
1443 * of bisecting from highest possible queue depth. If we have set
1444 * a limit other than td->o.iodepth, bisect between that.
1445 */
1446 if (td->latency_qd_high != o->iodepth)
1447 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1448 else
1449 td->latency_qd *= 2;
1450
1451 if (td->latency_qd > o->iodepth)
1452 td->latency_qd = o->iodepth;
1453
1454 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1455
1456 /*
1457 * Same as last one, we are done. Let it run a latency cycle, so
1458 * we get only the results from the targeted depth.
1459 */
1460 if (td->latency_qd == qd) {
1461 if (td->latency_end_run) {
1462 dprint(FD_RATE, "We are done\n");
1463 td->done = 1;
1464 } else {
1465 dprint(FD_RATE, "Quiesce and final run\n");
1466 io_u_quiesce(td);
1467 td->latency_end_run = 1;
1468 reset_all_stats(td);
1469 reset_io_stats(td);
1470 }
1471 }
1472
1473 lat_new_cycle(td);
1474}
1475
1476/*
1477 * Check if we can bump the queue depth
1478 */
1479void lat_target_check(struct thread_data *td)
1480{
1481 uint64_t usec_window;
1482 uint64_t ios;
1483 double success_ios;
1484
1485 usec_window = utime_since_now(&td->latency_ts);
1486 if (usec_window < td->o.latency_window)
1487 return;
1488
1489 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1490 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1491 success_ios *= 100.0;
1492
1493 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1494
1495 if (success_ios >= td->o.latency_percentile.u.f)
1496 lat_target_success(td);
1497 else
1498 __lat_target_failed(td);
1499}
1500
1501/*
1502 * If latency target is enabled, we might be ramping up or down and not
1503 * using the full queue depth available.
1504 */
1505bool queue_full(const struct thread_data *td)
1506{
1507 const int qempty = io_u_qempty(&td->io_u_freelist);
1508
1509 if (qempty)
1510 return true;
1511 if (!td->o.latency_target)
1512 return false;
1513
1514 return td->cur_depth >= td->latency_qd;
1515}
1516
1517struct io_u *__get_io_u(struct thread_data *td)
1518{
1519 const bool needs_lock = td_async_processing(td);
1520 struct io_u *io_u = NULL;
1521 int ret;
1522
1523 if (td->stop_io)
1524 return NULL;
1525
1526 if (needs_lock)
1527 __td_io_u_lock(td);
1528
1529again:
1530 if (!io_u_rempty(&td->io_u_requeues))
1531 io_u = io_u_rpop(&td->io_u_requeues);
1532 else if (!queue_full(td)) {
1533 io_u = io_u_qpop(&td->io_u_freelist);
1534
1535 io_u->file = NULL;
1536 io_u->buflen = 0;
1537 io_u->resid = 0;
1538 io_u->end_io = NULL;
1539 }
1540
1541 if (io_u) {
1542 assert(io_u->flags & IO_U_F_FREE);
1543 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
1544 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1545 IO_U_F_VER_LIST);
1546
1547 io_u->error = 0;
1548 io_u->acct_ddir = -1;
1549 td->cur_depth++;
1550 assert(!(td->flags & TD_F_CHILD));
1551 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
1552 io_u->ipo = NULL;
1553 } else if (td_async_processing(td)) {
1554 /*
1555 * We ran out, wait for async verify threads to finish and
1556 * return one
1557 */
1558 assert(!(td->flags & TD_F_CHILD));
1559 ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1560 assert(ret == 0);
1561 if (!td->error)
1562 goto again;
1563 }
1564
1565 if (needs_lock)
1566 __td_io_u_unlock(td);
1567
1568 return io_u;
1569}
1570
1571static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
1572{
1573 if (!(td->flags & TD_F_TRIM_BACKLOG))
1574 return false;
1575 if (!td->trim_entries)
1576 return false;
1577
1578 if (td->trim_batch) {
1579 td->trim_batch--;
1580 if (get_next_trim(td, io_u))
1581 return true;
1582 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1583 td->last_ddir != DDIR_READ) {
1584 td->trim_batch = td->o.trim_batch;
1585 if (!td->trim_batch)
1586 td->trim_batch = td->o.trim_backlog;
1587 if (get_next_trim(td, io_u))
1588 return true;
1589 }
1590
1591 return false;
1592}
1593
1594static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
1595{
1596 if (!(td->flags & TD_F_VER_BACKLOG))
1597 return false;
1598
1599 if (td->io_hist_len) {
1600 int get_verify = 0;
1601
1602 if (td->verify_batch)
1603 get_verify = 1;
1604 else if (!(td->io_hist_len % td->o.verify_backlog) &&
1605 td->last_ddir != DDIR_READ) {
1606 td->verify_batch = td->o.verify_batch;
1607 if (!td->verify_batch)
1608 td->verify_batch = td->o.verify_backlog;
1609 get_verify = 1;
1610 }
1611
1612 if (get_verify && !get_next_verify(td, io_u)) {
1613 td->verify_batch--;
1614 return true;
1615 }
1616 }
1617
1618 return false;
1619}
1620
1621/*
1622 * Fill offset and start time into the buffer content, to prevent too
1623 * easy compressible data for simple de-dupe attempts. Do this for every
1624 * 512b block in the range, since that should be the smallest block size
1625 * we can expect from a device.
1626 */
1627static void small_content_scramble(struct io_u *io_u)
1628{
1629 unsigned long long i, nr_blocks = io_u->buflen >> 9;
1630 unsigned int offset;
1631 uint64_t boffset, *iptr;
1632 char *p;
1633
1634 if (!nr_blocks)
1635 return;
1636
1637 p = io_u->xfer_buf;
1638 boffset = io_u->offset;
1639
1640 if (io_u->buf_filled_len)
1641 io_u->buf_filled_len = 0;
1642
1643 /*
1644 * Generate random index between 0..7. We do chunks of 512b, if
1645 * we assume a cacheline is 64 bytes, then we have 8 of those.
1646 * Scramble content within the blocks in the same cacheline to
1647 * speed things up.
1648 */
1649 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
1650
1651 for (i = 0; i < nr_blocks; i++) {
1652 /*
1653 * Fill offset into start of cacheline, time into end
1654 * of cacheline
1655 */
1656 iptr = (void *) p + (offset << 6);
1657 *iptr = boffset;
1658
1659 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1660 iptr[0] = io_u->start_time.tv_sec;
1661 iptr[1] = io_u->start_time.tv_nsec;
1662
1663 p += 512;
1664 boffset += 512;
1665 }
1666}
1667
1668/*
1669 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1670 * etc. The returned io_u is fully ready to be prepped, populated and submitted.
1671 */
1672struct io_u *get_io_u(struct thread_data *td)
1673{
1674 struct fio_file *f;
1675 struct io_u *io_u;
1676 int do_scramble = 0;
1677 long ret = 0;
1678
1679 io_u = __get_io_u(td);
1680 if (!io_u) {
1681 dprint(FD_IO, "__get_io_u failed\n");
1682 return NULL;
1683 }
1684
1685 if (check_get_verify(td, io_u))
1686 goto out;
1687 if (check_get_trim(td, io_u))
1688 goto out;
1689
1690 /*
1691 * from a requeue, io_u already setup
1692 */
1693 if (io_u->file)
1694 goto out;
1695
1696 /*
1697 * If using an iolog, grab next piece if any available.
1698 */
1699 if (td->flags & TD_F_READ_IOLOG) {
1700 if (read_iolog_get(td, io_u))
1701 goto err_put;
1702 } else if (set_io_u_file(td, io_u)) {
1703 ret = -EBUSY;
1704 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1705 goto err_put;
1706 }
1707
1708 f = io_u->file;
1709 if (!f) {
1710 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1711 goto err_put;
1712 }
1713
1714 assert(fio_file_open(f));
1715
1716 if (ddir_rw(io_u->ddir)) {
1717 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
1718 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
1719 goto err_put;
1720 }
1721
1722 f->last_start[io_u->ddir] = io_u->offset;
1723 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
1724
1725 if (io_u->ddir == DDIR_WRITE) {
1726 if (td->flags & TD_F_REFILL_BUFFERS) {
1727 io_u_fill_buffer(td, io_u,
1728 td->o.min_bs[DDIR_WRITE],
1729 io_u->buflen);
1730 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
1731 !(td->flags & TD_F_COMPRESS) &&
1732 !(td->flags & TD_F_DO_VERIFY))
1733 do_scramble = 1;
1734 } else if (io_u->ddir == DDIR_READ) {
1735 /*
1736 * Reset the buf_filled parameters so next time if the
1737 * buffer is used for writes it is refilled.
1738 */
1739 io_u->buf_filled_len = 0;
1740 }
1741 }
1742
1743 /*
1744 * Set io data pointers.
1745 */
1746 io_u->xfer_buf = io_u->buf;
1747 io_u->xfer_buflen = io_u->buflen;
1748
1749out:
1750 assert(io_u->file);
1751 if (!td_io_prep(td, io_u)) {
1752 if (!td->o.disable_lat)
1753 fio_gettime(&io_u->start_time, NULL);
1754
1755 if (do_scramble)
1756 small_content_scramble(io_u);
1757
1758 return io_u;
1759 }
1760err_put:
1761 dprint(FD_IO, "get_io_u failed\n");
1762 put_io_u(td, io_u);
1763 return ERR_PTR(ret);
1764}
1765
1766static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
1767{
1768 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
1769
1770 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1771 return;
1772
1773 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\n",
1774 io_u->file ? " on file " : "",
1775 io_u->file ? io_u->file->file_name : "",
1776 strerror(io_u->error),
1777 io_ddir_name(io_u->ddir),
1778 io_u->offset, io_u->xfer_buflen);
1779
1780 if (td->io_ops->errdetails) {
1781 char *err = td->io_ops->errdetails(io_u);
1782
1783 log_err("fio: %s\n", err);
1784 free(err);
1785 }
1786
1787 if (!td->error)
1788 td_verror(td, io_u->error, "io_u error");
1789}
1790
1791void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1792{
1793 __io_u_log_error(td, io_u);
1794 if (td->parent)
1795 __io_u_log_error(td->parent, io_u);
1796}
1797
1798static inline bool gtod_reduce(struct thread_data *td)
1799{
1800 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1801 || td->o.gtod_reduce;
1802}
1803
1804static void trim_block_info(struct thread_data *td, struct io_u *io_u)
1805{
1806 uint32_t *info = io_u_block_info(td, io_u);
1807
1808 if (BLOCK_INFO_STATE(*info) >= BLOCK_STATE_TRIM_FAILURE)
1809 return;
1810
1811 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, BLOCK_INFO_TRIMS(*info) + 1);
1812}
1813
1814static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1815 struct io_completion_data *icd,
1816 const enum fio_ddir idx, unsigned int bytes)
1817{
1818 const int no_reduce = !gtod_reduce(td);
1819 unsigned long long llnsec = 0;
1820
1821 if (td->parent)
1822 td = td->parent;
1823
1824 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
1825 return;
1826
1827 if (no_reduce)
1828 llnsec = ntime_since(&io_u->issue_time, &icd->time);
1829
1830 if (!td->o.disable_lat) {
1831 unsigned long long tnsec;
1832
1833 tnsec = ntime_since(&io_u->start_time, &icd->time);
1834 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
1835
1836 if (td->flags & TD_F_PROFILE_OPS) {
1837 struct prof_io_ops *ops = &td->prof_io_ops;
1838
1839 if (ops->io_u_lat)
1840 icd->error = ops->io_u_lat(td, tnsec);
1841 }
1842
1843 if (td->o.max_latency && tnsec > td->o.max_latency)
1844 lat_fatal(td, icd, tnsec, td->o.max_latency);
1845 if (td->o.latency_target && tnsec > td->o.latency_target) {
1846 if (lat_target_failed(td))
1847 lat_fatal(td, icd, tnsec, td->o.latency_target);
1848 }
1849 }
1850
1851 if (ddir_rw(idx)) {
1852 if (!td->o.disable_clat) {
1853 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1854 io_u_mark_latency(td, llnsec);
1855 }
1856
1857 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1858 add_bw_sample(td, io_u, bytes, llnsec);
1859
1860 if (no_reduce && per_unit_log(td->iops_log))
1861 add_iops_sample(td, io_u, bytes);
1862 } else if (ddir_sync(idx) && !td->o.disable_clat)
1863 add_sync_clat_sample(&td->ts, llnsec);
1864
1865 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM)
1866 trim_block_info(td, io_u);
1867}
1868
1869static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1870 uint64_t offset, unsigned int bytes)
1871{
1872 int idx;
1873
1874 if (!f)
1875 return;
1876
1877 if (f->first_write == -1ULL || offset < f->first_write)
1878 f->first_write = offset;
1879 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1880 f->last_write = offset + bytes;
1881
1882 if (!f->last_write_comp)
1883 return;
1884
1885 idx = f->last_write_idx++;
1886 f->last_write_comp[idx] = offset;
1887 if (f->last_write_idx == td->o.iodepth)
1888 f->last_write_idx = 0;
1889}
1890
1891static bool should_account(struct thread_data *td)
1892{
1893 return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1894 td->runstate == TD_VERIFYING);
1895}
1896
1897static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
1898 struct io_completion_data *icd)
1899{
1900 struct io_u *io_u = *io_u_ptr;
1901 enum fio_ddir ddir = io_u->ddir;
1902 struct fio_file *f = io_u->file;
1903
1904 dprint_io_u(io_u, "complete");
1905
1906 assert(io_u->flags & IO_U_F_FLIGHT);
1907 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
1908
1909 /*
1910 * Mark IO ok to verify
1911 */
1912 if (io_u->ipo) {
1913 /*
1914 * Remove errored entry from the verification list
1915 */
1916 if (io_u->error)
1917 unlog_io_piece(td, io_u);
1918 else {
1919 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1920 write_barrier();
1921 }
1922 }
1923
1924 if (ddir_sync(ddir)) {
1925 td->last_was_sync = true;
1926 if (f) {
1927 f->first_write = -1ULL;
1928 f->last_write = -1ULL;
1929 }
1930 if (should_account(td))
1931 account_io_completion(td, io_u, icd, ddir, io_u->buflen);
1932 return;
1933 }
1934
1935 td->last_was_sync = false;
1936 td->last_ddir = ddir;
1937
1938 if (!io_u->error && ddir_rw(ddir)) {
1939 unsigned long long bytes = io_u->buflen - io_u->resid;
1940 int ret;
1941
1942 td->io_blocks[ddir]++;
1943 td->io_bytes[ddir] += bytes;
1944
1945 if (!(io_u->flags & IO_U_F_VER_LIST)) {
1946 td->this_io_blocks[ddir]++;
1947 td->this_io_bytes[ddir] += bytes;
1948 }
1949
1950 if (ddir == DDIR_WRITE)
1951 file_log_write_comp(td, f, io_u->offset, bytes);
1952
1953 if (should_account(td))
1954 account_io_completion(td, io_u, icd, ddir, bytes);
1955
1956 icd->bytes_done[ddir] += bytes;
1957
1958 if (io_u->end_io) {
1959 ret = io_u->end_io(td, io_u_ptr);
1960 io_u = *io_u_ptr;
1961 if (ret && !icd->error)
1962 icd->error = ret;
1963 }
1964 } else if (io_u->error) {
1965 icd->error = io_u->error;
1966 io_u_log_error(td, io_u);
1967 }
1968 if (icd->error) {
1969 enum error_type_bit eb = td_error_type(ddir, icd->error);
1970
1971 if (!td_non_fatal_error(td, eb, icd->error))
1972 return;
1973
1974 /*
1975 * If there is a non_fatal error, then add to the error count
1976 * and clear all the errors.
1977 */
1978 update_error_count(td, icd->error);
1979 td_clear_error(td);
1980 icd->error = 0;
1981 if (io_u)
1982 io_u->error = 0;
1983 }
1984}
1985
1986static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1987 int nr)
1988{
1989 int ddir;
1990
1991 if (!gtod_reduce(td))
1992 fio_gettime(&icd->time, NULL);
1993
1994 icd->nr = nr;
1995
1996 icd->error = 0;
1997 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
1998 icd->bytes_done[ddir] = 0;
1999}
2000
2001static void ios_completed(struct thread_data *td,
2002 struct io_completion_data *icd)
2003{
2004 struct io_u *io_u;
2005 int i;
2006
2007 for (i = 0; i < icd->nr; i++) {
2008 io_u = td->io_ops->event(td, i);
2009
2010 io_completed(td, &io_u, icd);
2011
2012 if (io_u)
2013 put_io_u(td, io_u);
2014 }
2015}
2016
2017/*
2018 * Complete a single io_u for the sync engines.
2019 */
2020int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
2021{
2022 struct io_completion_data icd;
2023 int ddir;
2024
2025 init_icd(td, &icd, 1);
2026 io_completed(td, &io_u, &icd);
2027
2028 if (io_u)
2029 put_io_u(td, io_u);
2030
2031 if (icd.error) {
2032 td_verror(td, icd.error, "io_u_sync_complete");
2033 return -1;
2034 }
2035
2036 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2037 td->bytes_done[ddir] += icd.bytes_done[ddir];
2038
2039 return 0;
2040}
2041
2042/*
2043 * Called to complete min_events number of io for the async engines.
2044 */
2045int io_u_queued_complete(struct thread_data *td, int min_evts)
2046{
2047 struct io_completion_data icd;
2048 struct timespec *tvp = NULL;
2049 int ret, ddir;
2050 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
2051
2052 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
2053
2054 if (!min_evts)
2055 tvp = &ts;
2056 else if (min_evts > td->cur_depth)
2057 min_evts = td->cur_depth;
2058
2059 /* No worries, td_io_getevents fixes min and max if they are
2060 * set incorrectly */
2061 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
2062 if (ret < 0) {
2063 td_verror(td, -ret, "td_io_getevents");
2064 return ret;
2065 } else if (!ret)
2066 return ret;
2067
2068 init_icd(td, &icd, ret);
2069 ios_completed(td, &icd);
2070 if (icd.error) {
2071 td_verror(td, icd.error, "io_u_queued_complete");
2072 return -1;
2073 }
2074
2075 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
2076 td->bytes_done[ddir] += icd.bytes_done[ddir];
2077
2078 return ret;
2079}
2080
2081/*
2082 * Call when io_u is really queued, to update the submission latency.
2083 */
2084void io_u_queued(struct thread_data *td, struct io_u *io_u)
2085{
2086 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
2087 unsigned long slat_time;
2088
2089 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
2090
2091 if (td->parent)
2092 td = td->parent;
2093
2094 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2095 io_u->offset);
2096 }
2097}
2098
2099/*
2100 * See if we should reuse the last seed, if dedupe is enabled
2101 */
2102static struct frand_state *get_buf_state(struct thread_data *td)
2103{
2104 unsigned int v;
2105
2106 if (!td->o.dedupe_percentage)
2107 return &td->buf_state;
2108 else if (td->o.dedupe_percentage == 100) {
2109 frand_copy(&td->buf_state_prev, &td->buf_state);
2110 return &td->buf_state;
2111 }
2112
2113 v = rand_between(&td->dedupe_state, 1, 100);
2114
2115 if (v <= td->o.dedupe_percentage)
2116 return &td->buf_state_prev;
2117
2118 return &td->buf_state;
2119}
2120
2121static void save_buf_state(struct thread_data *td, struct frand_state *rs)
2122{
2123 if (td->o.dedupe_percentage == 100)
2124 frand_copy(rs, &td->buf_state_prev);
2125 else if (rs == &td->buf_state)
2126 frand_copy(&td->buf_state_prev, rs);
2127}
2128
2129void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write,
2130 unsigned long long max_bs)
2131{
2132 struct thread_options *o = &td->o;
2133
2134 if (o->mem_type == MEM_CUDA_MALLOC)
2135 return;
2136
2137 if (o->compress_percentage || o->dedupe_percentage) {
2138 unsigned int perc = td->o.compress_percentage;
2139 struct frand_state *rs;
2140 unsigned long long left = max_bs;
2141 unsigned long long this_write;
2142
2143 do {
2144 rs = get_buf_state(td);
2145
2146 min_write = min(min_write, left);
2147
2148 if (perc) {
2149 this_write = min_not_zero(min_write,
2150 (unsigned long long) td->o.compress_chunk);
2151
2152 fill_random_buf_percentage(rs, buf, perc,
2153 this_write, this_write,
2154 o->buffer_pattern,
2155 o->buffer_pattern_bytes);
2156 } else {
2157 fill_random_buf(rs, buf, min_write);
2158 this_write = min_write;
2159 }
2160
2161 buf += this_write;
2162 left -= this_write;
2163 save_buf_state(td, rs);
2164 } while (left);
2165 } else if (o->buffer_pattern_bytes)
2166 fill_buffer_pattern(td, buf, max_bs);
2167 else if (o->zero_buffers)
2168 memset(buf, 0, max_bs);
2169 else
2170 fill_random_buf(get_buf_state(td), buf, max_bs);
2171}
2172
2173/*
2174 * "randomly" fill the buffer contents
2175 */
2176void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2177 unsigned long long min_write, unsigned long long max_bs)
2178{
2179 io_u->buf_filled_len = 0;
2180 fill_io_buffer(td, io_u->buf, min_write, max_bs);
2181}
2182
2183static int do_sync_file_range(const struct thread_data *td,
2184 struct fio_file *f)
2185{
2186 off64_t offset, nbytes;
2187
2188 offset = f->first_write;
2189 nbytes = f->last_write - f->first_write;
2190
2191 if (!nbytes)
2192 return 0;
2193
2194 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2195}
2196
2197int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2198{
2199 int ret;
2200
2201 if (io_u->ddir == DDIR_SYNC) {
2202 ret = fsync(io_u->file->fd);
2203 } else if (io_u->ddir == DDIR_DATASYNC) {
2204#ifdef CONFIG_FDATASYNC
2205 ret = fdatasync(io_u->file->fd);
2206#else
2207 ret = io_u->xfer_buflen;
2208 io_u->error = EINVAL;
2209#endif
2210 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2211 ret = do_sync_file_range(td, io_u->file);
2212 else {
2213 ret = io_u->xfer_buflen;
2214 io_u->error = EINVAL;
2215 }
2216
2217 if (ret < 0)
2218 io_u->error = errno;
2219
2220 return ret;
2221}
2222
2223int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2224{
2225#ifndef FIO_HAVE_TRIM
2226 io_u->error = EINVAL;
2227 return 0;
2228#else
2229 struct fio_file *f = io_u->file;
2230 int ret;
2231
2232 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
2233 if (!ret)
2234 return io_u->xfer_buflen;
2235
2236 io_u->error = ret;
2237 return 0;
2238#endif
2239}
fio - Flexible IO Tester