Add a 'continue_on_error' option to fio
[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
f2bba182
RR
415void clear_io_u(struct thread_data *td, struct io_u *io_u)
416{
417 io_u->flags &= ~IO_U_F_FLIGHT;
418 put_io_u(td, io_u);
419}
420
755200a3
JA
421void requeue_io_u(struct thread_data *td, struct io_u **io_u)
422{
423 struct io_u *__io_u = *io_u;
424
465221b0
JA
425 dprint(FD_IO, "requeue %p\n", __io_u);
426
4d2e0f49 427 __io_u->flags |= IO_U_F_FREE;
e4f54adb
JA
428 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
429 td->io_issues[__io_u->ddir]--;
5ec10eaa 430
4d2e0f49
JA
431 __io_u->flags &= ~IO_U_F_FLIGHT;
432
01743ee1
JA
433 flist_del(&__io_u->list);
434 flist_add_tail(&__io_u->list, &td->io_u_requeues);
755200a3
JA
435 td->cur_depth--;
436 *io_u = NULL;
437}
438
9bf2061e 439static int fill_io_u(struct thread_data *td, struct io_u *io_u)
10ba535a 440{
b4c5e1ac
JA
441 if (td->io_ops->flags & FIO_NOIO)
442 goto out;
443
87dc1ab1
JA
444 /*
445 * see if it's time to sync
446 */
2dc1bbeb
JA
447 if (td->o.fsync_blocks &&
448 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
449 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
87dc1ab1 450 io_u->ddir = DDIR_SYNC;
c38e9468 451 goto out;
87dc1ab1
JA
452 }
453
a00735e6
JA
454 io_u->ddir = get_rw_ddir(td);
455
48f5abd3
JA
456 /*
457 * See if it's time to switch to a new zone
458 */
459 if (td->zone_bytes >= td->o.zone_size) {
460 td->zone_bytes = 0;
461 io_u->file->last_pos += td->o.zone_skip;
462 td->io_skip_bytes += td->o.zone_skip;
463 }
464
10ba535a 465 /*
c685b5b2
JA
466 * No log, let the seq/rand engine retrieve the next buflen and
467 * position.
10ba535a 468 */
2ba1c290
JA
469 if (get_next_offset(td, io_u)) {
470 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
bca4ed4d 471 return 1;
2ba1c290 472 }
10ba535a 473
9bf2061e 474 io_u->buflen = get_next_buflen(td, io_u);
2ba1c290
JA
475 if (!io_u->buflen) {
476 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
bca4ed4d 477 return 1;
2ba1c290 478 }
bca4ed4d 479
2ba1c290
JA
480 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
481 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
4ba66134
JA
482 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
483 io_u->buflen, io_u->file->real_file_size);
6a5e6884 484 return 1;
2ba1c290 485 }
6a5e6884 486
bca4ed4d
JA
487 /*
488 * mark entry before potentially trimming io_u
489 */
303032ae 490 if (td_random(td) && file_randommap(td, io_u->file))
9bf2061e 491 mark_random_map(td, io_u);
bca4ed4d
JA
492
493 /*
494 * If using a write iolog, store this entry.
495 */
c38e9468 496out:
2ba1c290 497 dprint_io_u(io_u, "fill_io_u");
d9d91e39 498 td->zone_bytes += io_u->buflen;
f29b25a3 499 log_io_u(td, io_u);
bca4ed4d 500 return 0;
10ba535a
JA
501}
502
838bc709
JA
503static void __io_u_mark_map(unsigned int *map, unsigned int nr)
504{
505 int index = 0;
506
507 switch (nr) {
508 default:
509 index = 6;
510 break;
511 case 33 ... 64:
512 index = 5;
513 break;
514 case 17 ... 32:
515 index = 4;
516 break;
517 case 9 ... 16:
518 index = 3;
519 break;
520 case 5 ... 8:
521 index = 2;
522 break;
523 case 1 ... 4:
524 index = 1;
525 case 0:
526 break;
527 }
528
529 map[index]++;
530}
531
532void io_u_mark_submit(struct thread_data *td, unsigned int nr)
533{
534 __io_u_mark_map(td->ts.io_u_submit, nr);
535 td->ts.total_submit++;
536}
537
538void io_u_mark_complete(struct thread_data *td, unsigned int nr)
539{
540 __io_u_mark_map(td->ts.io_u_complete, nr);
541 td->ts.total_complete++;
542}
543
d8005759 544void io_u_mark_depth(struct thread_data *td, unsigned int nr)
71619dc2
JA
545{
546 int index = 0;
547
548 switch (td->cur_depth) {
549 default:
a783e61a
JA
550 index = 6;
551 break;
71619dc2 552 case 32 ... 63:
a783e61a
JA
553 index = 5;
554 break;
71619dc2 555 case 16 ... 31:
a783e61a
JA
556 index = 4;
557 break;
71619dc2 558 case 8 ... 15:
a783e61a
JA
559 index = 3;
560 break;
71619dc2 561 case 4 ... 7:
a783e61a
JA
562 index = 2;
563 break;
71619dc2 564 case 2 ... 3:
a783e61a 565 index = 1;
71619dc2
JA
566 case 1:
567 break;
568 }
569
3bec7ae0 570 td->ts.io_u_map[index] += nr;
71619dc2
JA
571}
572
04a0feae
JA
573static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
574{
575 int index = 0;
576
577 assert(usec < 1000);
578
579 switch (usec) {
580 case 750 ... 999:
581 index = 9;
582 break;
583 case 500 ... 749:
584 index = 8;
585 break;
586 case 250 ... 499:
587 index = 7;
588 break;
589 case 100 ... 249:
590 index = 6;
591 break;
592 case 50 ... 99:
593 index = 5;
594 break;
595 case 20 ... 49:
596 index = 4;
597 break;
598 case 10 ... 19:
599 index = 3;
600 break;
601 case 4 ... 9:
602 index = 2;
603 break;
604 case 2 ... 3:
605 index = 1;
606 case 0 ... 1:
607 break;
608 }
609
610 assert(index < FIO_IO_U_LAT_U_NR);
611 td->ts.io_u_lat_u[index]++;
612}
613
614static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
ec118304
JA
615{
616 int index = 0;
617
618 switch (msec) {
619 default:
04a0feae
JA
620 index = 11;
621 break;
8abdce66 622 case 1000 ... 1999:
04a0feae
JA
623 index = 10;
624 break;
8abdce66 625 case 750 ... 999:
04a0feae
JA
626 index = 9;
627 break;
8abdce66 628 case 500 ... 749:
04a0feae
JA
629 index = 8;
630 break;
8abdce66 631 case 250 ... 499:
04a0feae
JA
632 index = 7;
633 break;
8abdce66 634 case 100 ... 249:
04a0feae
JA
635 index = 6;
636 break;
8abdce66 637 case 50 ... 99:
04a0feae
JA
638 index = 5;
639 break;
8abdce66 640 case 20 ... 49:
04a0feae
JA
641 index = 4;
642 break;
8abdce66 643 case 10 ... 19:
04a0feae
JA
644 index = 3;
645 break;
8abdce66 646 case 4 ... 9:
04a0feae
JA
647 index = 2;
648 break;
ec118304 649 case 2 ... 3:
04a0feae 650 index = 1;
ec118304
JA
651 case 0 ... 1:
652 break;
653 }
654
04a0feae
JA
655 assert(index < FIO_IO_U_LAT_M_NR);
656 td->ts.io_u_lat_m[index]++;
657}
658
659static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
660{
661 if (usec < 1000)
662 io_u_mark_lat_usec(td, usec);
663 else
664 io_u_mark_lat_msec(td, usec / 1000);
ec118304
JA
665}
666
0aabe160
JA
667/*
668 * Get next file to service by choosing one at random
669 */
2cc52930
JA
670static struct fio_file *get_next_file_rand(struct thread_data *td,
671 enum fio_file_flags goodf,
d6aed795 672 enum fio_file_flags badf)
0aabe160 673{
0aabe160 674 struct fio_file *f;
1c178180 675 int fno;
0aabe160
JA
676
677 do {
7c83c089 678 long r = os_random_long(&td->next_file_state);
87b10676 679 int opened = 0;
7c83c089 680
5ec10eaa 681 fno = (unsigned int) ((double) td->o.nr_files
dc873b6f 682 * (r / (OS_RAND_MAX + 1.0)));
126d65c6 683 f = td->files[fno];
d6aed795 684 if (fio_file_done(f))
059e63c0 685 continue;
1c178180 686
d6aed795 687 if (!fio_file_open(f)) {
87b10676
JA
688 int err;
689
690 err = td_io_open_file(td, f);
691 if (err)
692 continue;
693 opened = 1;
694 }
695
2ba1c290
JA
696 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
697 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
0aabe160 698 return f;
2ba1c290 699 }
87b10676
JA
700 if (opened)
701 td_io_close_file(td, f);
0aabe160
JA
702 } while (1);
703}
704
705/*
706 * Get next file to service by doing round robin between all available ones
707 */
1c178180
JA
708static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
709 int badf)
3d7c391d
JA
710{
711 unsigned int old_next_file = td->next_file;
712 struct fio_file *f;
713
714 do {
87b10676
JA
715 int opened = 0;
716
126d65c6 717 f = td->files[td->next_file];
3d7c391d
JA
718
719 td->next_file++;
2dc1bbeb 720 if (td->next_file >= td->o.nr_files)
3d7c391d
JA
721 td->next_file = 0;
722
87b10676 723 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
d6aed795 724 if (fio_file_done(f)) {
d5ed68ea 725 f = NULL;
059e63c0 726 continue;
d5ed68ea 727 }
059e63c0 728
d6aed795 729 if (!fio_file_open(f)) {
87b10676
JA
730 int err;
731
732 err = td_io_open_file(td, f);
b5696bfc
JA
733 if (err) {
734 dprint(FD_FILE, "error %d on open of %s\n",
735 err, f->file_name);
87c27b45 736 f = NULL;
87b10676 737 continue;
b5696bfc 738 }
87b10676
JA
739 opened = 1;
740 }
741
b5696bfc 742 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, f->flags);
1c178180 743 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
3d7c391d
JA
744 break;
745
87b10676
JA
746 if (opened)
747 td_io_close_file(td, f);
748
3d7c391d
JA
749 f = NULL;
750 } while (td->next_file != old_next_file);
751
2ba1c290 752 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
3d7c391d
JA
753 return f;
754}
755
bdb4e2e9
JA
756static struct fio_file *get_next_file(struct thread_data *td)
757{
1907dbc6
JA
758 struct fio_file *f;
759
2dc1bbeb 760 assert(td->o.nr_files <= td->files_index);
1c178180 761
b5696bfc 762 if (td->nr_done_files >= td->o.nr_files) {
5ec10eaa
JA
763 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
764 " nr_files=%d\n", td->nr_open_files,
765 td->nr_done_files,
766 td->o.nr_files);
bdb4e2e9 767 return NULL;
2ba1c290 768 }
bdb4e2e9 769
1907dbc6 770 f = td->file_service_file;
d6aed795 771 if (f && fio_file_open(f) && !fio_file_closing(f)) {
a086c257
JA
772 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
773 goto out;
774 if (td->file_service_left--)
775 goto out;
776 }
1907dbc6 777
a086c257
JA
778 if (td->o.file_service_type == FIO_FSERVICE_RR ||
779 td->o.file_service_type == FIO_FSERVICE_SEQ)
d6aed795 780 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
bdb4e2e9 781 else
d6aed795 782 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1907dbc6
JA
783
784 td->file_service_file = f;
785 td->file_service_left = td->file_service_nr - 1;
2ba1c290 786out:
683023e8 787 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1907dbc6 788 return f;
bdb4e2e9
JA
789}
790
429f6675
JA
791static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
792{
793 struct fio_file *f;
794
795 do {
796 f = get_next_file(td);
797 if (!f)
798 return 1;
799
429f6675
JA
800 io_u->file = f;
801 get_file(f);
802
803 if (!fill_io_u(td, io_u))
804 break;
805
b5696bfc 806 put_file_log(td, f);
429f6675 807 td_io_close_file(td, f);
b5696bfc 808 io_u->file = NULL;
d6aed795 809 fio_file_set_done(f);
429f6675 810 td->nr_done_files++;
87b10676 811 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, td->nr_done_files, td->o.nr_files);
429f6675
JA
812 } while (1);
813
814 return 0;
815}
816
817
10ba535a
JA
818struct io_u *__get_io_u(struct thread_data *td)
819{
820 struct io_u *io_u = NULL;
821
01743ee1
JA
822 if (!flist_empty(&td->io_u_requeues))
823 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
755200a3 824 else if (!queue_full(td)) {
01743ee1 825 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
10ba535a 826
6040dabc 827 io_u->buflen = 0;
10ba535a 828 io_u->resid = 0;
755200a3 829 io_u->file = NULL;
d7762cf8 830 io_u->end_io = NULL;
755200a3
JA
831 }
832
833 if (io_u) {
0c6e7517
JA
834 assert(io_u->flags & IO_U_F_FREE);
835 io_u->flags &= ~IO_U_F_FREE;
836
755200a3 837 io_u->error = 0;
01743ee1
JA
838 flist_del(&io_u->list);
839 flist_add(&io_u->list, &td->io_u_busylist);
10ba535a
JA
840 td->cur_depth++;
841 }
842
843 return io_u;
844}
845
846/*
847 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
848 * etc. The returned io_u is fully ready to be prepped and submitted.
849 */
3d7c391d 850struct io_u *get_io_u(struct thread_data *td)
10ba535a 851{
3d7c391d 852 struct fio_file *f;
10ba535a
JA
853 struct io_u *io_u;
854
855 io_u = __get_io_u(td);
2ba1c290
JA
856 if (!io_u) {
857 dprint(FD_IO, "__get_io_u failed\n");
10ba535a 858 return NULL;
2ba1c290 859 }
10ba535a 860
755200a3
JA
861 /*
862 * from a requeue, io_u already setup
863 */
864 if (io_u->file)
77f392bf 865 goto out;
755200a3 866
429f6675
JA
867 /*
868 * If using an iolog, grab next piece if any available.
869 */
870 if (td->o.read_iolog_file) {
871 if (read_iolog_get(td, io_u))
872 goto err_put;
2ba1c290
JA
873 } else if (set_io_u_file(td, io_u)) {
874 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
429f6675 875 goto err_put;
2ba1c290 876 }
5ec10eaa 877
429f6675 878 f = io_u->file;
d6aed795 879 assert(fio_file_open(f));
97af62ce 880
87dc1ab1 881 if (io_u->ddir != DDIR_SYNC) {
d0656a93 882 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
2ba1c290 883 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
429f6675 884 goto err_put;
2ba1c290 885 }
10ba535a 886
36167d82 887 f->last_pos = io_u->offset + io_u->buflen;
10ba535a 888
c311cd2a 889 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
87dc1ab1 890 populate_verify_io_u(td, io_u);
e4dad9c6
JA
891 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
892 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
87dc1ab1 893 }
10ba535a 894
165faf16
JA
895 /*
896 * Set io data pointers.
897 */
cec6b55d
JA
898 io_u->xfer_buf = io_u->buf;
899 io_u->xfer_buflen = io_u->buflen;
5973cafb 900
6ac7a331 901out:
429f6675 902 if (!td_io_prep(td, io_u)) {
993bf48b
JA
903 if (!td->o.disable_slat)
904 fio_gettime(&io_u->start_time, NULL);
429f6675 905 return io_u;
36167d82 906 }
429f6675 907err_put:
2ba1c290 908 dprint(FD_IO, "get_io_u failed\n");
429f6675
JA
909 put_io_u(td, io_u);
910 return NULL;
10ba535a
JA
911}
912
5451792e
JA
913void io_u_log_error(struct thread_data *td, struct io_u *io_u)
914{
915 const char *msg[] = { "read", "write", "sync" };
916
917 log_err("fio: io_u error");
918
919 if (io_u->file)
920 log_err(" on file %s", io_u->file->file_name);
921
922 log_err(": %s\n", strerror(io_u->error));
923
5ec10eaa
JA
924 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
925 io_u->offset, io_u->xfer_buflen);
5451792e
JA
926
927 if (!td->error)
928 td_verror(td, io_u->error, "io_u error");
929}
930
97601024
JA
931static void io_completed(struct thread_data *td, struct io_u *io_u,
932 struct io_completion_data *icd)
10ba535a 933{
dbad30b0
JA
934 /*
935 * Older gcc's are too dumb to realize that usec is always used
936 * initialized, silence that warning.
937 */
938 unsigned long uninitialized_var(usec);
10ba535a 939
2ba1c290
JA
940 dprint_io_u(io_u, "io complete");
941
0c6e7517
JA
942 assert(io_u->flags & IO_U_F_FLIGHT);
943 io_u->flags &= ~IO_U_F_FLIGHT;
944
87dc1ab1
JA
945 if (io_u->ddir == DDIR_SYNC) {
946 td->last_was_sync = 1;
947 return;
948 }
949
950 td->last_was_sync = 0;
951
10ba535a
JA
952 if (!io_u->error) {
953 unsigned int bytes = io_u->buflen - io_u->resid;
1e97cce9 954 const enum fio_ddir idx = io_u->ddir;
b29ee5b3 955 int ret;
10ba535a 956
b29ee5b3
JA
957 td->io_blocks[idx]++;
958 td->io_bytes[idx] += bytes;
959 td->this_io_bytes[idx] += bytes;
10ba535a 960
b29ee5b3 961 if (ramp_time_over(td)) {
40e1a6f0
JA
962 unsigned long uninitialized_var(lusec);
963 unsigned long uninitialized_var(rusec);
964
965 if (!td->o.disable_clat || !td->o.disable_bw)
966 lusec = utime_since(&io_u->issue_time,
967 &icd->time);
968 if (__should_check_rate(td, idx) ||
969 __should_check_rate(td, idx ^ 1))
970 rusec = utime_since(&io_u->start_time,
9520ebb9 971 &icd->time);
721938ae 972
9520ebb9 973 if (!td->o.disable_clat) {
306ddc97 974 add_clat_sample(td, idx, usec, bytes);
40e1a6f0 975 io_u_mark_latency(td, lusec);
9520ebb9
JA
976 }
977 if (!td->o.disable_bw)
306ddc97 978 add_bw_sample(td, idx, bytes, &icd->time);
b23b6a2f
JA
979 if (__should_check_rate(td, idx)) {
980 td->rate_pending_usleep[idx] +=
981 (long) td->rate_usec_cycle[idx] - rusec;
982 }
581e7141 983 if (__should_check_rate(td, idx ^ 1))
b23b6a2f 984 td->rate_pending_usleep[idx ^ 1] -= rusec;
721938ae 985 }
10ba535a 986
660a1cb5 987 if (td_write(td) && idx == DDIR_WRITE &&
8670579e 988 td->o.do_verify &&
41128405 989 td->o.verify != VERIFY_NONE)
10ba535a
JA
990 log_io_piece(td, io_u);
991
b29ee5b3 992 icd->bytes_done[idx] += bytes;
3af6ef39 993
d7762cf8 994 if (io_u->end_io) {
36690c9b 995 ret = io_u->end_io(td, io_u);
3af6ef39
JA
996 if (ret && !icd->error)
997 icd->error = ret;
998 }
5451792e 999 } else {
10ba535a 1000 icd->error = io_u->error;
5451792e
JA
1001 io_u_log_error(td, io_u);
1002 }
f2bba182
RR
1003 if (td->o.continue_on_error && icd->error &&
1004 td_non_fatal_error(icd->error)) {
1005 /*
1006 * If there is a non_fatal error, then add to the error count
1007 * and clear all the errors.
1008 */
1009 update_error_count(td, icd->error);
1010 td_clear_error(td);
1011 icd->error = 0;
1012 io_u->error = 0;
1013 }
10ba535a
JA
1014}
1015
9520ebb9
JA
1016static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1017 int nr)
10ba535a 1018{
9520ebb9
JA
1019 if (!td->o.disable_clat || !td->o.disable_bw)
1020 fio_gettime(&icd->time, NULL);
02bcaa8c 1021
3af6ef39
JA
1022 icd->nr = nr;
1023
10ba535a
JA
1024 icd->error = 0;
1025 icd->bytes_done[0] = icd->bytes_done[1] = 0;
36167d82
JA
1026}
1027
97601024
JA
1028static void ios_completed(struct thread_data *td,
1029 struct io_completion_data *icd)
36167d82
JA
1030{
1031 struct io_u *io_u;
1032 int i;
1033
10ba535a
JA
1034 for (i = 0; i < icd->nr; i++) {
1035 io_u = td->io_ops->event(td, i);
1036
1037 io_completed(td, io_u, icd);
1038 put_io_u(td, io_u);
1039 }
1040}
97601024 1041
e7e6cfb4
JA
1042/*
1043 * Complete a single io_u for the sync engines.
1044 */
581e7141
JA
1045int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1046 unsigned long *bytes)
97601024
JA
1047{
1048 struct io_completion_data icd;
1049
9520ebb9 1050 init_icd(td, &icd, 1);
97601024
JA
1051 io_completed(td, io_u, &icd);
1052 put_io_u(td, io_u);
1053
581e7141
JA
1054 if (icd.error) {
1055 td_verror(td, icd.error, "io_u_sync_complete");
1056 return -1;
1057 }
97601024 1058
581e7141
JA
1059 if (bytes) {
1060 bytes[0] += icd.bytes_done[0];
1061 bytes[1] += icd.bytes_done[1];
1062 }
1063
1064 return 0;
97601024
JA
1065}
1066
e7e6cfb4
JA
1067/*
1068 * Called to complete min_events number of io for the async engines.
1069 */
581e7141
JA
1070int io_u_queued_complete(struct thread_data *td, int min_evts,
1071 unsigned long *bytes)
97601024 1072{
97601024 1073 struct io_completion_data icd;
00de55ef 1074 struct timespec *tvp = NULL;
97601024 1075 int ret;
4d06a338 1076 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
97601024 1077
4950421a 1078 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
b271fe62 1079
4950421a 1080 if (!min_evts)
00de55ef 1081 tvp = &ts;
97601024 1082
4950421a 1083 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
97601024 1084 if (ret < 0) {
e1161c32 1085 td_verror(td, -ret, "td_io_getevents");
97601024
JA
1086 return ret;
1087 } else if (!ret)
1088 return ret;
1089
9520ebb9 1090 init_icd(td, &icd, ret);
97601024 1091 ios_completed(td, &icd);
581e7141
JA
1092 if (icd.error) {
1093 td_verror(td, icd.error, "io_u_queued_complete");
1094 return -1;
1095 }
97601024 1096
581e7141
JA
1097 if (bytes) {
1098 bytes[0] += icd.bytes_done[0];
1099 bytes[1] += icd.bytes_done[1];
1100 }
1101
1102 return 0;
97601024 1103}
7e77dd02
JA
1104
1105/*
1106 * Call when io_u is really queued, to update the submission latency.
1107 */
1108void io_u_queued(struct thread_data *td, struct io_u *io_u)
1109{
9520ebb9
JA
1110 if (!td->o.disable_slat) {
1111 unsigned long slat_time;
7e77dd02 1112
9520ebb9 1113 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
29a90ddb 1114 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
9520ebb9 1115 }
7e77dd02 1116}
433afcb4 1117
5973cafb
JA
1118/*
1119 * "randomly" fill the buffer contents
1120 */
1121void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1122 unsigned int max_bs)
1123{
1124 long *ptr = io_u->buf;
1125
1126 if (!td->o.zero_buffers) {
1127 while ((void *) ptr - io_u->buf < max_bs) {
1128 *ptr = rand() * GOLDEN_RATIO_PRIME;
1129 ptr++;
1130 }
1131 } else
1132 memset(ptr, 0, max_bs);
1133}