Add version 2 of the iolog format
[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 */
211097b2 58 if (!td->o.ddir_nr && !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 */
ec4015da 121 rb = *b + (f->file_offset / td->o.min_bs[ddir]);
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
JA
164 } else {
165 if (f->last_pos >= f->real_file_size)
166 return 1;
167
2dc1bbeb 168 b = f->last_pos / td->o.min_bs[ddir];
43063a1c 169 }
10ba535a 170
2dc1bbeb 171 io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
bca4ed4d 172 if (io_u->offset >= f->real_file_size)
10ba535a
JA
173 return 1;
174
175 return 0;
176}
177
9bf2061e 178static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
10ba535a 179{
bca4ed4d 180 const int ddir = io_u->ddir;
10ba535a
JA
181 unsigned int buflen;
182 long r;
183
2dc1bbeb
JA
184 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
185 buflen = td->o.min_bs[ddir];
10ba535a
JA
186 else {
187 r = os_random_long(&td->bsrange_state);
2dc1bbeb
JA
188 buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
189 if (!td->o.bs_unaligned)
190 buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
10ba535a
JA
191 }
192
10ba535a
JA
193 return buflen;
194}
195
afe24a5a
JA
196static void set_rwmix_bytes(struct thread_data *td)
197{
198 unsigned long long rbytes;
199 unsigned int diff;
200
201 /*
202 * we do time or byte based switch. this is needed because
203 * buffered writes may issue a lot quicker than they complete,
204 * whereas reads do not.
205 */
206 rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
e47f799f 207 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
afe24a5a 208
e47f799f
JA
209 td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
210}
211
212static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
213{
214 unsigned int v;
215 long r;
216
217 r = os_random_long(&td->rwmix_state);
218 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
219 if (v < td->o.rwmix[DDIR_READ])
220 return DDIR_READ;
221
222 return DDIR_WRITE;
afe24a5a
JA
223}
224
10ba535a
JA
225/*
226 * Return the data direction for the next io_u. If the job is a
227 * mixed read/write workload, check the rwmix cycle and switch if
228 * necessary.
229 */
1e97cce9 230static enum fio_ddir get_rw_ddir(struct thread_data *td)
10ba535a
JA
231{
232 if (td_rw(td)) {
233 struct timeval now;
234 unsigned long elapsed;
afe24a5a 235 unsigned int cycle;
10ba535a 236
02bcaa8c 237 fio_gettime(&now, NULL);
10ba535a
JA
238 elapsed = mtime_since_now(&td->rwmix_switch);
239
e47f799f
JA
240 /*
241 * if this is the first cycle, make it shorter
242 */
afe24a5a
JA
243 cycle = td->o.rwmixcycle;
244 if (!td->rwmix_bytes)
245 cycle /= 10;
246
10ba535a
JA
247 /*
248 * Check if it's time to seed a new data direction.
249 */
e47f799f 250 if (elapsed >= cycle ||
afe24a5a 251 td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
e47f799f
JA
252 unsigned long long max_bytes;
253 enum fio_ddir ddir;
254
255 /*
256 * Put a top limit on how many bytes we do for
257 * one data direction, to avoid overflowing the
258 * ranges too much
259 */
260 ddir = get_rand_ddir(td);
261 max_bytes = td->this_io_bytes[ddir];
7bb48f84 262 if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
38d77cae
JA
263 if (!td->rw_end_set[ddir]) {
264 td->rw_end_set[ddir] = 1;
265 memcpy(&td->rw_end[ddir], &now, sizeof(now));
266 }
e47f799f 267 ddir ^= 1;
38d77cae 268 }
e47f799f
JA
269
270 if (ddir != td->rwmix_ddir)
271 set_rwmix_bytes(td);
272
273 td->rwmix_ddir = ddir;
10ba535a
JA
274 memcpy(&td->rwmix_switch, &now, sizeof(now));
275 }
276 return td->rwmix_ddir;
277 } else if (td_read(td))
278 return DDIR_READ;
279 else
280 return DDIR_WRITE;
281}
282
10ba535a
JA
283void put_io_u(struct thread_data *td, struct io_u *io_u)
284{
0c6e7517
JA
285 assert((io_u->flags & IO_U_F_FREE) == 0);
286 io_u->flags |= IO_U_F_FREE;
287
2dbdab7e
JA
288 if (io_u->file)
289 put_file(td, io_u->file);
290
10ba535a
JA
291 io_u->file = NULL;
292 list_del(&io_u->list);
293 list_add(&io_u->list, &td->io_u_freelist);
294 td->cur_depth--;
295}
296
755200a3
JA
297void requeue_io_u(struct thread_data *td, struct io_u **io_u)
298{
299 struct io_u *__io_u = *io_u;
300
4d2e0f49
JA
301 __io_u->flags |= IO_U_F_FREE;
302 __io_u->flags &= ~IO_U_F_FLIGHT;
303
755200a3
JA
304 list_del(&__io_u->list);
305 list_add_tail(&__io_u->list, &td->io_u_requeues);
306 td->cur_depth--;
307 *io_u = NULL;
308}
309
9bf2061e 310static int fill_io_u(struct thread_data *td, struct io_u *io_u)
10ba535a
JA
311{
312 /*
313 * If using an iolog, grab next piece if any available.
314 */
fcb11708 315 if (td->o.read_iolog_file)
10ba535a
JA
316 return read_iolog_get(td, io_u);
317
87dc1ab1
JA
318 /*
319 * see if it's time to sync
320 */
2dc1bbeb
JA
321 if (td->o.fsync_blocks &&
322 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
323 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
87dc1ab1 324 io_u->ddir = DDIR_SYNC;
c38e9468 325 goto out;
87dc1ab1
JA
326 }
327
a00735e6
JA
328 io_u->ddir = get_rw_ddir(td);
329
48f5abd3
JA
330 /*
331 * See if it's time to switch to a new zone
332 */
333 if (td->zone_bytes >= td->o.zone_size) {
334 td->zone_bytes = 0;
335 io_u->file->last_pos += td->o.zone_skip;
336 td->io_skip_bytes += td->o.zone_skip;
337 }
338
10ba535a 339 /*
c685b5b2
JA
340 * No log, let the seq/rand engine retrieve the next buflen and
341 * position.
10ba535a 342 */
9bf2061e 343 if (get_next_offset(td, io_u))
bca4ed4d 344 return 1;
10ba535a 345
9bf2061e 346 io_u->buflen = get_next_buflen(td, io_u);
bca4ed4d
JA
347 if (!io_u->buflen)
348 return 1;
349
350 /*
351 * mark entry before potentially trimming io_u
352 */
8fd15a9a 353 if (td_random(td) && !td->o.norandommap)
9bf2061e 354 mark_random_map(td, io_u);
bca4ed4d
JA
355
356 /*
357 * If using a write iolog, store this entry.
358 */
c38e9468 359out:
f29b25a3 360 log_io_u(td, io_u);
bca4ed4d 361 return 0;
10ba535a
JA
362}
363
b3605062 364void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
71619dc2
JA
365{
366 int index = 0;
367
b3605062
JA
368 if (io_u->ddir == DDIR_SYNC)
369 return;
370
71619dc2
JA
371 switch (td->cur_depth) {
372 default:
a783e61a
JA
373 index = 6;
374 break;
71619dc2 375 case 32 ... 63:
a783e61a
JA
376 index = 5;
377 break;
71619dc2 378 case 16 ... 31:
a783e61a
JA
379 index = 4;
380 break;
71619dc2 381 case 8 ... 15:
a783e61a
JA
382 index = 3;
383 break;
71619dc2 384 case 4 ... 7:
a783e61a
JA
385 index = 2;
386 break;
71619dc2 387 case 2 ... 3:
a783e61a 388 index = 1;
71619dc2
JA
389 case 1:
390 break;
391 }
392
756867bd 393 td->ts.io_u_map[index]++;
b3605062 394 td->ts.total_io_u[io_u->ddir]++;
71619dc2
JA
395}
396
04a0feae
JA
397static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
398{
399 int index = 0;
400
401 assert(usec < 1000);
402
403 switch (usec) {
404 case 750 ... 999:
405 index = 9;
406 break;
407 case 500 ... 749:
408 index = 8;
409 break;
410 case 250 ... 499:
411 index = 7;
412 break;
413 case 100 ... 249:
414 index = 6;
415 break;
416 case 50 ... 99:
417 index = 5;
418 break;
419 case 20 ... 49:
420 index = 4;
421 break;
422 case 10 ... 19:
423 index = 3;
424 break;
425 case 4 ... 9:
426 index = 2;
427 break;
428 case 2 ... 3:
429 index = 1;
430 case 0 ... 1:
431 break;
432 }
433
434 assert(index < FIO_IO_U_LAT_U_NR);
435 td->ts.io_u_lat_u[index]++;
436}
437
438static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
ec118304
JA
439{
440 int index = 0;
441
442 switch (msec) {
443 default:
04a0feae
JA
444 index = 11;
445 break;
8abdce66 446 case 1000 ... 1999:
04a0feae
JA
447 index = 10;
448 break;
8abdce66 449 case 750 ... 999:
04a0feae
JA
450 index = 9;
451 break;
8abdce66 452 case 500 ... 749:
04a0feae
JA
453 index = 8;
454 break;
8abdce66 455 case 250 ... 499:
04a0feae
JA
456 index = 7;
457 break;
8abdce66 458 case 100 ... 249:
04a0feae
JA
459 index = 6;
460 break;
8abdce66 461 case 50 ... 99:
04a0feae
JA
462 index = 5;
463 break;
8abdce66 464 case 20 ... 49:
04a0feae
JA
465 index = 4;
466 break;
8abdce66 467 case 10 ... 19:
04a0feae
JA
468 index = 3;
469 break;
8abdce66 470 case 4 ... 9:
04a0feae
JA
471 index = 2;
472 break;
ec118304 473 case 2 ... 3:
04a0feae 474 index = 1;
ec118304
JA
475 case 0 ... 1:
476 break;
477 }
478
04a0feae
JA
479 assert(index < FIO_IO_U_LAT_M_NR);
480 td->ts.io_u_lat_m[index]++;
481}
482
483static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
484{
485 if (usec < 1000)
486 io_u_mark_lat_usec(td, usec);
487 else
488 io_u_mark_lat_msec(td, usec / 1000);
ec118304
JA
489}
490
0aabe160
JA
491/*
492 * Get next file to service by choosing one at random
493 */
1c178180
JA
494static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
495 int badf)
0aabe160 496{
0aabe160 497 struct fio_file *f;
1c178180 498 int fno;
0aabe160
JA
499
500 do {
7c83c089
JA
501 long r = os_random_long(&td->next_file_state);
502
2dc1bbeb 503 fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
1c178180 504 f = &td->files[fno];
059e63c0
JA
505 if (f->flags & FIO_FILE_DONE)
506 continue;
1c178180
JA
507
508 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
0aabe160
JA
509 return f;
510 } while (1);
511}
512
513/*
514 * Get next file to service by doing round robin between all available ones
515 */
1c178180
JA
516static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
517 int badf)
3d7c391d
JA
518{
519 unsigned int old_next_file = td->next_file;
520 struct fio_file *f;
521
522 do {
523 f = &td->files[td->next_file];
524
525 td->next_file++;
2dc1bbeb 526 if (td->next_file >= td->o.nr_files)
3d7c391d
JA
527 td->next_file = 0;
528
d5ed68ea
JA
529 if (f->flags & FIO_FILE_DONE) {
530 f = NULL;
059e63c0 531 continue;
d5ed68ea 532 }
059e63c0 533
1c178180 534 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
3d7c391d
JA
535 break;
536
537 f = NULL;
538 } while (td->next_file != old_next_file);
539
540 return f;
541}
542
bdb4e2e9
JA
543static struct fio_file *get_next_file(struct thread_data *td)
544{
1907dbc6
JA
545 struct fio_file *f;
546
2dc1bbeb 547 assert(td->o.nr_files <= td->files_index);
1c178180 548
1020a139 549 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
bdb4e2e9
JA
550 return NULL;
551
1907dbc6 552 f = td->file_service_file;
f11bd94d 553 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
1907dbc6
JA
554 return f;
555
2dc1bbeb 556 if (td->o.file_service_type == FIO_FSERVICE_RR)
1c178180 557 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
bdb4e2e9 558 else
1c178180 559 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
1907dbc6
JA
560
561 td->file_service_file = f;
562 td->file_service_left = td->file_service_nr - 1;
563 return f;
bdb4e2e9
JA
564}
565
1c178180
JA
566static struct fio_file *find_next_new_file(struct thread_data *td)
567{
568 struct fio_file *f;
569
1020a139
JA
570 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
571 return NULL;
572
2dc1bbeb 573 if (td->o.file_service_type == FIO_FSERVICE_RR)
1c178180
JA
574 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
575 else
576 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
577
578 return f;
579}
580
10ba535a
JA
581struct io_u *__get_io_u(struct thread_data *td)
582{
583 struct io_u *io_u = NULL;
584
755200a3
JA
585 if (!list_empty(&td->io_u_requeues))
586 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
587 else if (!queue_full(td)) {
10ba535a
JA
588 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
589
6040dabc 590 io_u->buflen = 0;
10ba535a 591 io_u->resid = 0;
755200a3 592 io_u->file = NULL;
d7762cf8 593 io_u->end_io = NULL;
755200a3
JA
594 }
595
596 if (io_u) {
0c6e7517
JA
597 assert(io_u->flags & IO_U_F_FREE);
598 io_u->flags &= ~IO_U_F_FREE;
599
755200a3 600 io_u->error = 0;
10ba535a
JA
601 list_del(&io_u->list);
602 list_add(&io_u->list, &td->io_u_busylist);
603 td->cur_depth++;
604 }
605
606 return io_u;
607}
608
609/*
610 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
611 * etc. The returned io_u is fully ready to be prepped and submitted.
612 */
3d7c391d 613struct io_u *get_io_u(struct thread_data *td)
10ba535a 614{
3d7c391d 615 struct fio_file *f;
10ba535a 616 struct io_u *io_u;
1c178180 617 int ret;
10ba535a
JA
618
619 io_u = __get_io_u(td);
620 if (!io_u)
621 return NULL;
622
755200a3
JA
623 /*
624 * from a requeue, io_u already setup
625 */
626 if (io_u->file)
77f392bf 627 goto out;
755200a3 628
bdb4e2e9
JA
629 do {
630 f = get_next_file(td);
631 if (!f) {
632 put_io_u(td, io_u);
633 return NULL;
634 }
3d7c391d 635
1c178180 636set_file:
bdb4e2e9 637 io_u->file = f;
2dbdab7e 638 get_file(f);
3d7c391d 639
bdb4e2e9
JA
640 if (!fill_io_u(td, io_u))
641 break;
10ba535a 642
bdb4e2e9 643 /*
2dbdab7e
JA
644 * td_io_close() does a put_file() as well, so no need to
645 * do that here.
bdb4e2e9
JA
646 */
647 io_u->file = NULL;
b5af8293 648 td_io_close_file(td, f);
059e63c0 649 f->flags |= FIO_FILE_DONE;
1020a139 650 td->nr_done_files++;
b5af8293
JA
651
652 /*
653 * probably not the right place to do this, but see
654 * if we need to open a new file
655 */
2dc1bbeb
JA
656 if (td->nr_open_files < td->o.open_files &&
657 td->o.open_files != td->o.nr_files) {
1c178180 658 f = find_next_new_file(td);
a978ba68 659
1c178180 660 if (!f || (ret = td_io_open_file(td, f))) {
a978ba68 661 put_io_u(td, io_u);
e291dc12 662 return NULL;
a978ba68 663 }
1c178180 664 goto set_file;
a978ba68 665 }
bdb4e2e9 666 } while (1);
10ba535a 667
97af62ce
JA
668 assert(io_u->file->flags & FIO_FILE_OPEN);
669
87dc1ab1
JA
670 if (io_u->ddir != DDIR_SYNC) {
671 if (!io_u->buflen) {
672 put_io_u(td, io_u);
673 return NULL;
674 }
10ba535a 675
36167d82 676 f->last_pos = io_u->offset + io_u->buflen;
10ba535a 677
2dc1bbeb 678 if (td->o.verify != VERIFY_NONE)
87dc1ab1
JA
679 populate_verify_io_u(td, io_u);
680 }
10ba535a 681
165faf16
JA
682 /*
683 * Set io data pointers.
684 */
d460eb31 685 io_u->endpos = io_u->offset + io_u->buflen;
77f392bf 686out:
cec6b55d
JA
687 io_u->xfer_buf = io_u->buf;
688 io_u->xfer_buflen = io_u->buflen;
165faf16 689
36167d82
JA
690 if (td_io_prep(td, io_u)) {
691 put_io_u(td, io_u);
692 return NULL;
693 }
694
02bcaa8c 695 fio_gettime(&io_u->start_time, NULL);
10ba535a
JA
696 return io_u;
697}
698
5451792e
JA
699void io_u_log_error(struct thread_data *td, struct io_u *io_u)
700{
701 const char *msg[] = { "read", "write", "sync" };
702
703 log_err("fio: io_u error");
704
705 if (io_u->file)
706 log_err(" on file %s", io_u->file->file_name);
707
708 log_err(": %s\n", strerror(io_u->error));
709
710 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
711
712 if (!td->error)
713 td_verror(td, io_u->error, "io_u error");
714}
715
97601024
JA
716static void io_completed(struct thread_data *td, struct io_u *io_u,
717 struct io_completion_data *icd)
10ba535a 718{
d85f5118 719 unsigned long usec;
10ba535a 720
0c6e7517
JA
721 assert(io_u->flags & IO_U_F_FLIGHT);
722 io_u->flags &= ~IO_U_F_FLIGHT;
723
87dc1ab1
JA
724 if (io_u->ddir == DDIR_SYNC) {
725 td->last_was_sync = 1;
726 return;
727 }
728
729 td->last_was_sync = 0;
730
10ba535a
JA
731 if (!io_u->error) {
732 unsigned int bytes = io_u->buflen - io_u->resid;
1e97cce9 733 const enum fio_ddir idx = io_u->ddir;
3af6ef39 734 int ret;
10ba535a
JA
735
736 td->io_blocks[idx]++;
737 td->io_bytes[idx] += bytes;
738 td->zone_bytes += bytes;
739 td->this_io_bytes[idx] += bytes;
740
d460eb31 741 io_u->file->last_completed_pos = io_u->endpos;
02bcaa8c 742
d85f5118 743 usec = utime_since(&io_u->issue_time, &icd->time);
10ba535a 744
d85f5118 745 add_clat_sample(td, idx, usec);
02bcaa8c 746 add_bw_sample(td, idx, &icd->time);
04a0feae 747 io_u_mark_latency(td, usec);
10ba535a 748
660a1cb5 749 if (td_write(td) && idx == DDIR_WRITE &&
41128405 750 td->o.verify != VERIFY_NONE)
10ba535a
JA
751 log_io_piece(td, io_u);
752
753 icd->bytes_done[idx] += bytes;
3af6ef39 754
d7762cf8 755 if (io_u->end_io) {
36690c9b 756 ret = io_u->end_io(td, io_u);
3af6ef39
JA
757 if (ret && !icd->error)
758 icd->error = ret;
759 }
5451792e 760 } else {
10ba535a 761 icd->error = io_u->error;
5451792e
JA
762 io_u_log_error(td, io_u);
763 }
10ba535a
JA
764}
765
d7762cf8 766static void init_icd(struct io_completion_data *icd, int nr)
10ba535a 767{
02bcaa8c
JA
768 fio_gettime(&icd->time, NULL);
769
3af6ef39
JA
770 icd->nr = nr;
771
10ba535a
JA
772 icd->error = 0;
773 icd->bytes_done[0] = icd->bytes_done[1] = 0;
36167d82
JA
774}
775
97601024
JA
776static void ios_completed(struct thread_data *td,
777 struct io_completion_data *icd)
36167d82
JA
778{
779 struct io_u *io_u;
780 int i;
781
10ba535a
JA
782 for (i = 0; i < icd->nr; i++) {
783 io_u = td->io_ops->event(td, i);
784
785 io_completed(td, io_u, icd);
786 put_io_u(td, io_u);
787 }
788}
97601024 789
e7e6cfb4
JA
790/*
791 * Complete a single io_u for the sync engines.
792 */
d7762cf8 793long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
97601024
JA
794{
795 struct io_completion_data icd;
796
d7762cf8 797 init_icd(&icd, 1);
97601024
JA
798 io_completed(td, io_u, &icd);
799 put_io_u(td, io_u);
800
801 if (!icd.error)
802 return icd.bytes_done[0] + icd.bytes_done[1];
803
37e974a5 804 td_verror(td, icd.error, "io_u_sync_complete");
97601024
JA
805 return -1;
806}
807
e7e6cfb4
JA
808/*
809 * Called to complete min_events number of io for the async engines.
810 */
d7762cf8 811long io_u_queued_complete(struct thread_data *td, int min_events)
97601024 812{
97601024 813 struct io_completion_data icd;
00de55ef 814 struct timespec *tvp = NULL;
97601024 815 int ret;
4d06a338 816 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
97601024 817
4d06a338 818 if (!min_events)
00de55ef 819 tvp = &ts;
97601024 820
00de55ef 821 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
97601024 822 if (ret < 0) {
e1161c32 823 td_verror(td, -ret, "td_io_getevents");
97601024
JA
824 return ret;
825 } else if (!ret)
826 return ret;
827
d7762cf8 828 init_icd(&icd, ret);
97601024
JA
829 ios_completed(td, &icd);
830 if (!icd.error)
831 return icd.bytes_done[0] + icd.bytes_done[1];
832
37e974a5 833 td_verror(td, icd.error, "io_u_queued_complete");
97601024
JA
834 return -1;
835}
7e77dd02
JA
836
837/*
838 * Call when io_u is really queued, to update the submission latency.
839 */
840void io_u_queued(struct thread_data *td, struct io_u *io_u)
841{
842 unsigned long slat_time;
843
d85f5118 844 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
7e77dd02
JA
845 add_slat_sample(td, io_u->ddir, slat_time);
846}
433afcb4 847
55bc9728 848#ifdef FIO_USE_TIMEOUT
433afcb4
JA
849void io_u_set_timeout(struct thread_data *td)
850{
851 assert(td->cur_depth);
852
853 td->timer.it_interval.tv_sec = 0;
854 td->timer.it_interval.tv_usec = 0;
855 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
856 td->timer.it_value.tv_usec = 0;
857 setitimer(ITIMER_REAL, &td->timer, NULL);
858 fio_gettime(&td->timeout_end, NULL);
859}
5945b9b4
JA
860
861static void io_u_dump(struct io_u *io_u)
862{
863 unsigned long t_start = mtime_since_now(&io_u->start_time);
864 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
865
866 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
867 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);
868 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
869}
55bc9728
JA
870#else
871void io_u_set_timeout(struct thread_data fio_unused *td)
872{
873}
874#endif
433afcb4 875
55bc9728 876#ifdef FIO_USE_TIMEOUT
433afcb4
JA
877static void io_u_timeout_handler(int fio_unused sig)
878{
879 struct thread_data *td, *__td;
880 pid_t pid = getpid();
5945b9b4
JA
881 struct list_head *entry;
882 struct io_u *io_u;
433afcb4
JA
883 int i;
884
885 log_err("fio: io_u timeout\n");
886
887 /*
888 * TLS would be nice...
889 */
890 td = NULL;
891 for_each_td(__td, i) {
892 if (__td->pid == pid) {
893 td = __td;
894 break;
895 }
896 }
897
898 if (!td) {
899 log_err("fio: io_u timeout, can't find job\n");
900 exit(1);
901 }
902
903 if (!td->cur_depth) {
904 log_err("fio: timeout without pending work?\n");
905 return;
906 }
907
15506d09 908 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
5945b9b4
JA
909
910 list_for_each(entry, &td->io_u_busylist) {
911 io_u = list_entry(entry, struct io_u, list);
912
913 io_u_dump(io_u);
914 }
915
916 td_verror(td, ETIMEDOUT, "io_u timeout");
433afcb4
JA
917 exit(1);
918}
55bc9728 919#endif
433afcb4
JA
920
921void io_u_init_timeout(void)
922{
55bc9728 923#ifdef FIO_USE_TIMEOUT
433afcb4 924 signal(SIGALRM, io_u_timeout_handler);
55bc9728 925#endif
433afcb4 926}