Add pending IO to the tail of the busy list
[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"
10ba535a 13
97601024
JA
14struct io_completion_data {
15 int nr; /* input */
97601024
JA
16
17 int error; /* output */
6eaf09d6 18 unsigned long bytes_done[DDIR_RWDIR_CNT]; /* output */
97601024
JA
19 struct timeval time; /* output */
20};
21
10ba535a
JA
22/*
23 * The ->file_map[] contains a map of blocks we have or have not done io
24 * to yet. Used to make sure we cover the entire range in a fair fashion.
25 */
aec2de20 26static int random_map_free(struct fio_file *f, const unsigned long long block)
10ba535a 27{
aec2de20
JA
28 unsigned int idx = RAND_MAP_IDX(f, block);
29 unsigned int bit = RAND_MAP_BIT(f, block);
10ba535a 30
84422acd
JA
31 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
32
0ce8b119 33 return (f->file_map[idx] & (1UL << bit)) == 0;
10ba535a
JA
34}
35
df415585
JA
36/*
37 * Mark a given offset as used in the map.
38 */
9bf2061e 39static void mark_random_map(struct thread_data *td, struct io_u *io_u)
df415585 40{
2dc1bbeb 41 unsigned int min_bs = td->o.rw_min_bs;
9bf2061e 42 struct fio_file *f = io_u->file;
a00735e6 43 unsigned long long block;
3e3357b1 44 unsigned int blocks, nr_blocks;
38dad62d 45 int busy_check;
df415585 46
b9c5b644 47 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
c685b5b2 48 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
3e3357b1 49 blocks = 0;
38dad62d 50 busy_check = !(io_u->flags & IO_U_F_BUSY_OK);
c685b5b2 51
3e3357b1 52 while (nr_blocks) {
df415585 53 unsigned int idx, bit;
0ce8b119 54 unsigned long mask, this_blocks;
df415585 55
1e3d53ac
JA
56 /*
57 * If we have a mixed random workload, we may
58 * encounter blocks we already did IO to.
59 */
38dad62d
JA
60 if (!busy_check) {
61 blocks = nr_blocks;
62 break;
63 }
5736c10d 64 if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block))
df415585
JA
65 break;
66
aec2de20
JA
67 idx = RAND_MAP_IDX(f, block);
68 bit = RAND_MAP_BIT(f, block);
df415585 69
0032bf9f 70 fio_assert(td, idx < f->num_maps);
df415585 71
3e3357b1
JA
72 this_blocks = nr_blocks;
73 if (this_blocks + bit > BLOCKS_PER_MAP)
74 this_blocks = BLOCKS_PER_MAP - bit;
75
da3758bf
JA
76 do {
77 if (this_blocks == BLOCKS_PER_MAP)
0ce8b119 78 mask = -1UL;
da3758bf 79 else
0ce8b119 80 mask = ((1UL << this_blocks) - 1) << bit;
da3758bf
JA
81
82 if (!(f->file_map[idx] & mask))
83 break;
84
85 this_blocks--;
86 } while (this_blocks);
87
88 if (!this_blocks)
89 break;
3e3357b1 90
3e3357b1
JA
91 f->file_map[idx] |= mask;
92 nr_blocks -= this_blocks;
93 blocks += this_blocks;
c9dd34b2 94 block += this_blocks;
df415585
JA
95 }
96
a00735e6
JA
97 if ((blocks * min_bs) < io_u->buflen)
98 io_u->buflen = blocks * min_bs;
df415585
JA
99}
100
3e3357b1
JA
101static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
102 enum fio_ddir ddir)
2ba1c290
JA
103{
104 unsigned long long max_blocks;
d9dd70f7 105 unsigned long long max_size;
2ba1c290 106
ff58fced
JA
107 assert(ddir_rw(ddir));
108
d9dd70f7
JA
109 /*
110 * Hmm, should we make sure that ->io_size <= ->real_file_size?
111 */
112 max_size = f->io_size;
113 if (max_size > f->real_file_size)
114 max_size = f->real_file_size;
115
ed335855
SN
116 if (td->o.zone_range)
117 max_size = td->o.zone_range;
118
2b7a01d0 119 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
2ba1c290
JA
120 if (!max_blocks)
121 return 0;
122
67778e88 123 return max_blocks;
2ba1c290
JA
124}
125
10ba535a
JA
126/*
127 * Return the next free block in the map.
128 */
129static int get_next_free_block(struct thread_data *td, struct fio_file *f,
4ba66134 130 enum fio_ddir ddir, unsigned long long *b)
10ba535a 131{
0ce8b119 132 unsigned long long block, min_bs = td->o.rw_min_bs, lastb;
10ba535a
JA
133 int i;
134
5e0baa7f
JA
135 lastb = last_block(td, f, ddir);
136 if (!lastb)
137 return 1;
138
c685b5b2 139 i = f->last_free_lookup;
0ce8b119
JA
140 block = i * BLOCKS_PER_MAP;
141 while (block * min_bs < f->real_file_size &&
142 block * min_bs < f->io_size) {
143 if (f->file_map[i] != -1UL) {
144 block += ffz(f->file_map[i]);
145 if (block > lastb)
2ba1c290 146 break;
c685b5b2 147 f->last_free_lookup = i;
0ce8b119 148 *b = block;
10ba535a
JA
149 return 0;
150 }
151
0ce8b119 152 block += BLOCKS_PER_MAP;
10ba535a
JA
153 i++;
154 }
155
2ba1c290 156 dprint(FD_IO, "failed finding a free block\n");
10ba535a
JA
157 return 1;
158}
159
ec4015da 160static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
4ba66134 161 enum fio_ddir ddir, unsigned long long *b)
ec4015da 162{
15b87723 163 unsigned long long rmax, r, lastb;
ec4015da
JA
164 int loops = 5;
165
5e0baa7f
JA
166 lastb = last_block(td, f, ddir);
167 if (!lastb)
168 return 1;
169
0ce8b119
JA
170 if (f->failed_rands >= 200)
171 goto ffz;
172
15b87723 173 rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
ec4015da 174 do {
15b87723 175 if (td->o.use_os_rand)
2615cc4b 176 r = os_random_long(&td->random_state);
15b87723 177 else
2615cc4b 178 r = __rand(&td->__random_state);
15b87723
JA
179
180 *b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0));
2615cc4b 181
84422acd 182 dprint(FD_RANDOM, "off rand %llu\n", r);
2615cc4b 183
2ba1c290 184
43c63a78
JA
185 /*
186 * if we are not maintaining a random map, we are done.
187 */
303032ae 188 if (!file_randommap(td, f))
0ce8b119 189 goto ret_good;
43c63a78
JA
190
191 /*
84422acd 192 * calculate map offset and check if it's free
43c63a78 193 */
aec2de20 194 if (random_map_free(f, *b))
0ce8b119 195 goto ret_good;
43c63a78 196
84422acd
JA
197 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
198 *b);
43c63a78 199 } while (--loops);
ec4015da 200
0ce8b119
JA
201 if (!f->failed_rands++)
202 f->last_free_lookup = 0;
203
ec4015da 204 /*
43c63a78
JA
205 * we get here, if we didn't suceed in looking up a block. generate
206 * a random start offset into the filemap, and find the first free
207 * block from there.
ec4015da 208 */
43c63a78
JA
209 loops = 10;
210 do {
dc873b6f 211 f->last_free_lookup = (f->num_maps - 1) *
15b87723 212 (r / ((unsigned long long) rmax + 1.0));
4ba66134 213 if (!get_next_free_block(td, f, ddir, b))
0ce8b119 214 goto ret;
ec4015da 215
d1cea521
JA
216 if (td->o.use_os_rand)
217 r = os_random_long(&td->random_state);
218 else
219 r = __rand(&td->__random_state);
43c63a78
JA
220 } while (--loops);
221
222 /*
223 * that didn't work either, try exhaustive search from the start
224 */
225 f->last_free_lookup = 0;
0ce8b119
JA
226ffz:
227 if (!get_next_free_block(td, f, ddir, b))
228 return 0;
229 f->last_free_lookup = 0;
4ba66134 230 return get_next_free_block(td, f, ddir, b);
0ce8b119
JA
231ret_good:
232 f->failed_rands = 0;
233ret:
234 return 0;
ec4015da
JA
235}
236
38dad62d
JA
237static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
238 enum fio_ddir ddir, unsigned long long *b)
239{
c04e4661
DE
240 if (!get_next_rand_offset(td, f, ddir, b))
241 return 0;
242
243 if (td->o.time_based) {
244 fio_file_reset(f);
245 if (!get_next_rand_offset(td, f, ddir, b))
246 return 0;
38dad62d
JA
247 }
248
c04e4661
DE
249 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
250 f->file_name, f->last_pos, f->real_file_size);
251 return 1;
38dad62d
JA
252}
253
37cf9e3c
JA
254static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
255 enum fio_ddir ddir, unsigned long long *offset)
38dad62d 256{
ff58fced
JA
257 assert(ddir_rw(ddir));
258
ce95d651 259 if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
c04e4661
DE
260 f->last_pos = f->last_pos - f->io_size;
261
38dad62d 262 if (f->last_pos < f->real_file_size) {
a66da7a2 263 unsigned long long pos;
059b0802 264
a66da7a2
JA
265 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
266 f->last_pos = f->real_file_size;
267
268 pos = f->last_pos - f->file_offset;
059b0802
JA
269 if (pos)
270 pos += td->o.ddir_seq_add;
271
37cf9e3c 272 *offset = pos;
38dad62d
JA
273 return 0;
274 }
275
276 return 1;
277}
278
279static int get_next_block(struct thread_data *td, struct io_u *io_u,
37cf9e3c 280 enum fio_ddir ddir, int rw_seq)
38dad62d
JA
281{
282 struct fio_file *f = io_u->file;
37cf9e3c 283 unsigned long long b, offset;
38dad62d
JA
284 int ret;
285
ff58fced
JA
286 assert(ddir_rw(ddir));
287
37cf9e3c
JA
288 b = offset = -1ULL;
289
38dad62d
JA
290 if (rw_seq) {
291 if (td_random(td))
37cf9e3c 292 ret = get_next_rand_block(td, f, ddir, &b);
38dad62d 293 else
37cf9e3c 294 ret = get_next_seq_offset(td, f, ddir, &offset);
38dad62d
JA
295 } else {
296 io_u->flags |= IO_U_F_BUSY_OK;
297
298 if (td->o.rw_seq == RW_SEQ_SEQ) {
37cf9e3c 299 ret = get_next_seq_offset(td, f, ddir, &offset);
38dad62d 300 if (ret)
37cf9e3c 301 ret = get_next_rand_block(td, f, ddir, &b);
38dad62d
JA
302 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
303 if (f->last_start != -1ULL)
37cf9e3c 304 offset = f->last_start - f->file_offset;
38dad62d 305 else
37cf9e3c 306 offset = 0;
38dad62d
JA
307 ret = 0;
308 } else {
309 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
310 ret = 1;
311 }
312 }
313
37cf9e3c
JA
314 if (!ret) {
315 if (offset != -1ULL)
316 io_u->offset = offset;
317 else if (b != -1ULL)
318 io_u->offset = b * td->o.ba[ddir];
319 else {
320 log_err("fio: bug in offset generation\n");
321 ret = 1;
322 }
323 }
324
38dad62d
JA
325 return ret;
326}
327
10ba535a
JA
328/*
329 * For random io, generate a random new block and see if it's used. Repeat
330 * until we find a free one. For sequential io, just return the end of
331 * the last io issued.
332 */
15dc1934 333static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
10ba535a 334{
9bf2061e 335 struct fio_file *f = io_u->file;
4ba66134 336 enum fio_ddir ddir = io_u->ddir;
38dad62d 337 int rw_seq_hit = 0;
10ba535a 338
ff58fced
JA
339 assert(ddir_rw(ddir));
340
38dad62d
JA
341 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
342 rw_seq_hit = 1;
5736c10d 343 td->ddir_seq_nr = td->o.ddir_seq_nr;
38dad62d 344 }
211097b2 345
37cf9e3c 346 if (get_next_block(td, io_u, ddir, rw_seq_hit))
38dad62d 347 return 1;
10ba535a 348
009bd847
JA
349 if (io_u->offset >= f->io_size) {
350 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
351 io_u->offset, f->io_size);
352 return 1;
353 }
354
355 io_u->offset += f->file_offset;
2ba1c290
JA
356 if (io_u->offset >= f->real_file_size) {
357 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
358 io_u->offset, f->real_file_size);
10ba535a 359 return 1;
2ba1c290 360 }
10ba535a
JA
361
362 return 0;
363}
364
15dc1934
JA
365static int get_next_offset(struct thread_data *td, struct io_u *io_u)
366{
7eb36574
JA
367 struct prof_io_ops *ops = &td->prof_io_ops;
368
369 if (ops->fill_io_u_off)
370 return ops->fill_io_u_off(td, io_u);
15dc1934
JA
371
372 return __get_next_offset(td, io_u);
373}
374
79944128
JA
375static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
376 unsigned int buflen)
377{
378 struct fio_file *f = io_u->file;
379
ce95d651 380 return io_u->offset + buflen <= f->io_size + get_start_offset(td);
79944128
JA
381}
382
15dc1934 383static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
10ba535a 384{
bca4ed4d 385 const int ddir = io_u->ddir;
f3f552b9 386 unsigned int uninitialized_var(buflen);
f3059de1 387 unsigned int minbs, maxbs;
1294c3ec 388 unsigned long r, rand_max;
10ba535a 389
ff58fced
JA
390 assert(ddir_rw(ddir));
391
f3059de1
JA
392 minbs = td->o.min_bs[ddir];
393 maxbs = td->o.max_bs[ddir];
394
79944128
JA
395 if (minbs == maxbs)
396 return minbs;
397
52c58027
JA
398 /*
399 * If we can't satisfy the min block size from here, then fail
400 */
401 if (!io_u_fits(td, io_u, minbs))
402 return 0;
403
4c07ad86
JA
404 if (td->o.use_os_rand)
405 rand_max = OS_RAND_MAX;
406 else
407 rand_max = FRAND_MAX;
408
79944128 409 do {
4c07ad86
JA
410 if (td->o.use_os_rand)
411 r = os_random_long(&td->bsrange_state);
412 else
413 r = __rand(&td->__bsrange_state);
414
720e84ad 415 if (!td->o.bssplit_nr[ddir]) {
f3059de1 416 buflen = 1 + (unsigned int) ((double) maxbs *
4c07ad86 417 (r / (rand_max + 1.0)));
f3059de1
JA
418 if (buflen < minbs)
419 buflen = minbs;
5ec10eaa 420 } else {
564ca972
JA
421 long perc = 0;
422 unsigned int i;
423
720e84ad
JA
424 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
425 struct bssplit *bsp = &td->o.bssplit[ddir][i];
564ca972
JA
426
427 buflen = bsp->bs;
428 perc += bsp->perc;
79944128
JA
429 if ((r <= ((rand_max / 100L) * perc)) &&
430 io_u_fits(td, io_u, buflen))
564ca972
JA
431 break;
432 }
433 }
79944128 434
f3059de1
JA
435 if (!td->o.bs_unaligned && is_power_of_2(minbs))
436 buflen = (buflen + minbs - 1) & ~(minbs - 1);
10ba535a 437
79944128 438 } while (!io_u_fits(td, io_u, buflen));
6a5e6884 439
10ba535a
JA
440 return buflen;
441}
442
15dc1934
JA
443static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
444{
7eb36574
JA
445 struct prof_io_ops *ops = &td->prof_io_ops;
446
447 if (ops->fill_io_u_size)
448 return ops->fill_io_u_size(td, io_u);
15dc1934
JA
449
450 return __get_next_buflen(td, io_u);
451}
452
afe24a5a
JA
453static void set_rwmix_bytes(struct thread_data *td)
454{
afe24a5a
JA
455 unsigned int diff;
456
457 /*
458 * we do time or byte based switch. this is needed because
459 * buffered writes may issue a lot quicker than they complete,
460 * whereas reads do not.
461 */
e47f799f 462 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
04c540d9 463 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
e47f799f
JA
464}
465
466static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
467{
468 unsigned int v;
1294c3ec 469 unsigned long r;
e47f799f 470
4c07ad86
JA
471 if (td->o.use_os_rand) {
472 r = os_random_long(&td->rwmix_state);
473 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
474 } else {
475 r = __rand(&td->__rwmix_state);
476 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
477 }
478
04c540d9 479 if (v <= td->o.rwmix[DDIR_READ])
e47f799f
JA
480 return DDIR_READ;
481
482 return DDIR_WRITE;
afe24a5a
JA
483}
484
581e7141
JA
485static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
486{
487 enum fio_ddir odir = ddir ^ 1;
488 struct timeval t;
489 long usec;
490
ff58fced
JA
491 assert(ddir_rw(ddir));
492
581e7141
JA
493 if (td->rate_pending_usleep[ddir] <= 0)
494 return ddir;
495
496 /*
497 * We have too much pending sleep in this direction. See if we
498 * should switch.
499 */
500 if (td_rw(td)) {
501 /*
502 * Other direction does not have too much pending, switch
503 */
504 if (td->rate_pending_usleep[odir] < 100000)
505 return odir;
506
507 /*
508 * Both directions have pending sleep. Sleep the minimum time
509 * and deduct from both.
510 */
511 if (td->rate_pending_usleep[ddir] <=
512 td->rate_pending_usleep[odir]) {
513 usec = td->rate_pending_usleep[ddir];
514 } else {
515 usec = td->rate_pending_usleep[odir];
516 ddir = odir;
517 }
518 } else
519 usec = td->rate_pending_usleep[ddir];
520
78c1eda5
JA
521 /*
522 * We are going to sleep, ensure that we flush anything pending as
422f9e4b
RM
523 * not to skew our latency numbers.
524 *
525 * Changed to only monitor 'in flight' requests here instead of the
526 * td->cur_depth, b/c td->cur_depth does not accurately represent
527 * io's that have been actually submitted to an async engine,
528 * and cur_depth is meaningless for sync engines.
78c1eda5 529 */
422f9e4b 530 if (td->io_u_in_flight) {
78c1eda5
JA
531 int fio_unused ret;
532
422f9e4b 533 ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
78c1eda5
JA
534 }
535
581e7141
JA
536 fio_gettime(&t, NULL);
537 usec_sleep(td, usec);
538 usec = utime_since_now(&t);
539
540 td->rate_pending_usleep[ddir] -= usec;
541
542 odir = ddir ^ 1;
543 if (td_rw(td) && __should_check_rate(td, odir))
544 td->rate_pending_usleep[odir] -= usec;
0b9d69ec 545
6eaf09d6
SL
546 if (ddir_trim(ddir))
547 return ddir;
581e7141
JA
548 return ddir;
549}
550
10ba535a
JA
551/*
552 * Return the data direction for the next io_u. If the job is a
553 * mixed read/write workload, check the rwmix cycle and switch if
554 * necessary.
555 */
1e97cce9 556static enum fio_ddir get_rw_ddir(struct thread_data *td)
10ba535a 557{
581e7141
JA
558 enum fio_ddir ddir;
559
5f9099ea
JA
560 /*
561 * see if it's time to fsync
562 */
563 if (td->o.fsync_blocks &&
564 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
565 td->io_issues[DDIR_WRITE] && should_fsync(td))
566 return DDIR_SYNC;
567
568 /*
569 * see if it's time to fdatasync
570 */
571 if (td->o.fdatasync_blocks &&
572 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
573 td->io_issues[DDIR_WRITE] && should_fsync(td))
574 return DDIR_DATASYNC;
575
44f29692
JA
576 /*
577 * see if it's time to sync_file_range
578 */
579 if (td->sync_file_range_nr &&
580 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
581 td->io_issues[DDIR_WRITE] && should_fsync(td))
582 return DDIR_SYNC_FILE_RANGE;
583
10ba535a 584 if (td_rw(td)) {
10ba535a
JA
585 /*
586 * Check if it's time to seed a new data direction.
587 */
e4928662 588 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
e47f799f
JA
589 /*
590 * Put a top limit on how many bytes we do for
591 * one data direction, to avoid overflowing the
592 * ranges too much
593 */
594 ddir = get_rand_ddir(td);
e47f799f
JA
595
596 if (ddir != td->rwmix_ddir)
597 set_rwmix_bytes(td);
598
599 td->rwmix_ddir = ddir;
10ba535a 600 }
581e7141 601 ddir = td->rwmix_ddir;
10ba535a 602 } else if (td_read(td))
581e7141 603 ddir = DDIR_READ;
6eaf09d6 604 else if (td_write(td))
581e7141 605 ddir = DDIR_WRITE;
6eaf09d6
SL
606 else
607 ddir = DDIR_TRIM;
581e7141
JA
608
609 td->rwmix_ddir = rate_ddir(td, ddir);
610 return td->rwmix_ddir;
10ba535a
JA
611}
612
1ef2b6be
JA
613static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
614{
615 io_u->ddir = get_rw_ddir(td);
616
617 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
618 td->o.barrier_blocks &&
619 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
620 td->io_issues[DDIR_WRITE])
621 io_u->flags |= IO_U_F_BARRIER;
622}
623
e8462bd8 624void put_file_log(struct thread_data *td, struct fio_file *f)
60f2c658
JA
625{
626 int ret = put_file(td, f);
627
628 if (ret)
629 td_verror(td, ret, "file close");
630}
631
10ba535a
JA
632void put_io_u(struct thread_data *td, struct io_u *io_u)
633{
e8462bd8
JA
634 td_io_u_lock(td);
635
d7ee2a7d 636 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
60f2c658 637 put_file_log(td, io_u->file);
10ba535a 638 io_u->file = NULL;
d7ee2a7d
SL
639 io_u->flags &= ~IO_U_F_FREE_DEF;
640 io_u->flags |= IO_U_F_FREE;
641
0c41214f
RR
642 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
643 td->cur_depth--;
e8462bd8 644 flist_del_init(&io_u->list);
01743ee1 645 flist_add(&io_u->list, &td->io_u_freelist);
e8462bd8
JA
646 td_io_u_unlock(td);
647 td_io_u_free_notify(td);
10ba535a
JA
648}
649
f2bba182
RR
650void clear_io_u(struct thread_data *td, struct io_u *io_u)
651{
652 io_u->flags &= ~IO_U_F_FLIGHT;
653 put_io_u(td, io_u);
654}
655
755200a3
JA
656void requeue_io_u(struct thread_data *td, struct io_u **io_u)
657{
658 struct io_u *__io_u = *io_u;
659
465221b0
JA
660 dprint(FD_IO, "requeue %p\n", __io_u);
661
e8462bd8
JA
662 td_io_u_lock(td);
663
4d2e0f49 664 __io_u->flags |= IO_U_F_FREE;
ff58fced 665 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
e4f54adb 666 td->io_issues[__io_u->ddir]--;
5ec10eaa 667
4d2e0f49 668 __io_u->flags &= ~IO_U_F_FLIGHT;
0c41214f
RR
669 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
670 td->cur_depth--;
01743ee1
JA
671 flist_del(&__io_u->list);
672 flist_add_tail(&__io_u->list, &td->io_u_requeues);
e8462bd8 673 td_io_u_unlock(td);
755200a3
JA
674 *io_u = NULL;
675}
676
9bf2061e 677static int fill_io_u(struct thread_data *td, struct io_u *io_u)
10ba535a 678{
b4c5e1ac
JA
679 if (td->io_ops->flags & FIO_NOIO)
680 goto out;
681
1ef2b6be 682 set_rw_ddir(td, io_u);
5f9099ea 683
87dc1ab1 684 /*
ff58fced 685 * fsync() or fdatasync() or trim etc, we are done
87dc1ab1 686 */
ff58fced 687 if (!ddir_rw(io_u->ddir))
c38e9468 688 goto out;
a00735e6 689
48f5abd3
JA
690 /*
691 * See if it's time to switch to a new zone
692 */
13af05ae 693 if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
48f5abd3 694 td->zone_bytes = 0;
ed335855
SN
695 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
696 io_u->file->last_pos = io_u->file->file_offset;
48f5abd3
JA
697 td->io_skip_bytes += td->o.zone_skip;
698 }
699
10ba535a 700 /*
c685b5b2
JA
701 * No log, let the seq/rand engine retrieve the next buflen and
702 * position.
10ba535a 703 */
2ba1c290
JA
704 if (get_next_offset(td, io_u)) {
705 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
bca4ed4d 706 return 1;
2ba1c290 707 }
10ba535a 708
9bf2061e 709 io_u->buflen = get_next_buflen(td, io_u);
2ba1c290
JA
710 if (!io_u->buflen) {
711 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
bca4ed4d 712 return 1;
2ba1c290 713 }
bca4ed4d 714
2ba1c290
JA
715 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
716 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
4ba66134
JA
717 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
718 io_u->buflen, io_u->file->real_file_size);
6a5e6884 719 return 1;
2ba1c290 720 }
6a5e6884 721
bca4ed4d
JA
722 /*
723 * mark entry before potentially trimming io_u
724 */
303032ae 725 if (td_random(td) && file_randommap(td, io_u->file))
9bf2061e 726 mark_random_map(td, io_u);
bca4ed4d
JA
727
728 /*
729 * If using a write iolog, store this entry.
730 */
c38e9468 731out:
2ba1c290 732 dprint_io_u(io_u, "fill_io_u");
d9d91e39 733 td->zone_bytes += io_u->buflen;
f29b25a3 734 log_io_u(td, io_u);
bca4ed4d 735 return 0;
10ba535a
JA
736}
737
838bc709
JA
738static void __io_u_mark_map(unsigned int *map, unsigned int nr)
739{
2b13e716 740 int idx = 0;
838bc709
JA
741
742 switch (nr) {
743 default:
2b13e716 744 idx = 6;
838bc709
JA
745 break;
746 case 33 ... 64:
2b13e716 747 idx = 5;
838bc709
JA
748 break;
749 case 17 ... 32:
2b13e716 750 idx = 4;
838bc709
JA
751 break;
752 case 9 ... 16:
2b13e716 753 idx = 3;
838bc709
JA
754 break;
755 case 5 ... 8:
2b13e716 756 idx = 2;
838bc709
JA
757 break;
758 case 1 ... 4:
2b13e716 759 idx = 1;
838bc709
JA
760 case 0:
761 break;
762 }
763
2b13e716 764 map[idx]++;
838bc709
JA
765}
766
767void io_u_mark_submit(struct thread_data *td, unsigned int nr)
768{
769 __io_u_mark_map(td->ts.io_u_submit, nr);
770 td->ts.total_submit++;
771}
772
773void io_u_mark_complete(struct thread_data *td, unsigned int nr)
774{
775 __io_u_mark_map(td->ts.io_u_complete, nr);
776 td->ts.total_complete++;
777}
778
d8005759 779void io_u_mark_depth(struct thread_data *td, unsigned int nr)
71619dc2 780{
2b13e716 781 int idx = 0;
71619dc2
JA
782
783 switch (td->cur_depth) {
784 default:
2b13e716 785 idx = 6;
a783e61a 786 break;
71619dc2 787 case 32 ... 63:
2b13e716 788 idx = 5;
a783e61a 789 break;
71619dc2 790 case 16 ... 31:
2b13e716 791 idx = 4;
a783e61a 792 break;
71619dc2 793 case 8 ... 15:
2b13e716 794 idx = 3;
a783e61a 795 break;
71619dc2 796 case 4 ... 7:
2b13e716 797 idx = 2;
a783e61a 798 break;
71619dc2 799 case 2 ... 3:
2b13e716 800 idx = 1;
71619dc2
JA
801 case 1:
802 break;
803 }
804
2b13e716 805 td->ts.io_u_map[idx] += nr;
71619dc2
JA
806}
807
04a0feae
JA
808static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
809{
2b13e716 810 int idx = 0;
04a0feae
JA
811
812 assert(usec < 1000);
813
814 switch (usec) {
815 case 750 ... 999:
2b13e716 816 idx = 9;
04a0feae
JA
817 break;
818 case 500 ... 749:
2b13e716 819 idx = 8;
04a0feae
JA
820 break;
821 case 250 ... 499:
2b13e716 822 idx = 7;
04a0feae
JA
823 break;
824 case 100 ... 249:
2b13e716 825 idx = 6;
04a0feae
JA
826 break;
827 case 50 ... 99:
2b13e716 828 idx = 5;
04a0feae
JA
829 break;
830 case 20 ... 49:
2b13e716 831 idx = 4;
04a0feae
JA
832 break;
833 case 10 ... 19:
2b13e716 834 idx = 3;
04a0feae
JA
835 break;
836 case 4 ... 9:
2b13e716 837 idx = 2;
04a0feae
JA
838 break;
839 case 2 ... 3:
2b13e716 840 idx = 1;
04a0feae
JA
841 case 0 ... 1:
842 break;
843 }
844
2b13e716
JA
845 assert(idx < FIO_IO_U_LAT_U_NR);
846 td->ts.io_u_lat_u[idx]++;
04a0feae
JA
847}
848
849static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
ec118304 850{
2b13e716 851 int idx = 0;
ec118304
JA
852
853 switch (msec) {
854 default:
2b13e716 855 idx = 11;
04a0feae 856 break;
8abdce66 857 case 1000 ... 1999:
2b13e716 858 idx = 10;
04a0feae 859 break;
8abdce66 860 case 750 ... 999:
2b13e716 861 idx = 9;
04a0feae 862 break;
8abdce66 863 case 500 ... 749:
2b13e716 864 idx = 8;
04a0feae 865 break;
8abdce66 866 case 250 ... 499:
2b13e716 867 idx = 7;
04a0feae 868 break;
8abdce66 869 case 100 ... 249:
2b13e716 870 idx = 6;
04a0feae 871 break;
8abdce66 872 case 50 ... 99:
2b13e716 873 idx = 5;
04a0feae 874 break;
8abdce66 875 case 20 ... 49:
2b13e716 876 idx = 4;
04a0feae 877 break;
8abdce66 878 case 10 ... 19:
2b13e716 879 idx = 3;
04a0feae 880 break;
8abdce66 881 case 4 ... 9:
2b13e716 882 idx = 2;
04a0feae 883 break;
ec118304 884 case 2 ... 3:
2b13e716 885 idx = 1;
ec118304
JA
886 case 0 ... 1:
887 break;
888 }
889
2b13e716
JA
890 assert(idx < FIO_IO_U_LAT_M_NR);
891 td->ts.io_u_lat_m[idx]++;
04a0feae
JA
892}
893
894static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
895{
896 if (usec < 1000)
897 io_u_mark_lat_usec(td, usec);
898 else
899 io_u_mark_lat_msec(td, usec / 1000);
ec118304
JA
900}
901
0aabe160
JA
902/*
903 * Get next file to service by choosing one at random
904 */
2cc52930
JA
905static struct fio_file *get_next_file_rand(struct thread_data *td,
906 enum fio_file_flags goodf,
d6aed795 907 enum fio_file_flags badf)
0aabe160 908{
0aabe160 909 struct fio_file *f;
1c178180 910 int fno;
0aabe160
JA
911
912 do {
87b10676 913 int opened = 0;
1294c3ec 914 unsigned long r;
4c07ad86
JA
915
916 if (td->o.use_os_rand) {
917 r = os_random_long(&td->next_file_state);
918 fno = (unsigned int) ((double) td->o.nr_files
919 * (r / (OS_RAND_MAX + 1.0)));
920 } else {
921 r = __rand(&td->__next_file_state);
922 fno = (unsigned int) ((double) td->o.nr_files
923 * (r / (FRAND_MAX + 1.0)));
924 }
7c83c089 925
126d65c6 926 f = td->files[fno];
d6aed795 927 if (fio_file_done(f))
059e63c0 928 continue;
1c178180 929
d6aed795 930 if (!fio_file_open(f)) {
87b10676
JA
931 int err;
932
933 err = td_io_open_file(td, f);
934 if (err)
935 continue;
936 opened = 1;
937 }
938
2ba1c290
JA
939 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
940 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
0aabe160 941 return f;
2ba1c290 942 }
87b10676
JA
943 if (opened)
944 td_io_close_file(td, f);
0aabe160
JA
945 } while (1);
946}
947
948/*
949 * Get next file to service by doing round robin between all available ones
950 */
1c178180
JA
951static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
952 int badf)
3d7c391d
JA
953{
954 unsigned int old_next_file = td->next_file;
955 struct fio_file *f;
956
957 do {
87b10676
JA
958 int opened = 0;
959
126d65c6 960 f = td->files[td->next_file];
3d7c391d
JA
961
962 td->next_file++;
2dc1bbeb 963 if (td->next_file >= td->o.nr_files)
3d7c391d
JA
964 td->next_file = 0;
965
87b10676 966 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
d6aed795 967 if (fio_file_done(f)) {
d5ed68ea 968 f = NULL;
059e63c0 969 continue;
d5ed68ea 970 }
059e63c0 971
d6aed795 972 if (!fio_file_open(f)) {
87b10676
JA
973 int err;
974
975 err = td_io_open_file(td, f);
b5696bfc
JA
976 if (err) {
977 dprint(FD_FILE, "error %d on open of %s\n",
978 err, f->file_name);
87c27b45 979 f = NULL;
87b10676 980 continue;
b5696bfc 981 }
87b10676
JA
982 opened = 1;
983 }
984
0b9d69ec
JA
985 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
986 f->flags);
1c178180 987 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
3d7c391d
JA
988 break;
989
87b10676
JA
990 if (opened)
991 td_io_close_file(td, f);
992
3d7c391d
JA
993 f = NULL;
994 } while (td->next_file != old_next_file);
995
2ba1c290 996 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
3d7c391d
JA
997 return f;
998}
999
7eb36574 1000static struct fio_file *__get_next_file(struct thread_data *td)
bdb4e2e9 1001{
1907dbc6
JA
1002 struct fio_file *f;
1003
2dc1bbeb 1004 assert(td->o.nr_files <= td->files_index);
1c178180 1005
b5696bfc 1006 if (td->nr_done_files >= td->o.nr_files) {
5ec10eaa
JA
1007 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
1008 " nr_files=%d\n", td->nr_open_files,
1009 td->nr_done_files,
1010 td->o.nr_files);
bdb4e2e9 1011 return NULL;
2ba1c290 1012 }
bdb4e2e9 1013
1907dbc6 1014 f = td->file_service_file;
d6aed795 1015 if (f && fio_file_open(f) && !fio_file_closing(f)) {
a086c257
JA
1016 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
1017 goto out;
1018 if (td->file_service_left--)
1019 goto out;
1020 }
1907dbc6 1021
a086c257
JA
1022 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1023 td->o.file_service_type == FIO_FSERVICE_SEQ)
d6aed795 1024 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
bdb4e2e9 1025 else
d6aed795 1026 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1907dbc6
JA
1027
1028 td->file_service_file = f;
1029 td->file_service_left = td->file_service_nr - 1;
2ba1c290 1030out:
683023e8 1031 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1907dbc6 1032 return f;
bdb4e2e9
JA
1033}
1034
7eb36574
JA
1035static struct fio_file *get_next_file(struct thread_data *td)
1036{
1037 struct prof_io_ops *ops = &td->prof_io_ops;
1038
1039 if (ops->get_next_file)
1040 return ops->get_next_file(td);
1041
1042 return __get_next_file(td);
1043}
1044
429f6675
JA
1045static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1046{
1047 struct fio_file *f;
1048
1049 do {
1050 f = get_next_file(td);
1051 if (!f)
1052 return 1;
1053
429f6675
JA
1054 io_u->file = f;
1055 get_file(f);
1056
1057 if (!fill_io_u(td, io_u))
1058 break;
1059
b5696bfc 1060 put_file_log(td, f);
429f6675 1061 td_io_close_file(td, f);
b5696bfc 1062 io_u->file = NULL;
d6aed795 1063 fio_file_set_done(f);
429f6675 1064 td->nr_done_files++;
0b9d69ec
JA
1065 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1066 td->nr_done_files, td->o.nr_files);
429f6675
JA
1067 } while (1);
1068
1069 return 0;
1070}
1071
1072
10ba535a
JA
1073struct io_u *__get_io_u(struct thread_data *td)
1074{
1075 struct io_u *io_u = NULL;
1076
e8462bd8
JA
1077 td_io_u_lock(td);
1078
1079again:
01743ee1
JA
1080 if (!flist_empty(&td->io_u_requeues))
1081 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
755200a3 1082 else if (!queue_full(td)) {
01743ee1 1083 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
10ba535a 1084
6040dabc 1085 io_u->buflen = 0;
10ba535a 1086 io_u->resid = 0;
755200a3 1087 io_u->file = NULL;
d7762cf8 1088 io_u->end_io = NULL;
755200a3
JA
1089 }
1090
1091 if (io_u) {
0c6e7517 1092 assert(io_u->flags & IO_U_F_FREE);
2ecc1b57 1093 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1ef2b6be 1094 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
82af2a7c 1095 io_u->flags &= ~IO_U_F_VER_LIST;
0c6e7517 1096
755200a3 1097 io_u->error = 0;
01743ee1 1098 flist_del(&io_u->list);
94c6372c 1099 flist_add_tail(&io_u->list, &td->io_u_busylist);
10ba535a 1100 td->cur_depth++;
0c41214f 1101 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1dec3e07
JA
1102 } else if (td->o.verify_async) {
1103 /*
1104 * We ran out, wait for async verify threads to finish and
1105 * return one
1106 */
1107 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1108 goto again;
10ba535a
JA
1109 }
1110
e8462bd8 1111 td_io_u_unlock(td);
10ba535a
JA
1112 return io_u;
1113}
1114
0d29de83 1115static int check_get_trim(struct thread_data *td, struct io_u *io_u)
10ba535a 1116{
0d29de83
JA
1117 if (td->o.trim_backlog && td->trim_entries) {
1118 int get_trim = 0;
10ba535a 1119
0d29de83
JA
1120 if (td->trim_batch) {
1121 td->trim_batch--;
1122 get_trim = 1;
1123 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1124 td->last_ddir != DDIR_READ) {
1125 td->trim_batch = td->o.trim_batch;
1126 if (!td->trim_batch)
1127 td->trim_batch = td->o.trim_backlog;
1128 get_trim = 1;
1129 }
1130
1131 if (get_trim && !get_next_trim(td, io_u))
1132 return 1;
2ba1c290 1133 }
10ba535a 1134
0d29de83
JA
1135 return 0;
1136}
1137
1138static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1139{
9e144189
JA
1140 if (td->o.verify_backlog && td->io_hist_len) {
1141 int get_verify = 0;
1142
d1ece0c7 1143 if (td->verify_batch)
9e144189 1144 get_verify = 1;
d1ece0c7 1145 else if (!(td->io_hist_len % td->o.verify_backlog) &&
9e144189
JA
1146 td->last_ddir != DDIR_READ) {
1147 td->verify_batch = td->o.verify_batch;
f8a75c99
JA
1148 if (!td->verify_batch)
1149 td->verify_batch = td->o.verify_backlog;
9e144189
JA
1150 get_verify = 1;
1151 }
1152
d1ece0c7
JA
1153 if (get_verify && !get_next_verify(td, io_u)) {
1154 td->verify_batch--;
0d29de83 1155 return 1;
d1ece0c7 1156 }
9e144189
JA
1157 }
1158
0d29de83
JA
1159 return 0;
1160}
1161
de789769
JA
1162/*
1163 * Fill offset and start time into the buffer content, to prevent too
23f394d5
JA
1164 * easy compressible data for simple de-dupe attempts. Do this for every
1165 * 512b block in the range, since that should be the smallest block size
1166 * we can expect from a device.
de789769
JA
1167 */
1168static void small_content_scramble(struct io_u *io_u)
1169{
23f394d5 1170 unsigned int i, nr_blocks = io_u->buflen / 512;
fad82f76 1171 unsigned long long boffset;
23f394d5
JA
1172 unsigned int offset;
1173 void *p, *end;
de789769 1174
23f394d5
JA
1175 if (!nr_blocks)
1176 return;
1177
1178 p = io_u->xfer_buf;
fba76ee8 1179 boffset = io_u->offset;
81f0366c 1180 io_u->buf_filled_len = 0;
fad82f76 1181
23f394d5
JA
1182 for (i = 0; i < nr_blocks; i++) {
1183 /*
1184 * Fill the byte offset into a "random" start offset of
1185 * the buffer, given by the product of the usec time
1186 * and the actual offset.
1187 */
fad82f76 1188 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
fba76ee8 1189 offset &= ~(sizeof(unsigned long long) - 1);
23f394d5
JA
1190 if (offset >= 512 - sizeof(unsigned long long))
1191 offset -= sizeof(unsigned long long);
fba76ee8 1192 memcpy(p + offset, &boffset, sizeof(boffset));
23f394d5
JA
1193
1194 end = p + 512 - sizeof(io_u->start_time);
1195 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1196 p += 512;
fad82f76 1197 boffset += 512;
23f394d5 1198 }
de789769
JA
1199}
1200
0d29de83
JA
1201/*
1202 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1203 * etc. The returned io_u is fully ready to be prepped and submitted.
1204 */
1205struct io_u *get_io_u(struct thread_data *td)
1206{
1207 struct fio_file *f;
1208 struct io_u *io_u;
de789769 1209 int do_scramble = 0;
0d29de83
JA
1210
1211 io_u = __get_io_u(td);
1212 if (!io_u) {
1213 dprint(FD_IO, "__get_io_u failed\n");
1214 return NULL;
1215 }
1216
1217 if (check_get_verify(td, io_u))
1218 goto out;
1219 if (check_get_trim(td, io_u))
1220 goto out;
1221
755200a3
JA
1222 /*
1223 * from a requeue, io_u already setup
1224 */
1225 if (io_u->file)
77f392bf 1226 goto out;
755200a3 1227
429f6675
JA
1228 /*
1229 * If using an iolog, grab next piece if any available.
1230 */
1231 if (td->o.read_iolog_file) {
1232 if (read_iolog_get(td, io_u))
1233 goto err_put;
2ba1c290
JA
1234 } else if (set_io_u_file(td, io_u)) {
1235 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
429f6675 1236 goto err_put;
2ba1c290 1237 }
5ec10eaa 1238
429f6675 1239 f = io_u->file;
d6aed795 1240 assert(fio_file_open(f));
97af62ce 1241
ff58fced 1242 if (ddir_rw(io_u->ddir)) {
d0656a93 1243 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
2ba1c290 1244 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
429f6675 1245 goto err_put;
2ba1c290 1246 }
10ba535a 1247
38dad62d 1248 f->last_start = io_u->offset;
36167d82 1249 f->last_pos = io_u->offset + io_u->buflen;
10ba535a 1250
fd68418e 1251 if (io_u->ddir == DDIR_WRITE) {
629f1d71 1252 if (td->o.refill_buffers) {
9c42684e
JA
1253 io_u_fill_buffer(td, io_u,
1254 io_u->xfer_buflen, io_u->xfer_buflen);
1255 } else if (td->o.scramble_buffers)
fd68418e 1256 do_scramble = 1;
629f1d71
JA
1257 if (td->o.verify != VERIFY_NONE) {
1258 populate_verify_io_u(td, io_u);
1259 do_scramble = 0;
1260 }
fd68418e 1261 } else if (io_u->ddir == DDIR_READ) {
cbe8d756
RR
1262 /*
1263 * Reset the buf_filled parameters so next time if the
1264 * buffer is used for writes it is refilled.
1265 */
cbe8d756
RR
1266 io_u->buf_filled_len = 0;
1267 }
87dc1ab1 1268 }
10ba535a 1269
165faf16
JA
1270 /*
1271 * Set io data pointers.
1272 */
cec6b55d
JA
1273 io_u->xfer_buf = io_u->buf;
1274 io_u->xfer_buflen = io_u->buflen;
5973cafb 1275
6ac7a331 1276out:
0d29de83 1277 assert(io_u->file);
429f6675 1278 if (!td_io_prep(td, io_u)) {
993bf48b
JA
1279 if (!td->o.disable_slat)
1280 fio_gettime(&io_u->start_time, NULL);
de789769
JA
1281 if (do_scramble)
1282 small_content_scramble(io_u);
429f6675 1283 return io_u;
36167d82 1284 }
429f6675 1285err_put:
2ba1c290 1286 dprint(FD_IO, "get_io_u failed\n");
429f6675
JA
1287 put_io_u(td, io_u);
1288 return NULL;
10ba535a
JA
1289}
1290
5451792e
JA
1291void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1292{
8b28bd41 1293 enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
825f818e
JA
1294 const char *msg[] = { "read", "write", "sync", "datasync",
1295 "sync_file_range", "wait", "trim" };
1296
8b28bd41
DM
1297 if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
1298 return;
5451792e
JA
1299
1300 log_err("fio: io_u error");
1301
1302 if (io_u->file)
1303 log_err(" on file %s", io_u->file->file_name);
1304
1305 log_err(": %s\n", strerror(io_u->error));
1306
5ec10eaa
JA
1307 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1308 io_u->offset, io_u->xfer_buflen);
5451792e
JA
1309
1310 if (!td->error)
1311 td_verror(td, io_u->error, "io_u error");
1312}
1313
c8eeb9df
JA
1314static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1315 struct io_completion_data *icd,
1316 const enum fio_ddir idx, unsigned int bytes)
1317{
1318 unsigned long uninitialized_var(lusec);
1319
c8eeb9df
JA
1320 if (!td->o.disable_clat || !td->o.disable_bw)
1321 lusec = utime_since(&io_u->issue_time, &icd->time);
1322
1323 if (!td->o.disable_lat) {
1324 unsigned long tusec;
1325
1326 tusec = utime_since(&io_u->start_time, &icd->time);
1327 add_lat_sample(td, idx, tusec, bytes);
1328 }
1329
1330 if (!td->o.disable_clat) {
1331 add_clat_sample(td, idx, lusec, bytes);
1332 io_u_mark_latency(td, lusec);
1333 }
1334
1335 if (!td->o.disable_bw)
1336 add_bw_sample(td, idx, bytes, &icd->time);
1337
1338 add_iops_sample(td, idx, &icd->time);
1339}
1340
1b8dbf25
SL
1341static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1342{
1343 unsigned long long secs, remainder, bps, bytes;
1344 bytes = td->this_io_bytes[ddir];
1345 bps = td->rate_bps[ddir];
1346 secs = bytes / bps;
1347 remainder = bytes % bps;
1348 return remainder * 1000000 / bps + secs * 1000000;
1349}
1350
97601024
JA
1351static void io_completed(struct thread_data *td, struct io_u *io_u,
1352 struct io_completion_data *icd)
10ba535a 1353{
dbad30b0
JA
1354 /*
1355 * Older gcc's are too dumb to realize that usec is always used
1356 * initialized, silence that warning.
1357 */
1358 unsigned long uninitialized_var(usec);
44f29692 1359 struct fio_file *f;
10ba535a 1360
2ba1c290
JA
1361 dprint_io_u(io_u, "io complete");
1362
2ecc1b57 1363 td_io_u_lock(td);
0c6e7517 1364 assert(io_u->flags & IO_U_F_FLIGHT);
38dad62d 1365 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
2ecc1b57 1366 td_io_u_unlock(td);
0c6e7517 1367
5f9099ea 1368 if (ddir_sync(io_u->ddir)) {
87dc1ab1 1369 td->last_was_sync = 1;
44f29692
JA
1370 f = io_u->file;
1371 if (f) {
1372 f->first_write = -1ULL;
1373 f->last_write = -1ULL;
1374 }
87dc1ab1
JA
1375 return;
1376 }
1377
1378 td->last_was_sync = 0;
9e144189 1379 td->last_ddir = io_u->ddir;
87dc1ab1 1380
ff58fced 1381 if (!io_u->error && ddir_rw(io_u->ddir)) {
10ba535a 1382 unsigned int bytes = io_u->buflen - io_u->resid;
1e97cce9 1383 const enum fio_ddir idx = io_u->ddir;
ba3e4e0c 1384 const enum fio_ddir odx = io_u->ddir ^ 1;
b29ee5b3 1385 int ret;
10ba535a 1386
b29ee5b3 1387 td->io_blocks[idx]++;
c8eeb9df 1388 td->this_io_blocks[idx]++;
b29ee5b3 1389 td->io_bytes[idx] += bytes;
ae2fafc8 1390
1391 if (!(io_u->flags & IO_U_F_VER_LIST))
1392 td->this_io_bytes[idx] += bytes;
10ba535a 1393
44f29692
JA
1394 if (idx == DDIR_WRITE) {
1395 f = io_u->file;
1396 if (f) {
1397 if (f->first_write == -1ULL ||
1398 io_u->offset < f->first_write)
1399 f->first_write = io_u->offset;
1400 if (f->last_write == -1ULL ||
1401 ((io_u->offset + bytes) > f->last_write))
1402 f->last_write = io_u->offset + bytes;
1403 }
1404 }
1405
6b1190fd
SL
1406 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1407 td->runstate == TD_VERIFYING)) {
c8eeb9df 1408 account_io_completion(td, io_u, icd, idx, bytes);
40e1a6f0 1409
b23b6a2f 1410 if (__should_check_rate(td, idx)) {
ba3e4e0c 1411 td->rate_pending_usleep[idx] =
1b8dbf25 1412 (usec_for_io(td, idx) -
ba3e4e0c 1413 utime_since_now(&td->start));
b23b6a2f 1414 }
6eaf09d6 1415 if (idx != DDIR_TRIM && __should_check_rate(td, odx))
ba3e4e0c 1416 td->rate_pending_usleep[odx] =
1b8dbf25 1417 (usec_for_io(td, odx) -
ba3e4e0c 1418 utime_since_now(&td->start));
721938ae 1419 }
10ba535a 1420
660a1cb5 1421 if (td_write(td) && idx == DDIR_WRITE &&
8670579e 1422 td->o.do_verify &&
41128405 1423 td->o.verify != VERIFY_NONE)
10ba535a
JA
1424 log_io_piece(td, io_u);
1425
b29ee5b3 1426 icd->bytes_done[idx] += bytes;
3af6ef39 1427
d7762cf8 1428 if (io_u->end_io) {
36690c9b 1429 ret = io_u->end_io(td, io_u);
3af6ef39
JA
1430 if (ret && !icd->error)
1431 icd->error = ret;
1432 }
ff58fced 1433 } else if (io_u->error) {
10ba535a 1434 icd->error = io_u->error;
5451792e
JA
1435 io_u_log_error(td, io_u);
1436 }
8b28bd41
DM
1437 if (icd->error) {
1438 enum error_type_bit eb = td_error_type(io_u->ddir, icd->error);
1439 if (!td_non_fatal_error(td, eb, icd->error))
1440 return;
f2bba182
RR
1441 /*
1442 * If there is a non_fatal error, then add to the error count
1443 * and clear all the errors.
1444 */
1445 update_error_count(td, icd->error);
1446 td_clear_error(td);
1447 icd->error = 0;
1448 io_u->error = 0;
1449 }
10ba535a
JA
1450}
1451
9520ebb9
JA
1452static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1453 int nr)
10ba535a 1454{
6eaf09d6 1455 int ddir;
9520ebb9
JA
1456 if (!td->o.disable_clat || !td->o.disable_bw)
1457 fio_gettime(&icd->time, NULL);
02bcaa8c 1458
3af6ef39
JA
1459 icd->nr = nr;
1460
10ba535a 1461 icd->error = 0;
6eaf09d6
SL
1462 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1463 icd->bytes_done[ddir] = 0;
36167d82
JA
1464}
1465
97601024
JA
1466static void ios_completed(struct thread_data *td,
1467 struct io_completion_data *icd)
36167d82
JA
1468{
1469 struct io_u *io_u;
1470 int i;
1471
10ba535a
JA
1472 for (i = 0; i < icd->nr; i++) {
1473 io_u = td->io_ops->event(td, i);
1474
1475 io_completed(td, io_u, icd);
e8462bd8
JA
1476
1477 if (!(io_u->flags & IO_U_F_FREE_DEF))
1478 put_io_u(td, io_u);
10ba535a
JA
1479 }
1480}
97601024 1481
e7e6cfb4
JA
1482/*
1483 * Complete a single io_u for the sync engines.
1484 */
581e7141
JA
1485int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1486 unsigned long *bytes)
97601024
JA
1487{
1488 struct io_completion_data icd;
1489
9520ebb9 1490 init_icd(td, &icd, 1);
97601024 1491 io_completed(td, io_u, &icd);
e8462bd8
JA
1492
1493 if (!(io_u->flags & IO_U_F_FREE_DEF))
1494 put_io_u(td, io_u);
97601024 1495
581e7141
JA
1496 if (icd.error) {
1497 td_verror(td, icd.error, "io_u_sync_complete");
1498 return -1;
1499 }
97601024 1500
581e7141 1501 if (bytes) {
6eaf09d6
SL
1502 int ddir;
1503
1504 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1505 bytes[ddir] += icd.bytes_done[ddir];
581e7141
JA
1506 }
1507
1508 return 0;
97601024
JA
1509}
1510
e7e6cfb4
JA
1511/*
1512 * Called to complete min_events number of io for the async engines.
1513 */
581e7141
JA
1514int io_u_queued_complete(struct thread_data *td, int min_evts,
1515 unsigned long *bytes)
97601024 1516{
97601024 1517 struct io_completion_data icd;
00de55ef 1518 struct timespec *tvp = NULL;
97601024 1519 int ret;
4d06a338 1520 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
97601024 1521
4950421a 1522 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
b271fe62 1523
4950421a 1524 if (!min_evts)
00de55ef 1525 tvp = &ts;
97601024 1526
4950421a 1527 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
97601024 1528 if (ret < 0) {
e1161c32 1529 td_verror(td, -ret, "td_io_getevents");
97601024
JA
1530 return ret;
1531 } else if (!ret)
1532 return ret;
1533
9520ebb9 1534 init_icd(td, &icd, ret);
97601024 1535 ios_completed(td, &icd);
581e7141
JA
1536 if (icd.error) {
1537 td_verror(td, icd.error, "io_u_queued_complete");
1538 return -1;
1539 }
97601024 1540
581e7141 1541 if (bytes) {
6eaf09d6
SL
1542 int ddir;
1543
1544 for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
1545 bytes[ddir] += icd.bytes_done[ddir];
581e7141
JA
1546 }
1547
1548 return 0;
97601024 1549}
7e77dd02
JA
1550
1551/*
1552 * Call when io_u is really queued, to update the submission latency.
1553 */
1554void io_u_queued(struct thread_data *td, struct io_u *io_u)
1555{
9520ebb9
JA
1556 if (!td->o.disable_slat) {
1557 unsigned long slat_time;
7e77dd02 1558
9520ebb9 1559 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
29a90ddb 1560 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
9520ebb9 1561 }
7e77dd02 1562}
433afcb4 1563
5973cafb
JA
1564/*
1565 * "randomly" fill the buffer contents
1566 */
1567void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
9c42684e 1568 unsigned int min_write, unsigned int max_bs)
5973cafb 1569{
e84b371e
JA
1570 io_u->buf_filled_len = 0;
1571
9c42684e
JA
1572 if (!td->o.zero_buffers) {
1573 unsigned int perc = td->o.compress_percentage;
1574
1575 if (perc) {
f97a43a1
JA
1576 unsigned int seg = min_write;
1577
1578 seg = min(min_write, td->o.compress_chunk);
9c42684e 1579 fill_random_buf_percentage(&td->buf_state, io_u->buf,
f97a43a1 1580 perc, seg, max_bs);
9c42684e
JA
1581 } else
1582 fill_random_buf(&td->buf_state, io_u->buf, max_bs);
1583 } else
637ef8d9 1584 memset(io_u->buf, 0, max_bs);
5973cafb 1585}