Missed int to pointer conversion check for ->ratemin
[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"
10ba535a 11
97601024
JA
12struct io_completion_data {
13 int nr; /* input */
97601024
JA
14
15 int error; /* output */
16 unsigned long bytes_done[2]; /* output */
17 struct timeval time; /* output */
18};
19
10ba535a
JA
20/*
21 * The ->file_map[] contains a map of blocks we have or have not done io
22 * to yet. Used to make sure we cover the entire range in a fair fashion.
23 */
aec2de20 24static int random_map_free(struct fio_file *f, const unsigned long long block)
10ba535a 25{
aec2de20
JA
26 unsigned int idx = RAND_MAP_IDX(f, block);
27 unsigned int bit = RAND_MAP_BIT(f, block);
10ba535a 28
84422acd
JA
29 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
30
de605666 31 return (f->file_map[idx] & (1 << bit)) == 0;
10ba535a
JA
32}
33
df415585
JA
34/*
35 * Mark a given offset as used in the map.
36 */
9bf2061e 37static void mark_random_map(struct thread_data *td, struct io_u *io_u)
df415585 38{
2dc1bbeb 39 unsigned int min_bs = td->o.rw_min_bs;
9bf2061e 40 struct fio_file *f = io_u->file;
a00735e6 41 unsigned long long block;
3e3357b1 42 unsigned int blocks, nr_blocks;
df415585 43
b9c5b644 44 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
c685b5b2 45 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
3e3357b1 46 blocks = 0;
c685b5b2 47
3e3357b1
JA
48 while (nr_blocks) {
49 unsigned int this_blocks, mask;
df415585
JA
50 unsigned int idx, bit;
51
1e3d53ac
JA
52 /*
53 * If we have a mixed random workload, we may
54 * encounter blocks we already did IO to.
55 */
8b113b4f
JA
56 if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) {
57 if (!blocks)
58 blocks = 1;
df415585 59 break;
8b113b4f 60 }
df415585 61
aec2de20
JA
62 idx = RAND_MAP_IDX(f, block);
63 bit = RAND_MAP_BIT(f, block);
df415585 64
0032bf9f 65 fio_assert(td, idx < f->num_maps);
df415585 66
3e3357b1
JA
67 this_blocks = nr_blocks;
68 if (this_blocks + bit > BLOCKS_PER_MAP)
69 this_blocks = BLOCKS_PER_MAP - bit;
70
71 if (this_blocks == BLOCKS_PER_MAP)
72 mask = -1U;
73 else
74 mask = ((1U << this_blocks) - 1) << bit;
75
3e3357b1
JA
76 f->file_map[idx] |= mask;
77 nr_blocks -= this_blocks;
78 blocks += this_blocks;
c9dd34b2 79 block += this_blocks;
df415585
JA
80 }
81
a00735e6
JA
82 if ((blocks * min_bs) < io_u->buflen)
83 io_u->buflen = blocks * min_bs;
df415585
JA
84}
85
3e3357b1
JA
86static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
87 enum fio_ddir ddir)
2ba1c290
JA
88{
89 unsigned long long max_blocks;
d9dd70f7 90 unsigned long long max_size;
2ba1c290 91
d9dd70f7
JA
92 /*
93 * Hmm, should we make sure that ->io_size <= ->real_file_size?
94 */
95 max_size = f->io_size;
96 if (max_size > f->real_file_size)
97 max_size = f->real_file_size;
98
2b7a01d0 99 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
2ba1c290
JA
100 if (!max_blocks)
101 return 0;
102
67778e88 103 return max_blocks;
2ba1c290
JA
104}
105
10ba535a
JA
106/*
107 * Return the next free block in the map.
108 */
109static int get_next_free_block(struct thread_data *td, struct fio_file *f,
4ba66134 110 enum fio_ddir ddir, unsigned long long *b)
10ba535a 111{
84422acd 112 unsigned long long min_bs = td->o.rw_min_bs;
10ba535a
JA
113 int i;
114
c685b5b2
JA
115 i = f->last_free_lookup;
116 *b = (i * BLOCKS_PER_MAP);
899c29f2
RR
117 while ((*b) * min_bs < f->real_file_size &&
118 (*b) * min_bs < f->io_size) {
de605666 119 if (f->file_map[i] != (unsigned int) -1) {
697a606c 120 *b += ffz(f->file_map[i]);
4ba66134 121 if (*b > last_block(td, f, ddir))
2ba1c290 122 break;
c685b5b2 123 f->last_free_lookup = i;
10ba535a
JA
124 return 0;
125 }
126
127 *b += BLOCKS_PER_MAP;
128 i++;
129 }
130
2ba1c290 131 dprint(FD_IO, "failed finding a free block\n");
10ba535a
JA
132 return 1;
133}
134
ec4015da 135static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
4ba66134 136 enum fio_ddir ddir, unsigned long long *b)
ec4015da 137{
84422acd 138 unsigned long long r;
ec4015da
JA
139 int loops = 5;
140
141 do {
142 r = os_random_long(&td->random_state);
84422acd 143 dprint(FD_RANDOM, "off rand %llu\n", r);
5ec10eaa 144 *b = (last_block(td, f, ddir) - 1)
dc873b6f 145 * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
2ba1c290 146
43c63a78
JA
147 /*
148 * if we are not maintaining a random map, we are done.
149 */
303032ae 150 if (!file_randommap(td, f))
43c63a78
JA
151 return 0;
152
153 /*
84422acd 154 * calculate map offset and check if it's free
43c63a78 155 */
aec2de20 156 if (random_map_free(f, *b))
43c63a78
JA
157 return 0;
158
84422acd
JA
159 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
160 *b);
43c63a78 161 } while (--loops);
ec4015da
JA
162
163 /*
43c63a78
JA
164 * we get here, if we didn't suceed in looking up a block. generate
165 * a random start offset into the filemap, and find the first free
166 * block from there.
ec4015da 167 */
43c63a78
JA
168 loops = 10;
169 do {
dc873b6f
JA
170 f->last_free_lookup = (f->num_maps - 1) *
171 (r / (OS_RAND_MAX + 1.0));
4ba66134 172 if (!get_next_free_block(td, f, ddir, b))
43c63a78 173 return 0;
ec4015da 174
43c63a78
JA
175 r = os_random_long(&td->random_state);
176 } while (--loops);
177
178 /*
179 * that didn't work either, try exhaustive search from the start
180 */
181 f->last_free_lookup = 0;
4ba66134 182 return get_next_free_block(td, f, ddir, b);
ec4015da
JA
183}
184
10ba535a
JA
185/*
186 * For random io, generate a random new block and see if it's used. Repeat
187 * until we find a free one. For sequential io, just return the end of
188 * the last io issued.
189 */
9bf2061e 190static int get_next_offset(struct thread_data *td, struct io_u *io_u)
10ba535a 191{
9bf2061e 192 struct fio_file *f = io_u->file;
ec4015da 193 unsigned long long b;
4ba66134 194 enum fio_ddir ddir = io_u->ddir;
10ba535a 195
ec4015da
JA
196 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
197 td->ddir_nr = td->o.ddir_nr;
211097b2 198
683023e8
JA
199 if (get_next_rand_offset(td, f, ddir, &b)) {
200 dprint(FD_IO, "%s: getting rand offset failed\n",
201 f->file_name);
bca4ed4d 202 return 1;
683023e8 203 }
43063a1c 204 } else {
0c3d768a 205 if (f->last_pos >= f->real_file_size) {
4ba66134 206 if (!td_random(td) ||
683023e8
JA
207 get_next_rand_offset(td, f, ddir, &b)) {
208 dprint(FD_IO, "%s: pos %llu > size %llu\n",
209 f->file_name, f->last_pos,
210 f->real_file_size);
0c3d768a 211 return 1;
683023e8 212 }
bcdedd0a 213 } else
4ba66134 214 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
43063a1c 215 }
10ba535a 216
2b7a01d0 217 io_u->offset = b * td->o.ba[ddir];
009bd847
JA
218 if (io_u->offset >= f->io_size) {
219 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
220 io_u->offset, f->io_size);
221 return 1;
222 }
223
224 io_u->offset += f->file_offset;
2ba1c290
JA
225 if (io_u->offset >= f->real_file_size) {
226 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
227 io_u->offset, f->real_file_size);
10ba535a 228 return 1;
2ba1c290 229 }
10ba535a
JA
230
231 return 0;
232}
233
f3059de1
JA
234static inline int is_power_of_2(unsigned int val)
235{
236 return (val != 0 && ((val & (val - 1)) == 0));
237}
238
9bf2061e 239static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
10ba535a 240{
bca4ed4d 241 const int ddir = io_u->ddir;
f3f552b9 242 unsigned int uninitialized_var(buflen);
f3059de1 243 unsigned int minbs, maxbs;
10ba535a
JA
244 long r;
245
f3059de1
JA
246 minbs = td->o.min_bs[ddir];
247 maxbs = td->o.max_bs[ddir];
248
249 if (minbs == maxbs)
250 buflen = minbs;
10ba535a
JA
251 else {
252 r = os_random_long(&td->bsrange_state);
720e84ad 253 if (!td->o.bssplit_nr[ddir]) {
f3059de1
JA
254 buflen = 1 + (unsigned int) ((double) maxbs *
255 (r / (OS_RAND_MAX + 1.0)));
256 if (buflen < minbs)
257 buflen = minbs;
5ec10eaa 258 } else {
564ca972
JA
259 long perc = 0;
260 unsigned int i;
261
720e84ad
JA
262 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
263 struct bssplit *bsp = &td->o.bssplit[ddir][i];
564ca972
JA
264
265 buflen = bsp->bs;
266 perc += bsp->perc;
f3059de1 267 if (r <= ((OS_RAND_MAX / 100L) * perc))
564ca972
JA
268 break;
269 }
270 }
f3059de1
JA
271 if (!td->o.bs_unaligned && is_power_of_2(minbs))
272 buflen = (buflen + minbs - 1) & ~(minbs - 1);
10ba535a
JA
273 }
274
4ba66134
JA
275 if (io_u->offset + buflen > io_u->file->real_file_size) {
276 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
f3059de1
JA
277 minbs, ddir);
278 buflen = minbs;
4ba66134 279 }
6a5e6884 280
10ba535a
JA
281 return buflen;
282}
283
afe24a5a
JA
284static void set_rwmix_bytes(struct thread_data *td)
285{
afe24a5a
JA
286 unsigned int diff;
287
288 /*
289 * we do time or byte based switch. this is needed because
290 * buffered writes may issue a lot quicker than they complete,
291 * whereas reads do not.
292 */
e47f799f 293 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
04c540d9 294 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
e47f799f
JA
295}
296
297static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
298{
299 unsigned int v;
300 long r;
301
302 r = os_random_long(&td->rwmix_state);
dc873b6f 303 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
04c540d9 304 if (v <= td->o.rwmix[DDIR_READ])
e47f799f
JA
305 return DDIR_READ;
306
307 return DDIR_WRITE;
afe24a5a
JA
308}
309
581e7141
JA
310static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
311{
312 enum fio_ddir odir = ddir ^ 1;
313 struct timeval t;
314 long usec;
315
316 if (td->rate_pending_usleep[ddir] <= 0)
317 return ddir;
318
319 /*
320 * We have too much pending sleep in this direction. See if we
321 * should switch.
322 */
323 if (td_rw(td)) {
324 /*
325 * Other direction does not have too much pending, switch
326 */
327 if (td->rate_pending_usleep[odir] < 100000)
328 return odir;
329
330 /*
331 * Both directions have pending sleep. Sleep the minimum time
332 * and deduct from both.
333 */
334 if (td->rate_pending_usleep[ddir] <=
335 td->rate_pending_usleep[odir]) {
336 usec = td->rate_pending_usleep[ddir];
337 } else {
338 usec = td->rate_pending_usleep[odir];
339 ddir = odir;
340 }
341 } else
342 usec = td->rate_pending_usleep[ddir];
343
344 fio_gettime(&t, NULL);
345 usec_sleep(td, usec);
346 usec = utime_since_now(&t);
347
348 td->rate_pending_usleep[ddir] -= usec;
349
350 odir = ddir ^ 1;
351 if (td_rw(td) && __should_check_rate(td, odir))
352 td->rate_pending_usleep[odir] -= usec;
353
354 return ddir;
355}
356
10ba535a
JA
357/*
358 * Return the data direction for the next io_u. If the job is a
359 * mixed read/write workload, check the rwmix cycle and switch if
360 * necessary.
361 */
1e97cce9 362static enum fio_ddir get_rw_ddir(struct thread_data *td)
10ba535a 363{
581e7141
JA
364 enum fio_ddir ddir;
365
10ba535a 366 if (td_rw(td)) {
10ba535a
JA
367 /*
368 * Check if it's time to seed a new data direction.
369 */
e4928662 370 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
e47f799f
JA
371 /*
372 * Put a top limit on how many bytes we do for
373 * one data direction, to avoid overflowing the
374 * ranges too much
375 */
376 ddir = get_rand_ddir(td);
e47f799f
JA
377
378 if (ddir != td->rwmix_ddir)
379 set_rwmix_bytes(td);
380
381 td->rwmix_ddir = ddir;
10ba535a 382 }
581e7141 383 ddir = td->rwmix_ddir;
10ba535a 384 } else if (td_read(td))
581e7141 385 ddir = DDIR_READ;
10ba535a 386 else
581e7141
JA
387 ddir = DDIR_WRITE;
388
389 td->rwmix_ddir = rate_ddir(td, ddir);
390 return td->rwmix_ddir;
10ba535a
JA
391}
392
60f2c658
JA
393static void put_file_log(struct thread_data *td, struct fio_file *f)
394{
395 int ret = put_file(td, f);
396
397 if (ret)
398 td_verror(td, ret, "file close");
399}
400
10ba535a
JA
401void put_io_u(struct thread_data *td, struct io_u *io_u)
402{
0c6e7517
JA
403 assert((io_u->flags & IO_U_F_FREE) == 0);
404 io_u->flags |= IO_U_F_FREE;
405
60f2c658
JA
406 if (io_u->file)
407 put_file_log(td, io_u->file);
2dbdab7e 408
10ba535a 409 io_u->file = NULL;
01743ee1
JA
410 flist_del(&io_u->list);
411 flist_add(&io_u->list, &td->io_u_freelist);
10ba535a
JA
412 td->cur_depth--;
413}
414
755200a3
JA
415void requeue_io_u(struct thread_data *td, struct io_u **io_u)
416{
417 struct io_u *__io_u = *io_u;
418
465221b0
JA
419 dprint(FD_IO, "requeue %p\n", __io_u);
420
4d2e0f49 421 __io_u->flags |= IO_U_F_FREE;
e4f54adb
JA
422 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
423 td->io_issues[__io_u->ddir]--;
5ec10eaa 424
4d2e0f49
JA
425 __io_u->flags &= ~IO_U_F_FLIGHT;
426
01743ee1
JA
427 flist_del(&__io_u->list);
428 flist_add_tail(&__io_u->list, &td->io_u_requeues);
755200a3
JA
429 td->cur_depth--;
430 *io_u = NULL;
431}
432
9bf2061e 433static int fill_io_u(struct thread_data *td, struct io_u *io_u)
10ba535a 434{
b4c5e1ac
JA
435 if (td->io_ops->flags & FIO_NOIO)
436 goto out;
437
87dc1ab1
JA
438 /*
439 * see if it's time to sync
440 */
2dc1bbeb
JA
441 if (td->o.fsync_blocks &&
442 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
443 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
87dc1ab1 444 io_u->ddir = DDIR_SYNC;
c38e9468 445 goto out;
87dc1ab1
JA
446 }
447
a00735e6
JA
448 io_u->ddir = get_rw_ddir(td);
449
48f5abd3
JA
450 /*
451 * See if it's time to switch to a new zone
452 */
453 if (td->zone_bytes >= td->o.zone_size) {
454 td->zone_bytes = 0;
455 io_u->file->last_pos += td->o.zone_skip;
456 td->io_skip_bytes += td->o.zone_skip;
457 }
458
10ba535a 459 /*
c685b5b2
JA
460 * No log, let the seq/rand engine retrieve the next buflen and
461 * position.
10ba535a 462 */
2ba1c290
JA
463 if (get_next_offset(td, io_u)) {
464 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
bca4ed4d 465 return 1;
2ba1c290 466 }
10ba535a 467
9bf2061e 468 io_u->buflen = get_next_buflen(td, io_u);
2ba1c290
JA
469 if (!io_u->buflen) {
470 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
bca4ed4d 471 return 1;
2ba1c290 472 }
bca4ed4d 473
2ba1c290
JA
474 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
475 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
4ba66134
JA
476 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
477 io_u->buflen, io_u->file->real_file_size);
6a5e6884 478 return 1;
2ba1c290 479 }
6a5e6884 480
bca4ed4d
JA
481 /*
482 * mark entry before potentially trimming io_u
483 */
303032ae 484 if (td_random(td) && file_randommap(td, io_u->file))
9bf2061e 485 mark_random_map(td, io_u);
bca4ed4d
JA
486
487 /*
488 * If using a write iolog, store this entry.
489 */
c38e9468 490out:
2ba1c290 491 dprint_io_u(io_u, "fill_io_u");
d9d91e39 492 td->zone_bytes += io_u->buflen;
f29b25a3 493 log_io_u(td, io_u);
bca4ed4d 494 return 0;
10ba535a
JA
495}
496
838bc709
JA
497static void __io_u_mark_map(unsigned int *map, unsigned int nr)
498{
499 int index = 0;
500
501 switch (nr) {
502 default:
503 index = 6;
504 break;
505 case 33 ... 64:
506 index = 5;
507 break;
508 case 17 ... 32:
509 index = 4;
510 break;
511 case 9 ... 16:
512 index = 3;
513 break;
514 case 5 ... 8:
515 index = 2;
516 break;
517 case 1 ... 4:
518 index = 1;
519 case 0:
520 break;
521 }
522
523 map[index]++;
524}
525
526void io_u_mark_submit(struct thread_data *td, unsigned int nr)
527{
528 __io_u_mark_map(td->ts.io_u_submit, nr);
529 td->ts.total_submit++;
530}
531
532void io_u_mark_complete(struct thread_data *td, unsigned int nr)
533{
534 __io_u_mark_map(td->ts.io_u_complete, nr);
535 td->ts.total_complete++;
536}
537
d8005759 538void io_u_mark_depth(struct thread_data *td, unsigned int nr)
71619dc2
JA
539{
540 int index = 0;
541
542 switch (td->cur_depth) {
543 default:
a783e61a
JA
544 index = 6;
545 break;
71619dc2 546 case 32 ... 63:
a783e61a
JA
547 index = 5;
548 break;
71619dc2 549 case 16 ... 31:
a783e61a
JA
550 index = 4;
551 break;
71619dc2 552 case 8 ... 15:
a783e61a
JA
553 index = 3;
554 break;
71619dc2 555 case 4 ... 7:
a783e61a
JA
556 index = 2;
557 break;
71619dc2 558 case 2 ... 3:
a783e61a 559 index = 1;
71619dc2
JA
560 case 1:
561 break;
562 }
563
3bec7ae0 564 td->ts.io_u_map[index] += nr;
71619dc2
JA
565}
566
04a0feae
JA
567static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
568{
569 int index = 0;
570
571 assert(usec < 1000);
572
573 switch (usec) {
574 case 750 ... 999:
575 index = 9;
576 break;
577 case 500 ... 749:
578 index = 8;
579 break;
580 case 250 ... 499:
581 index = 7;
582 break;
583 case 100 ... 249:
584 index = 6;
585 break;
586 case 50 ... 99:
587 index = 5;
588 break;
589 case 20 ... 49:
590 index = 4;
591 break;
592 case 10 ... 19:
593 index = 3;
594 break;
595 case 4 ... 9:
596 index = 2;
597 break;
598 case 2 ... 3:
599 index = 1;
600 case 0 ... 1:
601 break;
602 }
603
604 assert(index < FIO_IO_U_LAT_U_NR);
605 td->ts.io_u_lat_u[index]++;
606}
607
608static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
ec118304
JA
609{
610 int index = 0;
611
612 switch (msec) {
613 default:
04a0feae
JA
614 index = 11;
615 break;
8abdce66 616 case 1000 ... 1999:
04a0feae
JA
617 index = 10;
618 break;
8abdce66 619 case 750 ... 999:
04a0feae
JA
620 index = 9;
621 break;
8abdce66 622 case 500 ... 749:
04a0feae
JA
623 index = 8;
624 break;
8abdce66 625 case 250 ... 499:
04a0feae
JA
626 index = 7;
627 break;
8abdce66 628 case 100 ... 249:
04a0feae
JA
629 index = 6;
630 break;
8abdce66 631 case 50 ... 99:
04a0feae
JA
632 index = 5;
633 break;
8abdce66 634 case 20 ... 49:
04a0feae
JA
635 index = 4;
636 break;
8abdce66 637 case 10 ... 19:
04a0feae
JA
638 index = 3;
639 break;
8abdce66 640 case 4 ... 9:
04a0feae
JA
641 index = 2;
642 break;
ec118304 643 case 2 ... 3:
04a0feae 644 index = 1;
ec118304
JA
645 case 0 ... 1:
646 break;
647 }
648
04a0feae
JA
649 assert(index < FIO_IO_U_LAT_M_NR);
650 td->ts.io_u_lat_m[index]++;
651}
652
653static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
654{
655 if (usec < 1000)
656 io_u_mark_lat_usec(td, usec);
657 else
658 io_u_mark_lat_msec(td, usec / 1000);
ec118304
JA
659}
660
0aabe160
JA
661/*
662 * Get next file to service by choosing one at random
663 */
d6aed795
JA
664static struct fio_file *get_next_file_rand(struct thread_data *td, enum fio_file_flags goodf,
665 enum fio_file_flags badf)
0aabe160 666{
0aabe160 667 struct fio_file *f;
1c178180 668 int fno;
0aabe160
JA
669
670 do {
7c83c089 671 long r = os_random_long(&td->next_file_state);
87b10676 672 int opened = 0;
7c83c089 673
5ec10eaa 674 fno = (unsigned int) ((double) td->o.nr_files
dc873b6f 675 * (r / (OS_RAND_MAX + 1.0)));
126d65c6 676 f = td->files[fno];
d6aed795 677 if (fio_file_done(f))
059e63c0 678 continue;
1c178180 679
d6aed795 680 if (!fio_file_open(f)) {
87b10676
JA
681 int err;
682
683 err = td_io_open_file(td, f);
684 if (err)
685 continue;
686 opened = 1;
687 }
688
2ba1c290
JA
689 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
690 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
0aabe160 691 return f;
2ba1c290 692 }
87b10676
JA
693 if (opened)
694 td_io_close_file(td, f);
0aabe160
JA
695 } while (1);
696}
697
698/*
699 * Get next file to service by doing round robin between all available ones
700 */
1c178180
JA
701static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
702 int badf)
3d7c391d
JA
703{
704 unsigned int old_next_file = td->next_file;
705 struct fio_file *f;
706
707 do {
87b10676
JA
708 int opened = 0;
709
126d65c6 710 f = td->files[td->next_file];
3d7c391d
JA
711
712 td->next_file++;
2dc1bbeb 713 if (td->next_file >= td->o.nr_files)
3d7c391d
JA
714 td->next_file = 0;
715
87b10676 716 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
d6aed795 717 if (fio_file_done(f)) {
d5ed68ea 718 f = NULL;
059e63c0 719 continue;
d5ed68ea 720 }
059e63c0 721
d6aed795 722 if (!fio_file_open(f)) {
87b10676
JA
723 int err;
724
725 err = td_io_open_file(td, f);
b5696bfc
JA
726 if (err) {
727 dprint(FD_FILE, "error %d on open of %s\n",
728 err, f->file_name);
87c27b45 729 f = NULL;
87b10676 730 continue;
b5696bfc 731 }
87b10676
JA
732 opened = 1;
733 }
734
b5696bfc 735 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, f->flags);
1c178180 736 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
3d7c391d
JA
737 break;
738
87b10676
JA
739 if (opened)
740 td_io_close_file(td, f);
741
3d7c391d
JA
742 f = NULL;
743 } while (td->next_file != old_next_file);
744
2ba1c290 745 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
3d7c391d
JA
746 return f;
747}
748
bdb4e2e9
JA
749static struct fio_file *get_next_file(struct thread_data *td)
750{
1907dbc6
JA
751 struct fio_file *f;
752
2dc1bbeb 753 assert(td->o.nr_files <= td->files_index);
1c178180 754
b5696bfc 755 if (td->nr_done_files >= td->o.nr_files) {
5ec10eaa
JA
756 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
757 " nr_files=%d\n", td->nr_open_files,
758 td->nr_done_files,
759 td->o.nr_files);
bdb4e2e9 760 return NULL;
2ba1c290 761 }
bdb4e2e9 762
1907dbc6 763 f = td->file_service_file;
d6aed795 764 if (f && fio_file_open(f) && !fio_file_closing(f)) {
a086c257
JA
765 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
766 goto out;
767 if (td->file_service_left--)
768 goto out;
769 }
1907dbc6 770
a086c257
JA
771 if (td->o.file_service_type == FIO_FSERVICE_RR ||
772 td->o.file_service_type == FIO_FSERVICE_SEQ)
d6aed795 773 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
bdb4e2e9 774 else
d6aed795 775 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1907dbc6
JA
776
777 td->file_service_file = f;
778 td->file_service_left = td->file_service_nr - 1;
2ba1c290 779out:
683023e8 780 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1907dbc6 781 return f;
bdb4e2e9
JA
782}
783
429f6675
JA
784static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
785{
786 struct fio_file *f;
787
788 do {
789 f = get_next_file(td);
790 if (!f)
791 return 1;
792
429f6675
JA
793 io_u->file = f;
794 get_file(f);
795
796 if (!fill_io_u(td, io_u))
797 break;
798
b5696bfc 799 put_file_log(td, f);
429f6675 800 td_io_close_file(td, f);
b5696bfc 801 io_u->file = NULL;
d6aed795 802 fio_file_set_done(f);
429f6675 803 td->nr_done_files++;
87b10676 804 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, td->nr_done_files, td->o.nr_files);
429f6675
JA
805 } while (1);
806
807 return 0;
808}
809
810
10ba535a
JA
811struct io_u *__get_io_u(struct thread_data *td)
812{
813 struct io_u *io_u = NULL;
814
01743ee1
JA
815 if (!flist_empty(&td->io_u_requeues))
816 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
755200a3 817 else if (!queue_full(td)) {
01743ee1 818 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
10ba535a 819
6040dabc 820 io_u->buflen = 0;
10ba535a 821 io_u->resid = 0;
755200a3 822 io_u->file = NULL;
d7762cf8 823 io_u->end_io = NULL;
755200a3
JA
824 }
825
826 if (io_u) {
0c6e7517
JA
827 assert(io_u->flags & IO_U_F_FREE);
828 io_u->flags &= ~IO_U_F_FREE;
829
755200a3 830 io_u->error = 0;
01743ee1
JA
831 flist_del(&io_u->list);
832 flist_add(&io_u->list, &td->io_u_busylist);
10ba535a
JA
833 td->cur_depth++;
834 }
835
836 return io_u;
837}
838
839/*
840 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
841 * etc. The returned io_u is fully ready to be prepped and submitted.
842 */
3d7c391d 843struct io_u *get_io_u(struct thread_data *td)
10ba535a 844{
3d7c391d 845 struct fio_file *f;
10ba535a
JA
846 struct io_u *io_u;
847
848 io_u = __get_io_u(td);
2ba1c290
JA
849 if (!io_u) {
850 dprint(FD_IO, "__get_io_u failed\n");
10ba535a 851 return NULL;
2ba1c290 852 }
10ba535a 853
755200a3
JA
854 /*
855 * from a requeue, io_u already setup
856 */
857 if (io_u->file)
77f392bf 858 goto out;
755200a3 859
429f6675
JA
860 /*
861 * If using an iolog, grab next piece if any available.
862 */
863 if (td->o.read_iolog_file) {
864 if (read_iolog_get(td, io_u))
865 goto err_put;
2ba1c290
JA
866 } else if (set_io_u_file(td, io_u)) {
867 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
429f6675 868 goto err_put;
2ba1c290 869 }
5ec10eaa 870
429f6675 871 f = io_u->file;
d6aed795 872 assert(fio_file_open(f));
97af62ce 873
87dc1ab1 874 if (io_u->ddir != DDIR_SYNC) {
d0656a93 875 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
2ba1c290 876 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
429f6675 877 goto err_put;
2ba1c290 878 }
10ba535a 879
36167d82 880 f->last_pos = io_u->offset + io_u->buflen;
10ba535a 881
c311cd2a 882 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
87dc1ab1 883 populate_verify_io_u(td, io_u);
e4dad9c6
JA
884 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
885 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
87dc1ab1 886 }
10ba535a 887
165faf16
JA
888 /*
889 * Set io data pointers.
890 */
cec6b55d
JA
891 io_u->xfer_buf = io_u->buf;
892 io_u->xfer_buflen = io_u->buflen;
5973cafb 893
6ac7a331 894out:
429f6675 895 if (!td_io_prep(td, io_u)) {
993bf48b
JA
896 if (!td->o.disable_slat)
897 fio_gettime(&io_u->start_time, NULL);
429f6675 898 return io_u;
36167d82 899 }
429f6675 900err_put:
2ba1c290 901 dprint(FD_IO, "get_io_u failed\n");
429f6675
JA
902 put_io_u(td, io_u);
903 return NULL;
10ba535a
JA
904}
905
5451792e
JA
906void io_u_log_error(struct thread_data *td, struct io_u *io_u)
907{
908 const char *msg[] = { "read", "write", "sync" };
909
910 log_err("fio: io_u error");
911
912 if (io_u->file)
913 log_err(" on file %s", io_u->file->file_name);
914
915 log_err(": %s\n", strerror(io_u->error));
916
5ec10eaa
JA
917 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
918 io_u->offset, io_u->xfer_buflen);
5451792e
JA
919
920 if (!td->error)
921 td_verror(td, io_u->error, "io_u error");
922}
923
97601024
JA
924static void io_completed(struct thread_data *td, struct io_u *io_u,
925 struct io_completion_data *icd)
10ba535a 926{
dbad30b0
JA
927 /*
928 * Older gcc's are too dumb to realize that usec is always used
929 * initialized, silence that warning.
930 */
931 unsigned long uninitialized_var(usec);
10ba535a 932
2ba1c290
JA
933 dprint_io_u(io_u, "io complete");
934
0c6e7517
JA
935 assert(io_u->flags & IO_U_F_FLIGHT);
936 io_u->flags &= ~IO_U_F_FLIGHT;
937
87dc1ab1
JA
938 if (io_u->ddir == DDIR_SYNC) {
939 td->last_was_sync = 1;
940 return;
941 }
942
943 td->last_was_sync = 0;
944
10ba535a
JA
945 if (!io_u->error) {
946 unsigned int bytes = io_u->buflen - io_u->resid;
1e97cce9 947 const enum fio_ddir idx = io_u->ddir;
b29ee5b3 948 int ret;
10ba535a 949
b29ee5b3
JA
950 td->io_blocks[idx]++;
951 td->io_bytes[idx] += bytes;
952 td->this_io_bytes[idx] += bytes;
10ba535a 953
b29ee5b3 954 if (ramp_time_over(td)) {
581e7141
JA
955 if (!td->o.disable_clat || !td->o.disable_bw ||
956 __should_check_rate(td, idx))
9520ebb9
JA
957 usec = utime_since(&io_u->issue_time,
958 &icd->time);
721938ae 959
9520ebb9 960 if (!td->o.disable_clat) {
306ddc97 961 add_clat_sample(td, idx, usec, bytes);
9520ebb9
JA
962 io_u_mark_latency(td, usec);
963 }
964 if (!td->o.disable_bw)
306ddc97 965 add_bw_sample(td, idx, bytes, &icd->time);
581e7141
JA
966 if (__should_check_rate(td, idx))
967 td->rate_pending_usleep[idx] += (long) td->rate_usec_cycle[idx] - usec;
968 if (__should_check_rate(td, idx ^ 1))
969 td->rate_pending_usleep[idx ^ 1] -= usec;
721938ae 970 }
10ba535a 971
660a1cb5 972 if (td_write(td) && idx == DDIR_WRITE &&
8670579e 973 td->o.do_verify &&
41128405 974 td->o.verify != VERIFY_NONE)
10ba535a
JA
975 log_io_piece(td, io_u);
976
b29ee5b3 977 icd->bytes_done[idx] += bytes;
3af6ef39 978
d7762cf8 979 if (io_u->end_io) {
36690c9b 980 ret = io_u->end_io(td, io_u);
3af6ef39
JA
981 if (ret && !icd->error)
982 icd->error = ret;
983 }
5451792e 984 } else {
10ba535a 985 icd->error = io_u->error;
5451792e
JA
986 io_u_log_error(td, io_u);
987 }
10ba535a
JA
988}
989
9520ebb9
JA
990static void init_icd(struct thread_data *td, struct io_completion_data *icd,
991 int nr)
10ba535a 992{
9520ebb9
JA
993 if (!td->o.disable_clat || !td->o.disable_bw)
994 fio_gettime(&icd->time, NULL);
02bcaa8c 995
3af6ef39
JA
996 icd->nr = nr;
997
10ba535a
JA
998 icd->error = 0;
999 icd->bytes_done[0] = icd->bytes_done[1] = 0;
36167d82
JA
1000}
1001
97601024
JA
1002static void ios_completed(struct thread_data *td,
1003 struct io_completion_data *icd)
36167d82
JA
1004{
1005 struct io_u *io_u;
1006 int i;
1007
10ba535a
JA
1008 for (i = 0; i < icd->nr; i++) {
1009 io_u = td->io_ops->event(td, i);
1010
1011 io_completed(td, io_u, icd);
1012 put_io_u(td, io_u);
1013 }
1014}
97601024 1015
e7e6cfb4
JA
1016/*
1017 * Complete a single io_u for the sync engines.
1018 */
581e7141
JA
1019int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1020 unsigned long *bytes)
97601024
JA
1021{
1022 struct io_completion_data icd;
1023
9520ebb9 1024 init_icd(td, &icd, 1);
97601024
JA
1025 io_completed(td, io_u, &icd);
1026 put_io_u(td, io_u);
1027
581e7141
JA
1028 if (icd.error) {
1029 td_verror(td, icd.error, "io_u_sync_complete");
1030 return -1;
1031 }
97601024 1032
581e7141
JA
1033 if (bytes) {
1034 bytes[0] += icd.bytes_done[0];
1035 bytes[1] += icd.bytes_done[1];
1036 }
1037
1038 return 0;
97601024
JA
1039}
1040
e7e6cfb4
JA
1041/*
1042 * Called to complete min_events number of io for the async engines.
1043 */
581e7141
JA
1044int io_u_queued_complete(struct thread_data *td, int min_evts,
1045 unsigned long *bytes)
97601024 1046{
97601024 1047 struct io_completion_data icd;
00de55ef 1048 struct timespec *tvp = NULL;
97601024 1049 int ret;
4d06a338 1050 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
97601024 1051
4950421a 1052 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
b271fe62 1053
4950421a 1054 if (!min_evts)
00de55ef 1055 tvp = &ts;
97601024 1056
4950421a 1057 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
97601024 1058 if (ret < 0) {
e1161c32 1059 td_verror(td, -ret, "td_io_getevents");
97601024
JA
1060 return ret;
1061 } else if (!ret)
1062 return ret;
1063
9520ebb9 1064 init_icd(td, &icd, ret);
97601024 1065 ios_completed(td, &icd);
581e7141
JA
1066 if (icd.error) {
1067 td_verror(td, icd.error, "io_u_queued_complete");
1068 return -1;
1069 }
97601024 1070
581e7141
JA
1071 if (bytes) {
1072 bytes[0] += icd.bytes_done[0];
1073 bytes[1] += icd.bytes_done[1];
1074 }
1075
1076 return 0;
97601024 1077}
7e77dd02
JA
1078
1079/*
1080 * Call when io_u is really queued, to update the submission latency.
1081 */
1082void io_u_queued(struct thread_data *td, struct io_u *io_u)
1083{
9520ebb9
JA
1084 if (!td->o.disable_slat) {
1085 unsigned long slat_time;
7e77dd02 1086
9520ebb9 1087 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
306ddc97 1088 add_slat_sample(td, io_u->ddir, io_u->xfer_buflen, slat_time);
9520ebb9 1089 }
7e77dd02 1090}
433afcb4 1091
5973cafb
JA
1092/*
1093 * "randomly" fill the buffer contents
1094 */
1095void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1096 unsigned int max_bs)
1097{
1098 long *ptr = io_u->buf;
1099
1100 if (!td->o.zero_buffers) {
1101 while ((void *) ptr - io_u->buf < max_bs) {
1102 *ptr = rand() * GOLDEN_RATIO_PRIME;
1103 ptr++;
1104 }
1105 } else
1106 memset(ptr, 0, max_bs);
1107}