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