projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
When retrieving a requeued IO, don't setup data pointers again
[fio.git] / io_u.c
... / ...
CommitLineData
1#include <unistd.h>
2#include <fcntl.h>
3#include <string.h>
4#include <signal.h>
5#include <time.h>
6#include <assert.h>
7
8#include "fio.h"
9
10/*
11 * Change this define to play with the timeout handling
12 */
13#undef FIO_USE_TIMEOUT
14
15struct io_completion_data {
16 int nr; /* input */
17
18 int error; /* output */
19 unsigned long bytes_done[2]; /* output */
20 struct timeval time; /* output */
21};
22
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 const 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 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
34
35 return (f->file_map[idx] & (1UL << bit)) == 0;
36}
37
38/*
39 * Mark a given offset as used in the map.
40 */
41static void mark_random_map(struct thread_data *td, struct io_u *io_u)
42{
43 unsigned int min_bs = td->o.rw_min_bs;
44 struct fio_file *f = io_u->file;
45 unsigned long long block;
46 unsigned int blocks;
47 unsigned int nr_blocks;
48
49 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
50 blocks = 0;
51 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
52
53 while (blocks < nr_blocks) {
54 unsigned int idx, bit;
55
56 /*
57 * If we have a mixed random workload, we may
58 * encounter blocks we already did IO to.
59 */
60 if ((td->o.ddir_nr == 1) && !random_map_free(td, f, block))
61 break;
62
63 idx = RAND_MAP_IDX(td, f, block);
64 bit = RAND_MAP_BIT(td, f, block);
65
66 fio_assert(td, idx < f->num_maps);
67
68 f->file_map[idx] |= (1UL << bit);
69 block++;
70 blocks++;
71 }
72
73 if ((blocks * min_bs) < io_u->buflen)
74 io_u->buflen = blocks * min_bs;
75}
76
77static inline unsigned long long last_block(struct thread_data *td,
78 struct fio_file *f,
79 enum fio_ddir ddir)
80{
81 unsigned long long max_blocks;
82
83 max_blocks = f->io_size / (unsigned long long) td->o.min_bs[ddir];
84 if (!max_blocks)
85 return 0;
86
87 return max_blocks - 1;
88}
89
90/*
91 * Return the next free block in the map.
92 */
93static int get_next_free_block(struct thread_data *td, struct fio_file *f,
94 enum fio_ddir ddir, unsigned long long *b)
95{
96 unsigned long long min_bs = td->o.rw_min_bs;
97 int i;
98
99 i = f->last_free_lookup;
100 *b = (i * BLOCKS_PER_MAP);
101 while ((*b) * min_bs < f->real_file_size) {
102 if (f->file_map[i] != -1UL) {
103 *b += fio_ffz(f->file_map[i]);
104 if (*b > last_block(td, f, ddir))
105 break;
106 f->last_free_lookup = i;
107 return 0;
108 }
109
110 *b += BLOCKS_PER_MAP;
111 i++;
112 }
113
114 dprint(FD_IO, "failed finding a free block\n");
115 return 1;
116}
117
118static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
119 enum fio_ddir ddir, unsigned long long *b)
120{
121 unsigned long long r;
122 int loops = 5;
123
124 do {
125 r = os_random_long(&td->random_state);
126 dprint(FD_RANDOM, "off rand %llu\n", r);
127 *b = (last_block(td, f, ddir) - 1) * (r / ((unsigned long long) RAND_MAX + 1.0));
128
129 /*
130 * if we are not maintaining a random map, we are done.
131 */
132 if (td->o.norandommap)
133 return 0;
134
135 /*
136 * calculate map offset and check if it's free
137 */
138 if (random_map_free(td, f, *b))
139 return 0;
140
141 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
142 *b);
143 } while (--loops);
144
145 /*
146 * we get here, if we didn't suceed in looking up a block. generate
147 * a random start offset into the filemap, and find the first free
148 * block from there.
149 */
150 loops = 10;
151 do {
152 f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
153 if (!get_next_free_block(td, f, ddir, b))
154 return 0;
155
156 r = os_random_long(&td->random_state);
157 } while (--loops);
158
159 /*
160 * that didn't work either, try exhaustive search from the start
161 */
162 f->last_free_lookup = 0;
163 return get_next_free_block(td, f, ddir, b);
164}
165
166/*
167 * For random io, generate a random new block and see if it's used. Repeat
168 * until we find a free one. For sequential io, just return the end of
169 * the last io issued.
170 */
171static int get_next_offset(struct thread_data *td, struct io_u *io_u)
172{
173 struct fio_file *f = io_u->file;
174 unsigned long long b;
175 enum fio_ddir ddir = io_u->ddir;
176
177 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
178 td->ddir_nr = td->o.ddir_nr;
179
180 if (get_next_rand_offset(td, f, ddir, &b))
181 return 1;
182 } else {
183 if (f->last_pos >= f->real_file_size) {
184 if (!td_random(td) ||
185 get_next_rand_offset(td, f, ddir, &b))
186 return 1;
187 } else
188 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
189 }
190
191 io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
192 if (io_u->offset >= f->real_file_size) {
193 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
194 io_u->offset, f->real_file_size);
195 return 1;
196 }
197
198 return 0;
199}
200
201static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
202{
203 const int ddir = io_u->ddir;
204 unsigned int buflen;
205 long r;
206
207 if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
208 buflen = td->o.min_bs[ddir];
209 else {
210 r = os_random_long(&td->bsrange_state);
211 if (!td->o.bssplit_nr)
212 buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
213 else {
214 long perc = 0;
215 unsigned int i;
216
217 for (i = 0; i < td->o.bssplit_nr; i++) {
218 struct bssplit *bsp = &td->o.bssplit[i];
219
220 buflen = bsp->bs;
221 perc += bsp->perc;
222 if (r <= ((LONG_MAX / 100L) * perc))
223 break;
224 }
225 }
226 if (!td->o.bs_unaligned)
227 buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
228 }
229
230 if (io_u->offset + buflen > io_u->file->real_file_size) {
231 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
232 td->o.min_bs[ddir], ddir);
233 buflen = td->o.min_bs[ddir];
234 }
235
236 return buflen;
237}
238
239static void set_rwmix_bytes(struct thread_data *td)
240{
241 unsigned long long rbytes;
242 unsigned int diff;
243
244 /*
245 * we do time or byte based switch. this is needed because
246 * buffered writes may issue a lot quicker than they complete,
247 * whereas reads do not.
248 */
249 rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
250 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
251
252 td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
253}
254
255static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
256{
257 unsigned int v;
258 long r;
259
260 r = os_random_long(&td->rwmix_state);
261 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
262 if (v < td->o.rwmix[DDIR_READ])
263 return DDIR_READ;
264
265 return DDIR_WRITE;
266}
267
268/*
269 * Return the data direction for the next io_u. If the job is a
270 * mixed read/write workload, check the rwmix cycle and switch if
271 * necessary.
272 */
273static enum fio_ddir get_rw_ddir(struct thread_data *td)
274{
275 if (td_rw(td)) {
276 struct timeval now;
277 unsigned long elapsed;
278 unsigned int cycle;
279
280 fio_gettime(&now, NULL);
281 elapsed = mtime_since_now(&td->rwmix_switch);
282
283 /*
284 * if this is the first cycle, make it shorter
285 */
286 cycle = td->o.rwmixcycle;
287 if (!td->rwmix_bytes)
288 cycle /= 10;
289
290 /*
291 * Check if it's time to seed a new data direction.
292 */
293 if (elapsed >= cycle ||
294 td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
295 unsigned long long max_bytes;
296 enum fio_ddir ddir;
297
298 /*
299 * Put a top limit on how many bytes we do for
300 * one data direction, to avoid overflowing the
301 * ranges too much
302 */
303 ddir = get_rand_ddir(td);
304 max_bytes = td->this_io_bytes[ddir];
305 if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
306 if (!td->rw_end_set[ddir]) {
307 td->rw_end_set[ddir] = 1;
308 memcpy(&td->rw_end[ddir], &now, sizeof(now));
309 }
310 ddir ^= 1;
311 }
312
313 if (ddir != td->rwmix_ddir)
314 set_rwmix_bytes(td);
315
316 td->rwmix_ddir = ddir;
317 memcpy(&td->rwmix_switch, &now, sizeof(now));
318 }
319 return td->rwmix_ddir;
320 } else if (td_read(td))
321 return DDIR_READ;
322 else
323 return DDIR_WRITE;
324}
325
326void put_io_u(struct thread_data *td, struct io_u *io_u)
327{
328 assert((io_u->flags & IO_U_F_FREE) == 0);
329 io_u->flags |= IO_U_F_FREE;
330
331 if (io_u->file)
332 put_file(td, io_u->file);
333
334 io_u->file = NULL;
335 list_del(&io_u->list);
336 list_add(&io_u->list, &td->io_u_freelist);
337 td->cur_depth--;
338}
339
340void requeue_io_u(struct thread_data *td, struct io_u **io_u)
341{
342 struct io_u *__io_u = *io_u;
343
344 __io_u->flags |= IO_U_F_FREE;
345 if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
346 td->io_issues[__io_u->ddir]--;
347
348 __io_u->flags &= ~IO_U_F_FLIGHT;
349
350 list_del(&__io_u->list);
351 list_add_tail(&__io_u->list, &td->io_u_requeues);
352 td->cur_depth--;
353 *io_u = NULL;
354}
355
356static int fill_io_u(struct thread_data *td, struct io_u *io_u)
357{
358 if (td->io_ops->flags & FIO_NOIO)
359 goto out;
360
361 /*
362 * see if it's time to sync
363 */
364 if (td->o.fsync_blocks &&
365 !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
366 td->io_issues[DDIR_WRITE] && should_fsync(td)) {
367 io_u->ddir = DDIR_SYNC;
368 goto out;
369 }
370
371 io_u->ddir = get_rw_ddir(td);
372
373 /*
374 * See if it's time to switch to a new zone
375 */
376 if (td->zone_bytes >= td->o.zone_size) {
377 td->zone_bytes = 0;
378 io_u->file->last_pos += td->o.zone_skip;
379 td->io_skip_bytes += td->o.zone_skip;
380 }
381
382 /*
383 * No log, let the seq/rand engine retrieve the next buflen and
384 * position.
385 */
386 if (get_next_offset(td, io_u)) {
387 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
388 return 1;
389 }
390
391 io_u->buflen = get_next_buflen(td, io_u);
392 if (!io_u->buflen) {
393 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
394 return 1;
395 }
396
397 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
398 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
399 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
400 io_u->buflen, io_u->file->real_file_size);
401 return 1;
402 }
403
404 /*
405 * mark entry before potentially trimming io_u
406 */
407 if (td_random(td) && !td->o.norandommap)
408 mark_random_map(td, io_u);
409
410 /*
411 * If using a write iolog, store this entry.
412 */
413out:
414 dprint_io_u(io_u, "fill_io_u");
415 td->zone_bytes += io_u->buflen;
416 log_io_u(td, io_u);
417 return 0;
418}
419
420void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
421{
422 int index = 0;
423
424 if (io_u->ddir == DDIR_SYNC)
425 return;
426
427 switch (td->cur_depth) {
428 default:
429 index = 6;
430 break;
431 case 32 ... 63:
432 index = 5;
433 break;
434 case 16 ... 31:
435 index = 4;
436 break;
437 case 8 ... 15:
438 index = 3;
439 break;
440 case 4 ... 7:
441 index = 2;
442 break;
443 case 2 ... 3:
444 index = 1;
445 case 1:
446 break;
447 }
448
449 td->ts.io_u_map[index]++;
450 td->ts.total_io_u[io_u->ddir]++;
451}
452
453static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
454{
455 int index = 0;
456
457 assert(usec < 1000);
458
459 switch (usec) {
460 case 750 ... 999:
461 index = 9;
462 break;
463 case 500 ... 749:
464 index = 8;
465 break;
466 case 250 ... 499:
467 index = 7;
468 break;
469 case 100 ... 249:
470 index = 6;
471 break;
472 case 50 ... 99:
473 index = 5;
474 break;
475 case 20 ... 49:
476 index = 4;
477 break;
478 case 10 ... 19:
479 index = 3;
480 break;
481 case 4 ... 9:
482 index = 2;
483 break;
484 case 2 ... 3:
485 index = 1;
486 case 0 ... 1:
487 break;
488 }
489
490 assert(index < FIO_IO_U_LAT_U_NR);
491 td->ts.io_u_lat_u[index]++;
492}
493
494static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
495{
496 int index = 0;
497
498 switch (msec) {
499 default:
500 index = 11;
501 break;
502 case 1000 ... 1999:
503 index = 10;
504 break;
505 case 750 ... 999:
506 index = 9;
507 break;
508 case 500 ... 749:
509 index = 8;
510 break;
511 case 250 ... 499:
512 index = 7;
513 break;
514 case 100 ... 249:
515 index = 6;
516 break;
517 case 50 ... 99:
518 index = 5;
519 break;
520 case 20 ... 49:
521 index = 4;
522 break;
523 case 10 ... 19:
524 index = 3;
525 break;
526 case 4 ... 9:
527 index = 2;
528 break;
529 case 2 ... 3:
530 index = 1;
531 case 0 ... 1:
532 break;
533 }
534
535 assert(index < FIO_IO_U_LAT_M_NR);
536 td->ts.io_u_lat_m[index]++;
537}
538
539static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
540{
541 if (usec < 1000)
542 io_u_mark_lat_usec(td, usec);
543 else
544 io_u_mark_lat_msec(td, usec / 1000);
545}
546
547/*
548 * Get next file to service by choosing one at random
549 */
550static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
551 int badf)
552{
553 struct fio_file *f;
554 int fno;
555
556 do {
557 long r = os_random_long(&td->next_file_state);
558
559 fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
560 f = &td->files[fno];
561 if (f->flags & FIO_FILE_DONE)
562 continue;
563
564 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
565 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
566 return f;
567 }
568 } while (1);
569}
570
571/*
572 * Get next file to service by doing round robin between all available ones
573 */
574static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
575 int badf)
576{
577 unsigned int old_next_file = td->next_file;
578 struct fio_file *f;
579
580 do {
581 f = &td->files[td->next_file];
582
583 td->next_file++;
584 if (td->next_file >= td->o.nr_files)
585 td->next_file = 0;
586
587 if (f->flags & FIO_FILE_DONE) {
588 f = NULL;
589 continue;
590 }
591
592 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
593 break;
594
595 f = NULL;
596 } while (td->next_file != old_next_file);
597
598 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
599 return f;
600}
601
602static struct fio_file *get_next_file(struct thread_data *td)
603{
604 struct fio_file *f;
605
606 assert(td->o.nr_files <= td->files_index);
607
608 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
609 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);
610 return NULL;
611 }
612
613 f = td->file_service_file;
614 if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
615 goto out;
616
617 if (td->o.file_service_type == FIO_FSERVICE_RR)
618 f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
619 else
620 f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
621
622 td->file_service_file = f;
623 td->file_service_left = td->file_service_nr - 1;
624out:
625 dprint(FD_FILE, "get_next_file: %p\n", f);
626 return f;
627}
628
629static struct fio_file *find_next_new_file(struct thread_data *td)
630{
631 struct fio_file *f;
632
633 if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
634 return NULL;
635
636 if (td->o.file_service_type == FIO_FSERVICE_RR)
637 f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
638 else
639 f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
640
641 return f;
642}
643
644static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
645{
646 struct fio_file *f;
647
648 do {
649 f = get_next_file(td);
650 if (!f)
651 return 1;
652
653set_file:
654 io_u->file = f;
655 get_file(f);
656
657 if (!fill_io_u(td, io_u))
658 break;
659
660 /*
661 * td_io_close() does a put_file() as well, so no need to
662 * do that here.
663 */
664 io_u->file = NULL;
665 td_io_close_file(td, f);
666 f->flags |= FIO_FILE_DONE;
667 td->nr_done_files++;
668
669 /*
670 * probably not the right place to do this, but see
671 * if we need to open a new file
672 */
673 if (td->nr_open_files < td->o.open_files &&
674 td->o.open_files != td->o.nr_files) {
675 f = find_next_new_file(td);
676
677 if (!f || td_io_open_file(td, f))
678 return 1;
679
680 goto set_file;
681 }
682 } while (1);
683
684 return 0;
685}
686
687
688struct io_u *__get_io_u(struct thread_data *td)
689{
690 struct io_u *io_u = NULL;
691
692 if (!list_empty(&td->io_u_requeues))
693 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
694 else if (!queue_full(td)) {
695 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
696
697 io_u->buflen = 0;
698 io_u->resid = 0;
699 io_u->file = NULL;
700 io_u->end_io = NULL;
701 }
702
703 if (io_u) {
704 assert(io_u->flags & IO_U_F_FREE);
705 io_u->flags &= ~IO_U_F_FREE;
706
707 io_u->error = 0;
708 list_del(&io_u->list);
709 list_add(&io_u->list, &td->io_u_busylist);
710 td->cur_depth++;
711 }
712
713 return io_u;
714}
715
716/*
717 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
718 * etc. The returned io_u is fully ready to be prepped and submitted.
719 */
720struct io_u *get_io_u(struct thread_data *td)
721{
722 struct fio_file *f;
723 struct io_u *io_u;
724
725 io_u = __get_io_u(td);
726 if (!io_u) {
727 dprint(FD_IO, "__get_io_u failed\n");
728 return NULL;
729 }
730
731 /*
732 * from a requeue, io_u already setup
733 */
734 if (io_u->file)
735 goto out;
736
737 /*
738 * If using an iolog, grab next piece if any available.
739 */
740 if (td->o.read_iolog_file) {
741 if (read_iolog_get(td, io_u))
742 goto err_put;
743 } else if (set_io_u_file(td, io_u)) {
744 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
745 goto err_put;
746 }
747
748 f = io_u->file;
749 assert(f->flags & FIO_FILE_OPEN);
750
751 if (io_u->ddir != DDIR_SYNC) {
752 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
753 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
754 goto err_put;
755 }
756
757 f->last_pos = io_u->offset + io_u->buflen;
758
759 if (td->o.verify != VERIFY_NONE)
760 populate_verify_io_u(td, io_u);
761 }
762
763 /*
764 * Set io data pointers.
765 */
766 io_u->endpos = io_u->offset + io_u->buflen;
767 io_u->xfer_buf = io_u->buf;
768 io_u->xfer_buflen = io_u->buflen;
769out:
770 if (!td_io_prep(td, io_u)) {
771 fio_gettime(&io_u->start_time, NULL);
772 return io_u;
773 }
774err_put:
775 dprint(FD_IO, "get_io_u failed\n");
776 put_io_u(td, io_u);
777 return NULL;
778}
779
780void io_u_log_error(struct thread_data *td, struct io_u *io_u)
781{
782 const char *msg[] = { "read", "write", "sync" };
783
784 log_err("fio: io_u error");
785
786 if (io_u->file)
787 log_err(" on file %s", io_u->file->file_name);
788
789 log_err(": %s\n", strerror(io_u->error));
790
791 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
792
793 if (!td->error)
794 td_verror(td, io_u->error, "io_u error");
795}
796
797static void io_completed(struct thread_data *td, struct io_u *io_u,
798 struct io_completion_data *icd)
799{
800 unsigned long usec;
801
802 dprint_io_u(io_u, "io complete");
803
804 assert(io_u->flags & IO_U_F_FLIGHT);
805 io_u->flags &= ~IO_U_F_FLIGHT;
806
807 if (io_u->ddir == DDIR_SYNC) {
808 td->last_was_sync = 1;
809 return;
810 }
811
812 td->last_was_sync = 0;
813
814 if (!io_u->error) {
815 unsigned int bytes = io_u->buflen - io_u->resid;
816 const enum fio_ddir idx = io_u->ddir;
817 int ret;
818
819 td->io_blocks[idx]++;
820 td->io_bytes[idx] += bytes;
821 td->this_io_bytes[idx] += bytes;
822
823 io_u->file->last_completed_pos = io_u->endpos;
824
825 usec = utime_since(&io_u->issue_time, &icd->time);
826
827 add_clat_sample(td, idx, usec);
828 add_bw_sample(td, idx, &icd->time);
829 io_u_mark_latency(td, usec);
830
831 if (td_write(td) && idx == DDIR_WRITE &&
832 td->o.do_verify &&
833 td->o.verify != VERIFY_NONE)
834 log_io_piece(td, io_u);
835
836 icd->bytes_done[idx] += bytes;
837
838 if (io_u->end_io) {
839 ret = io_u->end_io(td, io_u);
840 if (ret && !icd->error)
841 icd->error = ret;
842 }
843 } else {
844 icd->error = io_u->error;
845 io_u_log_error(td, io_u);
846 }
847}
848
849static void init_icd(struct io_completion_data *icd, int nr)
850{
851 fio_gettime(&icd->time, NULL);
852
853 icd->nr = nr;
854
855 icd->error = 0;
856 icd->bytes_done[0] = icd->bytes_done[1] = 0;
857}
858
859static void ios_completed(struct thread_data *td,
860 struct io_completion_data *icd)
861{
862 struct io_u *io_u;
863 int i;
864
865 for (i = 0; i < icd->nr; i++) {
866 io_u = td->io_ops->event(td, i);
867
868 io_completed(td, io_u, icd);
869 put_io_u(td, io_u);
870 }
871}
872
873/*
874 * Complete a single io_u for the sync engines.
875 */
876long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
877{
878 struct io_completion_data icd;
879
880 init_icd(&icd, 1);
881 io_completed(td, io_u, &icd);
882 put_io_u(td, io_u);
883
884 if (!icd.error)
885 return icd.bytes_done[0] + icd.bytes_done[1];
886
887 td_verror(td, icd.error, "io_u_sync_complete");
888 return -1;
889}
890
891/*
892 * Called to complete min_events number of io for the async engines.
893 */
894long io_u_queued_complete(struct thread_data *td, int min_events)
895{
896 struct io_completion_data icd;
897 struct timespec *tvp = NULL;
898 int ret;
899 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
900
901 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
902
903 if (!min_events)
904 tvp = &ts;
905
906 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
907 if (ret < 0) {
908 td_verror(td, -ret, "td_io_getevents");
909 return ret;
910 } else if (!ret)
911 return ret;
912
913 init_icd(&icd, ret);
914 ios_completed(td, &icd);
915 if (!icd.error)
916 return icd.bytes_done[0] + icd.bytes_done[1];
917
918 td_verror(td, icd.error, "io_u_queued_complete");
919 return -1;
920}
921
922/*
923 * Call when io_u is really queued, to update the submission latency.
924 */
925void io_u_queued(struct thread_data *td, struct io_u *io_u)
926{
927 unsigned long slat_time;
928
929 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
930 add_slat_sample(td, io_u->ddir, slat_time);
931}
932
933#ifdef FIO_USE_TIMEOUT
934void io_u_set_timeout(struct thread_data *td)
935{
936 assert(td->cur_depth);
937
938 td->timer.it_interval.tv_sec = 0;
939 td->timer.it_interval.tv_usec = 0;
940 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
941 td->timer.it_value.tv_usec = 0;
942 setitimer(ITIMER_REAL, &td->timer, NULL);
943 fio_gettime(&td->timeout_end, NULL);
944}
945
946static void io_u_dump(struct io_u *io_u)
947{
948 unsigned long t_start = mtime_since_now(&io_u->start_time);
949 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
950
951 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
952 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);
953 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
954}
955#else
956void io_u_set_timeout(struct thread_data fio_unused *td)
957{
958}
959#endif
960
961#ifdef FIO_USE_TIMEOUT
962static void io_u_timeout_handler(int fio_unused sig)
963{
964 struct thread_data *td, *__td;
965 pid_t pid = getpid();
966 struct list_head *entry;
967 struct io_u *io_u;
968 int i;
969
970 log_err("fio: io_u timeout\n");
971
972 /*
973 * TLS would be nice...
974 */
975 td = NULL;
976 for_each_td(__td, i) {
977 if (__td->pid == pid) {
978 td = __td;
979 break;
980 }
981 }
982
983 if (!td) {
984 log_err("fio: io_u timeout, can't find job\n");
985 exit(1);
986 }
987
988 if (!td->cur_depth) {
989 log_err("fio: timeout without pending work?\n");
990 return;
991 }
992
993 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
994
995 list_for_each(entry, &td->io_u_busylist) {
996 io_u = list_entry(entry, struct io_u, list);
997
998 io_u_dump(io_u);
999 }
1000
1001 td_verror(td, ETIMEDOUT, "io_u timeout");
1002 exit(1);
1003}
1004#endif
1005
1006void io_u_init_timeout(void)
1007{
1008#ifdef FIO_USE_TIMEOUT
1009 signal(SIGALRM, io_u_timeout_handler);
1010#endif
1011}
fio - Flexible IO Tester