CPU burn engine fix
[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
2ba1c290
JA
75static inline unsigned long long last_block(struct thread_data *td,
76 struct fio_file *f)
77{
78 unsigned long long max_blocks;
79
80 max_blocks = f->io_size / td->o.rw_min_bs;
81 if (!max_blocks)
82 return 0;
83
84 return max_blocks - 1;
85}
86
10ba535a
JA
87/*
88 * Return the next free block in the map.
89 */
90static int get_next_free_block(struct thread_data *td, struct fio_file *f,
91 unsigned long long *b)
92{
93 int i;
94
c685b5b2
JA
95 i = f->last_free_lookup;
96 *b = (i * BLOCKS_PER_MAP);
2dc1bbeb 97 while ((*b) * td->o.rw_min_bs < f->real_file_size) {
10ba535a 98 if (f->file_map[i] != -1UL) {
b12ebc65 99 *b += fio_ffz(f->file_map[i]);
2ba1c290
JA
100 if (*b > last_block(td, f))
101 break;
c685b5b2 102 f->last_free_lookup = i;
10ba535a
JA
103 return 0;
104 }
105
106 *b += BLOCKS_PER_MAP;
107 i++;
108 }
109
2ba1c290 110 dprint(FD_IO, "failed finding a free block\n");
10ba535a
JA
111 return 1;
112}
113
ec4015da 114static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
2ba1c290 115 unsigned long long *b)
ec4015da 116{
ec4015da
JA
117 unsigned long long r, rb;
118 int loops = 5;
119
120 do {
121 r = os_random_long(&td->random_state);
2ba1c290
JA
122 *b = last_block(td, f);
123
43c63a78
JA
124 /*
125 * if we are not maintaining a random map, we are done.
126 */
ec4015da 127 if (td->o.norandommap)
43c63a78
JA
128 return 0;
129
130 /*
131 * calculate map offset and chec if it's free
132 */
273f8c91 133 rb = *b;
43c63a78
JA
134 if (random_map_free(td, f, rb))
135 return 0;
136
137 } while (--loops);
ec4015da
JA
138
139 /*
43c63a78
JA
140 * we get here, if we didn't suceed in looking up a block. generate
141 * a random start offset into the filemap, and find the first free
142 * block from there.
ec4015da 143 */
43c63a78
JA
144 loops = 10;
145 do {
146 f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
147 if (!get_next_free_block(td, f, b))
148 return 0;
ec4015da 149
43c63a78
JA
150 r = os_random_long(&td->random_state);
151 } while (--loops);
152
153 /*
154 * that didn't work either, try exhaustive search from the start
155 */
156 f->last_free_lookup = 0;
157 return get_next_free_block(td, f, b);
ec4015da
JA
158}
159
10ba535a
JA
160/*
161 * For random io, generate a random new block and see if it's used. Repeat
162 * until we find a free one. For sequential io, just return the end of
163 * the last io issued.
164 */
9bf2061e 165static int get_next_offset(struct thread_data *td, struct io_u *io_u)
10ba535a 166{
9bf2061e 167 struct fio_file *f = io_u->file;
ec4015da 168 unsigned long long b;
10ba535a 169
ec4015da
JA
170 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
171 td->ddir_nr = td->o.ddir_nr;
211097b2 172
2ba1c290 173 if (get_next_rand_offset(td, f, &b))
bca4ed4d 174 return 1;
43063a1c 175 } else {
0c3d768a 176 if (f->last_pos >= f->real_file_size) {
2ba1c290 177 if (!td_random(td) || get_next_rand_offset(td, f, &b))
0c3d768a 178 return 1;
bcdedd0a 179 } else
2ba1c290 180 b = (f->last_pos - f->file_offset) / td->o.rw_min_bs;
43063a1c 181 }
10ba535a 182
2ba1c290
JA
183 io_u->offset = (b * td->o.rw_min_bs) + f->file_offset;
184 if (io_u->offset >= f->real_file_size) {
185 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
186 io_u->offset, f->real_file_size);
10ba535a 187 return 1;
2ba1c290 188 }
10ba535a
JA
189
190 return 0;
191}
192
9bf2061e 193static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
10ba535a 194{
bca4ed4d 195 const int ddir = io_u->ddir;
10ba535a
JA
196 unsigned int buflen;
197 long r;
198
2dc1bbeb
JA
199 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
200 buflen = td->o.min_bs[ddir];
10ba535a
JA
201 else {
202 r = os_random_long(&td->bsrange_state);
564ca972
JA
203 if (!td->o.bssplit_nr)
204 buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
205 else {
206 long perc = 0;
207 unsigned int i;
208
209 for (i = 0; i < td->o.bssplit_nr; i++) {
210 struct bssplit *bsp = &td->o.bssplit[i];
211
212 buflen = bsp->bs;
213 perc += bsp->perc;
214 if (r <= ((LONG_MAX / 100L) * perc))
215 break;
216 }
217 }
2dc1bbeb
JA
218 if (!td->o.bs_unaligned)
219 buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
10ba535a
JA
220 }
221
6a5e6884
JA
222 if (io_u->offset + buflen > io_u->file->real_file_size)
223 buflen = td->o.min_bs[ddir];
224
10ba535a
JA
225 return buflen;
226}
227
afe24a5a
JA
228static void set_rwmix_bytes(struct thread_data *td)
229{
230 unsigned long long rbytes;
231 unsigned int diff;
232
233 /*
234 * we do time or byte based switch. this is needed because
235 * buffered writes may issue a lot quicker than they complete,
236 * whereas reads do not.
237 */
238 rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
e47f799f 239 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
afe24a5a 240
e47f799f
JA
241 td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
242}
243
244static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
245{
246 unsigned int v;
247 long r;
248
249 r = os_random_long(&td->rwmix_state);
250 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
251 if (v < td->o.rwmix[DDIR_READ])
252 return DDIR_READ;
253
254 return DDIR_WRITE;
afe24a5a
JA
255}
256
10ba535a
JA
257/*
258 * Return the data direction for the next io_u. If the job is a
259 * mixed read/write workload, check the rwmix cycle and switch if
260 * necessary.
261 */
1e97cce9 262static enum fio_ddir get_rw_ddir(struct thread_data *td)
10ba535a
JA
263{
264 if (td_rw(td)) {
265 struct timeval now;
266 unsigned long elapsed;
afe24a5a 267 unsigned int cycle;
10ba535a 268
02bcaa8c 269 fio_gettime(&now, NULL);
10ba535a
JA
270 elapsed = mtime_since_now(&td->rwmix_switch);
271
e47f799f
JA
272 /*
273 * if this is the first cycle, make it shorter
274 */
afe24a5a
JA
275 cycle = td->o.rwmixcycle;
276 if (!td->rwmix_bytes)
277 cycle /= 10;
278
10ba535a
JA
279 /*
280 * Check if it's time to seed a new data direction.
281 */
e47f799f 282 if (elapsed >= cycle ||
afe24a5a 283 td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
e47f799f
JA
284 unsigned long long max_bytes;
285 enum fio_ddir ddir;
286
287 /*
288 * Put a top limit on how many bytes we do for
289 * one data direction, to avoid overflowing the
290 * ranges too much
291 */
292 ddir = get_rand_ddir(td);
293 max_bytes = td->this_io_bytes[ddir];
7bb48f84 294 if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
38d77cae
JA
295 if (!td->rw_end_set[ddir]) {
296 td->rw_end_set[ddir] = 1;
297 memcpy(&td->rw_end[ddir], &now, sizeof(now));
298 }
e47f799f 299 ddir ^= 1;
38d77cae 300 }
e47f799f
JA
301
302 if (ddir != td->rwmix_ddir)
303 set_rwmix_bytes(td);
304
305 td->rwmix_ddir = ddir;
10ba535a
JA
306 memcpy(&td->rwmix_switch, &now, sizeof(now));
307 }
308 return td->rwmix_ddir;
309 } else if (td_read(td))
310 return DDIR_READ;
311 else
312 return DDIR_WRITE;
313}
314
10ba535a
JA
315void put_io_u(struct thread_data *td, struct io_u *io_u)
316{
0c6e7517
JA
317 assert((io_u->flags & IO_U_F_FREE) == 0);
318 io_u->flags |= IO_U_F_FREE;
319
2dbdab7e
JA
320 if (io_u->file)
321 put_file(td, io_u->file);
322
10ba535a
JA
323 io_u->file = NULL;
324 list_del(&io_u->list);
325 list_add(&io_u->list, &td->io_u_freelist);
326 td->cur_depth--;
327}
328
755200a3
JA
329void requeue_io_u(struct thread_data *td, struct io_u **io_u)
330{
331 struct io_u *__io_u = *io_u;
332
4d2e0f49
JA
333 __io_u->flags |= IO_U_F_FREE;
334 __io_u->flags &= ~IO_U_F_FLIGHT;
335
755200a3
JA
336 list_del(&__io_u->list);
337 list_add_tail(&__io_u->list, &td->io_u_requeues);
338 td->cur_depth--;
339 *io_u = NULL;
340}
341
9bf2061e 342static int fill_io_u(struct thread_data *td, struct io_u *io_u)
10ba535a 343{
b4c5e1ac
JA
344 if (td->io_ops->flags & FIO_NOIO)
345 goto out;
346
87dc1ab1
JA
347 /*
348 * see if it's time to sync
349 */
2dc1bbeb
JA
350 if (td->o.fsync_blocks &&
351 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
352 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
87dc1ab1 353 io_u->ddir = DDIR_SYNC;
c38e9468 354 goto out;
87dc1ab1
JA
355 }
356
a00735e6
JA
357 io_u->ddir = get_rw_ddir(td);
358
48f5abd3
JA
359 /*
360 * See if it's time to switch to a new zone
361 */
362 if (td->zone_bytes >= td->o.zone_size) {
363 td->zone_bytes = 0;
364 io_u->file->last_pos += td->o.zone_skip;
365 td->io_skip_bytes += td->o.zone_skip;
366 }
367
10ba535a 368 /*
c685b5b2
JA
369 * No log, let the seq/rand engine retrieve the next buflen and
370 * position.
10ba535a 371 */
2ba1c290
JA
372 if (get_next_offset(td, io_u)) {
373 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
bca4ed4d 374 return 1;
2ba1c290 375 }
10ba535a 376
9bf2061e 377 io_u->buflen = get_next_buflen(td, io_u);
2ba1c290
JA
378 if (!io_u->buflen) {
379 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
bca4ed4d 380 return 1;
2ba1c290 381 }
bca4ed4d 382
2ba1c290
JA
383 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
384 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
6a5e6884 385 return 1;
2ba1c290 386 }
6a5e6884 387
bca4ed4d
JA
388 /*
389 * mark entry before potentially trimming io_u
390 */
8fd15a9a 391 if (td_random(td) && !td->o.norandommap)
9bf2061e 392 mark_random_map(td, io_u);
bca4ed4d
JA
393
394 /*
395 * If using a write iolog, store this entry.
396 */
c38e9468 397out:
2ba1c290 398 dprint_io_u(io_u, "fill_io_u");
d9d91e39 399 td->zone_bytes += io_u->buflen;
f29b25a3 400 log_io_u(td, io_u);
bca4ed4d 401 return 0;
10ba535a
JA
402}
403
b3605062 404void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
71619dc2
JA
405{
406 int index = 0;
407
b3605062
JA
408 if (io_u->ddir == DDIR_SYNC)
409 return;
410
71619dc2
JA
411 switch (td->cur_depth) {
412 default:
a783e61a
JA
413 index = 6;
414 break;
71619dc2 415 case 32 ... 63:
a783e61a
JA
416 index = 5;
417 break;
71619dc2 418 case 16 ... 31:
a783e61a
JA
419 index = 4;
420 break;
71619dc2 421 case 8 ... 15:
a783e61a
JA
422 index = 3;
423 break;
71619dc2 424 case 4 ... 7:
a783e61a
JA
425 index = 2;
426 break;
71619dc2 427 case 2 ... 3:
a783e61a 428 index = 1;
71619dc2
JA
429 case 1:
430 break;
431 }
432
756867bd 433 td->ts.io_u_map[index]++;
b3605062 434 td->ts.total_io_u[io_u->ddir]++;
71619dc2
JA
435}
436
04a0feae
JA
437static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
438{
439 int index = 0;
440
441 assert(usec < 1000);
442
443 switch (usec) {
444 case 750 ... 999:
445 index = 9;
446 break;
447 case 500 ... 749:
448 index = 8;
449 break;
450 case 250 ... 499:
451 index = 7;
452 break;
453 case 100 ... 249:
454 index = 6;
455 break;
456 case 50 ... 99:
457 index = 5;
458 break;
459 case 20 ... 49:
460 index = 4;
461 break;
462 case 10 ... 19:
463 index = 3;
464 break;
465 case 4 ... 9:
466 index = 2;
467 break;
468 case 2 ... 3:
469 index = 1;
470 case 0 ... 1:
471 break;
472 }
473
474 assert(index < FIO_IO_U_LAT_U_NR);
475 td->ts.io_u_lat_u[index]++;
476}
477
478static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
ec118304
JA
479{
480 int index = 0;
481
482 switch (msec) {
483 default:
04a0feae
JA
484 index = 11;
485 break;
8abdce66 486 case 1000 ... 1999:
04a0feae
JA
487 index = 10;
488 break;
8abdce66 489 case 750 ... 999:
04a0feae
JA
490 index = 9;
491 break;
8abdce66 492 case 500 ... 749:
04a0feae
JA
493 index = 8;
494 break;
8abdce66 495 case 250 ... 499:
04a0feae
JA
496 index = 7;
497 break;
8abdce66 498 case 100 ... 249:
04a0feae
JA
499 index = 6;
500 break;
8abdce66 501 case 50 ... 99:
04a0feae
JA
502 index = 5;
503 break;
8abdce66 504 case 20 ... 49:
04a0feae
JA
505 index = 4;
506 break;
8abdce66 507 case 10 ... 19:
04a0feae
JA
508 index = 3;
509 break;
8abdce66 510 case 4 ... 9:
04a0feae
JA
511 index = 2;
512 break;
ec118304 513 case 2 ... 3:
04a0feae 514 index = 1;
ec118304
JA
515 case 0 ... 1:
516 break;
517 }
518
04a0feae
JA
519 assert(index < FIO_IO_U_LAT_M_NR);
520 td->ts.io_u_lat_m[index]++;
521}
522
523static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
524{
525 if (usec < 1000)
526 io_u_mark_lat_usec(td, usec);
527 else
528 io_u_mark_lat_msec(td, usec / 1000);
ec118304
JA
529}
530
0aabe160
JA
531/*
532 * Get next file to service by choosing one at random
533 */
1c178180
JA
534static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
535 int badf)
0aabe160 536{
0aabe160 537 struct fio_file *f;
1c178180 538 int fno;
0aabe160
JA
539
540 do {
7c83c089
JA
541 long r = os_random_long(&td->next_file_state);
542
2dc1bbeb 543 fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
1c178180 544 f = &td->files[fno];
059e63c0
JA
545 if (f->flags & FIO_FILE_DONE)
546 continue;
1c178180 547
2ba1c290
JA
548 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
549 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
0aabe160 550 return f;
2ba1c290 551 }
0aabe160
JA
552 } while (1);
553}
554
555/*
556 * Get next file to service by doing round robin between all available ones
557 */
1c178180
JA
558static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
559 int badf)
3d7c391d
JA
560{
561 unsigned int old_next_file = td->next_file;
562 struct fio_file *f;
563
564 do {
565 f = &td->files[td->next_file];
566
567 td->next_file++;
2dc1bbeb 568 if (td->next_file >= td->o.nr_files)
3d7c391d
JA
569 td->next_file = 0;
570
d5ed68ea
JA
571 if (f->flags & FIO_FILE_DONE) {
572 f = NULL;
059e63c0 573 continue;
d5ed68ea 574 }
059e63c0 575
1c178180 576 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
3d7c391d
JA
577 break;
578
579 f = NULL;
580 } while (td->next_file != old_next_file);
581
2ba1c290 582 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
3d7c391d
JA
583 return f;
584}
585
bdb4e2e9
JA
586static struct fio_file *get_next_file(struct thread_data *td)
587{
1907dbc6
JA
588 struct fio_file *f;
589
2dc1bbeb 590 assert(td->o.nr_files <= td->files_index);
1c178180 591
2ba1c290
JA
592 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
593 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d, nr_files=%d\n", td->nr_open_files, td->nr_done_files, td->o.nr_files);
bdb4e2e9 594 return NULL;
2ba1c290 595 }
bdb4e2e9 596
1907dbc6 597 f = td->file_service_file;
f11bd94d 598 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
2ba1c290 599 goto out;
1907dbc6 600
2dc1bbeb 601 if (td->o.file_service_type == FIO_FSERVICE_RR)
1c178180 602 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
bdb4e2e9 603 else
1c178180 604 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
1907dbc6
JA
605
606 td->file_service_file = f;
607 td->file_service_left = td->file_service_nr - 1;
2ba1c290
JA
608out:
609 dprint(FD_FILE, "get_next_file: %p\n", f);
1907dbc6 610 return f;
bdb4e2e9
JA
611}
612
1c178180
JA
613static struct fio_file *find_next_new_file(struct thread_data *td)
614{
615 struct fio_file *f;
616
1020a139
JA
617 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
618 return NULL;
619
2dc1bbeb 620 if (td->o.file_service_type == FIO_FSERVICE_RR)
1c178180
JA
621 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
622 else
623 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
624
625 return f;
626}
627
429f6675
JA
628static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
629{
630 struct fio_file *f;
631
632 do {
633 f = get_next_file(td);
634 if (!f)
635 return 1;
636
637set_file:
638 io_u->file = f;
639 get_file(f);
640
641 if (!fill_io_u(td, io_u))
642 break;
643
644 /*
645 * td_io_close() does a put_file() as well, so no need to
646 * do that here.
647 */
648 io_u->file = NULL;
649 td_io_close_file(td, f);
650 f->flags |= FIO_FILE_DONE;
651 td->nr_done_files++;
652
653 /*
654 * probably not the right place to do this, but see
655 * if we need to open a new file
656 */
657 if (td->nr_open_files < td->o.open_files &&
658 td->o.open_files != td->o.nr_files) {
659 f = find_next_new_file(td);
660
661 if (!f || td_io_open_file(td, f))
662 return 1;
663
664 goto set_file;
665 }
666 } while (1);
667
668 return 0;
669}
670
671
10ba535a
JA
672struct io_u *__get_io_u(struct thread_data *td)
673{
674 struct io_u *io_u = NULL;
675
755200a3
JA
676 if (!list_empty(&td->io_u_requeues))
677 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
678 else if (!queue_full(td)) {
10ba535a
JA
679 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
680
6040dabc 681 io_u->buflen = 0;
10ba535a 682 io_u->resid = 0;
755200a3 683 io_u->file = NULL;
d7762cf8 684 io_u->end_io = NULL;
755200a3
JA
685 }
686
687 if (io_u) {
0c6e7517
JA
688 assert(io_u->flags & IO_U_F_FREE);
689 io_u->flags &= ~IO_U_F_FREE;
690
755200a3 691 io_u->error = 0;
10ba535a
JA
692 list_del(&io_u->list);
693 list_add(&io_u->list, &td->io_u_busylist);
694 td->cur_depth++;
695 }
696
697 return io_u;
698}
699
700/*
701 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
702 * etc. The returned io_u is fully ready to be prepped and submitted.
703 */
3d7c391d 704struct io_u *get_io_u(struct thread_data *td)
10ba535a 705{
3d7c391d 706 struct fio_file *f;
10ba535a
JA
707 struct io_u *io_u;
708
709 io_u = __get_io_u(td);
2ba1c290
JA
710 if (!io_u) {
711 dprint(FD_IO, "__get_io_u failed\n");
10ba535a 712 return NULL;
2ba1c290 713 }
10ba535a 714
755200a3
JA
715 /*
716 * from a requeue, io_u already setup
717 */
718 if (io_u->file)
77f392bf 719 goto out;
755200a3 720
429f6675
JA
721 /*
722 * If using an iolog, grab next piece if any available.
723 */
724 if (td->o.read_iolog_file) {
725 if (read_iolog_get(td, io_u))
726 goto err_put;
2ba1c290
JA
727 } else if (set_io_u_file(td, io_u)) {
728 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
429f6675 729 goto err_put;
2ba1c290 730 }
429f6675
JA
731
732 f = io_u->file;
733 assert(f->flags & FIO_FILE_OPEN);
97af62ce 734
87dc1ab1 735 if (io_u->ddir != DDIR_SYNC) {
d0656a93 736 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
2ba1c290 737 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
429f6675 738 goto err_put;
2ba1c290 739 }
10ba535a 740
36167d82 741 f->last_pos = io_u->offset + io_u->buflen;
10ba535a 742
2dc1bbeb 743 if (td->o.verify != VERIFY_NONE)
87dc1ab1
JA
744 populate_verify_io_u(td, io_u);
745 }
10ba535a 746
165faf16
JA
747 /*
748 * Set io data pointers.
749 */
d460eb31 750 io_u->endpos = io_u->offset + io_u->buflen;
77f392bf 751out:
cec6b55d
JA
752 io_u->xfer_buf = io_u->buf;
753 io_u->xfer_buflen = io_u->buflen;
165faf16 754
429f6675
JA
755 if (!td_io_prep(td, io_u)) {
756 fio_gettime(&io_u->start_time, NULL);
757 return io_u;
36167d82 758 }
429f6675 759err_put:
2ba1c290 760 dprint(FD_IO, "get_io_u failed\n");
429f6675
JA
761 put_io_u(td, io_u);
762 return NULL;
10ba535a
JA
763}
764
5451792e
JA
765void io_u_log_error(struct thread_data *td, struct io_u *io_u)
766{
767 const char *msg[] = { "read", "write", "sync" };
768
769 log_err("fio: io_u error");
770
771 if (io_u->file)
772 log_err(" on file %s", io_u->file->file_name);
773
774 log_err(": %s\n", strerror(io_u->error));
775
776 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
777
778 if (!td->error)
779 td_verror(td, io_u->error, "io_u error");
780}
781
97601024
JA
782static void io_completed(struct thread_data *td, struct io_u *io_u,
783 struct io_completion_data *icd)
10ba535a 784{
d85f5118 785 unsigned long usec;
10ba535a 786
2ba1c290
JA
787 dprint_io_u(io_u, "io complete");
788
0c6e7517
JA
789 assert(io_u->flags & IO_U_F_FLIGHT);
790 io_u->flags &= ~IO_U_F_FLIGHT;
791
87dc1ab1
JA
792 if (io_u->ddir == DDIR_SYNC) {
793 td->last_was_sync = 1;
794 return;
795 }
796
797 td->last_was_sync = 0;
798
10ba535a
JA
799 if (!io_u->error) {
800 unsigned int bytes = io_u->buflen - io_u->resid;
1e97cce9 801 const enum fio_ddir idx = io_u->ddir;
3af6ef39 802 int ret;
10ba535a
JA
803
804 td->io_blocks[idx]++;
805 td->io_bytes[idx] += bytes;
10ba535a
JA
806 td->this_io_bytes[idx] += bytes;
807
d460eb31 808 io_u->file->last_completed_pos = io_u->endpos;
02bcaa8c 809
d85f5118 810 usec = utime_since(&io_u->issue_time, &icd->time);
10ba535a 811
d85f5118 812 add_clat_sample(td, idx, usec);
02bcaa8c 813 add_bw_sample(td, idx, &icd->time);
04a0feae 814 io_u_mark_latency(td, usec);
10ba535a 815
660a1cb5 816 if (td_write(td) && idx == DDIR_WRITE &&
8670579e 817 td->o.do_verify &&
41128405 818 td->o.verify != VERIFY_NONE)
10ba535a
JA
819 log_io_piece(td, io_u);
820
821 icd->bytes_done[idx] += bytes;
3af6ef39 822
d7762cf8 823 if (io_u->end_io) {
36690c9b 824 ret = io_u->end_io(td, io_u);
3af6ef39
JA
825 if (ret && !icd->error)
826 icd->error = ret;
827 }
5451792e 828 } else {
10ba535a 829 icd->error = io_u->error;
5451792e
JA
830 io_u_log_error(td, io_u);
831 }
10ba535a
JA
832}
833
d7762cf8 834static void init_icd(struct io_completion_data *icd, int nr)
10ba535a 835{
02bcaa8c
JA
836 fio_gettime(&icd->time, NULL);
837
3af6ef39
JA
838 icd->nr = nr;
839
10ba535a
JA
840 icd->error = 0;
841 icd->bytes_done[0] = icd->bytes_done[1] = 0;
36167d82
JA
842}
843
97601024
JA
844static void ios_completed(struct thread_data *td,
845 struct io_completion_data *icd)
36167d82
JA
846{
847 struct io_u *io_u;
848 int i;
849
10ba535a
JA
850 for (i = 0; i < icd->nr; i++) {
851 io_u = td->io_ops->event(td, i);
852
853 io_completed(td, io_u, icd);
854 put_io_u(td, io_u);
855 }
856}
97601024 857
e7e6cfb4
JA
858/*
859 * Complete a single io_u for the sync engines.
860 */
d7762cf8 861long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
97601024
JA
862{
863 struct io_completion_data icd;
864
d7762cf8 865 init_icd(&icd, 1);
97601024
JA
866 io_completed(td, io_u, &icd);
867 put_io_u(td, io_u);
868
869 if (!icd.error)
870 return icd.bytes_done[0] + icd.bytes_done[1];
871
37e974a5 872 td_verror(td, icd.error, "io_u_sync_complete");
97601024
JA
873 return -1;
874}
875
e7e6cfb4
JA
876/*
877 * Called to complete min_events number of io for the async engines.
878 */
d7762cf8 879long io_u_queued_complete(struct thread_data *td, int min_events)
97601024 880{
97601024 881 struct io_completion_data icd;
00de55ef 882 struct timespec *tvp = NULL;
97601024 883 int ret;
4d06a338 884 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
97601024 885
4d06a338 886 if (!min_events)
00de55ef 887 tvp = &ts;
97601024 888
00de55ef 889 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
97601024 890 if (ret < 0) {
e1161c32 891 td_verror(td, -ret, "td_io_getevents");
97601024
JA
892 return ret;
893 } else if (!ret)
894 return ret;
895
d7762cf8 896 init_icd(&icd, ret);
97601024
JA
897 ios_completed(td, &icd);
898 if (!icd.error)
899 return icd.bytes_done[0] + icd.bytes_done[1];
900
37e974a5 901 td_verror(td, icd.error, "io_u_queued_complete");
97601024
JA
902 return -1;
903}
7e77dd02
JA
904
905/*
906 * Call when io_u is really queued, to update the submission latency.
907 */
908void io_u_queued(struct thread_data *td, struct io_u *io_u)
909{
910 unsigned long slat_time;
911
d85f5118 912 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
7e77dd02
JA
913 add_slat_sample(td, io_u->ddir, slat_time);
914}
433afcb4 915
55bc9728 916#ifdef FIO_USE_TIMEOUT
433afcb4
JA
917void io_u_set_timeout(struct thread_data *td)
918{
919 assert(td->cur_depth);
920
921 td->timer.it_interval.tv_sec = 0;
922 td->timer.it_interval.tv_usec = 0;
923 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
924 td->timer.it_value.tv_usec = 0;
925 setitimer(ITIMER_REAL, &td->timer, NULL);
926 fio_gettime(&td->timeout_end, NULL);
927}
5945b9b4
JA
928
929static void io_u_dump(struct io_u *io_u)
930{
931 unsigned long t_start = mtime_since_now(&io_u->start_time);
932 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
933
934 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
935 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);
936 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
937}
55bc9728
JA
938#else
939void io_u_set_timeout(struct thread_data fio_unused *td)
940{
941}
942#endif
433afcb4 943
55bc9728 944#ifdef FIO_USE_TIMEOUT
433afcb4
JA
945static void io_u_timeout_handler(int fio_unused sig)
946{
947 struct thread_data *td, *__td;
948 pid_t pid = getpid();
5945b9b4
JA
949 struct list_head *entry;
950 struct io_u *io_u;
433afcb4
JA
951 int i;
952
953 log_err("fio: io_u timeout\n");
954
955 /*
956 * TLS would be nice...
957 */
958 td = NULL;
959 for_each_td(__td, i) {
960 if (__td->pid == pid) {
961 td = __td;
962 break;
963 }
964 }
965
966 if (!td) {
967 log_err("fio: io_u timeout, can't find job\n");
968 exit(1);
969 }
970
971 if (!td->cur_depth) {
972 log_err("fio: timeout without pending work?\n");
973 return;
974 }
975
15506d09 976 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
5945b9b4
JA
977
978 list_for_each(entry, &td->io_u_busylist) {
979 io_u = list_entry(entry, struct io_u, list);
980
981 io_u_dump(io_u);
982 }
983
984 td_verror(td, ETIMEDOUT, "io_u timeout");
433afcb4
JA
985 exit(1);
986}
55bc9728 987#endif
433afcb4
JA
988
989void io_u_init_timeout(void)
990{
55bc9728 991#ifdef FIO_USE_TIMEOUT
433afcb4 992 signal(SIGALRM, io_u_timeout_handler);
55bc9728 993#endif
433afcb4 994}