Enable completion latency stat collection on verify
[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 */
c8eeb9df 16 int account; /* input */
97601024
JA
17
18 int error; /* output */
19 unsigned long bytes_done[2]; /* output */
20 struct timeval time; /* output */
21};
22
10ba535a
JA
23/*
24 * The ->file_map[] contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
26 */
aec2de20 27static int random_map_free(struct fio_file *f, const unsigned long long block)
10ba535a 28{
aec2de20
JA
29 unsigned int idx = RAND_MAP_IDX(f, block);
30 unsigned int bit = RAND_MAP_BIT(f, block);
10ba535a 31
84422acd
JA
32 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
33
0ce8b119 34 return (f->file_map[idx] & (1UL << bit)) == 0;
10ba535a
JA
35}
36
df415585
JA
37/*
38 * Mark a given offset as used in the map.
39 */
9bf2061e 40static void mark_random_map(struct thread_data *td, struct io_u *io_u)
df415585 41{
2dc1bbeb 42 unsigned int min_bs = td->o.rw_min_bs;
9bf2061e 43 struct fio_file *f = io_u->file;
a00735e6 44 unsigned long long block;
3e3357b1 45 unsigned int blocks, nr_blocks;
38dad62d 46 int busy_check;
df415585 47
b9c5b644 48 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
c685b5b2 49 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
3e3357b1 50 blocks = 0;
38dad62d 51 busy_check = !(io_u->flags & IO_U_F_BUSY_OK);
c685b5b2 52
3e3357b1 53 while (nr_blocks) {
df415585 54 unsigned int idx, bit;
0ce8b119 55 unsigned long mask, this_blocks;
df415585 56
1e3d53ac
JA
57 /*
58 * If we have a mixed random workload, we may
59 * encounter blocks we already did IO to.
60 */
38dad62d
JA
61 if (!busy_check) {
62 blocks = nr_blocks;
63 break;
64 }
5736c10d 65 if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block))
df415585
JA
66 break;
67
aec2de20
JA
68 idx = RAND_MAP_IDX(f, block);
69 bit = RAND_MAP_BIT(f, block);
df415585 70
0032bf9f 71 fio_assert(td, idx < f->num_maps);
df415585 72
3e3357b1
JA
73 this_blocks = nr_blocks;
74 if (this_blocks + bit > BLOCKS_PER_MAP)
75 this_blocks = BLOCKS_PER_MAP - bit;
76
da3758bf
JA
77 do {
78 if (this_blocks == BLOCKS_PER_MAP)
0ce8b119 79 mask = -1UL;
da3758bf 80 else
0ce8b119 81 mask = ((1UL << this_blocks) - 1) << bit;
da3758bf
JA
82
83 if (!(f->file_map[idx] & mask))
84 break;
85
86 this_blocks--;
87 } while (this_blocks);
88
89 if (!this_blocks)
90 break;
3e3357b1 91
3e3357b1
JA
92 f->file_map[idx] |= mask;
93 nr_blocks -= this_blocks;
94 blocks += this_blocks;
c9dd34b2 95 block += this_blocks;
df415585
JA
96 }
97
a00735e6
JA
98 if ((blocks * min_bs) < io_u->buflen)
99 io_u->buflen = blocks * min_bs;
df415585
JA
100}
101
3e3357b1
JA
102static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
103 enum fio_ddir ddir)
2ba1c290
JA
104{
105 unsigned long long max_blocks;
d9dd70f7 106 unsigned long long max_size;
2ba1c290 107
ff58fced
JA
108 assert(ddir_rw(ddir));
109
d9dd70f7
JA
110 /*
111 * Hmm, should we make sure that ->io_size <= ->real_file_size?
112 */
113 max_size = f->io_size;
114 if (max_size > f->real_file_size)
115 max_size = f->real_file_size;
116
ed335855
SN
117 if (td->o.zone_range)
118 max_size = td->o.zone_range;
119
2b7a01d0 120 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
2ba1c290
JA
121 if (!max_blocks)
122 return 0;
123
67778e88 124 return max_blocks;
2ba1c290
JA
125}
126
10ba535a
JA
127/*
128 * Return the next free block in the map.
129 */
130static int get_next_free_block(struct thread_data *td, struct fio_file *f,
4ba66134 131 enum fio_ddir ddir, unsigned long long *b)
10ba535a 132{
0ce8b119 133 unsigned long long block, min_bs = td->o.rw_min_bs, lastb;
10ba535a
JA
134 int i;
135
5e0baa7f
JA
136 lastb = last_block(td, f, ddir);
137 if (!lastb)
138 return 1;
139
c685b5b2 140 i = f->last_free_lookup;
0ce8b119
JA
141 block = i * BLOCKS_PER_MAP;
142 while (block * min_bs < f->real_file_size &&
143 block * min_bs < f->io_size) {
144 if (f->file_map[i] != -1UL) {
145 block += ffz(f->file_map[i]);
146 if (block > lastb)
2ba1c290 147 break;
c685b5b2 148 f->last_free_lookup = i;
0ce8b119 149 *b = block;
10ba535a
JA
150 return 0;
151 }
152
0ce8b119 153 block += BLOCKS_PER_MAP;
10ba535a
JA
154 i++;
155 }
156
2ba1c290 157 dprint(FD_IO, "failed finding a free block\n");
10ba535a
JA
158 return 1;
159}
160
ec4015da 161static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
4ba66134 162 enum fio_ddir ddir, unsigned long long *b)
ec4015da 163{
15b87723 164 unsigned long long rmax, r, lastb;
ec4015da
JA
165 int loops = 5;
166
5e0baa7f
JA
167 lastb = last_block(td, f, ddir);
168 if (!lastb)
169 return 1;
170
0ce8b119
JA
171 if (f->failed_rands >= 200)
172 goto ffz;
173
15b87723 174 rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
ec4015da 175 do {
15b87723 176 if (td->o.use_os_rand)
2615cc4b 177 r = os_random_long(&td->random_state);
15b87723 178 else
2615cc4b 179 r = __rand(&td->__random_state);
15b87723
JA
180
181 *b = (lastb - 1) * (r / ((unsigned long long) rmax + 1.0));
2615cc4b 182
84422acd 183 dprint(FD_RANDOM, "off rand %llu\n", r);
2615cc4b 184
2ba1c290 185
43c63a78
JA
186 /*
187 * if we are not maintaining a random map, we are done.
188 */
303032ae 189 if (!file_randommap(td, f))
0ce8b119 190 goto ret_good;
43c63a78
JA
191
192 /*
84422acd 193 * calculate map offset and check if it's free
43c63a78 194 */
aec2de20 195 if (random_map_free(f, *b))
0ce8b119 196 goto ret_good;
43c63a78 197
84422acd
JA
198 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
199 *b);
43c63a78 200 } while (--loops);
ec4015da 201
0ce8b119
JA
202 if (!f->failed_rands++)
203 f->last_free_lookup = 0;
204
ec4015da 205 /*
43c63a78
JA
206 * we get here, if we didn't suceed in looking up a block. generate
207 * a random start offset into the filemap, and find the first free
208 * block from there.
ec4015da 209 */
43c63a78
JA
210 loops = 10;
211 do {
dc873b6f 212 f->last_free_lookup = (f->num_maps - 1) *
15b87723 213 (r / ((unsigned long long) rmax + 1.0));
4ba66134 214 if (!get_next_free_block(td, f, ddir, b))
0ce8b119 215 goto ret;
ec4015da 216
d1cea521
JA
217 if (td->o.use_os_rand)
218 r = os_random_long(&td->random_state);
219 else
220 r = __rand(&td->__random_state);
43c63a78
JA
221 } while (--loops);
222
223 /*
224 * that didn't work either, try exhaustive search from the start
225 */
226 f->last_free_lookup = 0;
0ce8b119
JA
227ffz:
228 if (!get_next_free_block(td, f, ddir, b))
229 return 0;
230 f->last_free_lookup = 0;
4ba66134 231 return get_next_free_block(td, f, ddir, b);
0ce8b119
JA
232ret_good:
233 f->failed_rands = 0;
234ret:
235 return 0;
ec4015da
JA
236}
237
38dad62d
JA
238static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
239 enum fio_ddir ddir, unsigned long long *b)
240{
241 if (get_next_rand_offset(td, f, ddir, b)) {
242 dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
243 f->file_name, f->last_pos, f->real_file_size);
244 return 1;
245 }
246
247 return 0;
248}
249
250static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
251 enum fio_ddir ddir, unsigned long long *b)
252{
ff58fced
JA
253 assert(ddir_rw(ddir));
254
38dad62d 255 if (f->last_pos < f->real_file_size) {
a66da7a2 256 unsigned long long pos;
059b0802 257
a66da7a2
JA
258 if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
259 f->last_pos = f->real_file_size;
260
261 pos = f->last_pos - f->file_offset;
059b0802
JA
262 if (pos)
263 pos += td->o.ddir_seq_add;
264
265 *b = pos / td->o.min_bs[ddir];
38dad62d
JA
266 return 0;
267 }
268
269 return 1;
270}
271
272static int get_next_block(struct thread_data *td, struct io_u *io_u,
273 enum fio_ddir ddir, int rw_seq, unsigned long long *b)
274{
275 struct fio_file *f = io_u->file;
276 int ret;
277
ff58fced
JA
278 assert(ddir_rw(ddir));
279
38dad62d
JA
280 if (rw_seq) {
281 if (td_random(td))
282 ret = get_next_rand_block(td, f, ddir, b);
283 else
284 ret = get_next_seq_block(td, f, ddir, b);
285 } else {
286 io_u->flags |= IO_U_F_BUSY_OK;
287
288 if (td->o.rw_seq == RW_SEQ_SEQ) {
289 ret = get_next_seq_block(td, f, ddir, b);
290 if (ret)
291 ret = get_next_rand_block(td, f, ddir, b);
292 } else if (td->o.rw_seq == RW_SEQ_IDENT) {
293 if (f->last_start != -1ULL)
0508320e
DN
294 *b = (f->last_start - f->file_offset)
295 / td->o.min_bs[ddir];
38dad62d
JA
296 else
297 *b = 0;
298 ret = 0;
299 } else {
300 log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
301 ret = 1;
302 }
303 }
304
305 return ret;
306}
307
10ba535a
JA
308/*
309 * For random io, generate a random new block and see if it's used. Repeat
310 * until we find a free one. For sequential io, just return the end of
311 * the last io issued.
312 */
15dc1934 313static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
10ba535a 314{
9bf2061e 315 struct fio_file *f = io_u->file;
ec4015da 316 unsigned long long b;
4ba66134 317 enum fio_ddir ddir = io_u->ddir;
38dad62d 318 int rw_seq_hit = 0;
10ba535a 319
ff58fced
JA
320 assert(ddir_rw(ddir));
321
38dad62d
JA
322 if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
323 rw_seq_hit = 1;
5736c10d 324 td->ddir_seq_nr = td->o.ddir_seq_nr;
38dad62d 325 }
211097b2 326
0508320e 327 if (get_next_block(td, io_u, ddir, rw_seq_hit, &b))
38dad62d 328 return 1;
10ba535a 329
2b7a01d0 330 io_u->offset = b * td->o.ba[ddir];
009bd847
JA
331 if (io_u->offset >= f->io_size) {
332 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
333 io_u->offset, f->io_size);
334 return 1;
335 }
336
337 io_u->offset += f->file_offset;
2ba1c290
JA
338 if (io_u->offset >= f->real_file_size) {
339 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
340 io_u->offset, f->real_file_size);
10ba535a 341 return 1;
2ba1c290 342 }
10ba535a
JA
343
344 return 0;
345}
346
15dc1934
JA
347static int get_next_offset(struct thread_data *td, struct io_u *io_u)
348{
7eb36574
JA
349 struct prof_io_ops *ops = &td->prof_io_ops;
350
351 if (ops->fill_io_u_off)
352 return ops->fill_io_u_off(td, io_u);
15dc1934
JA
353
354 return __get_next_offset(td, io_u);
355}
356
79944128
JA
357static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
358 unsigned int buflen)
359{
360 struct fio_file *f = io_u->file;
361
362 return io_u->offset + buflen <= f->io_size + td->o.start_offset;
363}
364
15dc1934 365static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
10ba535a 366{
bca4ed4d 367 const int ddir = io_u->ddir;
f3f552b9 368 unsigned int uninitialized_var(buflen);
f3059de1 369 unsigned int minbs, maxbs;
1294c3ec 370 unsigned long r, rand_max;
10ba535a 371
ff58fced
JA
372 assert(ddir_rw(ddir));
373
f3059de1
JA
374 minbs = td->o.min_bs[ddir];
375 maxbs = td->o.max_bs[ddir];
376
79944128
JA
377 if (minbs == maxbs)
378 return minbs;
379
52c58027
JA
380 /*
381 * If we can't satisfy the min block size from here, then fail
382 */
383 if (!io_u_fits(td, io_u, minbs))
384 return 0;
385
4c07ad86
JA
386 if (td->o.use_os_rand)
387 rand_max = OS_RAND_MAX;
388 else
389 rand_max = FRAND_MAX;
390
79944128 391 do {
4c07ad86
JA
392 if (td->o.use_os_rand)
393 r = os_random_long(&td->bsrange_state);
394 else
395 r = __rand(&td->__bsrange_state);
396
720e84ad 397 if (!td->o.bssplit_nr[ddir]) {
f3059de1 398 buflen = 1 + (unsigned int) ((double) maxbs *
4c07ad86 399 (r / (rand_max + 1.0)));
f3059de1
JA
400 if (buflen < minbs)
401 buflen = minbs;
5ec10eaa 402 } else {
564ca972
JA
403 long perc = 0;
404 unsigned int i;
405
720e84ad
JA
406 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
407 struct bssplit *bsp = &td->o.bssplit[ddir][i];
564ca972
JA
408
409 buflen = bsp->bs;
410 perc += bsp->perc;
79944128
JA
411 if ((r <= ((rand_max / 100L) * perc)) &&
412 io_u_fits(td, io_u, buflen))
564ca972
JA
413 break;
414 }
415 }
79944128 416
f3059de1
JA
417 if (!td->o.bs_unaligned && is_power_of_2(minbs))
418 buflen = (buflen + minbs - 1) & ~(minbs - 1);
10ba535a 419
79944128 420 } while (!io_u_fits(td, io_u, buflen));
6a5e6884 421
10ba535a
JA
422 return buflen;
423}
424
15dc1934
JA
425static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
426{
7eb36574
JA
427 struct prof_io_ops *ops = &td->prof_io_ops;
428
429 if (ops->fill_io_u_size)
430 return ops->fill_io_u_size(td, io_u);
15dc1934
JA
431
432 return __get_next_buflen(td, io_u);
433}
434
afe24a5a
JA
435static void set_rwmix_bytes(struct thread_data *td)
436{
afe24a5a
JA
437 unsigned int diff;
438
439 /*
440 * we do time or byte based switch. this is needed because
441 * buffered writes may issue a lot quicker than they complete,
442 * whereas reads do not.
443 */
e47f799f 444 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
04c540d9 445 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
e47f799f
JA
446}
447
448static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
449{
450 unsigned int v;
1294c3ec 451 unsigned long r;
e47f799f 452
4c07ad86
JA
453 if (td->o.use_os_rand) {
454 r = os_random_long(&td->rwmix_state);
455 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
456 } else {
457 r = __rand(&td->__rwmix_state);
458 v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
459 }
460
04c540d9 461 if (v <= td->o.rwmix[DDIR_READ])
e47f799f
JA
462 return DDIR_READ;
463
464 return DDIR_WRITE;
afe24a5a
JA
465}
466
581e7141
JA
467static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
468{
469 enum fio_ddir odir = ddir ^ 1;
470 struct timeval t;
471 long usec;
472
ff58fced
JA
473 assert(ddir_rw(ddir));
474
581e7141
JA
475 if (td->rate_pending_usleep[ddir] <= 0)
476 return ddir;
477
478 /*
479 * We have too much pending sleep in this direction. See if we
480 * should switch.
481 */
482 if (td_rw(td)) {
483 /*
484 * Other direction does not have too much pending, switch
485 */
486 if (td->rate_pending_usleep[odir] < 100000)
487 return odir;
488
489 /*
490 * Both directions have pending sleep. Sleep the minimum time
491 * and deduct from both.
492 */
493 if (td->rate_pending_usleep[ddir] <=
494 td->rate_pending_usleep[odir]) {
495 usec = td->rate_pending_usleep[ddir];
496 } else {
497 usec = td->rate_pending_usleep[odir];
498 ddir = odir;
499 }
500 } else
501 usec = td->rate_pending_usleep[ddir];
502
78c1eda5
JA
503 /*
504 * We are going to sleep, ensure that we flush anything pending as
422f9e4b
RM
505 * not to skew our latency numbers.
506 *
507 * Changed to only monitor 'in flight' requests here instead of the
508 * td->cur_depth, b/c td->cur_depth does not accurately represent
509 * io's that have been actually submitted to an async engine,
510 * and cur_depth is meaningless for sync engines.
78c1eda5 511 */
422f9e4b 512 if (td->io_u_in_flight) {
78c1eda5
JA
513 int fio_unused ret;
514
422f9e4b 515 ret = io_u_queued_complete(td, td->io_u_in_flight, NULL);
78c1eda5
JA
516 }
517
581e7141
JA
518 fio_gettime(&t, NULL);
519 usec_sleep(td, usec);
520 usec = utime_since_now(&t);
521
522 td->rate_pending_usleep[ddir] -= usec;
523
524 odir = ddir ^ 1;
525 if (td_rw(td) && __should_check_rate(td, odir))
526 td->rate_pending_usleep[odir] -= usec;
0b9d69ec 527
581e7141
JA
528 return ddir;
529}
530
10ba535a
JA
531/*
532 * Return the data direction for the next io_u. If the job is a
533 * mixed read/write workload, check the rwmix cycle and switch if
534 * necessary.
535 */
1e97cce9 536static enum fio_ddir get_rw_ddir(struct thread_data *td)
10ba535a 537{
581e7141
JA
538 enum fio_ddir ddir;
539
5f9099ea
JA
540 /*
541 * see if it's time to fsync
542 */
543 if (td->o.fsync_blocks &&
544 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
545 td->io_issues[DDIR_WRITE] && should_fsync(td))
546 return DDIR_SYNC;
547
548 /*
549 * see if it's time to fdatasync
550 */
551 if (td->o.fdatasync_blocks &&
552 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
553 td->io_issues[DDIR_WRITE] && should_fsync(td))
554 return DDIR_DATASYNC;
555
44f29692
JA
556 /*
557 * see if it's time to sync_file_range
558 */
559 if (td->sync_file_range_nr &&
560 !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
561 td->io_issues[DDIR_WRITE] && should_fsync(td))
562 return DDIR_SYNC_FILE_RANGE;
563
10ba535a 564 if (td_rw(td)) {
10ba535a
JA
565 /*
566 * Check if it's time to seed a new data direction.
567 */
e4928662 568 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
e47f799f
JA
569 /*
570 * Put a top limit on how many bytes we do for
571 * one data direction, to avoid overflowing the
572 * ranges too much
573 */
574 ddir = get_rand_ddir(td);
e47f799f
JA
575
576 if (ddir != td->rwmix_ddir)
577 set_rwmix_bytes(td);
578
579 td->rwmix_ddir = ddir;
10ba535a 580 }
581e7141 581 ddir = td->rwmix_ddir;
10ba535a 582 } else if (td_read(td))
581e7141 583 ddir = DDIR_READ;
10ba535a 584 else
581e7141
JA
585 ddir = DDIR_WRITE;
586
587 td->rwmix_ddir = rate_ddir(td, ddir);
588 return td->rwmix_ddir;
10ba535a
JA
589}
590
1ef2b6be
JA
591static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
592{
593 io_u->ddir = get_rw_ddir(td);
594
595 if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
596 td->o.barrier_blocks &&
597 !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
598 td->io_issues[DDIR_WRITE])
599 io_u->flags |= IO_U_F_BARRIER;
600}
601
e8462bd8 602void put_file_log(struct thread_data *td, struct fio_file *f)
60f2c658
JA
603{
604 int ret = put_file(td, f);
605
606 if (ret)
607 td_verror(td, ret, "file close");
608}
609
10ba535a
JA
610void put_io_u(struct thread_data *td, struct io_u *io_u)
611{
e8462bd8
JA
612 td_io_u_lock(td);
613
d7ee2a7d 614 if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
60f2c658 615 put_file_log(td, io_u->file);
10ba535a 616 io_u->file = NULL;
d7ee2a7d
SL
617 io_u->flags &= ~IO_U_F_FREE_DEF;
618 io_u->flags |= IO_U_F_FREE;
619
0c41214f
RR
620 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
621 td->cur_depth--;
e8462bd8 622 flist_del_init(&io_u->list);
01743ee1 623 flist_add(&io_u->list, &td->io_u_freelist);
e8462bd8
JA
624 td_io_u_unlock(td);
625 td_io_u_free_notify(td);
10ba535a
JA
626}
627
f2bba182
RR
628void clear_io_u(struct thread_data *td, struct io_u *io_u)
629{
630 io_u->flags &= ~IO_U_F_FLIGHT;
631 put_io_u(td, io_u);
632}
633
755200a3
JA
634void requeue_io_u(struct thread_data *td, struct io_u **io_u)
635{
636 struct io_u *__io_u = *io_u;
637
465221b0
JA
638 dprint(FD_IO, "requeue %p\n", __io_u);
639
e8462bd8
JA
640 td_io_u_lock(td);
641
4d2e0f49 642 __io_u->flags |= IO_U_F_FREE;
ff58fced 643 if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
e4f54adb 644 td->io_issues[__io_u->ddir]--;
5ec10eaa 645
4d2e0f49 646 __io_u->flags &= ~IO_U_F_FLIGHT;
0c41214f
RR
647 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
648 td->cur_depth--;
01743ee1
JA
649 flist_del(&__io_u->list);
650 flist_add_tail(&__io_u->list, &td->io_u_requeues);
e8462bd8 651 td_io_u_unlock(td);
755200a3
JA
652 *io_u = NULL;
653}
654
9bf2061e 655static int fill_io_u(struct thread_data *td, struct io_u *io_u)
10ba535a 656{
b4c5e1ac
JA
657 if (td->io_ops->flags & FIO_NOIO)
658 goto out;
659
1ef2b6be 660 set_rw_ddir(td, io_u);
5f9099ea 661
87dc1ab1 662 /*
ff58fced 663 * fsync() or fdatasync() or trim etc, we are done
87dc1ab1 664 */
ff58fced 665 if (!ddir_rw(io_u->ddir))
c38e9468 666 goto out;
a00735e6 667
48f5abd3
JA
668 /*
669 * See if it's time to switch to a new zone
670 */
671 if (td->zone_bytes >= td->o.zone_size) {
672 td->zone_bytes = 0;
ed335855
SN
673 io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
674 io_u->file->last_pos = io_u->file->file_offset;
48f5abd3
JA
675 td->io_skip_bytes += td->o.zone_skip;
676 }
677
10ba535a 678 /*
c685b5b2
JA
679 * No log, let the seq/rand engine retrieve the next buflen and
680 * position.
10ba535a 681 */
2ba1c290
JA
682 if (get_next_offset(td, io_u)) {
683 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
bca4ed4d 684 return 1;
2ba1c290 685 }
10ba535a 686
9bf2061e 687 io_u->buflen = get_next_buflen(td, io_u);
2ba1c290
JA
688 if (!io_u->buflen) {
689 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
bca4ed4d 690 return 1;
2ba1c290 691 }
bca4ed4d 692
2ba1c290
JA
693 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
694 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
4ba66134
JA
695 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
696 io_u->buflen, io_u->file->real_file_size);
6a5e6884 697 return 1;
2ba1c290 698 }
6a5e6884 699
bca4ed4d
JA
700 /*
701 * mark entry before potentially trimming io_u
702 */
303032ae 703 if (td_random(td) && file_randommap(td, io_u->file))
9bf2061e 704 mark_random_map(td, io_u);
bca4ed4d
JA
705
706 /*
707 * If using a write iolog, store this entry.
708 */
c38e9468 709out:
2ba1c290 710 dprint_io_u(io_u, "fill_io_u");
d9d91e39 711 td->zone_bytes += io_u->buflen;
f29b25a3 712 log_io_u(td, io_u);
bca4ed4d 713 return 0;
10ba535a
JA
714}
715
838bc709
JA
716static void __io_u_mark_map(unsigned int *map, unsigned int nr)
717{
2b13e716 718 int idx = 0;
838bc709
JA
719
720 switch (nr) {
721 default:
2b13e716 722 idx = 6;
838bc709
JA
723 break;
724 case 33 ... 64:
2b13e716 725 idx = 5;
838bc709
JA
726 break;
727 case 17 ... 32:
2b13e716 728 idx = 4;
838bc709
JA
729 break;
730 case 9 ... 16:
2b13e716 731 idx = 3;
838bc709
JA
732 break;
733 case 5 ... 8:
2b13e716 734 idx = 2;
838bc709
JA
735 break;
736 case 1 ... 4:
2b13e716 737 idx = 1;
838bc709
JA
738 case 0:
739 break;
740 }
741
2b13e716 742 map[idx]++;
838bc709
JA
743}
744
745void io_u_mark_submit(struct thread_data *td, unsigned int nr)
746{
747 __io_u_mark_map(td->ts.io_u_submit, nr);
748 td->ts.total_submit++;
749}
750
751void io_u_mark_complete(struct thread_data *td, unsigned int nr)
752{
753 __io_u_mark_map(td->ts.io_u_complete, nr);
754 td->ts.total_complete++;
755}
756
d8005759 757void io_u_mark_depth(struct thread_data *td, unsigned int nr)
71619dc2 758{
2b13e716 759 int idx = 0;
71619dc2
JA
760
761 switch (td->cur_depth) {
762 default:
2b13e716 763 idx = 6;
a783e61a 764 break;
71619dc2 765 case 32 ... 63:
2b13e716 766 idx = 5;
a783e61a 767 break;
71619dc2 768 case 16 ... 31:
2b13e716 769 idx = 4;
a783e61a 770 break;
71619dc2 771 case 8 ... 15:
2b13e716 772 idx = 3;
a783e61a 773 break;
71619dc2 774 case 4 ... 7:
2b13e716 775 idx = 2;
a783e61a 776 break;
71619dc2 777 case 2 ... 3:
2b13e716 778 idx = 1;
71619dc2
JA
779 case 1:
780 break;
781 }
782
2b13e716 783 td->ts.io_u_map[idx] += nr;
71619dc2
JA
784}
785
04a0feae
JA
786static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
787{
2b13e716 788 int idx = 0;
04a0feae
JA
789
790 assert(usec < 1000);
791
792 switch (usec) {
793 case 750 ... 999:
2b13e716 794 idx = 9;
04a0feae
JA
795 break;
796 case 500 ... 749:
2b13e716 797 idx = 8;
04a0feae
JA
798 break;
799 case 250 ... 499:
2b13e716 800 idx = 7;
04a0feae
JA
801 break;
802 case 100 ... 249:
2b13e716 803 idx = 6;
04a0feae
JA
804 break;
805 case 50 ... 99:
2b13e716 806 idx = 5;
04a0feae
JA
807 break;
808 case 20 ... 49:
2b13e716 809 idx = 4;
04a0feae
JA
810 break;
811 case 10 ... 19:
2b13e716 812 idx = 3;
04a0feae
JA
813 break;
814 case 4 ... 9:
2b13e716 815 idx = 2;
04a0feae
JA
816 break;
817 case 2 ... 3:
2b13e716 818 idx = 1;
04a0feae
JA
819 case 0 ... 1:
820 break;
821 }
822
2b13e716
JA
823 assert(idx < FIO_IO_U_LAT_U_NR);
824 td->ts.io_u_lat_u[idx]++;
04a0feae
JA
825}
826
827static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
ec118304 828{
2b13e716 829 int idx = 0;
ec118304
JA
830
831 switch (msec) {
832 default:
2b13e716 833 idx = 11;
04a0feae 834 break;
8abdce66 835 case 1000 ... 1999:
2b13e716 836 idx = 10;
04a0feae 837 break;
8abdce66 838 case 750 ... 999:
2b13e716 839 idx = 9;
04a0feae 840 break;
8abdce66 841 case 500 ... 749:
2b13e716 842 idx = 8;
04a0feae 843 break;
8abdce66 844 case 250 ... 499:
2b13e716 845 idx = 7;
04a0feae 846 break;
8abdce66 847 case 100 ... 249:
2b13e716 848 idx = 6;
04a0feae 849 break;
8abdce66 850 case 50 ... 99:
2b13e716 851 idx = 5;
04a0feae 852 break;
8abdce66 853 case 20 ... 49:
2b13e716 854 idx = 4;
04a0feae 855 break;
8abdce66 856 case 10 ... 19:
2b13e716 857 idx = 3;
04a0feae 858 break;
8abdce66 859 case 4 ... 9:
2b13e716 860 idx = 2;
04a0feae 861 break;
ec118304 862 case 2 ... 3:
2b13e716 863 idx = 1;
ec118304
JA
864 case 0 ... 1:
865 break;
866 }
867
2b13e716
JA
868 assert(idx < FIO_IO_U_LAT_M_NR);
869 td->ts.io_u_lat_m[idx]++;
04a0feae
JA
870}
871
872static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
873{
874 if (usec < 1000)
875 io_u_mark_lat_usec(td, usec);
876 else
877 io_u_mark_lat_msec(td, usec / 1000);
ec118304
JA
878}
879
0aabe160
JA
880/*
881 * Get next file to service by choosing one at random
882 */
2cc52930
JA
883static struct fio_file *get_next_file_rand(struct thread_data *td,
884 enum fio_file_flags goodf,
d6aed795 885 enum fio_file_flags badf)
0aabe160 886{
0aabe160 887 struct fio_file *f;
1c178180 888 int fno;
0aabe160
JA
889
890 do {
87b10676 891 int opened = 0;
1294c3ec 892 unsigned long r;
4c07ad86
JA
893
894 if (td->o.use_os_rand) {
895 r = os_random_long(&td->next_file_state);
896 fno = (unsigned int) ((double) td->o.nr_files
897 * (r / (OS_RAND_MAX + 1.0)));
898 } else {
899 r = __rand(&td->__next_file_state);
900 fno = (unsigned int) ((double) td->o.nr_files
901 * (r / (FRAND_MAX + 1.0)));
902 }
7c83c089 903
126d65c6 904 f = td->files[fno];
d6aed795 905 if (fio_file_done(f))
059e63c0 906 continue;
1c178180 907
d6aed795 908 if (!fio_file_open(f)) {
87b10676
JA
909 int err;
910
911 err = td_io_open_file(td, f);
912 if (err)
913 continue;
914 opened = 1;
915 }
916
2ba1c290
JA
917 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
918 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
0aabe160 919 return f;
2ba1c290 920 }
87b10676
JA
921 if (opened)
922 td_io_close_file(td, f);
0aabe160
JA
923 } while (1);
924}
925
926/*
927 * Get next file to service by doing round robin between all available ones
928 */
1c178180
JA
929static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
930 int badf)
3d7c391d
JA
931{
932 unsigned int old_next_file = td->next_file;
933 struct fio_file *f;
934
935 do {
87b10676
JA
936 int opened = 0;
937
126d65c6 938 f = td->files[td->next_file];
3d7c391d
JA
939
940 td->next_file++;
2dc1bbeb 941 if (td->next_file >= td->o.nr_files)
3d7c391d
JA
942 td->next_file = 0;
943
87b10676 944 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
d6aed795 945 if (fio_file_done(f)) {
d5ed68ea 946 f = NULL;
059e63c0 947 continue;
d5ed68ea 948 }
059e63c0 949
d6aed795 950 if (!fio_file_open(f)) {
87b10676
JA
951 int err;
952
953 err = td_io_open_file(td, f);
b5696bfc
JA
954 if (err) {
955 dprint(FD_FILE, "error %d on open of %s\n",
956 err, f->file_name);
87c27b45 957 f = NULL;
87b10676 958 continue;
b5696bfc 959 }
87b10676
JA
960 opened = 1;
961 }
962
0b9d69ec
JA
963 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
964 f->flags);
1c178180 965 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
3d7c391d
JA
966 break;
967
87b10676
JA
968 if (opened)
969 td_io_close_file(td, f);
970
3d7c391d
JA
971 f = NULL;
972 } while (td->next_file != old_next_file);
973
2ba1c290 974 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
3d7c391d
JA
975 return f;
976}
977
7eb36574 978static struct fio_file *__get_next_file(struct thread_data *td)
bdb4e2e9 979{
1907dbc6
JA
980 struct fio_file *f;
981
2dc1bbeb 982 assert(td->o.nr_files <= td->files_index);
1c178180 983
b5696bfc 984 if (td->nr_done_files >= td->o.nr_files) {
5ec10eaa
JA
985 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
986 " nr_files=%d\n", td->nr_open_files,
987 td->nr_done_files,
988 td->o.nr_files);
bdb4e2e9 989 return NULL;
2ba1c290 990 }
bdb4e2e9 991
1907dbc6 992 f = td->file_service_file;
d6aed795 993 if (f && fio_file_open(f) && !fio_file_closing(f)) {
a086c257
JA
994 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
995 goto out;
996 if (td->file_service_left--)
997 goto out;
998 }
1907dbc6 999
a086c257
JA
1000 if (td->o.file_service_type == FIO_FSERVICE_RR ||
1001 td->o.file_service_type == FIO_FSERVICE_SEQ)
d6aed795 1002 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
bdb4e2e9 1003 else
d6aed795 1004 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1907dbc6
JA
1005
1006 td->file_service_file = f;
1007 td->file_service_left = td->file_service_nr - 1;
2ba1c290 1008out:
683023e8 1009 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1907dbc6 1010 return f;
bdb4e2e9
JA
1011}
1012
7eb36574
JA
1013static struct fio_file *get_next_file(struct thread_data *td)
1014{
1015 struct prof_io_ops *ops = &td->prof_io_ops;
1016
1017 if (ops->get_next_file)
1018 return ops->get_next_file(td);
1019
1020 return __get_next_file(td);
1021}
1022
429f6675
JA
1023static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
1024{
1025 struct fio_file *f;
1026
1027 do {
1028 f = get_next_file(td);
1029 if (!f)
1030 return 1;
1031
429f6675
JA
1032 io_u->file = f;
1033 get_file(f);
1034
1035 if (!fill_io_u(td, io_u))
1036 break;
1037
b5696bfc 1038 put_file_log(td, f);
429f6675 1039 td_io_close_file(td, f);
b5696bfc 1040 io_u->file = NULL;
d6aed795 1041 fio_file_set_done(f);
429f6675 1042 td->nr_done_files++;
0b9d69ec
JA
1043 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
1044 td->nr_done_files, td->o.nr_files);
429f6675
JA
1045 } while (1);
1046
1047 return 0;
1048}
1049
1050
10ba535a
JA
1051struct io_u *__get_io_u(struct thread_data *td)
1052{
1053 struct io_u *io_u = NULL;
1054
e8462bd8
JA
1055 td_io_u_lock(td);
1056
1057again:
01743ee1
JA
1058 if (!flist_empty(&td->io_u_requeues))
1059 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
755200a3 1060 else if (!queue_full(td)) {
01743ee1 1061 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
10ba535a 1062
6040dabc 1063 io_u->buflen = 0;
10ba535a 1064 io_u->resid = 0;
755200a3 1065 io_u->file = NULL;
d7762cf8 1066 io_u->end_io = NULL;
755200a3
JA
1067 }
1068
1069 if (io_u) {
0c6e7517 1070 assert(io_u->flags & IO_U_F_FREE);
2ecc1b57 1071 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
1ef2b6be 1072 io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
0c6e7517 1073
755200a3 1074 io_u->error = 0;
01743ee1
JA
1075 flist_del(&io_u->list);
1076 flist_add(&io_u->list, &td->io_u_busylist);
10ba535a 1077 td->cur_depth++;
0c41214f 1078 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
1dec3e07
JA
1079 } else if (td->o.verify_async) {
1080 /*
1081 * We ran out, wait for async verify threads to finish and
1082 * return one
1083 */
1084 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
1085 goto again;
10ba535a
JA
1086 }
1087
e8462bd8 1088 td_io_u_unlock(td);
10ba535a
JA
1089 return io_u;
1090}
1091
0d29de83 1092static int check_get_trim(struct thread_data *td, struct io_u *io_u)
10ba535a 1093{
0d29de83
JA
1094 if (td->o.trim_backlog && td->trim_entries) {
1095 int get_trim = 0;
10ba535a 1096
0d29de83
JA
1097 if (td->trim_batch) {
1098 td->trim_batch--;
1099 get_trim = 1;
1100 } else if (!(td->io_hist_len % td->o.trim_backlog) &&
1101 td->last_ddir != DDIR_READ) {
1102 td->trim_batch = td->o.trim_batch;
1103 if (!td->trim_batch)
1104 td->trim_batch = td->o.trim_backlog;
1105 get_trim = 1;
1106 }
1107
1108 if (get_trim && !get_next_trim(td, io_u))
1109 return 1;
2ba1c290 1110 }
10ba535a 1111
0d29de83
JA
1112 return 0;
1113}
1114
1115static int check_get_verify(struct thread_data *td, struct io_u *io_u)
1116{
9e144189
JA
1117 if (td->o.verify_backlog && td->io_hist_len) {
1118 int get_verify = 0;
1119
1120 if (td->verify_batch) {
1121 td->verify_batch--;
1122 get_verify = 1;
1123 } else if (!(td->io_hist_len % td->o.verify_backlog) &&
1124 td->last_ddir != DDIR_READ) {
1125 td->verify_batch = td->o.verify_batch;
f8a75c99
JA
1126 if (!td->verify_batch)
1127 td->verify_batch = td->o.verify_backlog;
9e144189
JA
1128 get_verify = 1;
1129 }
1130
1131 if (get_verify && !get_next_verify(td, io_u))
0d29de83 1132 return 1;
9e144189
JA
1133 }
1134
0d29de83
JA
1135 return 0;
1136}
1137
de789769
JA
1138/*
1139 * Fill offset and start time into the buffer content, to prevent too
23f394d5
JA
1140 * easy compressible data for simple de-dupe attempts. Do this for every
1141 * 512b block in the range, since that should be the smallest block size
1142 * we can expect from a device.
de789769
JA
1143 */
1144static void small_content_scramble(struct io_u *io_u)
1145{
23f394d5 1146 unsigned int i, nr_blocks = io_u->buflen / 512;
fad82f76 1147 unsigned long long boffset;
23f394d5
JA
1148 unsigned int offset;
1149 void *p, *end;
de789769 1150
23f394d5
JA
1151 if (!nr_blocks)
1152 return;
1153
1154 p = io_u->xfer_buf;
fba76ee8 1155 boffset = io_u->offset;
81f0366c 1156 io_u->buf_filled_len = 0;
fad82f76 1157
23f394d5
JA
1158 for (i = 0; i < nr_blocks; i++) {
1159 /*
1160 * Fill the byte offset into a "random" start offset of
1161 * the buffer, given by the product of the usec time
1162 * and the actual offset.
1163 */
fad82f76 1164 offset = (io_u->start_time.tv_usec ^ boffset) & 511;
fba76ee8 1165 offset &= ~(sizeof(unsigned long long) - 1);
23f394d5
JA
1166 if (offset >= 512 - sizeof(unsigned long long))
1167 offset -= sizeof(unsigned long long);
fba76ee8 1168 memcpy(p + offset, &boffset, sizeof(boffset));
23f394d5
JA
1169
1170 end = p + 512 - sizeof(io_u->start_time);
1171 memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
1172 p += 512;
fad82f76 1173 boffset += 512;
23f394d5 1174 }
de789769
JA
1175}
1176
0d29de83
JA
1177/*
1178 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
1179 * etc. The returned io_u is fully ready to be prepped and submitted.
1180 */
1181struct io_u *get_io_u(struct thread_data *td)
1182{
1183 struct fio_file *f;
1184 struct io_u *io_u;
de789769 1185 int do_scramble = 0;
0d29de83
JA
1186
1187 io_u = __get_io_u(td);
1188 if (!io_u) {
1189 dprint(FD_IO, "__get_io_u failed\n");
1190 return NULL;
1191 }
1192
1193 if (check_get_verify(td, io_u))
1194 goto out;
1195 if (check_get_trim(td, io_u))
1196 goto out;
1197
755200a3
JA
1198 /*
1199 * from a requeue, io_u already setup
1200 */
1201 if (io_u->file)
77f392bf 1202 goto out;
755200a3 1203
429f6675
JA
1204 /*
1205 * If using an iolog, grab next piece if any available.
1206 */
1207 if (td->o.read_iolog_file) {
1208 if (read_iolog_get(td, io_u))
1209 goto err_put;
2ba1c290
JA
1210 } else if (set_io_u_file(td, io_u)) {
1211 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
429f6675 1212 goto err_put;
2ba1c290 1213 }
5ec10eaa 1214
429f6675 1215 f = io_u->file;
d6aed795 1216 assert(fio_file_open(f));
97af62ce 1217
ff58fced 1218 if (ddir_rw(io_u->ddir)) {
d0656a93 1219 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
2ba1c290 1220 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
429f6675 1221 goto err_put;
2ba1c290 1222 }
10ba535a 1223
38dad62d 1224 f->last_start = io_u->offset;
36167d82 1225 f->last_pos = io_u->offset + io_u->buflen;
10ba535a 1226
fd68418e
JA
1227 if (io_u->ddir == DDIR_WRITE) {
1228 if (td->o.verify != VERIFY_NONE)
1229 populate_verify_io_u(td, io_u);
1230 else if (td->o.refill_buffers)
1231 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
1232 else if (td->o.scramble_buffers)
1233 do_scramble = 1;
1234 } else if (io_u->ddir == DDIR_READ) {
cbe8d756
RR
1235 /*
1236 * Reset the buf_filled parameters so next time if the
1237 * buffer is used for writes it is refilled.
1238 */
cbe8d756
RR
1239 io_u->buf_filled_len = 0;
1240 }
87dc1ab1 1241 }
10ba535a 1242
165faf16
JA
1243 /*
1244 * Set io data pointers.
1245 */
cec6b55d
JA
1246 io_u->xfer_buf = io_u->buf;
1247 io_u->xfer_buflen = io_u->buflen;
5973cafb 1248
6ac7a331 1249out:
0d29de83 1250 assert(io_u->file);
429f6675 1251 if (!td_io_prep(td, io_u)) {
993bf48b
JA
1252 if (!td->o.disable_slat)
1253 fio_gettime(&io_u->start_time, NULL);
de789769
JA
1254 if (do_scramble)
1255 small_content_scramble(io_u);
429f6675 1256 return io_u;
36167d82 1257 }
429f6675 1258err_put:
2ba1c290 1259 dprint(FD_IO, "get_io_u failed\n");
429f6675
JA
1260 put_io_u(td, io_u);
1261 return NULL;
10ba535a
JA
1262}
1263
5451792e
JA
1264void io_u_log_error(struct thread_data *td, struct io_u *io_u)
1265{
825f818e
JA
1266 const char *msg[] = { "read", "write", "sync", "datasync",
1267 "sync_file_range", "wait", "trim" };
1268
1269
5451792e
JA
1270
1271 log_err("fio: io_u error");
1272
1273 if (io_u->file)
1274 log_err(" on file %s", io_u->file->file_name);
1275
1276 log_err(": %s\n", strerror(io_u->error));
1277
5ec10eaa
JA
1278 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
1279 io_u->offset, io_u->xfer_buflen);
5451792e
JA
1280
1281 if (!td->error)
1282 td_verror(td, io_u->error, "io_u error");
1283}
1284
c8eeb9df
JA
1285static void account_io_completion(struct thread_data *td, struct io_u *io_u,
1286 struct io_completion_data *icd,
1287 const enum fio_ddir idx, unsigned int bytes)
1288{
1289 unsigned long uninitialized_var(lusec);
1290
1291 if (!icd->account)
1292 return;
1293
1294 if (!td->o.disable_clat || !td->o.disable_bw)
1295 lusec = utime_since(&io_u->issue_time, &icd->time);
1296
1297 if (!td->o.disable_lat) {
1298 unsigned long tusec;
1299
1300 tusec = utime_since(&io_u->start_time, &icd->time);
1301 add_lat_sample(td, idx, tusec, bytes);
1302 }
1303
1304 if (!td->o.disable_clat) {
1305 add_clat_sample(td, idx, lusec, bytes);
1306 io_u_mark_latency(td, lusec);
1307 }
1308
1309 if (!td->o.disable_bw)
1310 add_bw_sample(td, idx, bytes, &icd->time);
1311
1312 add_iops_sample(td, idx, &icd->time);
1313}
1314
1b8dbf25
SL
1315static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
1316{
1317 unsigned long long secs, remainder, bps, bytes;
1318 bytes = td->this_io_bytes[ddir];
1319 bps = td->rate_bps[ddir];
1320 secs = bytes / bps;
1321 remainder = bytes % bps;
1322 return remainder * 1000000 / bps + secs * 1000000;
1323}
1324
97601024
JA
1325static void io_completed(struct thread_data *td, struct io_u *io_u,
1326 struct io_completion_data *icd)
10ba535a 1327{
dbad30b0
JA
1328 /*
1329 * Older gcc's are too dumb to realize that usec is always used
1330 * initialized, silence that warning.
1331 */
1332 unsigned long uninitialized_var(usec);
44f29692 1333 struct fio_file *f;
10ba535a 1334
2ba1c290
JA
1335 dprint_io_u(io_u, "io complete");
1336
2ecc1b57 1337 td_io_u_lock(td);
0c6e7517 1338 assert(io_u->flags & IO_U_F_FLIGHT);
38dad62d 1339 io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
2ecc1b57 1340 td_io_u_unlock(td);
0c6e7517 1341
5f9099ea 1342 if (ddir_sync(io_u->ddir)) {
87dc1ab1 1343 td->last_was_sync = 1;
44f29692
JA
1344 f = io_u->file;
1345 if (f) {
1346 f->first_write = -1ULL;
1347 f->last_write = -1ULL;
1348 }
87dc1ab1
JA
1349 return;
1350 }
1351
1352 td->last_was_sync = 0;
9e144189 1353 td->last_ddir = io_u->ddir;
87dc1ab1 1354
ff58fced 1355 if (!io_u->error && ddir_rw(io_u->ddir)) {
10ba535a 1356 unsigned int bytes = io_u->buflen - io_u->resid;
1e97cce9 1357 const enum fio_ddir idx = io_u->ddir;
ba3e4e0c 1358 const enum fio_ddir odx = io_u->ddir ^ 1;
b29ee5b3 1359 int ret;
10ba535a 1360
b29ee5b3 1361 td->io_blocks[idx]++;
c8eeb9df 1362 td->this_io_blocks[idx]++;
b29ee5b3
JA
1363 td->io_bytes[idx] += bytes;
1364 td->this_io_bytes[idx] += bytes;
10ba535a 1365
44f29692
JA
1366 if (idx == DDIR_WRITE) {
1367 f = io_u->file;
1368 if (f) {
1369 if (f->first_write == -1ULL ||
1370 io_u->offset < f->first_write)
1371 f->first_write = io_u->offset;
1372 if (f->last_write == -1ULL ||
1373 ((io_u->offset + bytes) > f->last_write))
1374 f->last_write = io_u->offset + bytes;
1375 }
1376 }
1377
6b1190fd
SL
1378 if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
1379 td->runstate == TD_VERIFYING)) {
c8eeb9df 1380 account_io_completion(td, io_u, icd, idx, bytes);
40e1a6f0 1381
b23b6a2f 1382 if (__should_check_rate(td, idx)) {
ba3e4e0c 1383 td->rate_pending_usleep[idx] =
1b8dbf25 1384 (usec_for_io(td, idx) -
ba3e4e0c 1385 utime_since_now(&td->start));
b23b6a2f 1386 }
1b8dbf25 1387 if (__should_check_rate(td, odx))
ba3e4e0c 1388 td->rate_pending_usleep[odx] =
1b8dbf25 1389 (usec_for_io(td, odx) -
ba3e4e0c 1390 utime_since_now(&td->start));
721938ae 1391 }
10ba535a 1392
660a1cb5 1393 if (td_write(td) && idx == DDIR_WRITE &&
8670579e 1394 td->o.do_verify &&
41128405 1395 td->o.verify != VERIFY_NONE)
10ba535a
JA
1396 log_io_piece(td, io_u);
1397
b29ee5b3 1398 icd->bytes_done[idx] += bytes;
3af6ef39 1399
d7762cf8 1400 if (io_u->end_io) {
36690c9b 1401 ret = io_u->end_io(td, io_u);
3af6ef39
JA
1402 if (ret && !icd->error)
1403 icd->error = ret;
1404 }
ff58fced 1405 } else if (io_u->error) {
10ba535a 1406 icd->error = io_u->error;
5451792e
JA
1407 io_u_log_error(td, io_u);
1408 }
06842027
SL
1409 if (icd->error && td_non_fatal_error(icd->error) &&
1410 (td->o.continue_on_error & td_error_type(io_u->ddir, icd->error))) {
f2bba182
RR
1411 /*
1412 * If there is a non_fatal error, then add to the error count
1413 * and clear all the errors.
1414 */
1415 update_error_count(td, icd->error);
1416 td_clear_error(td);
1417 icd->error = 0;
1418 io_u->error = 0;
1419 }
10ba535a
JA
1420}
1421
9520ebb9
JA
1422static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1423 int nr)
10ba535a 1424{
9520ebb9
JA
1425 if (!td->o.disable_clat || !td->o.disable_bw)
1426 fio_gettime(&icd->time, NULL);
02bcaa8c 1427
3af6ef39 1428 icd->nr = nr;
c8eeb9df 1429 icd->account = 1;
3af6ef39 1430
10ba535a
JA
1431 icd->error = 0;
1432 icd->bytes_done[0] = icd->bytes_done[1] = 0;
36167d82
JA
1433}
1434
97601024
JA
1435static void ios_completed(struct thread_data *td,
1436 struct io_completion_data *icd)
36167d82
JA
1437{
1438 struct io_u *io_u;
1439 int i;
1440
10ba535a
JA
1441 for (i = 0; i < icd->nr; i++) {
1442 io_u = td->io_ops->event(td, i);
1443
1444 io_completed(td, io_u, icd);
e8462bd8
JA
1445
1446 if (!(io_u->flags & IO_U_F_FREE_DEF))
1447 put_io_u(td, io_u);
c8eeb9df
JA
1448
1449 icd->account = 0;
10ba535a
JA
1450 }
1451}
97601024 1452
e7e6cfb4
JA
1453/*
1454 * Complete a single io_u for the sync engines.
1455 */
581e7141
JA
1456int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1457 unsigned long *bytes)
97601024
JA
1458{
1459 struct io_completion_data icd;
1460
9520ebb9 1461 init_icd(td, &icd, 1);
97601024 1462 io_completed(td, io_u, &icd);
e8462bd8
JA
1463
1464 if (!(io_u->flags & IO_U_F_FREE_DEF))
1465 put_io_u(td, io_u);
97601024 1466
581e7141
JA
1467 if (icd.error) {
1468 td_verror(td, icd.error, "io_u_sync_complete");
1469 return -1;
1470 }
97601024 1471
581e7141
JA
1472 if (bytes) {
1473 bytes[0] += icd.bytes_done[0];
1474 bytes[1] += icd.bytes_done[1];
1475 }
1476
1477 return 0;
97601024
JA
1478}
1479
e7e6cfb4
JA
1480/*
1481 * Called to complete min_events number of io for the async engines.
1482 */
581e7141
JA
1483int io_u_queued_complete(struct thread_data *td, int min_evts,
1484 unsigned long *bytes)
97601024 1485{
97601024 1486 struct io_completion_data icd;
00de55ef 1487 struct timespec *tvp = NULL;
97601024 1488 int ret;
4d06a338 1489 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
97601024 1490
4950421a 1491 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
b271fe62 1492
4950421a 1493 if (!min_evts)
00de55ef 1494 tvp = &ts;
97601024 1495
4950421a 1496 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
97601024 1497 if (ret < 0) {
e1161c32 1498 td_verror(td, -ret, "td_io_getevents");
97601024
JA
1499 return ret;
1500 } else if (!ret)
1501 return ret;
1502
9520ebb9 1503 init_icd(td, &icd, ret);
97601024 1504 ios_completed(td, &icd);
581e7141
JA
1505 if (icd.error) {
1506 td_verror(td, icd.error, "io_u_queued_complete");
1507 return -1;
1508 }
97601024 1509
581e7141
JA
1510 if (bytes) {
1511 bytes[0] += icd.bytes_done[0];
1512 bytes[1] += icd.bytes_done[1];
1513 }
1514
1515 return 0;
97601024 1516}
7e77dd02
JA
1517
1518/*
1519 * Call when io_u is really queued, to update the submission latency.
1520 */
1521void io_u_queued(struct thread_data *td, struct io_u *io_u)
1522{
9520ebb9
JA
1523 if (!td->o.disable_slat) {
1524 unsigned long slat_time;
7e77dd02 1525
9520ebb9 1526 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
29a90ddb 1527 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
9520ebb9 1528 }
7e77dd02 1529}
433afcb4 1530
5973cafb
JA
1531/*
1532 * "randomly" fill the buffer contents
1533 */
1534void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1535 unsigned int max_bs)
1536{
e84b371e
JA
1537 io_u->buf_filled_len = 0;
1538
637ef8d9 1539 if (!td->o.zero_buffers)
3545a109 1540 fill_random_buf(&td->buf_state, io_u->buf, max_bs);
637ef8d9
JA
1541 else
1542 memset(io_u->buf, 0, max_bs);
5973cafb 1543}