Random IO 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"
10ba535a 9
5945b9b4
JA
10/*
11 * Change this define to play with the timeout handling
12 */
13#undef FIO_USE_TIMEOUT
14
97601024
JA
15struct io_completion_data {
16 int nr; /* input */
97601024
JA
17
18 int error; /* output */
19 unsigned long bytes_done[2]; /* output */
20 struct timeval time; /* output */
21};
22
10ba535a
JA
23/*
24 * The ->file_map[] contains a map of blocks we have or have not done io
25 * to yet. Used to make sure we cover the entire range in a fair fashion.
26 */
27static int random_map_free(struct thread_data *td, struct fio_file *f,
28 unsigned long long block)
29{
30 unsigned int idx = RAND_MAP_IDX(td, f, block);
31 unsigned int bit = RAND_MAP_BIT(td, f, block);
32
33 return (f->file_map[idx] & (1UL << bit)) == 0;
34}
35
df415585
JA
36/*
37 * Mark a given offset as used in the map.
38 */
9bf2061e 39static void mark_random_map(struct thread_data *td, struct io_u *io_u)
df415585 40{
2dc1bbeb 41 unsigned int min_bs = td->o.rw_min_bs;
9bf2061e 42 struct fio_file *f = io_u->file;
a00735e6
JA
43 unsigned long long block;
44 unsigned int blocks;
c685b5b2 45 unsigned int nr_blocks;
df415585 46
a00735e6
JA
47 block = io_u->offset / (unsigned long long) min_bs;
48 blocks = 0;
c685b5b2
JA
49 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
50
51 while (blocks < nr_blocks) {
df415585
JA
52 unsigned int idx, bit;
53
1e3d53ac
JA
54 /*
55 * If we have a mixed random workload, we may
56 * encounter blocks we already did IO to.
57 */
b6a4c7d1 58 if (!td->o.ddir_nr == 1 && !random_map_free(td, f, block))
df415585
JA
59 break;
60
61 idx = RAND_MAP_IDX(td, f, block);
62 bit = RAND_MAP_BIT(td, f, block);
63
0032bf9f 64 fio_assert(td, idx < f->num_maps);
df415585
JA
65
66 f->file_map[idx] |= (1UL << bit);
67 block++;
68 blocks++;
69 }
70
a00735e6
JA
71 if ((blocks * min_bs) < io_u->buflen)
72 io_u->buflen = blocks * min_bs;
df415585
JA
73}
74
10ba535a
JA
75/*
76 * Return the next free block in the map.
77 */
78static int get_next_free_block(struct thread_data *td, struct fio_file *f,
79 unsigned long long *b)
80{
81 int i;
82
c685b5b2
JA
83 i = f->last_free_lookup;
84 *b = (i * BLOCKS_PER_MAP);
2dc1bbeb 85 while ((*b) * td->o.rw_min_bs < f->real_file_size) {
10ba535a 86 if (f->file_map[i] != -1UL) {
b12ebc65 87 *b += fio_ffz(f->file_map[i]);
c685b5b2 88 f->last_free_lookup = i;
10ba535a
JA
89 return 0;
90 }
91
92 *b += BLOCKS_PER_MAP;
93 i++;
94 }
95
96 return 1;
97}
98
ec4015da
JA
99static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
100 int ddir, unsigned long long *b)
101{
7bb48f84 102 unsigned long long max_blocks = f->io_size / td->o.min_bs[ddir];
ec4015da
JA
103 unsigned long long r, rb;
104 int loops = 5;
105
106 do {
107 r = os_random_long(&td->random_state);
108 if (!max_blocks)
109 *b = 0;
110 else
111 *b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
43c63a78
JA
112 /*
113 * if we are not maintaining a random map, we are done.
114 */
ec4015da 115 if (td->o.norandommap)
43c63a78
JA
116 return 0;
117
118 /*
119 * calculate map offset and chec if it's free
120 */
273f8c91 121 rb = *b;
43c63a78
JA
122 if (random_map_free(td, f, rb))
123 return 0;
124
125 } while (--loops);
ec4015da
JA
126
127 /*
43c63a78
JA
128 * we get here, if we didn't suceed in looking up a block. generate
129 * a random start offset into the filemap, and find the first free
130 * block from there.
ec4015da 131 */
43c63a78
JA
132 loops = 10;
133 do {
134 f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
135 if (!get_next_free_block(td, f, b))
136 return 0;
ec4015da 137
43c63a78
JA
138 r = os_random_long(&td->random_state);
139 } while (--loops);
140
141 /*
142 * that didn't work either, try exhaustive search from the start
143 */
144 f->last_free_lookup = 0;
145 return get_next_free_block(td, f, b);
ec4015da
JA
146}
147
10ba535a
JA
148/*
149 * For random io, generate a random new block and see if it's used. Repeat
150 * until we find a free one. For sequential io, just return the end of
151 * the last io issued.
152 */
9bf2061e 153static int get_next_offset(struct thread_data *td, struct io_u *io_u)
10ba535a 154{
9bf2061e 155 struct fio_file *f = io_u->file;
c685b5b2 156 const int ddir = io_u->ddir;
ec4015da 157 unsigned long long b;
10ba535a 158
ec4015da
JA
159 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
160 td->ddir_nr = td->o.ddir_nr;
211097b2 161
ec4015da 162 if (get_next_rand_offset(td, f, ddir, &b))
bca4ed4d 163 return 1;
43063a1c 164 } else {
0c3d768a
YHJT
165 if (f->last_pos >= f->real_file_size) {
166 if (!td_random(td) || get_next_rand_offset(td, f, ddir, &b))
167 return 1;
bcdedd0a
YHJT
168 } else
169 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
43063a1c 170 }
10ba535a 171
2dc1bbeb 172 io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
bca4ed4d 173 if (io_u->offset >= f->real_file_size)
10ba535a
JA
174 return 1;
175
176 return 0;
177}
178
9bf2061e 179static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
10ba535a 180{
bca4ed4d 181 const int ddir = io_u->ddir;
10ba535a
JA
182 unsigned int buflen;
183 long r;
184
2dc1bbeb
JA
185 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
186 buflen = td->o.min_bs[ddir];
10ba535a
JA
187 else {
188 r = os_random_long(&td->bsrange_state);
564ca972
JA
189 if (!td->o.bssplit_nr)
190 buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
191 else {
192 long perc = 0;
193 unsigned int i;
194
195 for (i = 0; i < td->o.bssplit_nr; i++) {
196 struct bssplit *bsp = &td->o.bssplit[i];
197
198 buflen = bsp->bs;
199 perc += bsp->perc;
200 if (r <= ((LONG_MAX / 100L) * perc))
201 break;
202 }
203 }
2dc1bbeb
JA
204 if (!td->o.bs_unaligned)
205 buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
10ba535a
JA
206 }
207
6a5e6884
JA
208 if (io_u->offset + buflen > io_u->file->real_file_size)
209 buflen = td->o.min_bs[ddir];
210
10ba535a
JA
211 return buflen;
212}
213
afe24a5a
JA
214static void set_rwmix_bytes(struct thread_data *td)
215{
216 unsigned long long rbytes;
217 unsigned int diff;
218
219 /*
220 * we do time or byte based switch. this is needed because
221 * buffered writes may issue a lot quicker than they complete,
222 * whereas reads do not.
223 */
224 rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
e47f799f 225 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
afe24a5a 226
e47f799f
JA
227 td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
228}
229
230static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
231{
232 unsigned int v;
233 long r;
234
235 r = os_random_long(&td->rwmix_state);
236 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
237 if (v < td->o.rwmix[DDIR_READ])
238 return DDIR_READ;
239
240 return DDIR_WRITE;
afe24a5a
JA
241}
242
10ba535a
JA
243/*
244 * Return the data direction for the next io_u. If the job is a
245 * mixed read/write workload, check the rwmix cycle and switch if
246 * necessary.
247 */
1e97cce9 248static enum fio_ddir get_rw_ddir(struct thread_data *td)
10ba535a
JA
249{
250 if (td_rw(td)) {
251 struct timeval now;
252 unsigned long elapsed;
afe24a5a 253 unsigned int cycle;
10ba535a 254
02bcaa8c 255 fio_gettime(&now, NULL);
10ba535a
JA
256 elapsed = mtime_since_now(&td->rwmix_switch);
257
e47f799f
JA
258 /*
259 * if this is the first cycle, make it shorter
260 */
afe24a5a
JA
261 cycle = td->o.rwmixcycle;
262 if (!td->rwmix_bytes)
263 cycle /= 10;
264
10ba535a
JA
265 /*
266 * Check if it's time to seed a new data direction.
267 */
e47f799f 268 if (elapsed >= cycle ||
afe24a5a 269 td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
e47f799f
JA
270 unsigned long long max_bytes;
271 enum fio_ddir ddir;
272
273 /*
274 * Put a top limit on how many bytes we do for
275 * one data direction, to avoid overflowing the
276 * ranges too much
277 */
278 ddir = get_rand_ddir(td);
279 max_bytes = td->this_io_bytes[ddir];
7bb48f84 280 if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
38d77cae
JA
281 if (!td->rw_end_set[ddir]) {
282 td->rw_end_set[ddir] = 1;
283 memcpy(&td->rw_end[ddir], &now, sizeof(now));
284 }
e47f799f 285 ddir ^= 1;
38d77cae 286 }
e47f799f
JA
287
288 if (ddir != td->rwmix_ddir)
289 set_rwmix_bytes(td);
290
291 td->rwmix_ddir = ddir;
10ba535a
JA
292 memcpy(&td->rwmix_switch, &now, sizeof(now));
293 }
294 return td->rwmix_ddir;
295 } else if (td_read(td))
296 return DDIR_READ;
297 else
298 return DDIR_WRITE;
299}
300
10ba535a
JA
301void put_io_u(struct thread_data *td, struct io_u *io_u)
302{
0c6e7517
JA
303 assert((io_u->flags & IO_U_F_FREE) == 0);
304 io_u->flags |= IO_U_F_FREE;
305
2dbdab7e
JA
306 if (io_u->file)
307 put_file(td, io_u->file);
308
10ba535a
JA
309 io_u->file = NULL;
310 list_del(&io_u->list);
311 list_add(&io_u->list, &td->io_u_freelist);
312 td->cur_depth--;
313}
314
755200a3
JA
315void requeue_io_u(struct thread_data *td, struct io_u **io_u)
316{
317 struct io_u *__io_u = *io_u;
318
4d2e0f49
JA
319 __io_u->flags |= IO_U_F_FREE;
320 __io_u->flags &= ~IO_U_F_FLIGHT;
321
755200a3
JA
322 list_del(&__io_u->list);
323 list_add_tail(&__io_u->list, &td->io_u_requeues);
324 td->cur_depth--;
325 *io_u = NULL;
326}
327
9bf2061e 328static int fill_io_u(struct thread_data *td, struct io_u *io_u)
10ba535a 329{
b4c5e1ac
JA
330 if (td->io_ops->flags & FIO_NOIO)
331 goto out;
332
87dc1ab1
JA
333 /*
334 * see if it's time to sync
335 */
2dc1bbeb
JA
336 if (td->o.fsync_blocks &&
337 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
338 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
87dc1ab1 339 io_u->ddir = DDIR_SYNC;
c38e9468 340 goto out;
87dc1ab1
JA
341 }
342
a00735e6
JA
343 io_u->ddir = get_rw_ddir(td);
344
48f5abd3
JA
345 /*
346 * See if it's time to switch to a new zone
347 */
348 if (td->zone_bytes >= td->o.zone_size) {
349 td->zone_bytes = 0;
350 io_u->file->last_pos += td->o.zone_skip;
351 td->io_skip_bytes += td->o.zone_skip;
352 }
353
10ba535a 354 /*
c685b5b2
JA
355 * No log, let the seq/rand engine retrieve the next buflen and
356 * position.
10ba535a 357 */
9bf2061e 358 if (get_next_offset(td, io_u))
bca4ed4d 359 return 1;
10ba535a 360
9bf2061e 361 io_u->buflen = get_next_buflen(td, io_u);
bca4ed4d
JA
362 if (!io_u->buflen)
363 return 1;
364
6a5e6884
JA
365 if (io_u->offset + io_u->buflen > io_u->file->real_file_size)
366 return 1;
367
bca4ed4d
JA
368 /*
369 * mark entry before potentially trimming io_u
370 */
8fd15a9a 371 if (td_random(td) && !td->o.norandommap)
9bf2061e 372 mark_random_map(td, io_u);
bca4ed4d
JA
373
374 /*
375 * If using a write iolog, store this entry.
376 */
c38e9468 377out:
f29b25a3 378 log_io_u(td, io_u);
bca4ed4d 379 return 0;
10ba535a
JA
380}
381
b3605062 382void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
71619dc2
JA
383{
384 int index = 0;
385
b3605062
JA
386 if (io_u->ddir == DDIR_SYNC)
387 return;
388
71619dc2
JA
389 switch (td->cur_depth) {
390 default:
a783e61a
JA
391 index = 6;
392 break;
71619dc2 393 case 32 ... 63:
a783e61a
JA
394 index = 5;
395 break;
71619dc2 396 case 16 ... 31:
a783e61a
JA
397 index = 4;
398 break;
71619dc2 399 case 8 ... 15:
a783e61a
JA
400 index = 3;
401 break;
71619dc2 402 case 4 ... 7:
a783e61a
JA
403 index = 2;
404 break;
71619dc2 405 case 2 ... 3:
a783e61a 406 index = 1;
71619dc2
JA
407 case 1:
408 break;
409 }
410
756867bd 411 td->ts.io_u_map[index]++;
b3605062 412 td->ts.total_io_u[io_u->ddir]++;
71619dc2
JA
413}
414
04a0feae
JA
415static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
416{
417 int index = 0;
418
419 assert(usec < 1000);
420
421 switch (usec) {
422 case 750 ... 999:
423 index = 9;
424 break;
425 case 500 ... 749:
426 index = 8;
427 break;
428 case 250 ... 499:
429 index = 7;
430 break;
431 case 100 ... 249:
432 index = 6;
433 break;
434 case 50 ... 99:
435 index = 5;
436 break;
437 case 20 ... 49:
438 index = 4;
439 break;
440 case 10 ... 19:
441 index = 3;
442 break;
443 case 4 ... 9:
444 index = 2;
445 break;
446 case 2 ... 3:
447 index = 1;
448 case 0 ... 1:
449 break;
450 }
451
452 assert(index < FIO_IO_U_LAT_U_NR);
453 td->ts.io_u_lat_u[index]++;
454}
455
456static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
ec118304
JA
457{
458 int index = 0;
459
460 switch (msec) {
461 default:
04a0feae
JA
462 index = 11;
463 break;
8abdce66 464 case 1000 ... 1999:
04a0feae
JA
465 index = 10;
466 break;
8abdce66 467 case 750 ... 999:
04a0feae
JA
468 index = 9;
469 break;
8abdce66 470 case 500 ... 749:
04a0feae
JA
471 index = 8;
472 break;
8abdce66 473 case 250 ... 499:
04a0feae
JA
474 index = 7;
475 break;
8abdce66 476 case 100 ... 249:
04a0feae
JA
477 index = 6;
478 break;
8abdce66 479 case 50 ... 99:
04a0feae
JA
480 index = 5;
481 break;
8abdce66 482 case 20 ... 49:
04a0feae
JA
483 index = 4;
484 break;
8abdce66 485 case 10 ... 19:
04a0feae
JA
486 index = 3;
487 break;
8abdce66 488 case 4 ... 9:
04a0feae
JA
489 index = 2;
490 break;
ec118304 491 case 2 ... 3:
04a0feae 492 index = 1;
ec118304
JA
493 case 0 ... 1:
494 break;
495 }
496
04a0feae
JA
497 assert(index < FIO_IO_U_LAT_M_NR);
498 td->ts.io_u_lat_m[index]++;
499}
500
501static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
502{
503 if (usec < 1000)
504 io_u_mark_lat_usec(td, usec);
505 else
506 io_u_mark_lat_msec(td, usec / 1000);
ec118304
JA
507}
508
0aabe160
JA
509/*
510 * Get next file to service by choosing one at random
511 */
1c178180
JA
512static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
513 int badf)
0aabe160 514{
0aabe160 515 struct fio_file *f;
1c178180 516 int fno;
0aabe160
JA
517
518 do {
7c83c089
JA
519 long r = os_random_long(&td->next_file_state);
520
2dc1bbeb 521 fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
1c178180 522 f = &td->files[fno];
059e63c0
JA
523 if (f->flags & FIO_FILE_DONE)
524 continue;
1c178180
JA
525
526 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
0aabe160
JA
527 return f;
528 } while (1);
529}
530
531/*
532 * Get next file to service by doing round robin between all available ones
533 */
1c178180
JA
534static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
535 int badf)
3d7c391d
JA
536{
537 unsigned int old_next_file = td->next_file;
538 struct fio_file *f;
539
540 do {
541 f = &td->files[td->next_file];
542
543 td->next_file++;
2dc1bbeb 544 if (td->next_file >= td->o.nr_files)
3d7c391d
JA
545 td->next_file = 0;
546
d5ed68ea
JA
547 if (f->flags & FIO_FILE_DONE) {
548 f = NULL;
059e63c0 549 continue;
d5ed68ea 550 }
059e63c0 551
1c178180 552 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
3d7c391d
JA
553 break;
554
555 f = NULL;
556 } while (td->next_file != old_next_file);
557
558 return f;
559}
560
bdb4e2e9
JA
561static struct fio_file *get_next_file(struct thread_data *td)
562{
1907dbc6
JA
563 struct fio_file *f;
564
2dc1bbeb 565 assert(td->o.nr_files <= td->files_index);
1c178180 566
1020a139 567 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
bdb4e2e9
JA
568 return NULL;
569
1907dbc6 570 f = td->file_service_file;
f11bd94d 571 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
1907dbc6
JA
572 return f;
573
2dc1bbeb 574 if (td->o.file_service_type == FIO_FSERVICE_RR)
1c178180 575 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
bdb4e2e9 576 else
1c178180 577 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
1907dbc6
JA
578
579 td->file_service_file = f;
580 td->file_service_left = td->file_service_nr - 1;
581 return f;
bdb4e2e9
JA
582}
583
1c178180
JA
584static struct fio_file *find_next_new_file(struct thread_data *td)
585{
586 struct fio_file *f;
587
1020a139
JA
588 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
589 return NULL;
590
2dc1bbeb 591 if (td->o.file_service_type == FIO_FSERVICE_RR)
1c178180
JA
592 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
593 else
594 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
595
596 return f;
597}
598
429f6675
JA
599static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
600{
601 struct fio_file *f;
602
603 do {
604 f = get_next_file(td);
605 if (!f)
606 return 1;
607
608set_file:
609 io_u->file = f;
610 get_file(f);
611
612 if (!fill_io_u(td, io_u))
613 break;
614
615 /*
616 * td_io_close() does a put_file() as well, so no need to
617 * do that here.
618 */
619 io_u->file = NULL;
620 td_io_close_file(td, f);
621 f->flags |= FIO_FILE_DONE;
622 td->nr_done_files++;
623
624 /*
625 * probably not the right place to do this, but see
626 * if we need to open a new file
627 */
628 if (td->nr_open_files < td->o.open_files &&
629 td->o.open_files != td->o.nr_files) {
630 f = find_next_new_file(td);
631
632 if (!f || td_io_open_file(td, f))
633 return 1;
634
635 goto set_file;
636 }
637 } while (1);
638
639 return 0;
640}
641
642
10ba535a
JA
643struct io_u *__get_io_u(struct thread_data *td)
644{
645 struct io_u *io_u = NULL;
646
755200a3
JA
647 if (!list_empty(&td->io_u_requeues))
648 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
649 else if (!queue_full(td)) {
10ba535a
JA
650 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
651
6040dabc 652 io_u->buflen = 0;
10ba535a 653 io_u->resid = 0;
755200a3 654 io_u->file = NULL;
d7762cf8 655 io_u->end_io = NULL;
755200a3
JA
656 }
657
658 if (io_u) {
0c6e7517
JA
659 assert(io_u->flags & IO_U_F_FREE);
660 io_u->flags &= ~IO_U_F_FREE;
661
755200a3 662 io_u->error = 0;
10ba535a
JA
663 list_del(&io_u->list);
664 list_add(&io_u->list, &td->io_u_busylist);
665 td->cur_depth++;
666 }
667
668 return io_u;
669}
670
671/*
672 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
673 * etc. The returned io_u is fully ready to be prepped and submitted.
674 */
3d7c391d 675struct io_u *get_io_u(struct thread_data *td)
10ba535a 676{
3d7c391d 677 struct fio_file *f;
10ba535a
JA
678 struct io_u *io_u;
679
680 io_u = __get_io_u(td);
681 if (!io_u)
682 return NULL;
683
755200a3
JA
684 /*
685 * from a requeue, io_u already setup
686 */
687 if (io_u->file)
77f392bf 688 goto out;
755200a3 689
429f6675
JA
690 /*
691 * If using an iolog, grab next piece if any available.
692 */
693 if (td->o.read_iolog_file) {
694 if (read_iolog_get(td, io_u))
695 goto err_put;
696 } else if (set_io_u_file(td, io_u))
697 goto err_put;
698
699 f = io_u->file;
700 assert(f->flags & FIO_FILE_OPEN);
97af62ce 701
87dc1ab1 702 if (io_u->ddir != DDIR_SYNC) {
429f6675
JA
703 if (!io_u->buflen)
704 goto err_put;
10ba535a 705
36167d82 706 f->last_pos = io_u->offset + io_u->buflen;
10ba535a 707
2dc1bbeb 708 if (td->o.verify != VERIFY_NONE)
87dc1ab1
JA
709 populate_verify_io_u(td, io_u);
710 }
10ba535a 711
165faf16
JA
712 /*
713 * Set io data pointers.
714 */
d460eb31 715 io_u->endpos = io_u->offset + io_u->buflen;
77f392bf 716out:
cec6b55d
JA
717 io_u->xfer_buf = io_u->buf;
718 io_u->xfer_buflen = io_u->buflen;
165faf16 719
429f6675
JA
720 if (!td_io_prep(td, io_u)) {
721 fio_gettime(&io_u->start_time, NULL);
722 return io_u;
36167d82 723 }
429f6675
JA
724err_put:
725 put_io_u(td, io_u);
726 return NULL;
10ba535a
JA
727}
728
5451792e
JA
729void io_u_log_error(struct thread_data *td, struct io_u *io_u)
730{
731 const char *msg[] = { "read", "write", "sync" };
732
733 log_err("fio: io_u error");
734
735 if (io_u->file)
736 log_err(" on file %s", io_u->file->file_name);
737
738 log_err(": %s\n", strerror(io_u->error));
739
740 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
741
742 if (!td->error)
743 td_verror(td, io_u->error, "io_u error");
744}
745
97601024
JA
746static void io_completed(struct thread_data *td, struct io_u *io_u,
747 struct io_completion_data *icd)
10ba535a 748{
d85f5118 749 unsigned long usec;
10ba535a 750
0c6e7517
JA
751 assert(io_u->flags & IO_U_F_FLIGHT);
752 io_u->flags &= ~IO_U_F_FLIGHT;
753
87dc1ab1
JA
754 if (io_u->ddir == DDIR_SYNC) {
755 td->last_was_sync = 1;
756 return;
757 }
758
759 td->last_was_sync = 0;
760
10ba535a
JA
761 if (!io_u->error) {
762 unsigned int bytes = io_u->buflen - io_u->resid;
1e97cce9 763 const enum fio_ddir idx = io_u->ddir;
3af6ef39 764 int ret;
10ba535a
JA
765
766 td->io_blocks[idx]++;
767 td->io_bytes[idx] += bytes;
768 td->zone_bytes += bytes;
769 td->this_io_bytes[idx] += bytes;
770
d460eb31 771 io_u->file->last_completed_pos = io_u->endpos;
02bcaa8c 772
d85f5118 773 usec = utime_since(&io_u->issue_time, &icd->time);
10ba535a 774
d85f5118 775 add_clat_sample(td, idx, usec);
02bcaa8c 776 add_bw_sample(td, idx, &icd->time);
04a0feae 777 io_u_mark_latency(td, usec);
10ba535a 778
660a1cb5 779 if (td_write(td) && idx == DDIR_WRITE &&
8670579e 780 td->o.do_verify &&
41128405 781 td->o.verify != VERIFY_NONE)
10ba535a
JA
782 log_io_piece(td, io_u);
783
784 icd->bytes_done[idx] += bytes;
3af6ef39 785
d7762cf8 786 if (io_u->end_io) {
36690c9b 787 ret = io_u->end_io(td, io_u);
3af6ef39
JA
788 if (ret && !icd->error)
789 icd->error = ret;
790 }
5451792e 791 } else {
10ba535a 792 icd->error = io_u->error;
5451792e
JA
793 io_u_log_error(td, io_u);
794 }
10ba535a
JA
795}
796
d7762cf8 797static void init_icd(struct io_completion_data *icd, int nr)
10ba535a 798{
02bcaa8c
JA
799 fio_gettime(&icd->time, NULL);
800
3af6ef39
JA
801 icd->nr = nr;
802
10ba535a
JA
803 icd->error = 0;
804 icd->bytes_done[0] = icd->bytes_done[1] = 0;
36167d82
JA
805}
806
97601024
JA
807static void ios_completed(struct thread_data *td,
808 struct io_completion_data *icd)
36167d82
JA
809{
810 struct io_u *io_u;
811 int i;
812
10ba535a
JA
813 for (i = 0; i < icd->nr; i++) {
814 io_u = td->io_ops->event(td, i);
815
816 io_completed(td, io_u, icd);
817 put_io_u(td, io_u);
818 }
819}
97601024 820
e7e6cfb4
JA
821/*
822 * Complete a single io_u for the sync engines.
823 */
d7762cf8 824long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
97601024
JA
825{
826 struct io_completion_data icd;
827
d7762cf8 828 init_icd(&icd, 1);
97601024
JA
829 io_completed(td, io_u, &icd);
830 put_io_u(td, io_u);
831
832 if (!icd.error)
833 return icd.bytes_done[0] + icd.bytes_done[1];
834
37e974a5 835 td_verror(td, icd.error, "io_u_sync_complete");
97601024
JA
836 return -1;
837}
838
e7e6cfb4
JA
839/*
840 * Called to complete min_events number of io for the async engines.
841 */
d7762cf8 842long io_u_queued_complete(struct thread_data *td, int min_events)
97601024 843{
97601024 844 struct io_completion_data icd;
00de55ef 845 struct timespec *tvp = NULL;
97601024 846 int ret;
4d06a338 847 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
97601024 848
4d06a338 849 if (!min_events)
00de55ef 850 tvp = &ts;
97601024 851
00de55ef 852 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
97601024 853 if (ret < 0) {
e1161c32 854 td_verror(td, -ret, "td_io_getevents");
97601024
JA
855 return ret;
856 } else if (!ret)
857 return ret;
858
d7762cf8 859 init_icd(&icd, ret);
97601024
JA
860 ios_completed(td, &icd);
861 if (!icd.error)
862 return icd.bytes_done[0] + icd.bytes_done[1];
863
37e974a5 864 td_verror(td, icd.error, "io_u_queued_complete");
97601024
JA
865 return -1;
866}
7e77dd02
JA
867
868/*
869 * Call when io_u is really queued, to update the submission latency.
870 */
871void io_u_queued(struct thread_data *td, struct io_u *io_u)
872{
873 unsigned long slat_time;
874
d85f5118 875 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
7e77dd02
JA
876 add_slat_sample(td, io_u->ddir, slat_time);
877}
433afcb4 878
55bc9728 879#ifdef FIO_USE_TIMEOUT
433afcb4
JA
880void io_u_set_timeout(struct thread_data *td)
881{
882 assert(td->cur_depth);
883
884 td->timer.it_interval.tv_sec = 0;
885 td->timer.it_interval.tv_usec = 0;
886 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
887 td->timer.it_value.tv_usec = 0;
888 setitimer(ITIMER_REAL, &td->timer, NULL);
889 fio_gettime(&td->timeout_end, NULL);
890}
5945b9b4
JA
891
892static void io_u_dump(struct io_u *io_u)
893{
894 unsigned long t_start = mtime_since_now(&io_u->start_time);
895 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
896
897 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
898 log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf, io_u->xfer_buf, io_u->buflen, io_u->xfer_buflen, io_u->offset);
899 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
900}
55bc9728
JA
901#else
902void io_u_set_timeout(struct thread_data fio_unused *td)
903{
904}
905#endif
433afcb4 906
55bc9728 907#ifdef FIO_USE_TIMEOUT
433afcb4
JA
908static void io_u_timeout_handler(int fio_unused sig)
909{
910 struct thread_data *td, *__td;
911 pid_t pid = getpid();
5945b9b4
JA
912 struct list_head *entry;
913 struct io_u *io_u;
433afcb4
JA
914 int i;
915
916 log_err("fio: io_u timeout\n");
917
918 /*
919 * TLS would be nice...
920 */
921 td = NULL;
922 for_each_td(__td, i) {
923 if (__td->pid == pid) {
924 td = __td;
925 break;
926 }
927 }
928
929 if (!td) {
930 log_err("fio: io_u timeout, can't find job\n");
931 exit(1);
932 }
933
934 if (!td->cur_depth) {
935 log_err("fio: timeout without pending work?\n");
936 return;
937 }
938
15506d09 939 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
5945b9b4
JA
940
941 list_for_each(entry, &td->io_u_busylist) {
942 io_u = list_entry(entry, struct io_u, list);
943
944 io_u_dump(io_u);
945 }
946
947 td_verror(td, ETIMEDOUT, "io_u timeout");
433afcb4
JA
948 exit(1);
949}
55bc9728 950#endif
433afcb4
JA
951
952void io_u_init_timeout(void)
953{
55bc9728 954#ifdef FIO_USE_TIMEOUT
433afcb4 955 signal(SIGALRM, io_u_timeout_handler);
55bc9728 956#endif
433afcb4 957}