debug: make debug=io readable with multiple threads
[fio.git] / io_u.c
CommitLineData
10ba535a
JA
1#include <unistd.h>
2#include <fcntl.h>
3#include <string.h>
4#include <signal.h>
5#include <time.h>
0c6e7517 6#include <assert.h>
10ba535a
JA
7
8#include "fio.h"
5973cafb 9#include "hash.h"
4f5af7b2 10#include "verify.h"
0d29de83 11#include "trim.h"
1fbbf72e 12#include "lib/rand.h"
7ebd796f 13#include "lib/axmap.h"
002fe734 14#include "err.h"
0f38bbef 15#include "lib/pow2.h"
4b157ac6 16#include "minmax.h"
10ba535a 17
97601024
JA
18struct io_completion_data {
19 int nr; /* input */
97601024
JA
20
21 int error; /* output */
100f49f1 22 uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */
8b6a404c 23 struct timespec time; /* output */
97601024
JA
24};
25
10ba535a 26/*
7ebd796f 27 * The ->io_axmap contains a map of blocks we have or have not done io
10ba535a
JA
28 * to yet. Used to make sure we cover the entire range in a fair fashion.
29 */
e39c0676 30static bool random_map_free(struct fio_file *f, const uint64_t block)
10ba535a 31{
7ebd796f 32 return !axmap_isset(f->io_axmap, block);
10ba535a
JA
33}
34
df415585
JA
35/*
36 * Mark a given offset as used in the map.
37 */
9bf2061e 38static void mark_random_map(struct thread_data *td, struct io_u *io_u)
df415585 39{
956e60ea 40 unsigned int min_bs = td->o.min_bs[io_u->ddir];
9bf2061e 41 struct fio_file *f = io_u->file;
51ede0b1 42 unsigned int nr_blocks;
1ae83d45 43 uint64_t block;
df415585 44
1ae83d45 45 block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
c685b5b2 46 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
df415585 47
2ab9e98b 48 if (!(io_u->flags & IO_U_F_BUSY_OK))
7ebd796f 49 nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
df415585 50
51ede0b1
JA
51 if ((nr_blocks * min_bs) < io_u->buflen)
52 io_u->buflen = nr_blocks * min_bs;
df415585
JA
53}
54
74776733
JA
55static uint64_t last_block(struct thread_data *td, struct fio_file *f,
56 enum fio_ddir ddir)
2ba1c290 57{
74776733
JA
58 uint64_t max_blocks;
59 uint64_t max_size;
2ba1c290 60
ff58fced
JA
61 assert(ddir_rw(ddir));
62
d9dd70f7
JA
63 /*
64 * Hmm, should we make sure that ->io_size <= ->real_file_size?
79591fa9 65 * -> not for now since there is code assuming it could go either.
d9dd70f7
JA
66 */
67 max_size = f->io_size;
68 if (max_size > f->real_file_size)
69 max_size = f->real_file_size;
70
ed335855
SN
71 if (td->o.zone_range)
72 max_size = td->o.zone_range;
73
0412d12e
JE
74 if (td->o.min_bs[ddir] > td->o.ba[ddir])
75 max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
76
1ae83d45 77 max_blocks = max_size / (uint64_t) td->o.ba[ddir];
2ba1c290
JA
78 if (!max_blocks)
79 return 0;
80
67778e88 81 return max_blocks;
2ba1c290
JA
82}
83
1ae83d45
JA
84struct rand_off {
85 struct flist_head list;
86 uint64_t off;
87};
88
e25839d4 89static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
e0a04ac1
JA
90 enum fio_ddir ddir, uint64_t *b,
91 uint64_t lastb)
ec4015da 92{
6f49f8bc 93 uint64_t r;
5e0baa7f 94
c3546b53
JA
95 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
96 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
6f49f8bc 97
d6b72507 98 r = __rand(&td->random_state);
8055e41d 99
4b91ee8f 100 dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
8055e41d 101
e0a04ac1 102 *b = lastb * (r / (rand_max(&td->random_state) + 1.0));
51ede0b1 103 } else {
8055e41d 104 uint64_t off = 0;
43c63a78 105
967d1b63
JA
106 assert(fio_file_lfsr(f));
107
6f49f8bc 108 if (lfsr_next(&f->lfsr, &off))
8055e41d 109 return 1;
ec4015da 110
8055e41d
JA
111 *b = off;
112 }
0ce8b119 113
ec4015da 114 /*
51ede0b1 115 * if we are not maintaining a random map, we are done.
ec4015da 116 */
51ede0b1
JA
117 if (!file_randommap(td, f))
118 goto ret;
43c63a78
JA
119
120 /*
51ede0b1 121 * calculate map offset and check if it's free
43c63a78 122 */
51ede0b1
JA
123 if (random_map_free(f, *b))
124 goto ret;
125
4b91ee8f
JA
126 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
127 (unsigned long long) *b);
51ede0b1 128
7ebd796f 129 *b = axmap_next_free(f->io_axmap, *b);
51ede0b1
JA
130 if (*b == (uint64_t) -1ULL)
131 return 1;
0ce8b119
JA
132ret:
133 return 0;
ec4015da
JA
134}
135
925fee33
JA
136static int __get_next_rand_offset_zipf(struct thread_data *td,
137 struct fio_file *f, enum fio_ddir ddir,
1ae83d45 138 uint64_t *b)
e25839d4 139{
9c6f6316 140 *b = zipf_next(&f->zipf);
e25839d4
JA
141 return 0;
142}
143
925fee33
JA
144static int __get_next_rand_offset_pareto(struct thread_data *td,
145 struct fio_file *f, enum fio_ddir ddir,
1ae83d45 146 uint64_t *b)
925fee33 147{
9c6f6316 148 *b = pareto_next(&f->zipf);
925fee33
JA
149 return 0;
150}
151
56d9fa4b
JA
152static int __get_next_rand_offset_gauss(struct thread_data *td,
153 struct fio_file *f, enum fio_ddir ddir,
154 uint64_t *b)
155{
156 *b = gauss_next(&f->gauss);
157 return 0;
158}
159
59466396
JA
160static int __get_next_rand_offset_zoned_abs(struct thread_data *td,
161 struct fio_file *f,
162 enum fio_ddir ddir, uint64_t *b)
163{
164 struct zone_split_index *zsi;
e3c8c108 165 uint64_t lastb, send, stotal;
59466396
JA
166 static int warned;
167 unsigned int v;
168
169 lastb = last_block(td, f, ddir);
170 if (!lastb)
171 return 1;
172
173 if (!td->o.zone_split_nr[ddir]) {
174bail:
175 return __get_next_rand_offset(td, f, ddir, b, lastb);
176 }
177
178 /*
179 * Generate a value, v, between 1 and 100, both inclusive
180 */
181 v = rand32_between(&td->zone_state, 1, 100);
182
e3c8c108
JA
183 /*
184 * Find our generated table. 'send' is the end block of this zone,
185 * 'stotal' is our start offset.
186 */
59466396
JA
187 zsi = &td->zone_state_index[ddir][v - 1];
188 stotal = zsi->size_prev / td->o.ba[ddir];
189 send = zsi->size / td->o.ba[ddir];
190
191 /*
192 * Should never happen
193 */
194 if (send == -1U) {
195 if (!warned) {
196 log_err("fio: bug in zoned generation\n");
197 warned = 1;
198 }
199 goto bail;
200 } else if (send > lastb) {
201 /*
202 * This happens if the user specifies ranges that exceed
203 * the file/device size. We can't handle that gracefully,
204 * so error and exit.
205 */
206 log_err("fio: zoned_abs sizes exceed file size\n");
207 return 1;
208 }
209
210 /*
e3c8c108 211 * Generate index from 0..send-stotal
59466396 212 */
e3c8c108 213 if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1)
59466396
JA
214 return 1;
215
e3c8c108 216 *b += stotal;
59466396
JA
217 return 0;
218}
219
e0a04ac1
JA
220static int __get_next_rand_offset_zoned(struct thread_data *td,
221 struct fio_file *f, enum fio_ddir ddir,
222 uint64_t *b)
223{
224 unsigned int v, send, stotal;
225 uint64_t offset, lastb;
226 static int warned;
227 struct zone_split_index *zsi;
228
229 lastb = last_block(td, f, ddir);
230 if (!lastb)
231 return 1;
232
233 if (!td->o.zone_split_nr[ddir]) {
234bail:
235 return __get_next_rand_offset(td, f, ddir, b, lastb);
236 }
237
238 /*
239 * Generate a value, v, between 1 and 100, both inclusive
240 */
36dd3379 241 v = rand32_between(&td->zone_state, 1, 100);
e0a04ac1
JA
242
243 zsi = &td->zone_state_index[ddir][v - 1];
244 stotal = zsi->size_perc_prev;
245 send = zsi->size_perc;
246
247 /*
248 * Should never happen
249 */
250 if (send == -1U) {
251 if (!warned) {
252 log_err("fio: bug in zoned generation\n");
253 warned = 1;
254 }
255 goto bail;
256 }
257
258 /*
259 * 'send' is some percentage below or equal to 100 that
260 * marks the end of the current IO range. 'stotal' marks
261 * the start, in percent.
262 */
263 if (stotal)
264 offset = stotal * lastb / 100ULL;
265 else
266 offset = 0;
267
268 lastb = lastb * (send - stotal) / 100ULL;
269
270 /*
271 * Generate index from 0..send-of-lastb
272 */
273 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
274 return 1;
275
276 /*
277 * Add our start offset, if any
278 */
279 if (offset)
280 *b += offset;
281
282 return 0;
283}
56d9fa4b 284
1ae83d45
JA
285static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
286{
287 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
288 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
289
290 return r1->off - r2->off;
291}
292
293static int get_off_from_method(struct thread_data *td, struct fio_file *f,
294 enum fio_ddir ddir, uint64_t *b)
e25839d4 295{
e0a04ac1
JA
296 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
297 uint64_t lastb;
298
299 lastb = last_block(td, f, ddir);
300 if (!lastb)
301 return 1;
302
303 return __get_next_rand_offset(td, f, ddir, b, lastb);
304 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
e25839d4 305 return __get_next_rand_offset_zipf(td, f, ddir, b);
925fee33
JA
306 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
307 return __get_next_rand_offset_pareto(td, f, ddir, b);
56d9fa4b
JA
308 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
309 return __get_next_rand_offset_gauss(td, f, ddir, b);
e0a04ac1
JA
310 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
311 return __get_next_rand_offset_zoned(td, f, ddir, b);
59466396
JA
312 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
313 return __get_next_rand_offset_zoned_abs(td, f, ddir, b);
e25839d4
JA
314
315 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
316 return 1;
317}
318
bcd5abfa
JA
319/*
320 * Sort the reads for a verify phase in batches of verifysort_nr, if
321 * specified.
322 */
e39c0676 323static inline bool should_sort_io(struct thread_data *td)
bcd5abfa
JA
324{
325 if (!td->o.verifysort_nr || !td->o.do_verify)
e39c0676 326 return false;
bcd5abfa 327 if (!td_random(td))
e39c0676 328 return false;
bcd5abfa 329 if (td->runstate != TD_VERIFYING)
e39c0676 330 return false;
c3546b53
JA
331 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
332 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
e39c0676 333 return false;
bcd5abfa 334
e39c0676 335 return true;
bcd5abfa
JA
336}
337
e39c0676 338static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
211c9b89
JA
339{
340 unsigned int v;
211c9b89 341
d9472271 342 if (td->o.perc_rand[ddir] == 100)
e39c0676 343 return true;
211c9b89 344
36dd3379 345 v = rand32_between(&td->seq_rand_state[ddir], 1, 100);
211c9b89 346
d9472271 347 return v <= td->o.perc_rand[ddir];
211c9b89
JA
348}
349
1ae83d45
JA
350static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
351 enum fio_ddir ddir, uint64_t *b)
352{
353 struct rand_off *r;
354 int i, ret = 1;
355
bcd5abfa 356 if (!should_sort_io(td))
1ae83d45
JA
357 return get_off_from_method(td, f, ddir, b);
358
359 if (!flist_empty(&td->next_rand_list)) {
1ae83d45 360fetch:
9342d5f8 361 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
1ae83d45
JA
362 flist_del(&r->list);
363 *b = r->off;
364 free(r);
365 return 0;
366 }
367
368 for (i = 0; i < td->o.verifysort_nr; i++) {
369 r = malloc(sizeof(*r));
370
371 ret = get_off_from_method(td, f, ddir, &r->off);
372 if (ret) {
373 free(r);
374 break;
375 }
376
377 flist_add(&r->list, &td->next_rand_list);
378 }
379
380 if (ret && !i)
381 return ret;
382
383 assert(!flist_empty(&td->next_rand_list));
384 flist_sort(NULL, &td->next_rand_list, flist_cmp);
385 goto fetch;
386}
387
0bcf41cd
JA
388static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f)
389{
390 struct thread_options *o = &td->o;
391
392 if (o->invalidate_cache && !o->odirect) {
393 int fio_unused ret;
394
395 ret = file_invalidate_cache(td, f);
396 }
397}
398
38dad62d 399static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
1ae83d45 400 enum fio_ddir ddir, uint64_t *b)
38dad62d 401{
c04e4661
DE
402 if (!get_next_rand_offset(td, f, ddir, b))
403 return 0;
404
8c07860d
JA
405 if (td->o.time_based ||
406 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
33c48814 407 fio_file_reset(td, f);
c04e4661
DE
408 if (!get_next_rand_offset(td, f, ddir, b))
409 return 0;
0bcf41cd 410 loop_cache_invalidate(td, f);
38dad62d
JA
411 }
412
c04e4661 413 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
f1dfb668 414 f->file_name, (unsigned long long) f->last_pos[ddir],
4b91ee8f 415 (unsigned long long) f->real_file_size);
c04e4661 416 return 1;
38dad62d
JA
417}
418
37cf9e3c 419static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
1ae83d45 420 enum fio_ddir ddir, uint64_t *offset)
38dad62d 421{
ac002339
JA
422 struct thread_options *o = &td->o;
423
ff58fced
JA
424 assert(ddir_rw(ddir));
425
17373ce2
JA
426 /*
427 * If we reach the end for a time based run, reset us back to 0
428 * and invalidate the cache, if we need to.
429 */
f1dfb668 430 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
19ddc35b 431 o->time_based) {
c89daa4a 432 f->last_pos[ddir] = f->file_offset;
0bcf41cd 433 loop_cache_invalidate(td, f);
19ddc35b 434 }
c04e4661 435
f1dfb668 436 if (f->last_pos[ddir] < f->real_file_size) {
1ae83d45 437 uint64_t pos;
059b0802 438
c22825bb
JA
439 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) {
440 if (f->real_file_size > f->io_size)
441 f->last_pos[ddir] = f->io_size;
442 else
443 f->last_pos[ddir] = f->real_file_size;
444 }
a66da7a2 445
f1dfb668 446 pos = f->last_pos[ddir] - f->file_offset;
ac002339
JA
447 if (pos && o->ddir_seq_add) {
448 pos += o->ddir_seq_add;
449
450 /*
451 * If we reach beyond the end of the file
452 * with holed IO, wrap around to the
b0a84f48
JA
453 * beginning again. If we're doing backwards IO,
454 * wrap to the end.
ac002339 455 */
b0a84f48
JA
456 if (pos >= f->real_file_size) {
457 if (o->ddir_seq_add > 0)
458 pos = f->file_offset;
c22825bb
JA
459 else {
460 if (f->real_file_size > f->io_size)
461 pos = f->io_size;
462 else
463 pos = f->real_file_size;
464
465 pos += o->ddir_seq_add;
466 }
b0a84f48 467 }
ac002339 468 }
059b0802 469
37cf9e3c 470 *offset = pos;
38dad62d
JA
471 return 0;
472 }
473
474 return 1;
475}
476
477static int get_next_block(struct thread_data *td, struct io_u *io_u,
6aca9b3d
JA
478 enum fio_ddir ddir, int rw_seq,
479 unsigned int *is_random)
38dad62d
JA
480{
481 struct fio_file *f = io_u->file;
1ae83d45 482 uint64_t b, offset;
38dad62d
JA
483 int ret;
484
ff58fced
JA
485 assert(ddir_rw(ddir));
486
37cf9e3c
JA
487 b = offset = -1ULL;
488
38dad62d 489 if (rw_seq) {
211c9b89 490 if (td_random(td)) {
6aca9b3d 491 if (should_do_random(td, ddir)) {
211c9b89 492 ret = get_next_rand_block(td, f, ddir, &b);
6aca9b3d
JA
493 *is_random = 1;
494 } else {
495 *is_random = 0;
1651e431 496 io_u_set(td, io_u, IO_U_F_BUSY_OK);
211c9b89
JA
497 ret = get_next_seq_offset(td, f, ddir, &offset);
498 if (ret)
499 ret = get_next_rand_block(td, f, ddir, &b);
500 }
6aca9b3d
JA
501 } else {
502 *is_random = 0;
37cf9e3c 503 ret = get_next_seq_offset(td, f, ddir, &offset);
6aca9b3d 504 }
38dad62d 505 } else {
1651e431 506 io_u_set(td, io_u, IO_U_F_BUSY_OK);
6aca9b3d 507 *is_random = 0;
38dad62d
JA
508
509 if (td->o.rw_seq == RW_SEQ_SEQ) {
37cf9e3c 510 ret = get_next_seq_offset(td, f, ddir, &offset);
6aca9b3d 511 if (ret) {
37cf9e3c 512 ret = get_next_rand_block(td, f, ddir, &b);
6aca9b3d
JA
513 *is_random = 0;
514 }
38dad62d 515 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
f1dfb668
JA
516 if (f->last_start[ddir] != -1ULL)
517 offset = f->last_start[ddir] - f->file_offset;
38dad62d 518 else
37cf9e3c 519 offset = 0;
38dad62d
JA
520 ret = 0;
521 } else {
522 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
523 ret = 1;
524 }
525 }
6d68b997 526
37cf9e3c
JA
527 if (!ret) {
528 if (offset != -1ULL)
529 io_u->offset = offset;
530 else if (b != -1ULL)
531 io_u->offset = b * td->o.ba[ddir];
532 else {
4e0a8fa2 533 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
37cf9e3c
JA
534 ret = 1;
535 }
536 }
537
38dad62d
JA
538 return ret;
539}
540
10ba535a
JA
541/*
542 * For random io, generate a random new block and see if it's used. Repeat
543 * until we find a free one. For sequential io, just return the end of
544 * the last io issued.
545 */
6aca9b3d
JA
546static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
547 unsigned int *is_random)
10ba535a 548{
9bf2061e 549 struct fio_file *f = io_u->file;
4ba66134 550 enum fio_ddir ddir = io_u->ddir;
38dad62d 551 int rw_seq_hit = 0;
10ba535a 552
ff58fced
JA
553 assert(ddir_rw(ddir));
554
38dad62d
JA
555 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
556 rw_seq_hit = 1;
5736c10d 557 td->ddir_seq_nr = td->o.ddir_seq_nr;
38dad62d 558 }
211097b2 559
6aca9b3d 560 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
38dad62d 561 return 1;
10ba535a 562
009bd847
JA
563 if (io_u->offset >= f->io_size) {
564 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
4b91ee8f
JA
565 (unsigned long long) io_u->offset,
566 (unsigned long long) f->io_size);
009bd847
JA
567 return 1;
568 }
569
570 io_u->offset += f->file_offset;
2ba1c290
JA
571 if (io_u->offset >= f->real_file_size) {
572 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
4b91ee8f
JA
573 (unsigned long long) io_u->offset,
574 (unsigned long long) f->real_file_size);
10ba535a 575 return 1;
2ba1c290 576 }
10ba535a
JA
577
578 return 0;
579}
580
6aca9b3d
JA
581static int get_next_offset(struct thread_data *td, struct io_u *io_u,
582 unsigned int *is_random)
15dc1934 583{
d72be545
JA
584 if (td->flags & TD_F_PROFILE_OPS) {
585 struct prof_io_ops *ops = &td->prof_io_ops;
7eb36574 586
d72be545 587 if (ops->fill_io_u_off)
6aca9b3d 588 return ops->fill_io_u_off(td, io_u, is_random);
d72be545 589 }
15dc1934 590
6aca9b3d 591 return __get_next_offset(td, io_u, is_random);
15dc1934
JA
592}
593
e39c0676
JA
594static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
595 unsigned int buflen)
79944128
JA
596{
597 struct fio_file *f = io_u->file;
598
bedc9dc2 599 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
79944128
JA
600}
601
6aca9b3d
JA
602static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
603 unsigned int is_random)
10ba535a 604{
6aca9b3d 605 int ddir = io_u->ddir;
24d23ca7 606 unsigned int buflen = 0;
f3059de1 607 unsigned int minbs, maxbs;
3dd29f7c 608 uint64_t frand_max, r;
7c961359 609 bool power_2;
10ba535a 610
9ee1c647 611 assert(ddir_rw(ddir));
6aca9b3d
JA
612
613 if (td->o.bs_is_seq_rand)
614 ddir = is_random ? DDIR_WRITE: DDIR_READ;
ff58fced 615
f3059de1
JA
616 minbs = td->o.min_bs[ddir];
617 maxbs = td->o.max_bs[ddir];
618
79944128
JA
619 if (minbs == maxbs)
620 return minbs;
621
52c58027
JA
622 /*
623 * If we can't satisfy the min block size from here, then fail
624 */
625 if (!io_u_fits(td, io_u, minbs))
626 return 0;
627
2f282cec 628 frand_max = rand_max(&td->bsrange_state[ddir]);
79944128 629 do {
2f282cec 630 r = __rand(&td->bsrange_state[ddir]);
4c07ad86 631
720e84ad 632 if (!td->o.bssplit_nr[ddir]) {
f3059de1 633 buflen = 1 + (unsigned int) ((double) maxbs *
c3546b53 634 (r / (frand_max + 1.0)));
f3059de1
JA
635 if (buflen < minbs)
636 buflen = minbs;
5ec10eaa 637 } else {
3dd29f7c 638 long long perc = 0;
564ca972
JA
639 unsigned int i;
640
720e84ad
JA
641 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
642 struct bssplit *bsp = &td->o.bssplit[ddir][i];
564ca972
JA
643
644 buflen = bsp->bs;
645 perc += bsp->perc;
3dd29f7c
JA
646 if (!perc)
647 break;
648 if ((r / perc <= frand_max / 100ULL) &&
79944128 649 io_u_fits(td, io_u, buflen))
564ca972
JA
650 break;
651 }
652 }
79944128 653
17a6b702
PL
654 power_2 = is_power_of_2(minbs);
655 if (!td->o.bs_unaligned && power_2)
7c306bb1 656 buflen &= ~(minbs - 1);
17a6b702
PL
657 else if (!td->o.bs_unaligned && !power_2)
658 buflen -= buflen % minbs;
79944128 659 } while (!io_u_fits(td, io_u, buflen));
6a5e6884 660
10ba535a
JA
661 return buflen;
662}
663
6aca9b3d
JA
664static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
665 unsigned int is_random)
15dc1934 666{
d72be545
JA
667 if (td->flags & TD_F_PROFILE_OPS) {
668 struct prof_io_ops *ops = &td->prof_io_ops;
7eb36574 669
d72be545 670 if (ops->fill_io_u_size)
6aca9b3d 671 return ops->fill_io_u_size(td, io_u, is_random);
d72be545 672 }
15dc1934 673
6aca9b3d 674 return __get_next_buflen(td, io_u, is_random);
15dc1934
JA
675}
676
afe24a5a
JA
677static void set_rwmix_bytes(struct thread_data *td)
678{
afe24a5a
JA
679 unsigned int diff;
680
681 /*
682 * we do time or byte based switch. this is needed because
683 * buffered writes may issue a lot quicker than they complete,
684 * whereas reads do not.
685 */
e47f799f 686 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
04c540d9 687 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
e47f799f
JA
688}
689
690static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
691{
692 unsigned int v;
e47f799f 693
36dd3379 694 v = rand32_between(&td->rwmix_state, 1, 100);
4c07ad86 695
04c540d9 696 if (v <= td->o.rwmix[DDIR_READ])
e47f799f
JA
697 return DDIR_READ;
698
699 return DDIR_WRITE;
afe24a5a
JA
700}
701
0d593542 702int io_u_quiesce(struct thread_data *td)
002e7183 703{
0d593542
JA
704 int completed = 0;
705
002e7183
JA
706 /*
707 * We are going to sleep, ensure that we flush anything pending as
708 * not to skew our latency numbers.
709 *
710 * Changed to only monitor 'in flight' requests here instead of the
711 * td->cur_depth, b/c td->cur_depth does not accurately represent
712 * io's that have been actually submitted to an async engine,
713 * and cur_depth is meaningless for sync engines.
714 */
9cc80b6d
JA
715 if (td->io_u_queued || td->cur_depth) {
716 int fio_unused ret;
717
718 ret = td_io_commit(td);
719 }
720
002e7183 721 while (td->io_u_in_flight) {
89fd0573 722 int ret;
002e7183 723
55312f9f 724 ret = io_u_queued_complete(td, 1);
0d593542
JA
725 if (ret > 0)
726 completed += ret;
002e7183 727 }
0d593542 728
6be06c46
JA
729 if (td->flags & TD_F_REGROW_LOGS)
730 regrow_logs(td);
731
0d593542 732 return completed;
002e7183
JA
733}
734
581e7141
JA
735static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
736{
737 enum fio_ddir odir = ddir ^ 1;
b5407f8b 738 uint64_t usec;
90eff1c9 739 uint64_t now;
581e7141 740
ff58fced 741 assert(ddir_rw(ddir));
50a8ce86 742 now = utime_since_now(&td->start);
ff58fced 743
50a8ce86
D
744 /*
745 * if rate_next_io_time is in the past, need to catch up to rate
746 */
747 if (td->rate_next_io_time[ddir] <= now)
581e7141
JA
748 return ddir;
749
750 /*
50a8ce86 751 * We are ahead of rate in this direction. See if we
581e7141
JA
752 * should switch.
753 */
315fcfec 754 if (td_rw(td) && td->o.rwmix[odir]) {
581e7141 755 /*
50a8ce86 756 * Other direction is behind rate, switch
581e7141 757 */
50a8ce86 758 if (td->rate_next_io_time[odir] <= now)
581e7141
JA
759 return odir;
760
761 /*
395feabb
JA
762 * Both directions are ahead of rate. sleep the min,
763 * switch if necessary
581e7141 764 */
50a8ce86 765 if (td->rate_next_io_time[ddir] <=
395feabb 766 td->rate_next_io_time[odir]) {
50a8ce86 767 usec = td->rate_next_io_time[ddir] - now;
581e7141 768 } else {
50a8ce86 769 usec = td->rate_next_io_time[odir] - now;
581e7141
JA
770 ddir = odir;
771 }
772 } else
50a8ce86 773 usec = td->rate_next_io_time[ddir] - now;
581e7141 774
a9da8ab2
JA
775 if (td->o.io_submit_mode == IO_MODE_INLINE)
776 io_u_quiesce(td);
78c1eda5 777
1a9bf814 778 usec_sleep(td, usec);
581e7141
JA
779 return ddir;
780}
781
10ba535a
JA
782/*
783 * Return the data direction for the next io_u. If the job is a
784 * mixed read/write workload, check the rwmix cycle and switch if
785 * necessary.
786 */
1e97cce9 787static enum fio_ddir get_rw_ddir(struct thread_data *td)
10ba535a 788{
581e7141
JA
789 enum fio_ddir ddir;
790
5f9099ea 791 /*
f6860972
TK
792 * See if it's time to fsync/fdatasync/sync_file_range first,
793 * and if not then move on to check regular I/Os.
5f9099ea 794 */
f6860972
TK
795 if (should_fsync(td)) {
796 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
797 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
798 return DDIR_SYNC;
799
800 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
801 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
802 return DDIR_DATASYNC;
803
804 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
805 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
806 return DDIR_SYNC_FILE_RANGE;
807 }
44f29692 808
10ba535a 809 if (td_rw(td)) {
10ba535a
JA
810 /*
811 * Check if it's time to seed a new data direction.
812 */
e4928662 813 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
e47f799f
JA
814 /*
815 * Put a top limit on how many bytes we do for
816 * one data direction, to avoid overflowing the
817 * ranges too much
818 */
819 ddir = get_rand_ddir(td);
e47f799f
JA
820
821 if (ddir != td->rwmix_ddir)
822 set_rwmix_bytes(td);
823
824 td->rwmix_ddir = ddir;
10ba535a 825 }
581e7141 826 ddir = td->rwmix_ddir;
10ba535a 827 } else if (td_read(td))
581e7141 828 ddir = DDIR_READ;
6eaf09d6 829 else if (td_write(td))
581e7141 830 ddir = DDIR_WRITE;
5c8e84ca 831 else if (td_trim(td))
6eaf09d6 832 ddir = DDIR_TRIM;
5c8e84ca
TK
833 else
834 ddir = DDIR_INVAL;
581e7141
JA
835
836 td->rwmix_ddir = rate_ddir(td, ddir);
837 return td->rwmix_ddir;
10ba535a
JA
838}
839
1ef2b6be
JA
840static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
841{
0e4dd95c
DE
842 enum fio_ddir ddir = get_rw_ddir(td);
843
82a90686 844 if (td_trimwrite(td)) {
0e4dd95c
DE
845 struct fio_file *f = io_u->file;
846 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
847 ddir = DDIR_TRIM;
848 else
849 ddir = DDIR_WRITE;
850 }
851
852 io_u->ddir = io_u->acct_ddir = ddir;
1ef2b6be 853
9b87f09b 854 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
1ef2b6be
JA
855 td->o.barrier_blocks &&
856 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
857 td->io_issues[DDIR_WRITE])
1651e431 858 io_u_set(td, io_u, IO_U_F_BARRIER);
1ef2b6be
JA
859}
860
e8462bd8 861void put_file_log(struct thread_data *td, struct fio_file *f)
60f2c658 862{
71b84caa 863 unsigned int ret = put_file(td, f);
60f2c658
JA
864
865 if (ret)
866 td_verror(td, ret, "file close");
867}
868
10ba535a
JA
869void put_io_u(struct thread_data *td, struct io_u *io_u)
870{
a9da8ab2
JA
871 if (td->parent)
872 td = td->parent;
873
e8462bd8
JA
874 td_io_u_lock(td);
875
f8b0bd10 876 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
60f2c658 877 put_file_log(td, io_u->file);
f8b0bd10 878
10ba535a 879 io_u->file = NULL;
1651e431 880 io_u_set(td, io_u, IO_U_F_FREE);
d7ee2a7d 881
a9da8ab2 882 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
0c41214f 883 td->cur_depth--;
a9da8ab2
JA
884 assert(!(td->flags & TD_F_CHILD));
885 }
2ae0b204 886 io_u_qpush(&td->io_u_freelist, io_u);
e8462bd8
JA
887 td_io_u_unlock(td);
888 td_io_u_free_notify(td);
10ba535a
JA
889}
890
f2bba182
RR
891void clear_io_u(struct thread_data *td, struct io_u *io_u)
892{
1651e431 893 io_u_clear(td, io_u, IO_U_F_FLIGHT);
f2bba182
RR
894 put_io_u(td, io_u);
895}
896
755200a3
JA
897void requeue_io_u(struct thread_data *td, struct io_u **io_u)
898{
899 struct io_u *__io_u = *io_u;
bcd5abfa 900 enum fio_ddir ddir = acct_ddir(__io_u);
755200a3 901
465221b0
JA
902 dprint(FD_IO, "requeue %p\n", __io_u);
903
a9da8ab2
JA
904 if (td->parent)
905 td = td->parent;
906
e8462bd8
JA
907 td_io_u_lock(td);
908
1651e431 909 io_u_set(td, __io_u, IO_U_F_FREE);
bcd5abfa
JA
910 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
911 td->io_issues[ddir]--;
5ec10eaa 912
1651e431 913 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
a9da8ab2 914 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
0c41214f 915 td->cur_depth--;
a9da8ab2
JA
916 assert(!(td->flags & TD_F_CHILD));
917 }
2ae0b204
JA
918
919 io_u_rpush(&td->io_u_requeues, __io_u);
e8462bd8 920 td_io_u_unlock(td);
a9da8ab2 921 td_io_u_free_notify(td);
755200a3
JA
922 *io_u = NULL;
923}
924
9bf2061e 925static int fill_io_u(struct thread_data *td, struct io_u *io_u)
10ba535a 926{
6aca9b3d
JA
927 unsigned int is_random;
928
9b87f09b 929 if (td_ioengine_flagged(td, FIO_NOIO))
b4c5e1ac
JA
930 goto out;
931
1ef2b6be 932 set_rw_ddir(td, io_u);
5f9099ea 933
87dc1ab1 934 /*
ff58fced 935 * fsync() or fdatasync() or trim etc, we are done
87dc1ab1 936 */
ff58fced 937 if (!ddir_rw(io_u->ddir))
c38e9468 938 goto out;
a00735e6 939
48f5abd3
JA
940 /*
941 * See if it's time to switch to a new zone
942 */
13af05ae 943 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
4c8be5b1
JA
944 struct fio_file *f = io_u->file;
945
48f5abd3 946 td->zone_bytes = 0;
4c8be5b1
JA
947 f->file_offset += td->o.zone_range + td->o.zone_skip;
948
949 /*
950 * Wrap from the beginning, if we exceed the file size
951 */
952 if (f->file_offset >= f->real_file_size)
953 f->file_offset = f->real_file_size - f->file_offset;
f1dfb668 954 f->last_pos[io_u->ddir] = f->file_offset;
48f5abd3
JA
955 td->io_skip_bytes += td->o.zone_skip;
956 }
957
10ba535a 958 /*
c685b5b2
JA
959 * No log, let the seq/rand engine retrieve the next buflen and
960 * position.
10ba535a 961 */
6aca9b3d 962 if (get_next_offset(td, io_u, &is_random)) {
2ba1c290 963 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
bca4ed4d 964 return 1;
2ba1c290 965 }
10ba535a 966
6aca9b3d 967 io_u->buflen = get_next_buflen(td, io_u, is_random);
2ba1c290
JA
968 if (!io_u->buflen) {
969 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
bca4ed4d 970 return 1;
2ba1c290 971 }
bca4ed4d 972
2ba1c290 973 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
e5f9a813
RE
974 dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%lx exceeds file size=0x%llx\n",
975 io_u,
4b91ee8f
JA
976 (unsigned long long) io_u->offset, io_u->buflen,
977 (unsigned long long) io_u->file->real_file_size);
6a5e6884 978 return 1;
2ba1c290 979 }
6a5e6884 980
bca4ed4d
JA
981 /*
982 * mark entry before potentially trimming io_u
983 */
303032ae 984 if (td_random(td) && file_randommap(td, io_u->file))
9bf2061e 985 mark_random_map(td, io_u);
bca4ed4d 986
c38e9468 987out:
e5f9a813 988 dprint_io_u(io_u, "fill");
d9d91e39 989 td->zone_bytes += io_u->buflen;
bca4ed4d 990 return 0;
10ba535a
JA
991}
992
838bc709
JA
993static void __io_u_mark_map(unsigned int *map, unsigned int nr)
994{
2b13e716 995 int idx = 0;
838bc709
JA
996
997 switch (nr) {
998 default:
2b13e716 999 idx = 6;
838bc709
JA
1000 break;
1001 case 33 ... 64:
2b13e716 1002 idx = 5;
838bc709
JA
1003 break;
1004 case 17 ... 32:
2b13e716 1005 idx = 4;
838bc709
JA
1006 break;
1007 case 9 ... 16:
2b13e716 1008 idx = 3;
838bc709
JA
1009 break;
1010 case 5 ... 8:
2b13e716 1011 idx = 2;
838bc709
JA
1012 break;
1013 case 1 ... 4:
2b13e716 1014 idx = 1;
838bc709
JA
1015 case 0:
1016 break;
1017 }
1018
2b13e716 1019 map[idx]++;
838bc709
JA
1020}
1021
1022void io_u_mark_submit(struct thread_data *td, unsigned int nr)
1023{
1024 __io_u_mark_map(td->ts.io_u_submit, nr);
1025 td->ts.total_submit++;
1026}
1027
1028void io_u_mark_complete(struct thread_data *td, unsigned int nr)
1029{
1030 __io_u_mark_map(td->ts.io_u_complete, nr);
1031 td->ts.total_complete++;
1032}
1033
d8005759 1034void io_u_mark_depth(struct thread_data *td, unsigned int nr)
71619dc2 1035{
2b13e716 1036 int idx = 0;
71619dc2
JA
1037
1038 switch (td->cur_depth) {
1039 default:
2b13e716 1040 idx = 6;
a783e61a 1041 break;
71619dc2 1042 case 32 ... 63:
2b13e716 1043 idx = 5;
a783e61a 1044 break;
71619dc2 1045 case 16 ... 31:
2b13e716 1046 idx = 4;
a783e61a 1047 break;
71619dc2 1048 case 8 ... 15:
2b13e716 1049 idx = 3;
a783e61a 1050 break;
71619dc2 1051 case 4 ... 7:
2b13e716 1052 idx = 2;
a783e61a 1053 break;
71619dc2 1054 case 2 ... 3:
2b13e716 1055 idx = 1;
71619dc2
JA
1056 case 1:
1057 break;
1058 }
1059
2b13e716 1060 td->ts.io_u_map[idx] += nr;
71619dc2
JA
1061}
1062
d6bb626e 1063static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
04a0feae 1064{
2b13e716 1065 int idx = 0;
04a0feae 1066
d6bb626e
VF
1067 assert(nsec < 1000);
1068
1069 switch (nsec) {
1070 case 750 ... 999:
1071 idx = 9;
1072 break;
1073 case 500 ... 749:
1074 idx = 8;
1075 break;
1076 case 250 ... 499:
1077 idx = 7;
1078 break;
1079 case 100 ... 249:
1080 idx = 6;
1081 break;
1082 case 50 ... 99:
1083 idx = 5;
1084 break;
1085 case 20 ... 49:
1086 idx = 4;
1087 break;
1088 case 10 ... 19:
1089 idx = 3;
1090 break;
1091 case 4 ... 9:
1092 idx = 2;
1093 break;
1094 case 2 ... 3:
1095 idx = 1;
1096 case 0 ... 1:
1097 break;
1098 }
1099
1100 assert(idx < FIO_IO_U_LAT_N_NR);
1101 td->ts.io_u_lat_n[idx]++;
1102}
1103
1104static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
1105{
1106 int idx = 0;
1107
1108 assert(usec < 1000 && usec >= 1);
04a0feae
JA
1109
1110 switch (usec) {
1111 case 750 ... 999:
2b13e716 1112 idx = 9;
04a0feae
JA
1113 break;
1114 case 500 ... 749:
2b13e716 1115 idx = 8;
04a0feae
JA
1116 break;
1117 case 250 ... 499:
2b13e716 1118 idx = 7;
04a0feae
JA
1119 break;
1120 case 100 ... 249:
2b13e716 1121 idx = 6;
04a0feae
JA
1122 break;
1123 case 50 ... 99:
2b13e716 1124 idx = 5;
04a0feae
JA
1125 break;
1126 case 20 ... 49:
2b13e716 1127 idx = 4;
04a0feae
JA
1128 break;
1129 case 10 ... 19:
2b13e716 1130 idx = 3;
04a0feae
JA
1131 break;
1132 case 4 ... 9:
2b13e716 1133 idx = 2;
04a0feae
JA
1134 break;
1135 case 2 ... 3:
2b13e716 1136 idx = 1;
04a0feae
JA
1137 case 0 ... 1:
1138 break;
1139 }
1140
2b13e716
JA
1141 assert(idx < FIO_IO_U_LAT_U_NR);
1142 td->ts.io_u_lat_u[idx]++;
04a0feae
JA
1143}
1144
d6bb626e 1145static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
ec118304 1146{
2b13e716 1147 int idx = 0;
ec118304 1148
d6bb626e
VF
1149 assert(msec >= 1);
1150
ec118304
JA
1151 switch (msec) {
1152 default:
2b13e716 1153 idx = 11;
04a0feae 1154 break;
8abdce66 1155 case 1000 ... 1999:
2b13e716 1156 idx = 10;
04a0feae 1157 break;
8abdce66 1158 case 750 ... 999:
2b13e716 1159 idx = 9;
04a0feae 1160 break;
8abdce66 1161 case 500 ... 749:
2b13e716 1162 idx = 8;
04a0feae 1163 break;
8abdce66 1164 case 250 ... 499:
2b13e716 1165 idx = 7;
04a0feae 1166 break;
8abdce66 1167 case 100 ... 249:
2b13e716 1168 idx = 6;
04a0feae 1169 break;
8abdce66 1170 case 50 ... 99:
2b13e716 1171 idx = 5;
04a0feae 1172 break;
8abdce66 1173 case 20 ... 49:
2b13e716 1174 idx = 4;
04a0feae 1175 break;
8abdce66 1176 case 10 ... 19:
2b13e716 1177 idx = 3;
04a0feae 1178 break;
8abdce66 1179 case 4 ... 9:
2b13e716 1180 idx = 2;
04a0feae 1181 break;
ec118304 1182 case 2 ... 3:
2b13e716 1183 idx = 1;
ec118304
JA
1184 case 0 ... 1:
1185 break;
1186 }
1187
2b13e716
JA
1188 assert(idx < FIO_IO_U_LAT_M_NR);
1189 td->ts.io_u_lat_m[idx]++;
04a0feae
JA
1190}
1191
d6bb626e 1192static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
04a0feae 1193{
d6bb626e
VF
1194 if (nsec < 1000)
1195 io_u_mark_lat_nsec(td, nsec);
1196 else if (nsec < 1000000)
1197 io_u_mark_lat_usec(td, nsec / 1000);
04a0feae 1198 else
d6bb626e 1199 io_u_mark_lat_msec(td, nsec / 1000000);
ec118304
JA
1200}
1201
8c07860d
JA
1202static unsigned int __get_next_fileno_rand(struct thread_data *td)
1203{
1204 unsigned long fileno;
1205
1206 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1207 uint64_t frand_max = rand_max(&td->next_file_state);
1208 unsigned long r;
1209
1210 r = __rand(&td->next_file_state);
1211 return (unsigned int) ((double) td->o.nr_files
1212 * (r / (frand_max + 1.0)));
1213 }
1214
1215 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1216 fileno = zipf_next(&td->next_file_zipf);
1217 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1218 fileno = pareto_next(&td->next_file_zipf);
1219 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1220 fileno = gauss_next(&td->next_file_gauss);
1221 else {
1222 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1223 assert(0);
1224 return 0;
1225 }
1226
1227 return fileno >> FIO_FSERVICE_SHIFT;
1228}
1229
0aabe160
JA
1230/*
1231 * Get next file to service by choosing one at random
1232 */
2cc52930
JA
1233static struct fio_file *get_next_file_rand(struct thread_data *td,
1234 enum fio_file_flags goodf,
d6aed795 1235 enum fio_file_flags badf)
0aabe160 1236{
0aabe160 1237 struct fio_file *f;
1c178180 1238 int fno;
0aabe160
JA
1239
1240 do {
87b10676 1241 int opened = 0;
4c07ad86 1242
8c07860d 1243 fno = __get_next_fileno_rand(td);
7c83c089 1244
126d65c6 1245 f = td->files[fno];
d6aed795 1246 if (fio_file_done(f))
059e63c0 1247 continue;
1c178180 1248
d6aed795 1249 if (!fio_file_open(f)) {
87b10676
JA
1250 int err;
1251
002fe734
JA
1252 if (td->nr_open_files >= td->o.open_files)
1253 return ERR_PTR(-EBUSY);
1254
87b10676
JA
1255 err = td_io_open_file(td, f);
1256 if (err)
1257 continue;
1258 opened = 1;
1259 }
1260
2ba1c290
JA
1261 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1262 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
0aabe160 1263 return f;
2ba1c290 1264 }
87b10676
JA
1265 if (opened)
1266 td_io_close_file(td, f);
0aabe160
JA
1267 } while (1);
1268}
1269
1270/*
1271 * Get next file to service by doing round robin between all available ones
1272 */
1c178180
JA
1273static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1274 int badf)
3d7c391d
JA
1275{
1276 unsigned int old_next_file = td->next_file;
1277 struct fio_file *f;
1278
1279 do {
87b10676
JA
1280 int opened = 0;
1281
126d65c6 1282 f = td->files[td->next_file];
3d7c391d
JA
1283
1284 td->next_file++;
2dc1bbeb 1285 if (td->next_file >= td->o.nr_files)
3d7c391d
JA
1286 td->next_file = 0;
1287
87b10676 1288 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
d6aed795 1289 if (fio_file_done(f)) {
d5ed68ea 1290 f = NULL;
059e63c0 1291 continue;
d5ed68ea 1292 }
059e63c0 1293
d6aed795 1294 if (!fio_file_open(f)) {
87b10676
JA
1295 int err;
1296
002fe734
JA
1297 if (td->nr_open_files >= td->o.open_files)
1298 return ERR_PTR(-EBUSY);
1299
87b10676 1300 err = td_io_open_file(td, f);
b5696bfc
JA
1301 if (err) {
1302 dprint(FD_FILE, "error %d on open of %s\n",
1303 err, f->file_name);
87c27b45 1304 f = NULL;
87b10676 1305 continue;
b5696bfc 1306 }
87b10676
JA
1307 opened = 1;
1308 }
1309
0b9d69ec
JA
1310 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1311 f->flags);
1c178180 1312 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
3d7c391d
JA
1313 break;
1314
87b10676
JA
1315 if (opened)
1316 td_io_close_file(td, f);
1317
3d7c391d
JA
1318 f = NULL;
1319 } while (td->next_file != old_next_file);
1320
2ba1c290 1321 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
3d7c391d
JA
1322 return f;
1323}
1324
7eb36574 1325static struct fio_file *__get_next_file(struct thread_data *td)
bdb4e2e9 1326{
1907dbc6
JA
1327 struct fio_file *f;
1328
2dc1bbeb 1329 assert(td->o.nr_files <= td->files_index);
1c178180 1330
b5696bfc 1331 if (td->nr_done_files >= td->o.nr_files) {
5ec10eaa
JA
1332 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1333 " nr_files=%d\n", td->nr_open_files,
1334 td->nr_done_files,
1335 td->o.nr_files);
bdb4e2e9 1336 return NULL;
2ba1c290 1337 }
bdb4e2e9 1338
1907dbc6 1339 f = td->file_service_file;
d6aed795 1340 if (f && fio_file_open(f) && !fio_file_closing(f)) {
a086c257
JA
1341 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1342 goto out;
1343 if (td->file_service_left--)
1344 goto out;
1345 }
1907dbc6 1346
a086c257
JA
1347 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1348 td->o.file_service_type == FIO_FSERVICE_SEQ)
d6aed795 1349 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
bdb4e2e9 1350 else
d6aed795 1351 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1907dbc6 1352
002fe734
JA
1353 if (IS_ERR(f))
1354 return f;
1355
1907dbc6
JA
1356 td->file_service_file = f;
1357 td->file_service_left = td->file_service_nr - 1;
2ba1c290 1358out:
0dac421f
JA
1359 if (f)
1360 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1361 else
1362 dprint(FD_FILE, "get_next_file: NULL\n");
1907dbc6 1363 return f;
bdb4e2e9
JA
1364}
1365
7eb36574
JA
1366static struct fio_file *get_next_file(struct thread_data *td)
1367{
372d8962 1368 if (td->flags & TD_F_PROFILE_OPS) {
d72be545 1369 struct prof_io_ops *ops = &td->prof_io_ops;
7eb36574 1370
d72be545
JA
1371 if (ops->get_next_file)
1372 return ops->get_next_file(td);
1373 }
7eb36574
JA
1374
1375 return __get_next_file(td);
1376}
1377
002fe734 1378static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
429f6675
JA
1379{
1380 struct fio_file *f;
1381
1382 do {
1383 f = get_next_file(td);
002fe734
JA
1384 if (IS_ERR_OR_NULL(f))
1385 return PTR_ERR(f);
429f6675 1386
429f6675
JA
1387 io_u->file = f;
1388 get_file(f);
1389
1390 if (!fill_io_u(td, io_u))
1391 break;
1392
b5696bfc 1393 put_file_log(td, f);
429f6675 1394 td_io_close_file(td, f);
b5696bfc 1395 io_u->file = NULL;
8c07860d
JA
1396 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1397 fio_file_reset(td, f);
1398 else {
1399 fio_file_set_done(f);
1400 td->nr_done_files++;
1401 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
0b9d69ec 1402 td->nr_done_files, td->o.nr_files);
8c07860d 1403 }
429f6675
JA
1404 } while (1);
1405
1406 return 0;
1407}
1408
3e260a46 1409static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
c3a32714 1410 unsigned long long tnsec, unsigned long long max_nsec)
3e260a46
JA
1411{
1412 if (!td->error)
c3a32714 1413 log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
3e260a46
JA
1414 td_verror(td, ETIMEDOUT, "max latency exceeded");
1415 icd->error = ETIMEDOUT;
1416}
1417
1418static void lat_new_cycle(struct thread_data *td)
1419{
1420 fio_gettime(&td->latency_ts, NULL);
1421 td->latency_ios = ddir_rw_sum(td->io_blocks);
1422 td->latency_failed = 0;
1423}
1424
1425/*
1426 * We had an IO outside the latency target. Reduce the queue depth. If we
1427 * are at QD=1, then it's time to give up.
1428 */
e39c0676 1429static bool __lat_target_failed(struct thread_data *td)
3e260a46
JA
1430{
1431 if (td->latency_qd == 1)
e39c0676 1432 return true;
3e260a46
JA
1433
1434 td->latency_qd_high = td->latency_qd;
6bb58215
JA
1435
1436 if (td->latency_qd == td->latency_qd_low)
1437 td->latency_qd_low--;
1438
3e260a46
JA
1439 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1440
1441 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1442
1443 /*
1444 * When we ramp QD down, quiesce existing IO to prevent
1445 * a storm of ramp downs due to pending higher depth.
1446 */
1447 io_u_quiesce(td);
1448 lat_new_cycle(td);
e39c0676 1449 return false;
3e260a46
JA
1450}
1451
e39c0676 1452static bool lat_target_failed(struct thread_data *td)
3e260a46
JA
1453{
1454 if (td->o.latency_percentile.u.f == 100.0)
1455 return __lat_target_failed(td);
1456
1457 td->latency_failed++;
e39c0676 1458 return false;
3e260a46
JA
1459}
1460
1461void lat_target_init(struct thread_data *td)
1462{
6bb58215
JA
1463 td->latency_end_run = 0;
1464
3e260a46
JA
1465 if (td->o.latency_target) {
1466 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1467 fio_gettime(&td->latency_ts, NULL);
1468 td->latency_qd = 1;
1469 td->latency_qd_high = td->o.iodepth;
1470 td->latency_qd_low = 1;
1471 td->latency_ios = ddir_rw_sum(td->io_blocks);
1472 } else
1473 td->latency_qd = td->o.iodepth;
1474}
1475
6bb58215
JA
1476void lat_target_reset(struct thread_data *td)
1477{
1478 if (!td->latency_end_run)
1479 lat_target_init(td);
1480}
1481
3e260a46
JA
1482static void lat_target_success(struct thread_data *td)
1483{
1484 const unsigned int qd = td->latency_qd;
6bb58215 1485 struct thread_options *o = &td->o;
3e260a46
JA
1486
1487 td->latency_qd_low = td->latency_qd;
1488
1489 /*
1490 * If we haven't failed yet, we double up to a failing value instead
1491 * of bisecting from highest possible queue depth. If we have set
1492 * a limit other than td->o.iodepth, bisect between that.
1493 */
6bb58215 1494 if (td->latency_qd_high != o->iodepth)
3e260a46
JA
1495 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1496 else
1497 td->latency_qd *= 2;
1498
6bb58215
JA
1499 if (td->latency_qd > o->iodepth)
1500 td->latency_qd = o->iodepth;
3e260a46
JA
1501
1502 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
6bb58215 1503
3e260a46 1504 /*
6bb58215
JA
1505 * Same as last one, we are done. Let it run a latency cycle, so
1506 * we get only the results from the targeted depth.
3e260a46 1507 */
6bb58215
JA
1508 if (td->latency_qd == qd) {
1509 if (td->latency_end_run) {
1510 dprint(FD_RATE, "We are done\n");
1511 td->done = 1;
1512 } else {
1513 dprint(FD_RATE, "Quiesce and final run\n");
1514 io_u_quiesce(td);
1515 td->latency_end_run = 1;
1516 reset_all_stats(td);
1517 reset_io_stats(td);
1518 }
1519 }
3e260a46
JA
1520
1521 lat_new_cycle(td);
1522}
1523
1524/*
1525 * Check if we can bump the queue depth
1526 */
1527void lat_target_check(struct thread_data *td)
1528{
1529 uint64_t usec_window;
1530 uint64_t ios;
1531 double success_ios;
1532
1533 usec_window = utime_since_now(&td->latency_ts);
1534 if (usec_window < td->o.latency_window)
1535 return;
1536
1537 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1538 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1539 success_ios *= 100.0;
1540
1541 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1542
1543 if (success_ios >= td->o.latency_percentile.u.f)
1544 lat_target_success(td);
1545 else
1546 __lat_target_failed(td);
1547}
1548
1549/*
1550 * If latency target is enabled, we might be ramping up or down and not
1551 * using the full queue depth available.
1552 */
e39c0676 1553bool queue_full(const struct thread_data *td)
3e260a46
JA
1554{
1555 const int qempty = io_u_qempty(&td->io_u_freelist);
1556
1557 if (qempty)
e39c0676 1558 return true;
3e260a46 1559 if (!td->o.latency_target)
e39c0676 1560 return false;
3e260a46
JA
1561
1562 return td->cur_depth >= td->latency_qd;
1563}
429f6675 1564
10ba535a
JA
1565struct io_u *__get_io_u(struct thread_data *td)
1566{
0cae66f6 1567 struct io_u *io_u = NULL;
10ba535a 1568
ca09be4b
JA
1569 if (td->stop_io)
1570 return NULL;
1571
e8462bd8
JA
1572 td_io_u_lock(td);
1573
1574again:
2ae0b204
JA
1575 if (!io_u_rempty(&td->io_u_requeues))
1576 io_u = io_u_rpop(&td->io_u_requeues);
3e260a46 1577 else if (!queue_full(td)) {
2ae0b204 1578 io_u = io_u_qpop(&td->io_u_freelist);
10ba535a 1579
225ba9e3 1580 io_u->file = NULL;
6040dabc 1581 io_u->buflen = 0;
10ba535a 1582 io_u->resid = 0;
d7762cf8 1583 io_u->end_io = NULL;
755200a3
JA
1584 }
1585
1586 if (io_u) {
0c6e7517 1587 assert(io_u->flags & IO_U_F_FREE);
1651e431 1588 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
f8b0bd10
JA
1589 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1590 IO_U_F_VER_LIST);
0c6e7517 1591
755200a3 1592 io_u->error = 0;
bcd5abfa 1593 io_u->acct_ddir = -1;
10ba535a 1594 td->cur_depth++;
a9da8ab2 1595 assert(!(td->flags & TD_F_CHILD));
1651e431 1596 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
f9401285 1597 io_u->ipo = NULL;
a9da8ab2 1598 } else if (td_async_processing(td)) {
1dec3e07
JA
1599 /*
1600 * We ran out, wait for async verify threads to finish and
1601 * return one
1602 */
a9da8ab2
JA
1603 assert(!(td->flags & TD_F_CHILD));
1604 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1dec3e07 1605 goto again;
10ba535a
JA
1606 }
1607
e8462bd8 1608 td_io_u_unlock(td);
10ba535a
JA
1609 return io_u;
1610}
1611
e39c0676 1612static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
10ba535a 1613{
d72be545 1614 if (!(td->flags & TD_F_TRIM_BACKLOG))
e39c0676 1615 return false;
c9a73054
JA
1616 if (!td->trim_entries)
1617 return false;
d72be545 1618
c9a73054
JA
1619 if (td->trim_batch) {
1620 td->trim_batch--;
1621 if (get_next_trim(td, io_u))
1622 return true;
1623 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1624 td->last_ddir != DDIR_READ) {
1625 td->trim_batch = td->o.trim_batch;
1626 if (!td->trim_batch)
1627 td->trim_batch = td->o.trim_backlog;
1628 if (get_next_trim(td, io_u))
e39c0676 1629 return true;
2ba1c290 1630 }
10ba535a 1631
e39c0676 1632 return false;
0d29de83
JA
1633}
1634
e39c0676 1635static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
0d29de83 1636{
d72be545 1637 if (!(td->flags & TD_F_VER_BACKLOG))
e39c0676 1638 return false;
d72be545
JA
1639
1640 if (td->io_hist_len) {
9e144189
JA
1641 int get_verify = 0;
1642
d1ece0c7 1643 if (td->verify_batch)
9e144189 1644 get_verify = 1;
d1ece0c7 1645 else if (!(td->io_hist_len % td->o.verify_backlog) &&
9e144189
JA
1646 td->last_ddir != DDIR_READ) {
1647 td->verify_batch = td->o.verify_batch;
f8a75c99
JA
1648 if (!td->verify_batch)
1649 td->verify_batch = td->o.verify_backlog;
9e144189
JA
1650 get_verify = 1;
1651 }
1652
d1ece0c7
JA
1653 if (get_verify && !get_next_verify(td, io_u)) {
1654 td->verify_batch--;
e39c0676 1655 return true;
d1ece0c7 1656 }
9e144189
JA
1657 }
1658
e39c0676 1659 return false;
0d29de83
JA
1660}
1661
de789769
JA
1662/*
1663 * Fill offset and start time into the buffer content, to prevent too
23f394d5
JA
1664 * easy compressible data for simple de-dupe attempts. Do this for every
1665 * 512b block in the range, since that should be the smallest block size
1666 * we can expect from a device.
de789769
JA
1667 */
1668static void small_content_scramble(struct io_u *io_u)
1669{
319b073f 1670 unsigned int i, nr_blocks = io_u->buflen >> 9;
23f394d5 1671 unsigned int offset;
319b073f
JA
1672 uint64_t boffset, *iptr;
1673 char *p;
de789769 1674
23f394d5
JA
1675 if (!nr_blocks)
1676 return;
1677
1678 p = io_u->xfer_buf;
fba76ee8 1679 boffset = io_u->offset;
319b073f
JA
1680
1681 if (io_u->buf_filled_len)
1682 io_u->buf_filled_len = 0;
1683
1684 /*
1685 * Generate random index between 0..7. We do chunks of 512b, if
1686 * we assume a cacheline is 64 bytes, then we have 8 of those.
1687 * Scramble content within the blocks in the same cacheline to
1688 * speed things up.
1689 */
1690 offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
fad82f76 1691
23f394d5
JA
1692 for (i = 0; i < nr_blocks; i++) {
1693 /*
319b073f
JA
1694 * Fill offset into start of cacheline, time into end
1695 * of cacheline
23f394d5 1696 */
319b073f
JA
1697 iptr = (void *) p + (offset << 6);
1698 *iptr = boffset;
1699
1700 iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
1701 iptr[0] = io_u->start_time.tv_sec;
1702 iptr[1] = io_u->start_time.tv_nsec;
23f394d5 1703
23f394d5 1704 p += 512;
fad82f76 1705 boffset += 512;
23f394d5 1706 }
de789769
JA
1707}
1708
0d29de83
JA
1709/*
1710 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1711 * etc. The returned io_u is fully ready to be prepped and submitted.
1712 */
1713struct io_u *get_io_u(struct thread_data *td)
1714{
1715 struct fio_file *f;
1716 struct io_u *io_u;
de789769 1717 int do_scramble = 0;
002fe734 1718 long ret = 0;
0d29de83
JA
1719
1720 io_u = __get_io_u(td);
1721 if (!io_u) {
1722 dprint(FD_IO, "__get_io_u failed\n");
1723 return NULL;
1724 }
1725
1726 if (check_get_verify(td, io_u))
1727 goto out;
1728 if (check_get_trim(td, io_u))
1729 goto out;
1730
755200a3
JA
1731 /*
1732 * from a requeue, io_u already setup
1733 */
1734 if (io_u->file)
77f392bf 1735 goto out;
755200a3 1736
429f6675
JA
1737 /*
1738 * If using an iolog, grab next piece if any available.
1739 */
d72be545 1740 if (td->flags & TD_F_READ_IOLOG) {
429f6675
JA
1741 if (read_iolog_get(td, io_u))
1742 goto err_put;
2ba1c290 1743 } else if (set_io_u_file(td, io_u)) {
002fe734 1744 ret = -EBUSY;
2ba1c290 1745 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
429f6675 1746 goto err_put;
2ba1c290 1747 }
5ec10eaa 1748
429f6675 1749 f = io_u->file;
002fe734
JA
1750 if (!f) {
1751 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1752 goto err_put;
1753 }
1754
d6aed795 1755 assert(fio_file_open(f));
97af62ce 1756
ff58fced 1757 if (ddir_rw(io_u->ddir)) {
9b87f09b 1758 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
2ba1c290 1759 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
429f6675 1760 goto err_put;
2ba1c290 1761 }
10ba535a 1762
f1dfb668
JA
1763 f->last_start[io_u->ddir] = io_u->offset;
1764 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
10ba535a 1765
fd68418e 1766 if (io_u->ddir == DDIR_WRITE) {
d72be545 1767 if (td->flags & TD_F_REFILL_BUFFERS) {
9c42684e 1768 io_u_fill_buffer(td, io_u,
1066358a 1769 td->o.min_bs[DDIR_WRITE],
9e129577 1770 io_u->buflen);
ff441ae8 1771 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
bedc9dc2 1772 !(td->flags & TD_F_COMPRESS))
fd68418e 1773 do_scramble = 1;
d72be545 1774 if (td->flags & TD_F_VER_NONE) {
629f1d71
JA
1775 populate_verify_io_u(td, io_u);
1776 do_scramble = 0;
1777 }
fd68418e 1778 } else if (io_u->ddir == DDIR_READ) {
cbe8d756
RR
1779 /*
1780 * Reset the buf_filled parameters so next time if the
1781 * buffer is used for writes it is refilled.
1782 */
cbe8d756
RR
1783 io_u->buf_filled_len = 0;
1784 }
87dc1ab1 1785 }
10ba535a 1786
165faf16
JA
1787 /*
1788 * Set io data pointers.
1789 */
cec6b55d
JA
1790 io_u->xfer_buf = io_u->buf;
1791 io_u->xfer_buflen = io_u->buflen;
5973cafb 1792
6ac7a331 1793out:
0d29de83 1794 assert(io_u->file);
429f6675 1795 if (!td_io_prep(td, io_u)) {
3ecc8c67 1796 if (!td->o.disable_lat)
993bf48b 1797 fio_gettime(&io_u->start_time, NULL);
03553853 1798
de789769
JA
1799 if (do_scramble)
1800 small_content_scramble(io_u);
03553853 1801
429f6675 1802 return io_u;
36167d82 1803 }
429f6675 1804err_put:
2ba1c290 1805 dprint(FD_IO, "get_io_u failed\n");
429f6675 1806 put_io_u(td, io_u);
002fe734 1807 return ERR_PTR(ret);
10ba535a
JA
1808}
1809
a9da8ab2 1810static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
5451792e 1811{
8b28bd41 1812 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
825f818e 1813
8b28bd41
DM
1814 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1815 return;
5451792e 1816
709c8313
RE
1817 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1818 io_u->file ? " on file " : "",
1819 io_u->file ? io_u->file->file_name : "",
1820 strerror(io_u->error),
1821 io_ddir_name(io_u->ddir),
1822 io_u->offset, io_u->xfer_buflen);
5451792e 1823
5ad7be56
KD
1824 if (td->io_ops->errdetails) {
1825 char *err = td->io_ops->errdetails(io_u);
1826
1827 log_err("fio: %s\n", err);
1828 free(err);
1829 }
1830
5451792e
JA
1831 if (!td->error)
1832 td_verror(td, io_u->error, "io_u error");
1833}
1834
a9da8ab2
JA
1835void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1836{
1837 __io_u_log_error(td, io_u);
1838 if (td->parent)
094e66cb 1839 __io_u_log_error(td->parent, io_u);
a9da8ab2
JA
1840}
1841
e39c0676 1842static inline bool gtod_reduce(struct thread_data *td)
aba6c951 1843{
3ecc8c67
JA
1844 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1845 || td->o.gtod_reduce;
aba6c951
JA
1846}
1847
c8eeb9df
JA
1848static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1849 struct io_completion_data *icd,
1850 const enum fio_ddir idx, unsigned int bytes)
1851{
a9da8ab2 1852 const int no_reduce = !gtod_reduce(td);
d6bb626e 1853 unsigned long long llnsec = 0;
c8eeb9df 1854
75dc383e
JA
1855 if (td->parent)
1856 td = td->parent;
1857
132b1ee4 1858 if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
8243be59
JA
1859 return;
1860
a9da8ab2 1861 if (no_reduce)
d6bb626e 1862 llnsec = ntime_since(&io_u->issue_time, &icd->time);
c8eeb9df
JA
1863
1864 if (!td->o.disable_lat) {
c3a32714 1865 unsigned long long tnsec;
c8eeb9df 1866
d6bb626e
VF
1867 tnsec = ntime_since(&io_u->start_time, &icd->time);
1868 add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
15501535 1869
d4afedfd
JA
1870 if (td->flags & TD_F_PROFILE_OPS) {
1871 struct prof_io_ops *ops = &td->prof_io_ops;
1872
1873 if (ops->io_u_lat)
c3a32714 1874 icd->error = ops->io_u_lat(td, tnsec);
d4afedfd
JA
1875 }
1876
c3a32714
JA
1877 if (td->o.max_latency && tnsec > td->o.max_latency)
1878 lat_fatal(td, icd, tnsec, td->o.max_latency);
1879 if (td->o.latency_target && tnsec > td->o.latency_target) {
3e260a46 1880 if (lat_target_failed(td))
c3a32714 1881 lat_fatal(td, icd, tnsec, td->o.latency_target);
15501535 1882 }
c8eeb9df
JA
1883 }
1884
a47591e4
JA
1885 if (ddir_rw(idx)) {
1886 if (!td->o.disable_clat) {
d6bb626e
VF
1887 add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
1888 io_u_mark_latency(td, llnsec);
a47591e4 1889 }
c8eeb9df 1890
a47591e4 1891 if (!td->o.disable_bw && per_unit_log(td->bw_log))
d6bb626e 1892 add_bw_sample(td, io_u, bytes, llnsec);
c8eeb9df 1893
a47591e4
JA
1894 if (no_reduce && per_unit_log(td->iops_log))
1895 add_iops_sample(td, io_u, bytes);
1896 }
66347cfa
DE
1897
1898 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1899 uint32_t *info = io_u_block_info(td, io_u);
1900 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1901 if (io_u->ddir == DDIR_TRIM) {
1902 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1903 BLOCK_INFO_TRIMS(*info) + 1);
1904 } else if (io_u->ddir == DDIR_WRITE) {
1905 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1906 *info);
1907 }
1908 }
1909 }
c8eeb9df
JA
1910}
1911
94a6e1bb
JA
1912static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1913 uint64_t offset, unsigned int bytes)
1914{
1915 int idx;
1916
639ad1ea
JA
1917 if (!f)
1918 return;
1919
94a6e1bb
JA
1920 if (f->first_write == -1ULL || offset < f->first_write)
1921 f->first_write = offset;
1922 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1923 f->last_write = offset + bytes;
1924
1925 if (!f->last_write_comp)
1926 return;
1927
1928 idx = f->last_write_idx++;
1929 f->last_write_comp[idx] = offset;
1930 if (f->last_write_idx == td->o.iodepth)
1931 f->last_write_idx = 0;
1932}
1933
f8b0bd10 1934static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
97601024 1935 struct io_completion_data *icd)
10ba535a 1936{
f8b0bd10
JA
1937 struct io_u *io_u = *io_u_ptr;
1938 enum fio_ddir ddir = io_u->ddir;
1939 struct fio_file *f = io_u->file;
10ba535a 1940
e5f9a813 1941 dprint_io_u(io_u, "complete");
2ba1c290 1942
0c6e7517 1943 assert(io_u->flags & IO_U_F_FLIGHT);
1651e431 1944 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
f9401285
JA
1945
1946 /*
1947 * Mark IO ok to verify
1948 */
1949 if (io_u->ipo) {
890b6656
JA
1950 /*
1951 * Remove errored entry from the verification list
1952 */
1953 if (io_u->error)
1954 unlog_io_piece(td, io_u);
1955 else {
1956 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1957 write_barrier();
1958 }
f9401285
JA
1959 }
1960
f8b0bd10 1961 if (ddir_sync(ddir)) {
87dc1ab1 1962 td->last_was_sync = 1;
44f29692
JA
1963 if (f) {
1964 f->first_write = -1ULL;
1965 f->last_write = -1ULL;
1966 }
87dc1ab1
JA
1967 return;
1968 }
1969
1970 td->last_was_sync = 0;
f8b0bd10 1971 td->last_ddir = ddir;
87dc1ab1 1972
f8b0bd10 1973 if (!io_u->error && ddir_rw(ddir)) {
10ba535a 1974 unsigned int bytes = io_u->buflen - io_u->resid;
b29ee5b3 1975 int ret;
10ba535a 1976
f8b0bd10 1977 td->io_blocks[ddir]++;
f8b0bd10 1978 td->io_bytes[ddir] += bytes;
ae2fafc8 1979
e1c325d2
JA
1980 if (!(io_u->flags & IO_U_F_VER_LIST)) {
1981 td->this_io_blocks[ddir]++;
f8b0bd10 1982 td->this_io_bytes[ddir] += bytes;
e1c325d2 1983 }
f8b0bd10 1984
639ad1ea 1985 if (ddir == DDIR_WRITE)
94a6e1bb 1986 file_log_write_comp(td, f, io_u->offset, bytes);
44f29692 1987
6b1190fd 1988 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
50a8ce86 1989 td->runstate == TD_VERIFYING))
f8b0bd10 1990 account_io_completion(td, io_u, icd, ddir, bytes);
40e1a6f0 1991
f8b0bd10 1992 icd->bytes_done[ddir] += bytes;
3af6ef39 1993
d7762cf8 1994 if (io_u->end_io) {
f8b0bd10
JA
1995 ret = io_u->end_io(td, io_u_ptr);
1996 io_u = *io_u_ptr;
3af6ef39
JA
1997 if (ret && !icd->error)
1998 icd->error = ret;
1999 }
ff58fced 2000 } else if (io_u->error) {
10ba535a 2001 icd->error = io_u->error;
5451792e
JA
2002 io_u_log_error(td, io_u);
2003 }
8b28bd41 2004 if (icd->error) {
f8b0bd10
JA
2005 enum error_type_bit eb = td_error_type(ddir, icd->error);
2006
8b28bd41
DM
2007 if (!td_non_fatal_error(td, eb, icd->error))
2008 return;
f8b0bd10 2009
f2bba182
RR
2010 /*
2011 * If there is a non_fatal error, then add to the error count
2012 * and clear all the errors.
2013 */
2014 update_error_count(td, icd->error);
2015 td_clear_error(td);
2016 icd->error = 0;
f8b0bd10
JA
2017 if (io_u)
2018 io_u->error = 0;
f2bba182 2019 }
10ba535a
JA
2020}
2021
9520ebb9
JA
2022static void init_icd(struct thread_data *td, struct io_completion_data *icd,
2023 int nr)
10ba535a 2024{
6eaf09d6 2025 int ddir;
aba6c951
JA
2026
2027 if (!gtod_reduce(td))
9520ebb9 2028 fio_gettime(&icd->time, NULL);
02bcaa8c 2029
3af6ef39
JA
2030 icd->nr = nr;
2031
10ba535a 2032 icd->error = 0;
c1f50f76 2033 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
6eaf09d6 2034 icd->bytes_done[ddir] = 0;
36167d82
JA
2035}
2036
97601024
JA
2037static void ios_completed(struct thread_data *td,
2038 struct io_completion_data *icd)
36167d82
JA
2039{
2040 struct io_u *io_u;
2041 int i;
2042
10ba535a
JA
2043 for (i = 0; i < icd->nr; i++) {
2044 io_u = td->io_ops->event(td, i);
2045
f8b0bd10 2046 io_completed(td, &io_u, icd);
e8462bd8 2047
f8b0bd10 2048 if (io_u)
e8462bd8 2049 put_io_u(td, io_u);
10ba535a
JA
2050 }
2051}
97601024 2052
e7e6cfb4
JA
2053/*
2054 * Complete a single io_u for the sync engines.
2055 */
55312f9f 2056int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
97601024
JA
2057{
2058 struct io_completion_data icd;
55312f9f 2059 int ddir;
97601024 2060
9520ebb9 2061 init_icd(td, &icd, 1);
f8b0bd10 2062 io_completed(td, &io_u, &icd);
e8462bd8 2063
f8b0bd10 2064 if (io_u)
e8462bd8 2065 put_io_u(td, io_u);
97601024 2066
581e7141
JA
2067 if (icd.error) {
2068 td_verror(td, icd.error, "io_u_sync_complete");
2069 return -1;
2070 }
97601024 2071
c1f50f76 2072 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
55312f9f 2073 td->bytes_done[ddir] += icd.bytes_done[ddir];
581e7141
JA
2074
2075 return 0;
97601024
JA
2076}
2077
e7e6cfb4
JA
2078/*
2079 * Called to complete min_events number of io for the async engines.
2080 */
55312f9f 2081int io_u_queued_complete(struct thread_data *td, int min_evts)
97601024 2082{
97601024 2083 struct io_completion_data icd;
00de55ef 2084 struct timespec *tvp = NULL;
55312f9f 2085 int ret, ddir;
4d06a338 2086 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
97601024 2087
12bb8569 2088 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
b271fe62 2089
4950421a 2090 if (!min_evts)
00de55ef 2091 tvp = &ts;
5fb4b366
RE
2092 else if (min_evts > td->cur_depth)
2093 min_evts = td->cur_depth;
97601024 2094
82407585
RP
2095 /* No worries, td_io_getevents fixes min and max if they are
2096 * set incorrectly */
2097 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
97601024 2098 if (ret < 0) {
e1161c32 2099 td_verror(td, -ret, "td_io_getevents");
97601024
JA
2100 return ret;
2101 } else if (!ret)
2102 return ret;
2103
9520ebb9 2104 init_icd(td, &icd, ret);
97601024 2105 ios_completed(td, &icd);
581e7141
JA
2106 if (icd.error) {
2107 td_verror(td, icd.error, "io_u_queued_complete");
2108 return -1;
2109 }
97601024 2110
c1f50f76 2111 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
55312f9f 2112 td->bytes_done[ddir] += icd.bytes_done[ddir];
581e7141 2113
0d593542 2114 return ret;
97601024 2115}
7e77dd02
JA
2116
2117/*
2118 * Call when io_u is really queued, to update the submission latency.
2119 */
2120void io_u_queued(struct thread_data *td, struct io_u *io_u)
2121{
8243be59 2122 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
9520ebb9 2123 unsigned long slat_time;
7e77dd02 2124
d6bb626e 2125 slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
75dc383e
JA
2126
2127 if (td->parent)
2128 td = td->parent;
2129
ae588852
JA
2130 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2131 io_u->offset);
9520ebb9 2132 }
7e77dd02 2133}
433afcb4 2134
5c94b008
JA
2135/*
2136 * See if we should reuse the last seed, if dedupe is enabled
2137 */
9451b93e 2138static struct frand_state *get_buf_state(struct thread_data *td)
5c94b008
JA
2139{
2140 unsigned int v;
5c94b008
JA
2141
2142 if (!td->o.dedupe_percentage)
2143 return &td->buf_state;
732eedd0 2144 else if (td->o.dedupe_percentage == 100) {
9451b93e
JA
2145 frand_copy(&td->buf_state_prev, &td->buf_state);
2146 return &td->buf_state;
732eedd0 2147 }
5c94b008 2148
36dd3379 2149 v = rand32_between(&td->dedupe_state, 1, 100);
5c94b008
JA
2150
2151 if (v <= td->o.dedupe_percentage)
2152 return &td->buf_state_prev;
2153
2154 return &td->buf_state;
2155}
2156
9451b93e 2157static void save_buf_state(struct thread_data *td, struct frand_state *rs)
5c94b008 2158{
9451b93e
JA
2159 if (td->o.dedupe_percentage == 100)
2160 frand_copy(rs, &td->buf_state_prev);
2161 else if (rs == &td->buf_state)
5c94b008
JA
2162 frand_copy(&td->buf_state_prev, rs);
2163}
2164
cc86c395
JA
2165void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2166 unsigned int max_bs)
5973cafb 2167{
d1af2894
JA
2168 struct thread_options *o = &td->o;
2169
15600335
JA
2170 if (o->mem_type == MEM_CUDA_MALLOC)
2171 return;
03553853 2172
4eff3e57 2173 if (o->compress_percentage || o->dedupe_percentage) {
9c42684e 2174 unsigned int perc = td->o.compress_percentage;
5c94b008 2175 struct frand_state *rs;
1066358a 2176 unsigned int left = max_bs;
1e7f82e2 2177 unsigned int this_write;
5c94b008 2178
1066358a 2179 do {
9451b93e 2180 rs = get_buf_state(td);
9c42684e 2181
1066358a 2182 min_write = min(min_write, left);
f97a43a1 2183
1066358a 2184 if (perc) {
4b157ac6
JA
2185 this_write = min_not_zero(min_write,
2186 td->o.compress_chunk);
1e7f82e2
JA
2187
2188 fill_random_buf_percentage(rs, buf, perc,
2189 this_write, this_write,
2190 o->buffer_pattern,
2191 o->buffer_pattern_bytes);
2192 } else {
1066358a 2193 fill_random_buf(rs, buf, min_write);
1e7f82e2
JA
2194 this_write = min_write;
2195 }
1066358a 2196
1e7f82e2
JA
2197 buf += this_write;
2198 left -= this_write;
9451b93e 2199 save_buf_state(td, rs);
1066358a 2200 } while (left);
d1af2894
JA
2201 } else if (o->buffer_pattern_bytes)
2202 fill_buffer_pattern(td, buf, max_bs);
999d245e 2203 else if (o->zero_buffers)
cc86c395 2204 memset(buf, 0, max_bs);
999d245e 2205 else
9451b93e 2206 fill_random_buf(get_buf_state(td), buf, max_bs);
cc86c395
JA
2207}
2208
2209/*
2210 * "randomly" fill the buffer contents
2211 */
2212void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2213 unsigned int min_write, unsigned int max_bs)
2214{
2215 io_u->buf_filled_len = 0;
2216 fill_io_buffer(td, io_u->buf, min_write, max_bs);
5973cafb 2217}
e2c75fc4
TK
2218
2219static int do_sync_file_range(const struct thread_data *td,
2220 struct fio_file *f)
2221{
2222 off64_t offset, nbytes;
2223
2224 offset = f->first_write;
2225 nbytes = f->last_write - f->first_write;
2226
2227 if (!nbytes)
2228 return 0;
2229
2230 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2231}
2232
2233int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2234{
2235 int ret;
2236
2237 if (io_u->ddir == DDIR_SYNC) {
2238 ret = fsync(io_u->file->fd);
2239 } else if (io_u->ddir == DDIR_DATASYNC) {
2240#ifdef CONFIG_FDATASYNC
2241 ret = fdatasync(io_u->file->fd);
2242#else
2243 ret = io_u->xfer_buflen;
2244 io_u->error = EINVAL;
2245#endif
2246 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2247 ret = do_sync_file_range(td, io_u->file);
2248 else {
2249 ret = io_u->xfer_buflen;
2250 io_u->error = EINVAL;
2251 }
2252
2253 if (ret < 0)
2254 io_u->error = errno;
2255
2256 return ret;
2257}
2258
2259int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2260{
2261#ifndef FIO_HAVE_TRIM
2262 io_u->error = EINVAL;
2263 return 0;
2264#else
2265 struct fio_file *f = io_u->file;
2266 int ret;
2267
496b1f9e 2268 ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
e2c75fc4
TK
2269 if (!ret)
2270 return io_u->xfer_buflen;
2271
2272 io_u->error = ret;
2273 return 0;
2274#endif
2275}