pre_read fixes
[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"
10ba535a 10
97601024
JA
11struct io_completion_data {
12 int nr; /* input */
97601024
JA
13
14 int error; /* output */
15 unsigned long bytes_done[2]; /* output */
16 struct timeval time; /* output */
17};
18
10ba535a
JA
19/*
20 * The ->file_map[] contains a map of blocks we have or have not done io
21 * to yet. Used to make sure we cover the entire range in a fair fashion.
22 */
aec2de20 23static int random_map_free(struct fio_file *f, const unsigned long long block)
10ba535a 24{
aec2de20
JA
25 unsigned int idx = RAND_MAP_IDX(f, block);
26 unsigned int bit = RAND_MAP_BIT(f, block);
10ba535a 27
84422acd
JA
28 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
29
de605666 30 return (f->file_map[idx] & (1 << bit)) == 0;
10ba535a
JA
31}
32
df415585
JA
33/*
34 * Mark a given offset as used in the map.
35 */
9bf2061e 36static void mark_random_map(struct thread_data *td, struct io_u *io_u)
df415585 37{
2dc1bbeb 38 unsigned int min_bs = td->o.rw_min_bs;
9bf2061e 39 struct fio_file *f = io_u->file;
a00735e6 40 unsigned long long block;
3e3357b1 41 unsigned int blocks, nr_blocks;
df415585 42
b9c5b644 43 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
c685b5b2 44 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
3e3357b1 45 blocks = 0;
c685b5b2 46
3e3357b1
JA
47 while (nr_blocks) {
48 unsigned int this_blocks, mask;
df415585
JA
49 unsigned int idx, bit;
50
1e3d53ac
JA
51 /*
52 * If we have a mixed random workload, we may
53 * encounter blocks we already did IO to.
54 */
8b113b4f
JA
55 if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) {
56 if (!blocks)
57 blocks = 1;
df415585 58 break;
8b113b4f 59 }
df415585 60
aec2de20
JA
61 idx = RAND_MAP_IDX(f, block);
62 bit = RAND_MAP_BIT(f, block);
df415585 63
0032bf9f 64 fio_assert(td, idx < f->num_maps);
df415585 65
3e3357b1
JA
66 this_blocks = nr_blocks;
67 if (this_blocks + bit > BLOCKS_PER_MAP)
68 this_blocks = BLOCKS_PER_MAP - bit;
69
70 if (this_blocks == BLOCKS_PER_MAP)
71 mask = -1U;
72 else
73 mask = ((1U << this_blocks) - 1) << bit;
74
3e3357b1
JA
75 f->file_map[idx] |= mask;
76 nr_blocks -= this_blocks;
77 blocks += this_blocks;
c9dd34b2 78 block += this_blocks;
df415585
JA
79 }
80
a00735e6
JA
81 if ((blocks * min_bs) < io_u->buflen)
82 io_u->buflen = blocks * min_bs;
df415585
JA
83}
84
3e3357b1
JA
85static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
86 enum fio_ddir ddir)
2ba1c290
JA
87{
88 unsigned long long max_blocks;
d9dd70f7 89 unsigned long long max_size;
2ba1c290 90
d9dd70f7
JA
91 /*
92 * Hmm, should we make sure that ->io_size <= ->real_file_size?
93 */
94 max_size = f->io_size;
95 if (max_size > f->real_file_size)
96 max_size = f->real_file_size;
97
2b7a01d0 98 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
2ba1c290
JA
99 if (!max_blocks)
100 return 0;
101
67778e88 102 return max_blocks;
2ba1c290
JA
103}
104
10ba535a
JA
105/*
106 * Return the next free block in the map.
107 */
108static int get_next_free_block(struct thread_data *td, struct fio_file *f,
4ba66134 109 enum fio_ddir ddir, unsigned long long *b)
10ba535a 110{
84422acd 111 unsigned long long min_bs = td->o.rw_min_bs;
10ba535a
JA
112 int i;
113
c685b5b2
JA
114 i = f->last_free_lookup;
115 *b = (i * BLOCKS_PER_MAP);
899c29f2
RR
116 while ((*b) * min_bs < f->real_file_size &&
117 (*b) * min_bs < f->io_size) {
de605666 118 if (f->file_map[i] != (unsigned int) -1) {
697a606c 119 *b += ffz(f->file_map[i]);
4ba66134 120 if (*b > last_block(td, f, ddir))
2ba1c290 121 break;
c685b5b2 122 f->last_free_lookup = i;
10ba535a
JA
123 return 0;
124 }
125
126 *b += BLOCKS_PER_MAP;
127 i++;
128 }
129
2ba1c290 130 dprint(FD_IO, "failed finding a free block\n");
10ba535a
JA
131 return 1;
132}
133
ec4015da 134static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
4ba66134 135 enum fio_ddir ddir, unsigned long long *b)
ec4015da 136{
84422acd 137 unsigned long long r;
ec4015da
JA
138 int loops = 5;
139
140 do {
141 r = os_random_long(&td->random_state);
84422acd 142 dprint(FD_RANDOM, "off rand %llu\n", r);
5ec10eaa 143 *b = (last_block(td, f, ddir) - 1)
dc873b6f 144 * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
2ba1c290 145
43c63a78
JA
146 /*
147 * if we are not maintaining a random map, we are done.
148 */
303032ae 149 if (!file_randommap(td, f))
43c63a78
JA
150 return 0;
151
152 /*
84422acd 153 * calculate map offset and check if it's free
43c63a78 154 */
aec2de20 155 if (random_map_free(f, *b))
43c63a78
JA
156 return 0;
157
84422acd
JA
158 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
159 *b);
43c63a78 160 } while (--loops);
ec4015da
JA
161
162 /*
43c63a78
JA
163 * we get here, if we didn't suceed in looking up a block. generate
164 * a random start offset into the filemap, and find the first free
165 * block from there.
ec4015da 166 */
43c63a78
JA
167 loops = 10;
168 do {
dc873b6f
JA
169 f->last_free_lookup = (f->num_maps - 1) *
170 (r / (OS_RAND_MAX + 1.0));
4ba66134 171 if (!get_next_free_block(td, f, ddir, b))
43c63a78 172 return 0;
ec4015da 173
43c63a78
JA
174 r = os_random_long(&td->random_state);
175 } while (--loops);
176
177 /*
178 * that didn't work either, try exhaustive search from the start
179 */
180 f->last_free_lookup = 0;
4ba66134 181 return get_next_free_block(td, f, ddir, b);
ec4015da
JA
182}
183
10ba535a
JA
184/*
185 * For random io, generate a random new block and see if it's used. Repeat
186 * until we find a free one. For sequential io, just return the end of
187 * the last io issued.
188 */
9bf2061e 189static int get_next_offset(struct thread_data *td, struct io_u *io_u)
10ba535a 190{
9bf2061e 191 struct fio_file *f = io_u->file;
ec4015da 192 unsigned long long b;
4ba66134 193 enum fio_ddir ddir = io_u->ddir;
10ba535a 194
ec4015da
JA
195 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
196 td->ddir_nr = td->o.ddir_nr;
211097b2 197
683023e8
JA
198 if (get_next_rand_offset(td, f, ddir, &b)) {
199 dprint(FD_IO, "%s: getting rand offset failed\n",
200 f->file_name);
bca4ed4d 201 return 1;
683023e8 202 }
43063a1c 203 } else {
0c3d768a 204 if (f->last_pos >= f->real_file_size) {
4ba66134 205 if (!td_random(td) ||
683023e8
JA
206 get_next_rand_offset(td, f, ddir, &b)) {
207 dprint(FD_IO, "%s: pos %llu > size %llu\n",
208 f->file_name, f->last_pos,
209 f->real_file_size);
0c3d768a 210 return 1;
683023e8 211 }
bcdedd0a 212 } else
4ba66134 213 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
43063a1c 214 }
10ba535a 215
2b7a01d0 216 io_u->offset = b * td->o.ba[ddir];
009bd847
JA
217 if (io_u->offset >= f->io_size) {
218 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
219 io_u->offset, f->io_size);
220 return 1;
221 }
222
223 io_u->offset += f->file_offset;
2ba1c290
JA
224 if (io_u->offset >= f->real_file_size) {
225 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
226 io_u->offset, f->real_file_size);
10ba535a 227 return 1;
2ba1c290 228 }
10ba535a
JA
229
230 return 0;
231}
232
f3059de1
JA
233static inline int is_power_of_2(unsigned int val)
234{
235 return (val != 0 && ((val & (val - 1)) == 0));
236}
237
9bf2061e 238static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
10ba535a 239{
bca4ed4d 240 const int ddir = io_u->ddir;
f3f552b9 241 unsigned int uninitialized_var(buflen);
f3059de1 242 unsigned int minbs, maxbs;
10ba535a
JA
243 long r;
244
f3059de1
JA
245 minbs = td->o.min_bs[ddir];
246 maxbs = td->o.max_bs[ddir];
247
248 if (minbs == maxbs)
249 buflen = minbs;
10ba535a
JA
250 else {
251 r = os_random_long(&td->bsrange_state);
720e84ad 252 if (!td->o.bssplit_nr[ddir]) {
f3059de1
JA
253 buflen = 1 + (unsigned int) ((double) maxbs *
254 (r / (OS_RAND_MAX + 1.0)));
255 if (buflen < minbs)
256 buflen = minbs;
5ec10eaa 257 } else {
564ca972
JA
258 long perc = 0;
259 unsigned int i;
260
720e84ad
JA
261 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
262 struct bssplit *bsp = &td->o.bssplit[ddir][i];
564ca972
JA
263
264 buflen = bsp->bs;
265 perc += bsp->perc;
f3059de1 266 if (r <= ((OS_RAND_MAX / 100L) * perc))
564ca972
JA
267 break;
268 }
269 }
f3059de1
JA
270 if (!td->o.bs_unaligned && is_power_of_2(minbs))
271 buflen = (buflen + minbs - 1) & ~(minbs - 1);
10ba535a
JA
272 }
273
4ba66134
JA
274 if (io_u->offset + buflen > io_u->file->real_file_size) {
275 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
f3059de1
JA
276 minbs, ddir);
277 buflen = minbs;
4ba66134 278 }
6a5e6884 279
10ba535a
JA
280 return buflen;
281}
282
afe24a5a
JA
283static void set_rwmix_bytes(struct thread_data *td)
284{
afe24a5a
JA
285 unsigned int diff;
286
287 /*
288 * we do time or byte based switch. this is needed because
289 * buffered writes may issue a lot quicker than they complete,
290 * whereas reads do not.
291 */
e47f799f 292 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
04c540d9 293 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
e47f799f
JA
294}
295
296static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
297{
298 unsigned int v;
299 long r;
300
301 r = os_random_long(&td->rwmix_state);
dc873b6f 302 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
04c540d9 303 if (v <= td->o.rwmix[DDIR_READ])
e47f799f
JA
304 return DDIR_READ;
305
306 return DDIR_WRITE;
afe24a5a
JA
307}
308
10ba535a
JA
309/*
310 * Return the data direction for the next io_u. If the job is a
311 * mixed read/write workload, check the rwmix cycle and switch if
312 * necessary.
313 */
1e97cce9 314static enum fio_ddir get_rw_ddir(struct thread_data *td)
10ba535a
JA
315{
316 if (td_rw(td)) {
10ba535a
JA
317 /*
318 * Check if it's time to seed a new data direction.
319 */
e4928662 320 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
e47f799f 321 unsigned long long max_bytes;
5ec10eaa 322 enum fio_ddir ddir;
e47f799f
JA
323
324 /*
325 * Put a top limit on how many bytes we do for
326 * one data direction, to avoid overflowing the
327 * ranges too much
328 */
329 ddir = get_rand_ddir(td);
330 max_bytes = td->this_io_bytes[ddir];
5ec10eaa
JA
331 if (max_bytes >=
332 (td->o.size * td->o.rwmix[ddir] / 100)) {
6162f524 333 if (!td->rw_end_set[ddir]) {
38d77cae 334 td->rw_end_set[ddir] = 1;
6162f524
JA
335 fio_gettime(&td->rw_end[ddir], NULL);
336 }
e4928662 337
e47f799f 338 ddir ^= 1;
38d77cae 339 }
e47f799f
JA
340
341 if (ddir != td->rwmix_ddir)
342 set_rwmix_bytes(td);
343
344 td->rwmix_ddir = ddir;
10ba535a
JA
345 }
346 return td->rwmix_ddir;
347 } else if (td_read(td))
348 return DDIR_READ;
349 else
350 return DDIR_WRITE;
351}
352
60f2c658
JA
353static void put_file_log(struct thread_data *td, struct fio_file *f)
354{
355 int ret = put_file(td, f);
356
357 if (ret)
358 td_verror(td, ret, "file close");
359}
360
10ba535a
JA
361void put_io_u(struct thread_data *td, struct io_u *io_u)
362{
0c6e7517
JA
363 assert((io_u->flags & IO_U_F_FREE) == 0);
364 io_u->flags |= IO_U_F_FREE;
365
60f2c658
JA
366 if (io_u->file)
367 put_file_log(td, io_u->file);
2dbdab7e 368
10ba535a 369 io_u->file = NULL;
01743ee1
JA
370 flist_del(&io_u->list);
371 flist_add(&io_u->list, &td->io_u_freelist);
10ba535a
JA
372 td->cur_depth--;
373}
374
755200a3
JA
375void requeue_io_u(struct thread_data *td, struct io_u **io_u)
376{
377 struct io_u *__io_u = *io_u;
378
465221b0
JA
379 dprint(FD_IO, "requeue %p\n", __io_u);
380
4d2e0f49 381 __io_u->flags |= IO_U_F_FREE;
e4f54adb
JA
382 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
383 td->io_issues[__io_u->ddir]--;
5ec10eaa 384
4d2e0f49
JA
385 __io_u->flags &= ~IO_U_F_FLIGHT;
386
01743ee1
JA
387 flist_del(&__io_u->list);
388 flist_add_tail(&__io_u->list, &td->io_u_requeues);
755200a3
JA
389 td->cur_depth--;
390 *io_u = NULL;
391}
392
9bf2061e 393static int fill_io_u(struct thread_data *td, struct io_u *io_u)
10ba535a 394{
b4c5e1ac
JA
395 if (td->io_ops->flags & FIO_NOIO)
396 goto out;
397
87dc1ab1
JA
398 /*
399 * see if it's time to sync
400 */
2dc1bbeb
JA
401 if (td->o.fsync_blocks &&
402 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
403 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
87dc1ab1 404 io_u->ddir = DDIR_SYNC;
c38e9468 405 goto out;
87dc1ab1
JA
406 }
407
a00735e6
JA
408 io_u->ddir = get_rw_ddir(td);
409
48f5abd3
JA
410 /*
411 * See if it's time to switch to a new zone
412 */
413 if (td->zone_bytes >= td->o.zone_size) {
414 td->zone_bytes = 0;
415 io_u->file->last_pos += td->o.zone_skip;
416 td->io_skip_bytes += td->o.zone_skip;
417 }
418
10ba535a 419 /*
c685b5b2
JA
420 * No log, let the seq/rand engine retrieve the next buflen and
421 * position.
10ba535a 422 */
2ba1c290
JA
423 if (get_next_offset(td, io_u)) {
424 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
bca4ed4d 425 return 1;
2ba1c290 426 }
10ba535a 427
9bf2061e 428 io_u->buflen = get_next_buflen(td, io_u);
2ba1c290
JA
429 if (!io_u->buflen) {
430 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
bca4ed4d 431 return 1;
2ba1c290 432 }
bca4ed4d 433
2ba1c290
JA
434 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
435 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
4ba66134
JA
436 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
437 io_u->buflen, io_u->file->real_file_size);
6a5e6884 438 return 1;
2ba1c290 439 }
6a5e6884 440
bca4ed4d
JA
441 /*
442 * mark entry before potentially trimming io_u
443 */
303032ae 444 if (td_random(td) && file_randommap(td, io_u->file))
9bf2061e 445 mark_random_map(td, io_u);
bca4ed4d
JA
446
447 /*
448 * If using a write iolog, store this entry.
449 */
c38e9468 450out:
2ba1c290 451 dprint_io_u(io_u, "fill_io_u");
d9d91e39 452 td->zone_bytes += io_u->buflen;
f29b25a3 453 log_io_u(td, io_u);
bca4ed4d 454 return 0;
10ba535a
JA
455}
456
838bc709
JA
457static void __io_u_mark_map(unsigned int *map, unsigned int nr)
458{
459 int index = 0;
460
461 switch (nr) {
462 default:
463 index = 6;
464 break;
465 case 33 ... 64:
466 index = 5;
467 break;
468 case 17 ... 32:
469 index = 4;
470 break;
471 case 9 ... 16:
472 index = 3;
473 break;
474 case 5 ... 8:
475 index = 2;
476 break;
477 case 1 ... 4:
478 index = 1;
479 case 0:
480 break;
481 }
482
483 map[index]++;
484}
485
486void io_u_mark_submit(struct thread_data *td, unsigned int nr)
487{
488 __io_u_mark_map(td->ts.io_u_submit, nr);
489 td->ts.total_submit++;
490}
491
492void io_u_mark_complete(struct thread_data *td, unsigned int nr)
493{
494 __io_u_mark_map(td->ts.io_u_complete, nr);
495 td->ts.total_complete++;
496}
497
d8005759 498void io_u_mark_depth(struct thread_data *td, unsigned int nr)
71619dc2
JA
499{
500 int index = 0;
501
502 switch (td->cur_depth) {
503 default:
a783e61a
JA
504 index = 6;
505 break;
71619dc2 506 case 32 ... 63:
a783e61a
JA
507 index = 5;
508 break;
71619dc2 509 case 16 ... 31:
a783e61a
JA
510 index = 4;
511 break;
71619dc2 512 case 8 ... 15:
a783e61a
JA
513 index = 3;
514 break;
71619dc2 515 case 4 ... 7:
a783e61a
JA
516 index = 2;
517 break;
71619dc2 518 case 2 ... 3:
a783e61a 519 index = 1;
71619dc2
JA
520 case 1:
521 break;
522 }
523
3bec7ae0 524 td->ts.io_u_map[index] += nr;
71619dc2
JA
525}
526
04a0feae
JA
527static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
528{
529 int index = 0;
530
531 assert(usec < 1000);
532
533 switch (usec) {
534 case 750 ... 999:
535 index = 9;
536 break;
537 case 500 ... 749:
538 index = 8;
539 break;
540 case 250 ... 499:
541 index = 7;
542 break;
543 case 100 ... 249:
544 index = 6;
545 break;
546 case 50 ... 99:
547 index = 5;
548 break;
549 case 20 ... 49:
550 index = 4;
551 break;
552 case 10 ... 19:
553 index = 3;
554 break;
555 case 4 ... 9:
556 index = 2;
557 break;
558 case 2 ... 3:
559 index = 1;
560 case 0 ... 1:
561 break;
562 }
563
564 assert(index < FIO_IO_U_LAT_U_NR);
565 td->ts.io_u_lat_u[index]++;
566}
567
568static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
ec118304
JA
569{
570 int index = 0;
571
572 switch (msec) {
573 default:
04a0feae
JA
574 index = 11;
575 break;
8abdce66 576 case 1000 ... 1999:
04a0feae
JA
577 index = 10;
578 break;
8abdce66 579 case 750 ... 999:
04a0feae
JA
580 index = 9;
581 break;
8abdce66 582 case 500 ... 749:
04a0feae
JA
583 index = 8;
584 break;
8abdce66 585 case 250 ... 499:
04a0feae
JA
586 index = 7;
587 break;
8abdce66 588 case 100 ... 249:
04a0feae
JA
589 index = 6;
590 break;
8abdce66 591 case 50 ... 99:
04a0feae
JA
592 index = 5;
593 break;
8abdce66 594 case 20 ... 49:
04a0feae
JA
595 index = 4;
596 break;
8abdce66 597 case 10 ... 19:
04a0feae
JA
598 index = 3;
599 break;
8abdce66 600 case 4 ... 9:
04a0feae
JA
601 index = 2;
602 break;
ec118304 603 case 2 ... 3:
04a0feae 604 index = 1;
ec118304
JA
605 case 0 ... 1:
606 break;
607 }
608
04a0feae
JA
609 assert(index < FIO_IO_U_LAT_M_NR);
610 td->ts.io_u_lat_m[index]++;
611}
612
613static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
614{
615 if (usec < 1000)
616 io_u_mark_lat_usec(td, usec);
617 else
618 io_u_mark_lat_msec(td, usec / 1000);
ec118304
JA
619}
620
0aabe160
JA
621/*
622 * Get next file to service by choosing one at random
623 */
1c178180
JA
624static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
625 int badf)
0aabe160 626{
0aabe160 627 struct fio_file *f;
1c178180 628 int fno;
0aabe160
JA
629
630 do {
7c83c089 631 long r = os_random_long(&td->next_file_state);
87b10676 632 int opened = 0;
7c83c089 633
5ec10eaa 634 fno = (unsigned int) ((double) td->o.nr_files
dc873b6f 635 * (r / (OS_RAND_MAX + 1.0)));
126d65c6 636 f = td->files[fno];
059e63c0
JA
637 if (f->flags & FIO_FILE_DONE)
638 continue;
1c178180 639
87b10676
JA
640 if (!(f->flags & FIO_FILE_OPEN)) {
641 int err;
642
643 err = td_io_open_file(td, f);
644 if (err)
645 continue;
646 opened = 1;
647 }
648
2ba1c290
JA
649 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
650 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
0aabe160 651 return f;
2ba1c290 652 }
87b10676
JA
653 if (opened)
654 td_io_close_file(td, f);
0aabe160
JA
655 } while (1);
656}
657
658/*
659 * Get next file to service by doing round robin between all available ones
660 */
1c178180
JA
661static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
662 int badf)
3d7c391d
JA
663{
664 unsigned int old_next_file = td->next_file;
665 struct fio_file *f;
666
667 do {
87b10676
JA
668 int opened = 0;
669
126d65c6 670 f = td->files[td->next_file];
3d7c391d
JA
671
672 td->next_file++;
2dc1bbeb 673 if (td->next_file >= td->o.nr_files)
3d7c391d
JA
674 td->next_file = 0;
675
87b10676 676 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
d5ed68ea
JA
677 if (f->flags & FIO_FILE_DONE) {
678 f = NULL;
059e63c0 679 continue;
d5ed68ea 680 }
059e63c0 681
87b10676
JA
682 if (!(f->flags & FIO_FILE_OPEN)) {
683 int err;
684
685 err = td_io_open_file(td, f);
b5696bfc
JA
686 if (err) {
687 dprint(FD_FILE, "error %d on open of %s\n",
688 err, f->file_name);
87c27b45 689 f = NULL;
87b10676 690 continue;
b5696bfc 691 }
87b10676
JA
692 opened = 1;
693 }
694
b5696bfc 695 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, f->flags);
1c178180 696 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
3d7c391d
JA
697 break;
698
87b10676
JA
699 if (opened)
700 td_io_close_file(td, f);
701
3d7c391d
JA
702 f = NULL;
703 } while (td->next_file != old_next_file);
704
2ba1c290 705 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
3d7c391d
JA
706 return f;
707}
708
bdb4e2e9
JA
709static struct fio_file *get_next_file(struct thread_data *td)
710{
1907dbc6
JA
711 struct fio_file *f;
712
2dc1bbeb 713 assert(td->o.nr_files <= td->files_index);
1c178180 714
b5696bfc 715 if (td->nr_done_files >= td->o.nr_files) {
5ec10eaa
JA
716 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
717 " nr_files=%d\n", td->nr_open_files,
718 td->nr_done_files,
719 td->o.nr_files);
bdb4e2e9 720 return NULL;
2ba1c290 721 }
bdb4e2e9 722
1907dbc6 723 f = td->file_service_file;
683023e8 724 if (f && (f->flags & FIO_FILE_OPEN) && !(f->flags & FIO_FILE_CLOSING)) {
a086c257
JA
725 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
726 goto out;
727 if (td->file_service_left--)
728 goto out;
729 }
1907dbc6 730
a086c257
JA
731 if (td->o.file_service_type == FIO_FSERVICE_RR ||
732 td->o.file_service_type == FIO_FSERVICE_SEQ)
1c178180 733 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
bdb4e2e9 734 else
1c178180 735 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
1907dbc6
JA
736
737 td->file_service_file = f;
738 td->file_service_left = td->file_service_nr - 1;
2ba1c290 739out:
683023e8 740 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1907dbc6 741 return f;
bdb4e2e9
JA
742}
743
429f6675
JA
744static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
745{
746 struct fio_file *f;
747
748 do {
749 f = get_next_file(td);
750 if (!f)
751 return 1;
752
429f6675
JA
753 io_u->file = f;
754 get_file(f);
755
756 if (!fill_io_u(td, io_u))
757 break;
758
b5696bfc 759 put_file_log(td, f);
429f6675 760 td_io_close_file(td, f);
b5696bfc 761 io_u->file = NULL;
429f6675
JA
762 f->flags |= FIO_FILE_DONE;
763 td->nr_done_files++;
87b10676 764 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, td->nr_done_files, td->o.nr_files);
429f6675
JA
765 } while (1);
766
767 return 0;
768}
769
770
10ba535a
JA
771struct io_u *__get_io_u(struct thread_data *td)
772{
773 struct io_u *io_u = NULL;
774
01743ee1
JA
775 if (!flist_empty(&td->io_u_requeues))
776 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
755200a3 777 else if (!queue_full(td)) {
01743ee1 778 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
10ba535a 779
6040dabc 780 io_u->buflen = 0;
10ba535a 781 io_u->resid = 0;
755200a3 782 io_u->file = NULL;
d7762cf8 783 io_u->end_io = NULL;
755200a3
JA
784 }
785
786 if (io_u) {
0c6e7517
JA
787 assert(io_u->flags & IO_U_F_FREE);
788 io_u->flags &= ~IO_U_F_FREE;
789
755200a3 790 io_u->error = 0;
01743ee1
JA
791 flist_del(&io_u->list);
792 flist_add(&io_u->list, &td->io_u_busylist);
10ba535a
JA
793 td->cur_depth++;
794 }
795
796 return io_u;
797}
798
799/*
800 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
801 * etc. The returned io_u is fully ready to be prepped and submitted.
802 */
3d7c391d 803struct io_u *get_io_u(struct thread_data *td)
10ba535a 804{
3d7c391d 805 struct fio_file *f;
10ba535a
JA
806 struct io_u *io_u;
807
808 io_u = __get_io_u(td);
2ba1c290
JA
809 if (!io_u) {
810 dprint(FD_IO, "__get_io_u failed\n");
10ba535a 811 return NULL;
2ba1c290 812 }
10ba535a 813
755200a3
JA
814 /*
815 * from a requeue, io_u already setup
816 */
817 if (io_u->file)
77f392bf 818 goto out;
755200a3 819
429f6675
JA
820 /*
821 * If using an iolog, grab next piece if any available.
822 */
823 if (td->o.read_iolog_file) {
824 if (read_iolog_get(td, io_u))
825 goto err_put;
2ba1c290
JA
826 } else if (set_io_u_file(td, io_u)) {
827 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
429f6675 828 goto err_put;
2ba1c290 829 }
5ec10eaa 830
429f6675
JA
831 f = io_u->file;
832 assert(f->flags & FIO_FILE_OPEN);
97af62ce 833
87dc1ab1 834 if (io_u->ddir != DDIR_SYNC) {
d0656a93 835 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
2ba1c290 836 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
429f6675 837 goto err_put;
2ba1c290 838 }
10ba535a 839
36167d82 840 f->last_pos = io_u->offset + io_u->buflen;
10ba535a 841
c311cd2a 842 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
87dc1ab1 843 populate_verify_io_u(td, io_u);
e4dad9c6
JA
844 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
845 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
87dc1ab1 846 }
10ba535a 847
165faf16
JA
848 /*
849 * Set io data pointers.
850 */
d460eb31 851 io_u->endpos = io_u->offset + io_u->buflen;
cec6b55d
JA
852 io_u->xfer_buf = io_u->buf;
853 io_u->xfer_buflen = io_u->buflen;
5973cafb 854
6ac7a331 855out:
429f6675 856 if (!td_io_prep(td, io_u)) {
993bf48b
JA
857 if (!td->o.disable_slat)
858 fio_gettime(&io_u->start_time, NULL);
429f6675 859 return io_u;
36167d82 860 }
429f6675 861err_put:
2ba1c290 862 dprint(FD_IO, "get_io_u failed\n");
429f6675
JA
863 put_io_u(td, io_u);
864 return NULL;
10ba535a
JA
865}
866
5451792e
JA
867void io_u_log_error(struct thread_data *td, struct io_u *io_u)
868{
869 const char *msg[] = { "read", "write", "sync" };
870
871 log_err("fio: io_u error");
872
873 if (io_u->file)
874 log_err(" on file %s", io_u->file->file_name);
875
876 log_err(": %s\n", strerror(io_u->error));
877
5ec10eaa
JA
878 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
879 io_u->offset, io_u->xfer_buflen);
5451792e
JA
880
881 if (!td->error)
882 td_verror(td, io_u->error, "io_u error");
883}
884
97601024
JA
885static void io_completed(struct thread_data *td, struct io_u *io_u,
886 struct io_completion_data *icd)
10ba535a 887{
dbad30b0
JA
888 /*
889 * Older gcc's are too dumb to realize that usec is always used
890 * initialized, silence that warning.
891 */
892 unsigned long uninitialized_var(usec);
10ba535a 893
2ba1c290
JA
894 dprint_io_u(io_u, "io complete");
895
0c6e7517
JA
896 assert(io_u->flags & IO_U_F_FLIGHT);
897 io_u->flags &= ~IO_U_F_FLIGHT;
898
87dc1ab1
JA
899 if (io_u->ddir == DDIR_SYNC) {
900 td->last_was_sync = 1;
901 return;
902 }
903
904 td->last_was_sync = 0;
905
10ba535a
JA
906 if (!io_u->error) {
907 unsigned int bytes = io_u->buflen - io_u->resid;
1e97cce9 908 const enum fio_ddir idx = io_u->ddir;
b29ee5b3 909 int ret;
10ba535a 910
b29ee5b3
JA
911 td->io_blocks[idx]++;
912 td->io_bytes[idx] += bytes;
913 td->this_io_bytes[idx] += bytes;
10ba535a 914
b29ee5b3 915 if (ramp_time_over(td)) {
9520ebb9
JA
916 if (!td->o.disable_clat || !td->o.disable_bw)
917 usec = utime_since(&io_u->issue_time,
918 &icd->time);
721938ae 919
9520ebb9 920 if (!td->o.disable_clat) {
306ddc97 921 add_clat_sample(td, idx, usec, bytes);
9520ebb9
JA
922 io_u_mark_latency(td, usec);
923 }
924 if (!td->o.disable_bw)
306ddc97 925 add_bw_sample(td, idx, bytes, &icd->time);
721938ae 926 }
10ba535a 927
660a1cb5 928 if (td_write(td) && idx == DDIR_WRITE &&
8670579e 929 td->o.do_verify &&
41128405 930 td->o.verify != VERIFY_NONE)
10ba535a
JA
931 log_io_piece(td, io_u);
932
b29ee5b3 933 icd->bytes_done[idx] += bytes;
3af6ef39 934
d7762cf8 935 if (io_u->end_io) {
36690c9b 936 ret = io_u->end_io(td, io_u);
3af6ef39
JA
937 if (ret && !icd->error)
938 icd->error = ret;
939 }
5451792e 940 } else {
10ba535a 941 icd->error = io_u->error;
5451792e
JA
942 io_u_log_error(td, io_u);
943 }
10ba535a
JA
944}
945
9520ebb9
JA
946static void init_icd(struct thread_data *td, struct io_completion_data *icd,
947 int nr)
10ba535a 948{
9520ebb9
JA
949 if (!td->o.disable_clat || !td->o.disable_bw)
950 fio_gettime(&icd->time, NULL);
02bcaa8c 951
3af6ef39
JA
952 icd->nr = nr;
953
10ba535a
JA
954 icd->error = 0;
955 icd->bytes_done[0] = icd->bytes_done[1] = 0;
36167d82
JA
956}
957
97601024
JA
958static void ios_completed(struct thread_data *td,
959 struct io_completion_data *icd)
36167d82
JA
960{
961 struct io_u *io_u;
962 int i;
963
10ba535a
JA
964 for (i = 0; i < icd->nr; i++) {
965 io_u = td->io_ops->event(td, i);
966
967 io_completed(td, io_u, icd);
968 put_io_u(td, io_u);
969 }
970}
97601024 971
e7e6cfb4
JA
972/*
973 * Complete a single io_u for the sync engines.
974 */
d7762cf8 975long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
97601024
JA
976{
977 struct io_completion_data icd;
978
9520ebb9 979 init_icd(td, &icd, 1);
97601024
JA
980 io_completed(td, io_u, &icd);
981 put_io_u(td, io_u);
982
983 if (!icd.error)
984 return icd.bytes_done[0] + icd.bytes_done[1];
985
37e974a5 986 td_verror(td, icd.error, "io_u_sync_complete");
97601024
JA
987 return -1;
988}
989
e7e6cfb4
JA
990/*
991 * Called to complete min_events number of io for the async engines.
992 */
4950421a 993long io_u_queued_complete(struct thread_data *td, int min_evts)
97601024 994{
97601024 995 struct io_completion_data icd;
00de55ef 996 struct timespec *tvp = NULL;
97601024 997 int ret;
4d06a338 998 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
97601024 999
4950421a 1000 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
b271fe62 1001
4950421a 1002 if (!min_evts)
00de55ef 1003 tvp = &ts;
97601024 1004
4950421a 1005 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
97601024 1006 if (ret < 0) {
e1161c32 1007 td_verror(td, -ret, "td_io_getevents");
97601024
JA
1008 return ret;
1009 } else if (!ret)
1010 return ret;
1011
9520ebb9 1012 init_icd(td, &icd, ret);
97601024
JA
1013 ios_completed(td, &icd);
1014 if (!icd.error)
1015 return icd.bytes_done[0] + icd.bytes_done[1];
1016
37e974a5 1017 td_verror(td, icd.error, "io_u_queued_complete");
97601024
JA
1018 return -1;
1019}
7e77dd02
JA
1020
1021/*
1022 * Call when io_u is really queued, to update the submission latency.
1023 */
1024void io_u_queued(struct thread_data *td, struct io_u *io_u)
1025{
9520ebb9
JA
1026 if (!td->o.disable_slat) {
1027 unsigned long slat_time;
7e77dd02 1028
9520ebb9 1029 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
306ddc97 1030 add_slat_sample(td, io_u->ddir, io_u->xfer_buflen, slat_time);
9520ebb9 1031 }
7e77dd02 1032}
433afcb4 1033
5973cafb
JA
1034/*
1035 * "randomly" fill the buffer contents
1036 */
1037void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1038 unsigned int max_bs)
1039{
1040 long *ptr = io_u->buf;
1041
1042 if (!td->o.zero_buffers) {
1043 while ((void *) ptr - io_u->buf < max_bs) {
1044 *ptr = rand() * GOLDEN_RATIO_PRIME;
1045 ptr++;
1046 }
1047 } else
1048 memset(ptr, 0, max_bs);
1049}