Separate io_u from ioengine [2/3] - move io_u functions
[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 */
97601024
JA
23 struct timeval time; /* output */
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{
2dc1bbeb 40 unsigned int min_bs = td->o.rw_min_bs;
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
e0a04ac1
JA
160static int __get_next_rand_offset_zoned(struct thread_data *td,
161 struct fio_file *f, enum fio_ddir ddir,
162 uint64_t *b)
163{
164 unsigned int v, send, stotal;
165 uint64_t offset, lastb;
166 static int warned;
167 struct zone_split_index *zsi;
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 */
36dd3379 181 v = rand32_between(&td->zone_state, 1, 100);
e0a04ac1
JA
182
183 zsi = &td->zone_state_index[ddir][v - 1];
184 stotal = zsi->size_perc_prev;
185 send = zsi->size_perc;
186
187 /*
188 * Should never happen
189 */
190 if (send == -1U) {
191 if (!warned) {
192 log_err("fio: bug in zoned generation\n");
193 warned = 1;
194 }
195 goto bail;
196 }
197
198 /*
199 * 'send' is some percentage below or equal to 100 that
200 * marks the end of the current IO range. 'stotal' marks
201 * the start, in percent.
202 */
203 if (stotal)
204 offset = stotal * lastb / 100ULL;
205 else
206 offset = 0;
207
208 lastb = lastb * (send - stotal) / 100ULL;
209
210 /*
211 * Generate index from 0..send-of-lastb
212 */
213 if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
214 return 1;
215
216 /*
217 * Add our start offset, if any
218 */
219 if (offset)
220 *b += offset;
221
222 return 0;
223}
56d9fa4b 224
1ae83d45
JA
225static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
226{
227 struct rand_off *r1 = flist_entry(a, struct rand_off, list);
228 struct rand_off *r2 = flist_entry(b, struct rand_off, list);
229
230 return r1->off - r2->off;
231}
232
233static int get_off_from_method(struct thread_data *td, struct fio_file *f,
234 enum fio_ddir ddir, uint64_t *b)
e25839d4 235{
e0a04ac1
JA
236 if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
237 uint64_t lastb;
238
239 lastb = last_block(td, f, ddir);
240 if (!lastb)
241 return 1;
242
243 return __get_next_rand_offset(td, f, ddir, b, lastb);
244 } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
e25839d4 245 return __get_next_rand_offset_zipf(td, f, ddir, b);
925fee33
JA
246 else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
247 return __get_next_rand_offset_pareto(td, f, ddir, b);
56d9fa4b
JA
248 else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
249 return __get_next_rand_offset_gauss(td, f, ddir, b);
e0a04ac1
JA
250 else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
251 return __get_next_rand_offset_zoned(td, f, ddir, b);
e25839d4
JA
252
253 log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
254 return 1;
255}
256
bcd5abfa
JA
257/*
258 * Sort the reads for a verify phase in batches of verifysort_nr, if
259 * specified.
260 */
e39c0676 261static inline bool should_sort_io(struct thread_data *td)
bcd5abfa
JA
262{
263 if (!td->o.verifysort_nr || !td->o.do_verify)
e39c0676 264 return false;
bcd5abfa 265 if (!td_random(td))
e39c0676 266 return false;
bcd5abfa 267 if (td->runstate != TD_VERIFYING)
e39c0676 268 return false;
c3546b53
JA
269 if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
270 td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
e39c0676 271 return false;
bcd5abfa 272
e39c0676 273 return true;
bcd5abfa
JA
274}
275
e39c0676 276static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
211c9b89
JA
277{
278 unsigned int v;
211c9b89 279
d9472271 280 if (td->o.perc_rand[ddir] == 100)
e39c0676 281 return true;
211c9b89 282
36dd3379 283 v = rand32_between(&td->seq_rand_state[ddir], 1, 100);
211c9b89 284
d9472271 285 return v <= td->o.perc_rand[ddir];
211c9b89
JA
286}
287
1ae83d45
JA
288static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
289 enum fio_ddir ddir, uint64_t *b)
290{
291 struct rand_off *r;
292 int i, ret = 1;
293
bcd5abfa 294 if (!should_sort_io(td))
1ae83d45
JA
295 return get_off_from_method(td, f, ddir, b);
296
297 if (!flist_empty(&td->next_rand_list)) {
1ae83d45 298fetch:
9342d5f8 299 r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
1ae83d45
JA
300 flist_del(&r->list);
301 *b = r->off;
302 free(r);
303 return 0;
304 }
305
306 for (i = 0; i < td->o.verifysort_nr; i++) {
307 r = malloc(sizeof(*r));
308
309 ret = get_off_from_method(td, f, ddir, &r->off);
310 if (ret) {
311 free(r);
312 break;
313 }
314
315 flist_add(&r->list, &td->next_rand_list);
316 }
317
318 if (ret && !i)
319 return ret;
320
321 assert(!flist_empty(&td->next_rand_list));
322 flist_sort(NULL, &td->next_rand_list, flist_cmp);
323 goto fetch;
324}
325
38dad62d 326static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
1ae83d45 327 enum fio_ddir ddir, uint64_t *b)
38dad62d 328{
c04e4661
DE
329 if (!get_next_rand_offset(td, f, ddir, b))
330 return 0;
331
8c07860d
JA
332 if (td->o.time_based ||
333 (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
33c48814 334 fio_file_reset(td, f);
c04e4661
DE
335 if (!get_next_rand_offset(td, f, ddir, b))
336 return 0;
38dad62d
JA
337 }
338
c04e4661 339 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
f1dfb668 340 f->file_name, (unsigned long long) f->last_pos[ddir],
4b91ee8f 341 (unsigned long long) f->real_file_size);
c04e4661 342 return 1;
38dad62d
JA
343}
344
37cf9e3c 345static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
1ae83d45 346 enum fio_ddir ddir, uint64_t *offset)
38dad62d 347{
ac002339
JA
348 struct thread_options *o = &td->o;
349
ff58fced
JA
350 assert(ddir_rw(ddir));
351
f1dfb668 352 if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
19ddc35b
JA
353 o->time_based) {
354 struct thread_options *o = &td->o;
355 uint64_t io_size = f->io_size + (f->io_size % o->min_bs[ddir]);
356
357 if (io_size > f->last_pos[ddir])
358 f->last_pos[ddir] = 0;
359 else
360 f->last_pos[ddir] = f->last_pos[ddir] - io_size;
361 }
c04e4661 362
f1dfb668 363 if (f->last_pos[ddir] < f->real_file_size) {
1ae83d45 364 uint64_t pos;
059b0802 365
c22825bb
JA
366 if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) {
367 if (f->real_file_size > f->io_size)
368 f->last_pos[ddir] = f->io_size;
369 else
370 f->last_pos[ddir] = f->real_file_size;
371 }
a66da7a2 372
f1dfb668 373 pos = f->last_pos[ddir] - f->file_offset;
ac002339
JA
374 if (pos && o->ddir_seq_add) {
375 pos += o->ddir_seq_add;
376
377 /*
378 * If we reach beyond the end of the file
379 * with holed IO, wrap around to the
b0a84f48
JA
380 * beginning again. If we're doing backwards IO,
381 * wrap to the end.
ac002339 382 */
b0a84f48
JA
383 if (pos >= f->real_file_size) {
384 if (o->ddir_seq_add > 0)
385 pos = f->file_offset;
c22825bb
JA
386 else {
387 if (f->real_file_size > f->io_size)
388 pos = f->io_size;
389 else
390 pos = f->real_file_size;
391
392 pos += o->ddir_seq_add;
393 }
b0a84f48 394 }
ac002339 395 }
059b0802 396
37cf9e3c 397 *offset = pos;
38dad62d
JA
398 return 0;
399 }
400
401 return 1;
402}
403
404static int get_next_block(struct thread_data *td, struct io_u *io_u,
6aca9b3d
JA
405 enum fio_ddir ddir, int rw_seq,
406 unsigned int *is_random)
38dad62d
JA
407{
408 struct fio_file *f = io_u->file;
1ae83d45 409 uint64_t b, offset;
38dad62d
JA
410 int ret;
411
ff58fced
JA
412 assert(ddir_rw(ddir));
413
37cf9e3c
JA
414 b = offset = -1ULL;
415
38dad62d 416 if (rw_seq) {
211c9b89 417 if (td_random(td)) {
6aca9b3d 418 if (should_do_random(td, ddir)) {
211c9b89 419 ret = get_next_rand_block(td, f, ddir, &b);
6aca9b3d
JA
420 *is_random = 1;
421 } else {
422 *is_random = 0;
1651e431 423 io_u_set(td, io_u, IO_U_F_BUSY_OK);
211c9b89
JA
424 ret = get_next_seq_offset(td, f, ddir, &offset);
425 if (ret)
426 ret = get_next_rand_block(td, f, ddir, &b);
427 }
6aca9b3d
JA
428 } else {
429 *is_random = 0;
37cf9e3c 430 ret = get_next_seq_offset(td, f, ddir, &offset);
6aca9b3d 431 }
38dad62d 432 } else {
1651e431 433 io_u_set(td, io_u, IO_U_F_BUSY_OK);
6aca9b3d 434 *is_random = 0;
38dad62d
JA
435
436 if (td->o.rw_seq == RW_SEQ_SEQ) {
37cf9e3c 437 ret = get_next_seq_offset(td, f, ddir, &offset);
6aca9b3d 438 if (ret) {
37cf9e3c 439 ret = get_next_rand_block(td, f, ddir, &b);
6aca9b3d
JA
440 *is_random = 0;
441 }
38dad62d 442 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
f1dfb668
JA
443 if (f->last_start[ddir] != -1ULL)
444 offset = f->last_start[ddir] - f->file_offset;
38dad62d 445 else
37cf9e3c 446 offset = 0;
38dad62d
JA
447 ret = 0;
448 } else {
449 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
450 ret = 1;
451 }
452 }
6d68b997 453
37cf9e3c
JA
454 if (!ret) {
455 if (offset != -1ULL)
456 io_u->offset = offset;
457 else if (b != -1ULL)
458 io_u->offset = b * td->o.ba[ddir];
459 else {
4e0a8fa2 460 log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
37cf9e3c
JA
461 ret = 1;
462 }
463 }
464
38dad62d
JA
465 return ret;
466}
467
10ba535a
JA
468/*
469 * For random io, generate a random new block and see if it's used. Repeat
470 * until we find a free one. For sequential io, just return the end of
471 * the last io issued.
472 */
6aca9b3d
JA
473static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
474 unsigned int *is_random)
10ba535a 475{
9bf2061e 476 struct fio_file *f = io_u->file;
4ba66134 477 enum fio_ddir ddir = io_u->ddir;
38dad62d 478 int rw_seq_hit = 0;
10ba535a 479
ff58fced
JA
480 assert(ddir_rw(ddir));
481
38dad62d
JA
482 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
483 rw_seq_hit = 1;
5736c10d 484 td->ddir_seq_nr = td->o.ddir_seq_nr;
38dad62d 485 }
211097b2 486
6aca9b3d 487 if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
38dad62d 488 return 1;
10ba535a 489
009bd847
JA
490 if (io_u->offset >= f->io_size) {
491 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
4b91ee8f
JA
492 (unsigned long long) io_u->offset,
493 (unsigned long long) f->io_size);
009bd847
JA
494 return 1;
495 }
496
497 io_u->offset += f->file_offset;
2ba1c290
JA
498 if (io_u->offset >= f->real_file_size) {
499 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
4b91ee8f
JA
500 (unsigned long long) io_u->offset,
501 (unsigned long long) f->real_file_size);
10ba535a 502 return 1;
2ba1c290 503 }
10ba535a
JA
504
505 return 0;
506}
507
6aca9b3d
JA
508static int get_next_offset(struct thread_data *td, struct io_u *io_u,
509 unsigned int *is_random)
15dc1934 510{
d72be545
JA
511 if (td->flags & TD_F_PROFILE_OPS) {
512 struct prof_io_ops *ops = &td->prof_io_ops;
7eb36574 513
d72be545 514 if (ops->fill_io_u_off)
6aca9b3d 515 return ops->fill_io_u_off(td, io_u, is_random);
d72be545 516 }
15dc1934 517
6aca9b3d 518 return __get_next_offset(td, io_u, is_random);
15dc1934
JA
519}
520
e39c0676
JA
521static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
522 unsigned int buflen)
79944128
JA
523{
524 struct fio_file *f = io_u->file;
525
bedc9dc2 526 return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
79944128
JA
527}
528
6aca9b3d
JA
529static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
530 unsigned int is_random)
10ba535a 531{
6aca9b3d 532 int ddir = io_u->ddir;
24d23ca7 533 unsigned int buflen = 0;
f3059de1 534 unsigned int minbs, maxbs;
3dd29f7c 535 uint64_t frand_max, r;
7c961359 536 bool power_2;
10ba535a 537
9ee1c647 538 assert(ddir_rw(ddir));
6aca9b3d
JA
539
540 if (td->o.bs_is_seq_rand)
541 ddir = is_random ? DDIR_WRITE: DDIR_READ;
ff58fced 542
f3059de1
JA
543 minbs = td->o.min_bs[ddir];
544 maxbs = td->o.max_bs[ddir];
545
79944128
JA
546 if (minbs == maxbs)
547 return minbs;
548
52c58027
JA
549 /*
550 * If we can't satisfy the min block size from here, then fail
551 */
552 if (!io_u_fits(td, io_u, minbs))
553 return 0;
554
c3546b53 555 frand_max = rand_max(&td->bsrange_state);
79944128 556 do {
d6b72507 557 r = __rand(&td->bsrange_state);
4c07ad86 558
720e84ad 559 if (!td->o.bssplit_nr[ddir]) {
f3059de1 560 buflen = 1 + (unsigned int) ((double) maxbs *
c3546b53 561 (r / (frand_max + 1.0)));
f3059de1
JA
562 if (buflen < minbs)
563 buflen = minbs;
5ec10eaa 564 } else {
3dd29f7c 565 long long perc = 0;
564ca972
JA
566 unsigned int i;
567
720e84ad
JA
568 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
569 struct bssplit *bsp = &td->o.bssplit[ddir][i];
564ca972
JA
570
571 buflen = bsp->bs;
572 perc += bsp->perc;
3dd29f7c
JA
573 if (!perc)
574 break;
575 if ((r / perc <= frand_max / 100ULL) &&
79944128 576 io_u_fits(td, io_u, buflen))
564ca972
JA
577 break;
578 }
579 }
79944128 580
17a6b702
PL
581 power_2 = is_power_of_2(minbs);
582 if (!td->o.bs_unaligned && power_2)
7c306bb1 583 buflen &= ~(minbs - 1);
17a6b702
PL
584 else if (!td->o.bs_unaligned && !power_2)
585 buflen -= buflen % minbs;
79944128 586 } while (!io_u_fits(td, io_u, buflen));
6a5e6884 587
10ba535a
JA
588 return buflen;
589}
590
6aca9b3d
JA
591static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
592 unsigned int is_random)
15dc1934 593{
d72be545
JA
594 if (td->flags & TD_F_PROFILE_OPS) {
595 struct prof_io_ops *ops = &td->prof_io_ops;
7eb36574 596
d72be545 597 if (ops->fill_io_u_size)
6aca9b3d 598 return ops->fill_io_u_size(td, io_u, is_random);
d72be545 599 }
15dc1934 600
6aca9b3d 601 return __get_next_buflen(td, io_u, is_random);
15dc1934
JA
602}
603
afe24a5a
JA
604static void set_rwmix_bytes(struct thread_data *td)
605{
afe24a5a
JA
606 unsigned int diff;
607
608 /*
609 * we do time or byte based switch. this is needed because
610 * buffered writes may issue a lot quicker than they complete,
611 * whereas reads do not.
612 */
e47f799f 613 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
04c540d9 614 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
e47f799f
JA
615}
616
617static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
618{
619 unsigned int v;
e47f799f 620
36dd3379 621 v = rand32_between(&td->rwmix_state, 1, 100);
4c07ad86 622
04c540d9 623 if (v <= td->o.rwmix[DDIR_READ])
e47f799f
JA
624 return DDIR_READ;
625
626 return DDIR_WRITE;
afe24a5a
JA
627}
628
0d593542 629int io_u_quiesce(struct thread_data *td)
002e7183 630{
0d593542
JA
631 int completed = 0;
632
002e7183
JA
633 /*
634 * We are going to sleep, ensure that we flush anything pending as
635 * not to skew our latency numbers.
636 *
637 * Changed to only monitor 'in flight' requests here instead of the
638 * td->cur_depth, b/c td->cur_depth does not accurately represent
639 * io's that have been actually submitted to an async engine,
640 * and cur_depth is meaningless for sync engines.
641 */
9cc80b6d
JA
642 if (td->io_u_queued || td->cur_depth) {
643 int fio_unused ret;
644
645 ret = td_io_commit(td);
646 }
647
002e7183 648 while (td->io_u_in_flight) {
89fd0573 649 int ret;
002e7183 650
55312f9f 651 ret = io_u_queued_complete(td, 1);
0d593542
JA
652 if (ret > 0)
653 completed += ret;
002e7183 654 }
0d593542 655
6be06c46
JA
656 if (td->flags & TD_F_REGROW_LOGS)
657 regrow_logs(td);
658
0d593542 659 return completed;
002e7183
JA
660}
661
581e7141
JA
662static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
663{
664 enum fio_ddir odir = ddir ^ 1;
90eff1c9
SW
665 long usec;
666 uint64_t now;
581e7141 667
ff58fced 668 assert(ddir_rw(ddir));
50a8ce86 669 now = utime_since_now(&td->start);
ff58fced 670
50a8ce86
D
671 /*
672 * if rate_next_io_time is in the past, need to catch up to rate
673 */
674 if (td->rate_next_io_time[ddir] <= now)
581e7141
JA
675 return ddir;
676
677 /*
50a8ce86 678 * We are ahead of rate in this direction. See if we
581e7141
JA
679 * should switch.
680 */
315fcfec 681 if (td_rw(td) && td->o.rwmix[odir]) {
581e7141 682 /*
50a8ce86 683 * Other direction is behind rate, switch
581e7141 684 */
50a8ce86 685 if (td->rate_next_io_time[odir] <= now)
581e7141
JA
686 return odir;
687
688 /*
50a8ce86
D
689 * Both directions are ahead of rate. sleep the min
690 * switch if necissary
581e7141 691 */
50a8ce86
D
692 if (td->rate_next_io_time[ddir] <=
693 td->rate_next_io_time[odir]) {
694 usec = td->rate_next_io_time[ddir] - now;
581e7141 695 } else {
50a8ce86 696 usec = td->rate_next_io_time[odir] - now;
581e7141
JA
697 ddir = odir;
698 }
699 } else
50a8ce86 700 usec = td->rate_next_io_time[ddir] - now;
581e7141 701
a9da8ab2
JA
702 if (td->o.io_submit_mode == IO_MODE_INLINE)
703 io_u_quiesce(td);
78c1eda5 704
9cc80b6d 705 usec = usec_sleep(td, usec);
581e7141 706
581e7141
JA
707 return ddir;
708}
709
10ba535a
JA
710/*
711 * Return the data direction for the next io_u. If the job is a
712 * mixed read/write workload, check the rwmix cycle and switch if
713 * necessary.
714 */
1e97cce9 715static enum fio_ddir get_rw_ddir(struct thread_data *td)
10ba535a 716{
581e7141
JA
717 enum fio_ddir ddir;
718
5f9099ea 719 /*
f6860972
TK
720 * See if it's time to fsync/fdatasync/sync_file_range first,
721 * and if not then move on to check regular I/Os.
5f9099ea 722 */
f6860972
TK
723 if (should_fsync(td)) {
724 if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
725 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
726 return DDIR_SYNC;
727
728 if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
729 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
730 return DDIR_DATASYNC;
731
732 if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
733 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
734 return DDIR_SYNC_FILE_RANGE;
735 }
44f29692 736
10ba535a 737 if (td_rw(td)) {
10ba535a
JA
738 /*
739 * Check if it's time to seed a new data direction.
740 */
e4928662 741 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
e47f799f
JA
742 /*
743 * Put a top limit on how many bytes we do for
744 * one data direction, to avoid overflowing the
745 * ranges too much
746 */
747 ddir = get_rand_ddir(td);
e47f799f
JA
748
749 if (ddir != td->rwmix_ddir)
750 set_rwmix_bytes(td);
751
752 td->rwmix_ddir = ddir;
10ba535a 753 }
581e7141 754 ddir = td->rwmix_ddir;
10ba535a 755 } else if (td_read(td))
581e7141 756 ddir = DDIR_READ;
6eaf09d6 757 else if (td_write(td))
581e7141 758 ddir = DDIR_WRITE;
5c8e84ca 759 else if (td_trim(td))
6eaf09d6 760 ddir = DDIR_TRIM;
5c8e84ca
TK
761 else
762 ddir = DDIR_INVAL;
581e7141
JA
763
764 td->rwmix_ddir = rate_ddir(td, ddir);
765 return td->rwmix_ddir;
10ba535a
JA
766}
767
1ef2b6be
JA
768static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
769{
0e4dd95c
DE
770 enum fio_ddir ddir = get_rw_ddir(td);
771
82a90686 772 if (td_trimwrite(td)) {
0e4dd95c
DE
773 struct fio_file *f = io_u->file;
774 if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
775 ddir = DDIR_TRIM;
776 else
777 ddir = DDIR_WRITE;
778 }
779
780 io_u->ddir = io_u->acct_ddir = ddir;
1ef2b6be 781
9b87f09b 782 if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
1ef2b6be
JA
783 td->o.barrier_blocks &&
784 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
785 td->io_issues[DDIR_WRITE])
1651e431 786 io_u_set(td, io_u, IO_U_F_BARRIER);
1ef2b6be
JA
787}
788
e8462bd8 789void put_file_log(struct thread_data *td, struct fio_file *f)
60f2c658 790{
71b84caa 791 unsigned int ret = put_file(td, f);
60f2c658
JA
792
793 if (ret)
794 td_verror(td, ret, "file close");
795}
796
10ba535a
JA
797void put_io_u(struct thread_data *td, struct io_u *io_u)
798{
a9da8ab2
JA
799 if (td->parent)
800 td = td->parent;
801
e8462bd8
JA
802 td_io_u_lock(td);
803
f8b0bd10 804 if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
60f2c658 805 put_file_log(td, io_u->file);
f8b0bd10 806
10ba535a 807 io_u->file = NULL;
1651e431 808 io_u_set(td, io_u, IO_U_F_FREE);
d7ee2a7d 809
a9da8ab2 810 if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
0c41214f 811 td->cur_depth--;
a9da8ab2
JA
812 assert(!(td->flags & TD_F_CHILD));
813 }
2ae0b204 814 io_u_qpush(&td->io_u_freelist, io_u);
e8462bd8
JA
815 td_io_u_unlock(td);
816 td_io_u_free_notify(td);
10ba535a
JA
817}
818
f2bba182
RR
819void clear_io_u(struct thread_data *td, struct io_u *io_u)
820{
1651e431 821 io_u_clear(td, io_u, IO_U_F_FLIGHT);
f2bba182
RR
822 put_io_u(td, io_u);
823}
824
755200a3
JA
825void requeue_io_u(struct thread_data *td, struct io_u **io_u)
826{
827 struct io_u *__io_u = *io_u;
bcd5abfa 828 enum fio_ddir ddir = acct_ddir(__io_u);
755200a3 829
465221b0
JA
830 dprint(FD_IO, "requeue %p\n", __io_u);
831
a9da8ab2
JA
832 if (td->parent)
833 td = td->parent;
834
e8462bd8
JA
835 td_io_u_lock(td);
836
1651e431 837 io_u_set(td, __io_u, IO_U_F_FREE);
bcd5abfa
JA
838 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
839 td->io_issues[ddir]--;
5ec10eaa 840
1651e431 841 io_u_clear(td, __io_u, IO_U_F_FLIGHT);
a9da8ab2 842 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
0c41214f 843 td->cur_depth--;
a9da8ab2
JA
844 assert(!(td->flags & TD_F_CHILD));
845 }
2ae0b204
JA
846
847 io_u_rpush(&td->io_u_requeues, __io_u);
e8462bd8 848 td_io_u_unlock(td);
a9da8ab2 849 td_io_u_free_notify(td);
755200a3
JA
850 *io_u = NULL;
851}
852
9bf2061e 853static int fill_io_u(struct thread_data *td, struct io_u *io_u)
10ba535a 854{
6aca9b3d
JA
855 unsigned int is_random;
856
9b87f09b 857 if (td_ioengine_flagged(td, FIO_NOIO))
b4c5e1ac
JA
858 goto out;
859
1ef2b6be 860 set_rw_ddir(td, io_u);
5f9099ea 861
87dc1ab1 862 /*
ff58fced 863 * fsync() or fdatasync() or trim etc, we are done
87dc1ab1 864 */
ff58fced 865 if (!ddir_rw(io_u->ddir))
c38e9468 866 goto out;
a00735e6 867
48f5abd3
JA
868 /*
869 * See if it's time to switch to a new zone
870 */
13af05ae 871 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
4c8be5b1
JA
872 struct fio_file *f = io_u->file;
873
48f5abd3 874 td->zone_bytes = 0;
4c8be5b1
JA
875 f->file_offset += td->o.zone_range + td->o.zone_skip;
876
877 /*
878 * Wrap from the beginning, if we exceed the file size
879 */
880 if (f->file_offset >= f->real_file_size)
881 f->file_offset = f->real_file_size - f->file_offset;
f1dfb668 882 f->last_pos[io_u->ddir] = f->file_offset;
48f5abd3
JA
883 td->io_skip_bytes += td->o.zone_skip;
884 }
885
10ba535a 886 /*
c685b5b2
JA
887 * No log, let the seq/rand engine retrieve the next buflen and
888 * position.
10ba535a 889 */
6aca9b3d 890 if (get_next_offset(td, io_u, &is_random)) {
2ba1c290 891 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
bca4ed4d 892 return 1;
2ba1c290 893 }
10ba535a 894
6aca9b3d 895 io_u->buflen = get_next_buflen(td, io_u, is_random);
2ba1c290
JA
896 if (!io_u->buflen) {
897 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
bca4ed4d 898 return 1;
2ba1c290 899 }
bca4ed4d 900
2ba1c290
JA
901 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
902 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
4b91ee8f
JA
903 dprint(FD_IO, " off=%llu/%lu > %llu\n",
904 (unsigned long long) io_u->offset, io_u->buflen,
905 (unsigned long long) io_u->file->real_file_size);
6a5e6884 906 return 1;
2ba1c290 907 }
6a5e6884 908
bca4ed4d
JA
909 /*
910 * mark entry before potentially trimming io_u
911 */
303032ae 912 if (td_random(td) && file_randommap(td, io_u->file))
9bf2061e 913 mark_random_map(td, io_u);
bca4ed4d 914
c38e9468 915out:
2ba1c290 916 dprint_io_u(io_u, "fill_io_u");
d9d91e39 917 td->zone_bytes += io_u->buflen;
bca4ed4d 918 return 0;
10ba535a
JA
919}
920
838bc709
JA
921static void __io_u_mark_map(unsigned int *map, unsigned int nr)
922{
2b13e716 923 int idx = 0;
838bc709
JA
924
925 switch (nr) {
926 default:
2b13e716 927 idx = 6;
838bc709
JA
928 break;
929 case 33 ... 64:
2b13e716 930 idx = 5;
838bc709
JA
931 break;
932 case 17 ... 32:
2b13e716 933 idx = 4;
838bc709
JA
934 break;
935 case 9 ... 16:
2b13e716 936 idx = 3;
838bc709
JA
937 break;
938 case 5 ... 8:
2b13e716 939 idx = 2;
838bc709
JA
940 break;
941 case 1 ... 4:
2b13e716 942 idx = 1;
838bc709
JA
943 case 0:
944 break;
945 }
946
2b13e716 947 map[idx]++;
838bc709
JA
948}
949
950void io_u_mark_submit(struct thread_data *td, unsigned int nr)
951{
952 __io_u_mark_map(td->ts.io_u_submit, nr);
953 td->ts.total_submit++;
954}
955
956void io_u_mark_complete(struct thread_data *td, unsigned int nr)
957{
958 __io_u_mark_map(td->ts.io_u_complete, nr);
959 td->ts.total_complete++;
960}
961
d8005759 962void io_u_mark_depth(struct thread_data *td, unsigned int nr)
71619dc2 963{
2b13e716 964 int idx = 0;
71619dc2
JA
965
966 switch (td->cur_depth) {
967 default:
2b13e716 968 idx = 6;
a783e61a 969 break;
71619dc2 970 case 32 ... 63:
2b13e716 971 idx = 5;
a783e61a 972 break;
71619dc2 973 case 16 ... 31:
2b13e716 974 idx = 4;
a783e61a 975 break;
71619dc2 976 case 8 ... 15:
2b13e716 977 idx = 3;
a783e61a 978 break;
71619dc2 979 case 4 ... 7:
2b13e716 980 idx = 2;
a783e61a 981 break;
71619dc2 982 case 2 ... 3:
2b13e716 983 idx = 1;
71619dc2
JA
984 case 1:
985 break;
986 }
987
2b13e716 988 td->ts.io_u_map[idx] += nr;
71619dc2
JA
989}
990
04a0feae
JA
991static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
992{
2b13e716 993 int idx = 0;
04a0feae
JA
994
995 assert(usec < 1000);
996
997 switch (usec) {
998 case 750 ... 999:
2b13e716 999 idx = 9;
04a0feae
JA
1000 break;
1001 case 500 ... 749:
2b13e716 1002 idx = 8;
04a0feae
JA
1003 break;
1004 case 250 ... 499:
2b13e716 1005 idx = 7;
04a0feae
JA
1006 break;
1007 case 100 ... 249:
2b13e716 1008 idx = 6;
04a0feae
JA
1009 break;
1010 case 50 ... 99:
2b13e716 1011 idx = 5;
04a0feae
JA
1012 break;
1013 case 20 ... 49:
2b13e716 1014 idx = 4;
04a0feae
JA
1015 break;
1016 case 10 ... 19:
2b13e716 1017 idx = 3;
04a0feae
JA
1018 break;
1019 case 4 ... 9:
2b13e716 1020 idx = 2;
04a0feae
JA
1021 break;
1022 case 2 ... 3:
2b13e716 1023 idx = 1;
04a0feae
JA
1024 case 0 ... 1:
1025 break;
1026 }
1027
2b13e716
JA
1028 assert(idx < FIO_IO_U_LAT_U_NR);
1029 td->ts.io_u_lat_u[idx]++;
04a0feae
JA
1030}
1031
1032static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
ec118304 1033{
2b13e716 1034 int idx = 0;
ec118304
JA
1035
1036 switch (msec) {
1037 default:
2b13e716 1038 idx = 11;
04a0feae 1039 break;
8abdce66 1040 case 1000 ... 1999:
2b13e716 1041 idx = 10;
04a0feae 1042 break;
8abdce66 1043 case 750 ... 999:
2b13e716 1044 idx = 9;
04a0feae 1045 break;
8abdce66 1046 case 500 ... 749:
2b13e716 1047 idx = 8;
04a0feae 1048 break;
8abdce66 1049 case 250 ... 499:
2b13e716 1050 idx = 7;
04a0feae 1051 break;
8abdce66 1052 case 100 ... 249:
2b13e716 1053 idx = 6;
04a0feae 1054 break;
8abdce66 1055 case 50 ... 99:
2b13e716 1056 idx = 5;
04a0feae 1057 break;
8abdce66 1058 case 20 ... 49:
2b13e716 1059 idx = 4;
04a0feae 1060 break;
8abdce66 1061 case 10 ... 19:
2b13e716 1062 idx = 3;
04a0feae 1063 break;
8abdce66 1064 case 4 ... 9:
2b13e716 1065 idx = 2;
04a0feae 1066 break;
ec118304 1067 case 2 ... 3:
2b13e716 1068 idx = 1;
ec118304
JA
1069 case 0 ... 1:
1070 break;
1071 }
1072
2b13e716
JA
1073 assert(idx < FIO_IO_U_LAT_M_NR);
1074 td->ts.io_u_lat_m[idx]++;
04a0feae
JA
1075}
1076
1077static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
1078{
1079 if (usec < 1000)
1080 io_u_mark_lat_usec(td, usec);
1081 else
1082 io_u_mark_lat_msec(td, usec / 1000);
ec118304
JA
1083}
1084
8c07860d
JA
1085static unsigned int __get_next_fileno_rand(struct thread_data *td)
1086{
1087 unsigned long fileno;
1088
1089 if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
1090 uint64_t frand_max = rand_max(&td->next_file_state);
1091 unsigned long r;
1092
1093 r = __rand(&td->next_file_state);
1094 return (unsigned int) ((double) td->o.nr_files
1095 * (r / (frand_max + 1.0)));
1096 }
1097
1098 if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
1099 fileno = zipf_next(&td->next_file_zipf);
1100 else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
1101 fileno = pareto_next(&td->next_file_zipf);
1102 else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
1103 fileno = gauss_next(&td->next_file_gauss);
1104 else {
1105 log_err("fio: bad file service type: %d\n", td->o.file_service_type);
1106 assert(0);
1107 return 0;
1108 }
1109
1110 return fileno >> FIO_FSERVICE_SHIFT;
1111}
1112
0aabe160
JA
1113/*
1114 * Get next file to service by choosing one at random
1115 */
2cc52930
JA
1116static struct fio_file *get_next_file_rand(struct thread_data *td,
1117 enum fio_file_flags goodf,
d6aed795 1118 enum fio_file_flags badf)
0aabe160 1119{
0aabe160 1120 struct fio_file *f;
1c178180 1121 int fno;
0aabe160
JA
1122
1123 do {
87b10676 1124 int opened = 0;
4c07ad86 1125
8c07860d 1126 fno = __get_next_fileno_rand(td);
7c83c089 1127
126d65c6 1128 f = td->files[fno];
d6aed795 1129 if (fio_file_done(f))
059e63c0 1130 continue;
1c178180 1131
d6aed795 1132 if (!fio_file_open(f)) {
87b10676
JA
1133 int err;
1134
002fe734
JA
1135 if (td->nr_open_files >= td->o.open_files)
1136 return ERR_PTR(-EBUSY);
1137
87b10676
JA
1138 err = td_io_open_file(td, f);
1139 if (err)
1140 continue;
1141 opened = 1;
1142 }
1143
2ba1c290
JA
1144 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
1145 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
0aabe160 1146 return f;
2ba1c290 1147 }
87b10676
JA
1148 if (opened)
1149 td_io_close_file(td, f);
0aabe160
JA
1150 } while (1);
1151}
1152
1153/*
1154 * Get next file to service by doing round robin between all available ones
1155 */
1c178180
JA
1156static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
1157 int badf)
3d7c391d
JA
1158{
1159 unsigned int old_next_file = td->next_file;
1160 struct fio_file *f;
1161
1162 do {
87b10676
JA
1163 int opened = 0;
1164
126d65c6 1165 f = td->files[td->next_file];
3d7c391d
JA
1166
1167 td->next_file++;
2dc1bbeb 1168 if (td->next_file >= td->o.nr_files)
3d7c391d
JA
1169 td->next_file = 0;
1170
87b10676 1171 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
d6aed795 1172 if (fio_file_done(f)) {
d5ed68ea 1173 f = NULL;
059e63c0 1174 continue;
d5ed68ea 1175 }
059e63c0 1176
d6aed795 1177 if (!fio_file_open(f)) {
87b10676
JA
1178 int err;
1179
002fe734
JA
1180 if (td->nr_open_files >= td->o.open_files)
1181 return ERR_PTR(-EBUSY);
1182
87b10676 1183 err = td_io_open_file(td, f);
b5696bfc
JA
1184 if (err) {
1185 dprint(FD_FILE, "error %d on open of %s\n",
1186 err, f->file_name);
87c27b45 1187 f = NULL;
87b10676 1188 continue;
b5696bfc 1189 }
87b10676
JA
1190 opened = 1;
1191 }
1192
0b9d69ec
JA
1193 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
1194 f->flags);
1c178180 1195 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
3d7c391d
JA
1196 break;
1197
87b10676
JA
1198 if (opened)
1199 td_io_close_file(td, f);
1200
3d7c391d
JA
1201 f = NULL;
1202 } while (td->next_file != old_next_file);
1203
2ba1c290 1204 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
3d7c391d
JA
1205 return f;
1206}
1207
7eb36574 1208static struct fio_file *__get_next_file(struct thread_data *td)
bdb4e2e9 1209{
1907dbc6
JA
1210 struct fio_file *f;
1211
2dc1bbeb 1212 assert(td->o.nr_files <= td->files_index);
1c178180 1213
b5696bfc 1214 if (td->nr_done_files >= td->o.nr_files) {
5ec10eaa
JA
1215 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1216 " nr_files=%d\n", td->nr_open_files,
1217 td->nr_done_files,
1218 td->o.nr_files);
bdb4e2e9 1219 return NULL;
2ba1c290 1220 }
bdb4e2e9 1221
1907dbc6 1222 f = td->file_service_file;
d6aed795 1223 if (f && fio_file_open(f) && !fio_file_closing(f)) {
a086c257
JA
1224 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1225 goto out;
1226 if (td->file_service_left--)
1227 goto out;
1228 }
1907dbc6 1229
a086c257
JA
1230 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1231 td->o.file_service_type == FIO_FSERVICE_SEQ)
d6aed795 1232 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
bdb4e2e9 1233 else
d6aed795 1234 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1907dbc6 1235
002fe734
JA
1236 if (IS_ERR(f))
1237 return f;
1238
1907dbc6
JA
1239 td->file_service_file = f;
1240 td->file_service_left = td->file_service_nr - 1;
2ba1c290 1241out:
0dac421f
JA
1242 if (f)
1243 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1244 else
1245 dprint(FD_FILE, "get_next_file: NULL\n");
1907dbc6 1246 return f;
bdb4e2e9
JA
1247}
1248
7eb36574
JA
1249static struct fio_file *get_next_file(struct thread_data *td)
1250{
372d8962 1251 if (td->flags & TD_F_PROFILE_OPS) {
d72be545 1252 struct prof_io_ops *ops = &td->prof_io_ops;
7eb36574 1253
d72be545
JA
1254 if (ops->get_next_file)
1255 return ops->get_next_file(td);
1256 }
7eb36574
JA
1257
1258 return __get_next_file(td);
1259}
1260
002fe734 1261static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
429f6675
JA
1262{
1263 struct fio_file *f;
1264
1265 do {
1266 f = get_next_file(td);
002fe734
JA
1267 if (IS_ERR_OR_NULL(f))
1268 return PTR_ERR(f);
429f6675 1269
429f6675
JA
1270 io_u->file = f;
1271 get_file(f);
1272
1273 if (!fill_io_u(td, io_u))
1274 break;
1275
b5696bfc 1276 put_file_log(td, f);
429f6675 1277 td_io_close_file(td, f);
b5696bfc 1278 io_u->file = NULL;
8c07860d
JA
1279 if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
1280 fio_file_reset(td, f);
1281 else {
1282 fio_file_set_done(f);
1283 td->nr_done_files++;
1284 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
0b9d69ec 1285 td->nr_done_files, td->o.nr_files);
8c07860d 1286 }
429f6675
JA
1287 } while (1);
1288
1289 return 0;
1290}
1291
3e260a46
JA
1292static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
1293 unsigned long tusec, unsigned long max_usec)
1294{
1295 if (!td->error)
1296 log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
1297 td_verror(td, ETIMEDOUT, "max latency exceeded");
1298 icd->error = ETIMEDOUT;
1299}
1300
1301static void lat_new_cycle(struct thread_data *td)
1302{
1303 fio_gettime(&td->latency_ts, NULL);
1304 td->latency_ios = ddir_rw_sum(td->io_blocks);
1305 td->latency_failed = 0;
1306}
1307
1308/*
1309 * We had an IO outside the latency target. Reduce the queue depth. If we
1310 * are at QD=1, then it's time to give up.
1311 */
e39c0676 1312static bool __lat_target_failed(struct thread_data *td)
3e260a46
JA
1313{
1314 if (td->latency_qd == 1)
e39c0676 1315 return true;
3e260a46
JA
1316
1317 td->latency_qd_high = td->latency_qd;
6bb58215
JA
1318
1319 if (td->latency_qd == td->latency_qd_low)
1320 td->latency_qd_low--;
1321
3e260a46
JA
1322 td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
1323
1324 dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
1325
1326 /*
1327 * When we ramp QD down, quiesce existing IO to prevent
1328 * a storm of ramp downs due to pending higher depth.
1329 */
1330 io_u_quiesce(td);
1331 lat_new_cycle(td);
e39c0676 1332 return false;
3e260a46
JA
1333}
1334
e39c0676 1335static bool lat_target_failed(struct thread_data *td)
3e260a46
JA
1336{
1337 if (td->o.latency_percentile.u.f == 100.0)
1338 return __lat_target_failed(td);
1339
1340 td->latency_failed++;
e39c0676 1341 return false;
3e260a46
JA
1342}
1343
1344void lat_target_init(struct thread_data *td)
1345{
6bb58215
JA
1346 td->latency_end_run = 0;
1347
3e260a46
JA
1348 if (td->o.latency_target) {
1349 dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
1350 fio_gettime(&td->latency_ts, NULL);
1351 td->latency_qd = 1;
1352 td->latency_qd_high = td->o.iodepth;
1353 td->latency_qd_low = 1;
1354 td->latency_ios = ddir_rw_sum(td->io_blocks);
1355 } else
1356 td->latency_qd = td->o.iodepth;
1357}
1358
6bb58215
JA
1359void lat_target_reset(struct thread_data *td)
1360{
1361 if (!td->latency_end_run)
1362 lat_target_init(td);
1363}
1364
3e260a46
JA
1365static void lat_target_success(struct thread_data *td)
1366{
1367 const unsigned int qd = td->latency_qd;
6bb58215 1368 struct thread_options *o = &td->o;
3e260a46
JA
1369
1370 td->latency_qd_low = td->latency_qd;
1371
1372 /*
1373 * If we haven't failed yet, we double up to a failing value instead
1374 * of bisecting from highest possible queue depth. If we have set
1375 * a limit other than td->o.iodepth, bisect between that.
1376 */
6bb58215 1377 if (td->latency_qd_high != o->iodepth)
3e260a46
JA
1378 td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
1379 else
1380 td->latency_qd *= 2;
1381
6bb58215
JA
1382 if (td->latency_qd > o->iodepth)
1383 td->latency_qd = o->iodepth;
3e260a46
JA
1384
1385 dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
6bb58215 1386
3e260a46 1387 /*
6bb58215
JA
1388 * Same as last one, we are done. Let it run a latency cycle, so
1389 * we get only the results from the targeted depth.
3e260a46 1390 */
6bb58215
JA
1391 if (td->latency_qd == qd) {
1392 if (td->latency_end_run) {
1393 dprint(FD_RATE, "We are done\n");
1394 td->done = 1;
1395 } else {
1396 dprint(FD_RATE, "Quiesce and final run\n");
1397 io_u_quiesce(td);
1398 td->latency_end_run = 1;
1399 reset_all_stats(td);
1400 reset_io_stats(td);
1401 }
1402 }
3e260a46
JA
1403
1404 lat_new_cycle(td);
1405}
1406
1407/*
1408 * Check if we can bump the queue depth
1409 */
1410void lat_target_check(struct thread_data *td)
1411{
1412 uint64_t usec_window;
1413 uint64_t ios;
1414 double success_ios;
1415
1416 usec_window = utime_since_now(&td->latency_ts);
1417 if (usec_window < td->o.latency_window)
1418 return;
1419
1420 ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
1421 success_ios = (double) (ios - td->latency_failed) / (double) ios;
1422 success_ios *= 100.0;
1423
1424 dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
1425
1426 if (success_ios >= td->o.latency_percentile.u.f)
1427 lat_target_success(td);
1428 else
1429 __lat_target_failed(td);
1430}
1431
1432/*
1433 * If latency target is enabled, we might be ramping up or down and not
1434 * using the full queue depth available.
1435 */
e39c0676 1436bool queue_full(const struct thread_data *td)
3e260a46
JA
1437{
1438 const int qempty = io_u_qempty(&td->io_u_freelist);
1439
1440 if (qempty)
e39c0676 1441 return true;
3e260a46 1442 if (!td->o.latency_target)
e39c0676 1443 return false;
3e260a46
JA
1444
1445 return td->cur_depth >= td->latency_qd;
1446}
429f6675 1447
10ba535a
JA
1448struct io_u *__get_io_u(struct thread_data *td)
1449{
0cae66f6 1450 struct io_u *io_u = NULL;
10ba535a 1451
ca09be4b
JA
1452 if (td->stop_io)
1453 return NULL;
1454
e8462bd8
JA
1455 td_io_u_lock(td);
1456
1457again:
2ae0b204
JA
1458 if (!io_u_rempty(&td->io_u_requeues))
1459 io_u = io_u_rpop(&td->io_u_requeues);
3e260a46 1460 else if (!queue_full(td)) {
2ae0b204 1461 io_u = io_u_qpop(&td->io_u_freelist);
10ba535a 1462
225ba9e3 1463 io_u->file = NULL;
6040dabc 1464 io_u->buflen = 0;
10ba535a 1465 io_u->resid = 0;
d7762cf8 1466 io_u->end_io = NULL;
755200a3
JA
1467 }
1468
1469 if (io_u) {
0c6e7517 1470 assert(io_u->flags & IO_U_F_FREE);
1651e431 1471 io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
f8b0bd10
JA
1472 IO_U_F_TRIMMED | IO_U_F_BARRIER |
1473 IO_U_F_VER_LIST);
0c6e7517 1474
755200a3 1475 io_u->error = 0;
bcd5abfa 1476 io_u->acct_ddir = -1;
10ba535a 1477 td->cur_depth++;
a9da8ab2 1478 assert(!(td->flags & TD_F_CHILD));
1651e431 1479 io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
f9401285 1480 io_u->ipo = NULL;
a9da8ab2 1481 } else if (td_async_processing(td)) {
1dec3e07
JA
1482 /*
1483 * We ran out, wait for async verify threads to finish and
1484 * return one
1485 */
a9da8ab2
JA
1486 assert(!(td->flags & TD_F_CHILD));
1487 assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
1dec3e07 1488 goto again;
10ba535a
JA
1489 }
1490
e8462bd8 1491 td_io_u_unlock(td);
10ba535a
JA
1492 return io_u;
1493}
1494
e39c0676 1495static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
10ba535a 1496{
d72be545 1497 if (!(td->flags & TD_F_TRIM_BACKLOG))
e39c0676 1498 return false;
d72be545
JA
1499
1500 if (td->trim_entries) {
0d29de83 1501 int get_trim = 0;
10ba535a 1502
0d29de83
JA
1503 if (td->trim_batch) {
1504 td->trim_batch--;
1505 get_trim = 1;
1506 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1507 td->last_ddir != DDIR_READ) {
1508 td->trim_batch = td->o.trim_batch;
1509 if (!td->trim_batch)
1510 td->trim_batch = td->o.trim_backlog;
1511 get_trim = 1;
1512 }
1513
917c64b3 1514 if (get_trim && get_next_trim(td, io_u))
e39c0676 1515 return true;
2ba1c290 1516 }
10ba535a 1517
e39c0676 1518 return false;
0d29de83
JA
1519}
1520
e39c0676 1521static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
0d29de83 1522{
d72be545 1523 if (!(td->flags & TD_F_VER_BACKLOG))
e39c0676 1524 return false;
d72be545
JA
1525
1526 if (td->io_hist_len) {
9e144189
JA
1527 int get_verify = 0;
1528
d1ece0c7 1529 if (td->verify_batch)
9e144189 1530 get_verify = 1;
d1ece0c7 1531 else if (!(td->io_hist_len % td->o.verify_backlog) &&
9e144189
JA
1532 td->last_ddir != DDIR_READ) {
1533 td->verify_batch = td->o.verify_batch;
f8a75c99
JA
1534 if (!td->verify_batch)
1535 td->verify_batch = td->o.verify_backlog;
9e144189
JA
1536 get_verify = 1;
1537 }
1538
d1ece0c7
JA
1539 if (get_verify && !get_next_verify(td, io_u)) {
1540 td->verify_batch--;
e39c0676 1541 return true;
d1ece0c7 1542 }
9e144189
JA
1543 }
1544
e39c0676 1545 return false;
0d29de83
JA
1546}
1547
de789769
JA
1548/*
1549 * Fill offset and start time into the buffer content, to prevent too
23f394d5
JA
1550 * easy compressible data for simple de-dupe attempts. Do this for every
1551 * 512b block in the range, since that should be the smallest block size
1552 * we can expect from a device.
de789769
JA
1553 */
1554static void small_content_scramble(struct io_u *io_u)
1555{
23f394d5 1556 unsigned int i, nr_blocks = io_u->buflen / 512;
1ae83d45 1557 uint64_t boffset;
23f394d5
JA
1558 unsigned int offset;
1559 void *p, *end;
de789769 1560
23f394d5
JA
1561 if (!nr_blocks)
1562 return;
1563
1564 p = io_u->xfer_buf;
fba76ee8 1565 boffset = io_u->offset;
81f0366c 1566 io_u->buf_filled_len = 0;
fad82f76 1567
23f394d5
JA
1568 for (i = 0; i < nr_blocks; i++) {
1569 /*
1570 * Fill the byte offset into a "random" start offset of
1571 * the buffer, given by the product of the usec time
1572 * and the actual offset.
1573 */
fad82f76 1574 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
1ae83d45
JA
1575 offset &= ~(sizeof(uint64_t) - 1);
1576 if (offset >= 512 - sizeof(uint64_t))
1577 offset -= sizeof(uint64_t);
fba76ee8 1578 memcpy(p + offset, &boffset, sizeof(boffset));
23f394d5
JA
1579
1580 end = p + 512 - sizeof(io_u->start_time);
1581 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1582 p += 512;
fad82f76 1583 boffset += 512;
23f394d5 1584 }
de789769
JA
1585}
1586
0d29de83
JA
1587/*
1588 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1589 * etc. The returned io_u is fully ready to be prepped and submitted.
1590 */
1591struct io_u *get_io_u(struct thread_data *td)
1592{
1593 struct fio_file *f;
1594 struct io_u *io_u;
de789769 1595 int do_scramble = 0;
002fe734 1596 long ret = 0;
0d29de83
JA
1597
1598 io_u = __get_io_u(td);
1599 if (!io_u) {
1600 dprint(FD_IO, "__get_io_u failed\n");
1601 return NULL;
1602 }
1603
1604 if (check_get_verify(td, io_u))
1605 goto out;
1606 if (check_get_trim(td, io_u))
1607 goto out;
1608
755200a3
JA
1609 /*
1610 * from a requeue, io_u already setup
1611 */
1612 if (io_u->file)
77f392bf 1613 goto out;
755200a3 1614
429f6675
JA
1615 /*
1616 * If using an iolog, grab next piece if any available.
1617 */
d72be545 1618 if (td->flags & TD_F_READ_IOLOG) {
429f6675
JA
1619 if (read_iolog_get(td, io_u))
1620 goto err_put;
2ba1c290 1621 } else if (set_io_u_file(td, io_u)) {
002fe734 1622 ret = -EBUSY;
2ba1c290 1623 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
429f6675 1624 goto err_put;
2ba1c290 1625 }
5ec10eaa 1626
429f6675 1627 f = io_u->file;
002fe734
JA
1628 if (!f) {
1629 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
1630 goto err_put;
1631 }
1632
d6aed795 1633 assert(fio_file_open(f));
97af62ce 1634
ff58fced 1635 if (ddir_rw(io_u->ddir)) {
9b87f09b 1636 if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
2ba1c290 1637 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
429f6675 1638 goto err_put;
2ba1c290 1639 }
10ba535a 1640
f1dfb668
JA
1641 f->last_start[io_u->ddir] = io_u->offset;
1642 f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
10ba535a 1643
fd68418e 1644 if (io_u->ddir == DDIR_WRITE) {
d72be545 1645 if (td->flags & TD_F_REFILL_BUFFERS) {
9c42684e 1646 io_u_fill_buffer(td, io_u,
1066358a 1647 td->o.min_bs[DDIR_WRITE],
9e129577 1648 io_u->buflen);
ff441ae8 1649 } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
bedc9dc2 1650 !(td->flags & TD_F_COMPRESS))
fd68418e 1651 do_scramble = 1;
d72be545 1652 if (td->flags & TD_F_VER_NONE) {
629f1d71
JA
1653 populate_verify_io_u(td, io_u);
1654 do_scramble = 0;
1655 }
fd68418e 1656 } else if (io_u->ddir == DDIR_READ) {
cbe8d756
RR
1657 /*
1658 * Reset the buf_filled parameters so next time if the
1659 * buffer is used for writes it is refilled.
1660 */
cbe8d756
RR
1661 io_u->buf_filled_len = 0;
1662 }
87dc1ab1 1663 }
10ba535a 1664
165faf16
JA
1665 /*
1666 * Set io data pointers.
1667 */
cec6b55d
JA
1668 io_u->xfer_buf = io_u->buf;
1669 io_u->xfer_buflen = io_u->buflen;
5973cafb 1670
6ac7a331 1671out:
0d29de83 1672 assert(io_u->file);
429f6675 1673 if (!td_io_prep(td, io_u)) {
3ecc8c67 1674 if (!td->o.disable_lat)
993bf48b 1675 fio_gettime(&io_u->start_time, NULL);
de789769
JA
1676 if (do_scramble)
1677 small_content_scramble(io_u);
429f6675 1678 return io_u;
36167d82 1679 }
429f6675 1680err_put:
2ba1c290 1681 dprint(FD_IO, "get_io_u failed\n");
429f6675 1682 put_io_u(td, io_u);
002fe734 1683 return ERR_PTR(ret);
10ba535a
JA
1684}
1685
a9da8ab2 1686static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
5451792e 1687{
8b28bd41 1688 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
825f818e 1689
8b28bd41
DM
1690 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1691 return;
5451792e 1692
709c8313
RE
1693 log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
1694 io_u->file ? " on file " : "",
1695 io_u->file ? io_u->file->file_name : "",
1696 strerror(io_u->error),
1697 io_ddir_name(io_u->ddir),
1698 io_u->offset, io_u->xfer_buflen);
5451792e 1699
5ad7be56
KD
1700 if (td->io_ops->errdetails) {
1701 char *err = td->io_ops->errdetails(io_u);
1702
1703 log_err("fio: %s\n", err);
1704 free(err);
1705 }
1706
5451792e
JA
1707 if (!td->error)
1708 td_verror(td, io_u->error, "io_u error");
1709}
1710
a9da8ab2
JA
1711void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1712{
1713 __io_u_log_error(td, io_u);
1714 if (td->parent)
094e66cb 1715 __io_u_log_error(td->parent, io_u);
a9da8ab2
JA
1716}
1717
e39c0676 1718static inline bool gtod_reduce(struct thread_data *td)
aba6c951 1719{
3ecc8c67
JA
1720 return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
1721 || td->o.gtod_reduce;
aba6c951
JA
1722}
1723
c8eeb9df
JA
1724static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1725 struct io_completion_data *icd,
1726 const enum fio_ddir idx, unsigned int bytes)
1727{
a9da8ab2 1728 const int no_reduce = !gtod_reduce(td);
24d23ca7 1729 unsigned long lusec = 0;
c8eeb9df 1730
75dc383e
JA
1731 if (td->parent)
1732 td = td->parent;
1733
8243be59
JA
1734 if (!td->o.stats)
1735 return;
1736
a9da8ab2 1737 if (no_reduce)
c8eeb9df
JA
1738 lusec = utime_since(&io_u->issue_time, &icd->time);
1739
1740 if (!td->o.disable_lat) {
1741 unsigned long tusec;
1742
1743 tusec = utime_since(&io_u->start_time, &icd->time);
ae588852 1744 add_lat_sample(td, idx, tusec, bytes, io_u->offset);
15501535 1745
d4afedfd
JA
1746 if (td->flags & TD_F_PROFILE_OPS) {
1747 struct prof_io_ops *ops = &td->prof_io_ops;
1748
1749 if (ops->io_u_lat)
1750 icd->error = ops->io_u_lat(td, tusec);
1751 }
1752
3e260a46
JA
1753 if (td->o.max_latency && tusec > td->o.max_latency)
1754 lat_fatal(td, icd, tusec, td->o.max_latency);
1755 if (td->o.latency_target && tusec > td->o.latency_target) {
1756 if (lat_target_failed(td))
1757 lat_fatal(td, icd, tusec, td->o.latency_target);
15501535 1758 }
c8eeb9df
JA
1759 }
1760
a47591e4
JA
1761 if (ddir_rw(idx)) {
1762 if (!td->o.disable_clat) {
1763 add_clat_sample(td, idx, lusec, bytes, io_u->offset);
1764 io_u_mark_latency(td, lusec);
1765 }
c8eeb9df 1766
a47591e4
JA
1767 if (!td->o.disable_bw && per_unit_log(td->bw_log))
1768 add_bw_sample(td, io_u, bytes, lusec);
c8eeb9df 1769
a47591e4
JA
1770 if (no_reduce && per_unit_log(td->iops_log))
1771 add_iops_sample(td, io_u, bytes);
1772 }
66347cfa
DE
1773
1774 if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
1775 uint32_t *info = io_u_block_info(td, io_u);
1776 if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
1777 if (io_u->ddir == DDIR_TRIM) {
1778 *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
1779 BLOCK_INFO_TRIMS(*info) + 1);
1780 } else if (io_u->ddir == DDIR_WRITE) {
1781 *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
1782 *info);
1783 }
1784 }
1785 }
c8eeb9df
JA
1786}
1787
94a6e1bb
JA
1788static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
1789 uint64_t offset, unsigned int bytes)
1790{
1791 int idx;
1792
639ad1ea
JA
1793 if (!f)
1794 return;
1795
94a6e1bb
JA
1796 if (f->first_write == -1ULL || offset < f->first_write)
1797 f->first_write = offset;
1798 if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
1799 f->last_write = offset + bytes;
1800
1801 if (!f->last_write_comp)
1802 return;
1803
1804 idx = f->last_write_idx++;
1805 f->last_write_comp[idx] = offset;
1806 if (f->last_write_idx == td->o.iodepth)
1807 f->last_write_idx = 0;
1808}
1809
f8b0bd10 1810static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
97601024 1811 struct io_completion_data *icd)
10ba535a 1812{
f8b0bd10
JA
1813 struct io_u *io_u = *io_u_ptr;
1814 enum fio_ddir ddir = io_u->ddir;
1815 struct fio_file *f = io_u->file;
10ba535a 1816
2ba1c290
JA
1817 dprint_io_u(io_u, "io complete");
1818
0c6e7517 1819 assert(io_u->flags & IO_U_F_FLIGHT);
1651e431 1820 io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
f9401285
JA
1821
1822 /*
1823 * Mark IO ok to verify
1824 */
1825 if (io_u->ipo) {
890b6656
JA
1826 /*
1827 * Remove errored entry from the verification list
1828 */
1829 if (io_u->error)
1830 unlog_io_piece(td, io_u);
1831 else {
1832 io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
1833 write_barrier();
1834 }
f9401285
JA
1835 }
1836
f8b0bd10 1837 if (ddir_sync(ddir)) {
87dc1ab1 1838 td->last_was_sync = 1;
44f29692
JA
1839 if (f) {
1840 f->first_write = -1ULL;
1841 f->last_write = -1ULL;
1842 }
87dc1ab1
JA
1843 return;
1844 }
1845
1846 td->last_was_sync = 0;
f8b0bd10 1847 td->last_ddir = ddir;
87dc1ab1 1848
f8b0bd10 1849 if (!io_u->error && ddir_rw(ddir)) {
10ba535a 1850 unsigned int bytes = io_u->buflen - io_u->resid;
b29ee5b3 1851 int ret;
10ba535a 1852
f8b0bd10
JA
1853 td->io_blocks[ddir]++;
1854 td->this_io_blocks[ddir]++;
1855 td->io_bytes[ddir] += bytes;
ae2fafc8 1856
1857 if (!(io_u->flags & IO_U_F_VER_LIST))
f8b0bd10
JA
1858 td->this_io_bytes[ddir] += bytes;
1859
639ad1ea 1860 if (ddir == DDIR_WRITE)
94a6e1bb 1861 file_log_write_comp(td, f, io_u->offset, bytes);
44f29692 1862
6b1190fd 1863 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
50a8ce86 1864 td->runstate == TD_VERIFYING))
f8b0bd10 1865 account_io_completion(td, io_u, icd, ddir, bytes);
40e1a6f0 1866
f8b0bd10 1867 icd->bytes_done[ddir] += bytes;
3af6ef39 1868
d7762cf8 1869 if (io_u->end_io) {
f8b0bd10
JA
1870 ret = io_u->end_io(td, io_u_ptr);
1871 io_u = *io_u_ptr;
3af6ef39
JA
1872 if (ret && !icd->error)
1873 icd->error = ret;
1874 }
ff58fced 1875 } else if (io_u->error) {
10ba535a 1876 icd->error = io_u->error;
5451792e
JA
1877 io_u_log_error(td, io_u);
1878 }
8b28bd41 1879 if (icd->error) {
f8b0bd10
JA
1880 enum error_type_bit eb = td_error_type(ddir, icd->error);
1881
8b28bd41
DM
1882 if (!td_non_fatal_error(td, eb, icd->error))
1883 return;
f8b0bd10 1884
f2bba182
RR
1885 /*
1886 * If there is a non_fatal error, then add to the error count
1887 * and clear all the errors.
1888 */
1889 update_error_count(td, icd->error);
1890 td_clear_error(td);
1891 icd->error = 0;
f8b0bd10
JA
1892 if (io_u)
1893 io_u->error = 0;
f2bba182 1894 }
10ba535a
JA
1895}
1896
9520ebb9
JA
1897static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1898 int nr)
10ba535a 1899{
6eaf09d6 1900 int ddir;
aba6c951
JA
1901
1902 if (!gtod_reduce(td))
9520ebb9 1903 fio_gettime(&icd->time, NULL);
02bcaa8c 1904
3af6ef39
JA
1905 icd->nr = nr;
1906
10ba535a 1907 icd->error = 0;
c1f50f76 1908 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
6eaf09d6 1909 icd->bytes_done[ddir] = 0;
36167d82
JA
1910}
1911
97601024
JA
1912static void ios_completed(struct thread_data *td,
1913 struct io_completion_data *icd)
36167d82
JA
1914{
1915 struct io_u *io_u;
1916 int i;
1917
10ba535a
JA
1918 for (i = 0; i < icd->nr; i++) {
1919 io_u = td->io_ops->event(td, i);
1920
f8b0bd10 1921 io_completed(td, &io_u, icd);
e8462bd8 1922
f8b0bd10 1923 if (io_u)
e8462bd8 1924 put_io_u(td, io_u);
10ba535a
JA
1925 }
1926}
97601024 1927
e7e6cfb4
JA
1928/*
1929 * Complete a single io_u for the sync engines.
1930 */
55312f9f 1931int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
97601024
JA
1932{
1933 struct io_completion_data icd;
55312f9f 1934 int ddir;
97601024 1935
9520ebb9 1936 init_icd(td, &icd, 1);
f8b0bd10 1937 io_completed(td, &io_u, &icd);
e8462bd8 1938
f8b0bd10 1939 if (io_u)
e8462bd8 1940 put_io_u(td, io_u);
97601024 1941
581e7141
JA
1942 if (icd.error) {
1943 td_verror(td, icd.error, "io_u_sync_complete");
1944 return -1;
1945 }
97601024 1946
c1f50f76 1947 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
55312f9f 1948 td->bytes_done[ddir] += icd.bytes_done[ddir];
581e7141
JA
1949
1950 return 0;
97601024
JA
1951}
1952
e7e6cfb4
JA
1953/*
1954 * Called to complete min_events number of io for the async engines.
1955 */
55312f9f 1956int io_u_queued_complete(struct thread_data *td, int min_evts)
97601024 1957{
97601024 1958 struct io_completion_data icd;
00de55ef 1959 struct timespec *tvp = NULL;
55312f9f 1960 int ret, ddir;
4d06a338 1961 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
97601024 1962
12bb8569 1963 dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
b271fe62 1964
4950421a 1965 if (!min_evts)
00de55ef 1966 tvp = &ts;
5fb4b366
RE
1967 else if (min_evts > td->cur_depth)
1968 min_evts = td->cur_depth;
97601024 1969
82407585
RP
1970 /* No worries, td_io_getevents fixes min and max if they are
1971 * set incorrectly */
1972 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
97601024 1973 if (ret < 0) {
e1161c32 1974 td_verror(td, -ret, "td_io_getevents");
97601024
JA
1975 return ret;
1976 } else if (!ret)
1977 return ret;
1978
9520ebb9 1979 init_icd(td, &icd, ret);
97601024 1980 ios_completed(td, &icd);
581e7141
JA
1981 if (icd.error) {
1982 td_verror(td, icd.error, "io_u_queued_complete");
1983 return -1;
1984 }
97601024 1985
c1f50f76 1986 for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
55312f9f 1987 td->bytes_done[ddir] += icd.bytes_done[ddir];
581e7141 1988
0d593542 1989 return ret;
97601024 1990}
7e77dd02
JA
1991
1992/*
1993 * Call when io_u is really queued, to update the submission latency.
1994 */
1995void io_u_queued(struct thread_data *td, struct io_u *io_u)
1996{
8243be59 1997 if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
9520ebb9 1998 unsigned long slat_time;
7e77dd02 1999
9520ebb9 2000 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
75dc383e
JA
2001
2002 if (td->parent)
2003 td = td->parent;
2004
ae588852
JA
2005 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
2006 io_u->offset);
9520ebb9 2007 }
7e77dd02 2008}
433afcb4 2009
5c94b008
JA
2010/*
2011 * See if we should reuse the last seed, if dedupe is enabled
2012 */
9451b93e 2013static struct frand_state *get_buf_state(struct thread_data *td)
5c94b008
JA
2014{
2015 unsigned int v;
5c94b008
JA
2016
2017 if (!td->o.dedupe_percentage)
2018 return &td->buf_state;
732eedd0 2019 else if (td->o.dedupe_percentage == 100) {
9451b93e
JA
2020 frand_copy(&td->buf_state_prev, &td->buf_state);
2021 return &td->buf_state;
732eedd0 2022 }
5c94b008 2023
36dd3379 2024 v = rand32_between(&td->dedupe_state, 1, 100);
5c94b008
JA
2025
2026 if (v <= td->o.dedupe_percentage)
2027 return &td->buf_state_prev;
2028
2029 return &td->buf_state;
2030}
2031
9451b93e 2032static void save_buf_state(struct thread_data *td, struct frand_state *rs)
5c94b008 2033{
9451b93e
JA
2034 if (td->o.dedupe_percentage == 100)
2035 frand_copy(rs, &td->buf_state_prev);
2036 else if (rs == &td->buf_state)
5c94b008
JA
2037 frand_copy(&td->buf_state_prev, rs);
2038}
2039
cc86c395
JA
2040void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
2041 unsigned int max_bs)
5973cafb 2042{
d1af2894
JA
2043 struct thread_options *o = &td->o;
2044
4eff3e57 2045 if (o->compress_percentage || o->dedupe_percentage) {
9c42684e 2046 unsigned int perc = td->o.compress_percentage;
5c94b008 2047 struct frand_state *rs;
1066358a 2048 unsigned int left = max_bs;
1e7f82e2 2049 unsigned int this_write;
5c94b008 2050
1066358a 2051 do {
9451b93e 2052 rs = get_buf_state(td);
9c42684e 2053
1066358a 2054 min_write = min(min_write, left);
f97a43a1 2055
1066358a 2056 if (perc) {
4b157ac6
JA
2057 this_write = min_not_zero(min_write,
2058 td->o.compress_chunk);
1e7f82e2
JA
2059
2060 fill_random_buf_percentage(rs, buf, perc,
2061 this_write, this_write,
2062 o->buffer_pattern,
2063 o->buffer_pattern_bytes);
2064 } else {
1066358a 2065 fill_random_buf(rs, buf, min_write);
1e7f82e2
JA
2066 this_write = min_write;
2067 }
1066358a 2068
1e7f82e2
JA
2069 buf += this_write;
2070 left -= this_write;
9451b93e 2071 save_buf_state(td, rs);
1066358a 2072 } while (left);
d1af2894
JA
2073 } else if (o->buffer_pattern_bytes)
2074 fill_buffer_pattern(td, buf, max_bs);
999d245e 2075 else if (o->zero_buffers)
cc86c395 2076 memset(buf, 0, max_bs);
999d245e 2077 else
9451b93e 2078 fill_random_buf(get_buf_state(td), buf, max_bs);
cc86c395
JA
2079}
2080
2081/*
2082 * "randomly" fill the buffer contents
2083 */
2084void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
2085 unsigned int min_write, unsigned int max_bs)
2086{
2087 io_u->buf_filled_len = 0;
2088 fill_io_buffer(td, io_u->buf, min_write, max_bs);
5973cafb 2089}
e2c75fc4
TK
2090
2091static int do_sync_file_range(const struct thread_data *td,
2092 struct fio_file *f)
2093{
2094 off64_t offset, nbytes;
2095
2096 offset = f->first_write;
2097 nbytes = f->last_write - f->first_write;
2098
2099 if (!nbytes)
2100 return 0;
2101
2102 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
2103}
2104
2105int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
2106{
2107 int ret;
2108
2109 if (io_u->ddir == DDIR_SYNC) {
2110 ret = fsync(io_u->file->fd);
2111 } else if (io_u->ddir == DDIR_DATASYNC) {
2112#ifdef CONFIG_FDATASYNC
2113 ret = fdatasync(io_u->file->fd);
2114#else
2115 ret = io_u->xfer_buflen;
2116 io_u->error = EINVAL;
2117#endif
2118 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
2119 ret = do_sync_file_range(td, io_u->file);
2120 else {
2121 ret = io_u->xfer_buflen;
2122 io_u->error = EINVAL;
2123 }
2124
2125 if (ret < 0)
2126 io_u->error = errno;
2127
2128 return ret;
2129}
2130
2131int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
2132{
2133#ifndef FIO_HAVE_TRIM
2134 io_u->error = EINVAL;
2135 return 0;
2136#else
2137 struct fio_file *f = io_u->file;
2138 int ret;
2139
2140 ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
2141 if (!ret)
2142 return io_u->xfer_buflen;
2143
2144 io_u->error = ret;
2145 return 0;
2146#endif
2147}