posixaio engine: residual data count
[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"
9#include "os.h"
10
5945b9b4
JA
11/*
12 * Change this define to play with the timeout handling
13 */
14#undef FIO_USE_TIMEOUT
15
97601024
JA
16struct io_completion_data {
17 int nr; /* input */
97601024
JA
18
19 int error; /* output */
20 unsigned long bytes_done[2]; /* output */
21 struct timeval time; /* output */
22};
23
10ba535a
JA
24/*
25 * The ->file_map[] contains a map of blocks we have or have not done io
26 * to yet. Used to make sure we cover the entire range in a fair fashion.
27 */
28static int random_map_free(struct thread_data *td, struct fio_file *f,
29 unsigned long long block)
30{
31 unsigned int idx = RAND_MAP_IDX(td, f, block);
32 unsigned int bit = RAND_MAP_BIT(td, f, block);
33
34 return (f->file_map[idx] & (1UL << bit)) == 0;
35}
36
df415585
JA
37/*
38 * Mark a given offset as used in the map.
39 */
9bf2061e 40static void mark_random_map(struct thread_data *td, struct io_u *io_u)
df415585 41{
afdbe580 42 unsigned int min_bs = td->rw_min_bs;
9bf2061e 43 struct fio_file *f = io_u->file;
a00735e6
JA
44 unsigned long long block;
45 unsigned int blocks;
c685b5b2 46 unsigned int nr_blocks;
df415585 47
a00735e6
JA
48 block = io_u->offset / (unsigned long long) min_bs;
49 blocks = 0;
c685b5b2
JA
50 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
51
52 while (blocks < nr_blocks) {
df415585
JA
53 unsigned int idx, bit;
54
55 if (!random_map_free(td, f, block))
56 break;
57
58 idx = RAND_MAP_IDX(td, f, block);
59 bit = RAND_MAP_BIT(td, f, block);
60
0032bf9f 61 fio_assert(td, idx < f->num_maps);
df415585
JA
62
63 f->file_map[idx] |= (1UL << bit);
64 block++;
65 blocks++;
66 }
67
a00735e6
JA
68 if ((blocks * min_bs) < io_u->buflen)
69 io_u->buflen = blocks * min_bs;
df415585
JA
70}
71
10ba535a
JA
72/*
73 * Return the next free block in the map.
74 */
75static int get_next_free_block(struct thread_data *td, struct fio_file *f,
76 unsigned long long *b)
77{
78 int i;
79
c685b5b2
JA
80 i = f->last_free_lookup;
81 *b = (i * BLOCKS_PER_MAP);
ee88470c 82 while ((*b) * td->rw_min_bs < f->real_file_size) {
10ba535a
JA
83 if (f->file_map[i] != -1UL) {
84 *b += ffz(f->file_map[i]);
c685b5b2 85 f->last_free_lookup = i;
10ba535a
JA
86 return 0;
87 }
88
89 *b += BLOCKS_PER_MAP;
90 i++;
91 }
92
93 return 1;
94}
95
96/*
97 * For random io, generate a random new block and see if it's used. Repeat
98 * until we find a free one. For sequential io, just return the end of
99 * the last io issued.
100 */
9bf2061e 101static int get_next_offset(struct thread_data *td, struct io_u *io_u)
10ba535a 102{
9bf2061e 103 struct fio_file *f = io_u->file;
c685b5b2 104 const int ddir = io_u->ddir;
10ba535a
JA
105 unsigned long long b, rb;
106 long r;
107
413dd459 108 if (td_random(td)) {
f697125a 109 unsigned long long max_blocks = f->file_size / td->min_bs[ddir];
c685b5b2 110 int loops = 5;
10ba535a
JA
111
112 do {
113 r = os_random_long(&td->random_state);
114 b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
bb8895e0
JA
115 if (td->norandommap)
116 break;
a00735e6 117 rb = b + (f->file_offset / td->min_bs[ddir]);
10ba535a
JA
118 loops--;
119 } while (!random_map_free(td, f, rb) && loops);
120
bca4ed4d
JA
121 /*
122 * if we failed to retrieve a truly random offset within
123 * the loops assigned, see if there are free ones left at all
124 */
125 if (!loops && get_next_free_block(td, f, &b))
126 return 1;
10ba535a 127 } else
a00735e6 128 b = f->last_pos / td->min_bs[ddir];
10ba535a 129
c685b5b2 130 io_u->offset = (b * td->min_bs[ddir]) + f->file_offset;
bca4ed4d 131 if (io_u->offset >= f->real_file_size)
10ba535a
JA
132 return 1;
133
134 return 0;
135}
136
9bf2061e 137static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
10ba535a 138{
9bf2061e 139 struct fio_file *f = io_u->file;
bca4ed4d 140 const int ddir = io_u->ddir;
10ba535a
JA
141 unsigned int buflen;
142 long r;
143
a00735e6
JA
144 if (td->min_bs[ddir] == td->max_bs[ddir])
145 buflen = td->min_bs[ddir];
10ba535a
JA
146 else {
147 r = os_random_long(&td->bsrange_state);
1e97cce9 148 buflen = (unsigned int) (1 + (double) (td->max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
690adba3 149 if (!td->bs_unaligned)
a00735e6 150 buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1);
10ba535a
JA
151 }
152
bca4ed4d 153 while (buflen + io_u->offset > f->real_file_size) {
bca4ed4d
JA
154 if (buflen == td->min_bs[ddir])
155 return 0;
156
157 buflen = td->min_bs[ddir];
10ba535a
JA
158 }
159
160 return buflen;
161}
162
163/*
164 * Return the data direction for the next io_u. If the job is a
165 * mixed read/write workload, check the rwmix cycle and switch if
166 * necessary.
167 */
1e97cce9 168static enum fio_ddir get_rw_ddir(struct thread_data *td)
10ba535a
JA
169{
170 if (td_rw(td)) {
171 struct timeval now;
172 unsigned long elapsed;
173
02bcaa8c 174 fio_gettime(&now, NULL);
10ba535a
JA
175 elapsed = mtime_since_now(&td->rwmix_switch);
176
177 /*
178 * Check if it's time to seed a new data direction.
179 */
180 if (elapsed >= td->rwmixcycle) {
181 unsigned int v;
182 long r;
183
184 r = os_random_long(&td->rwmix_state);
185 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
186 if (v < td->rwmixread)
187 td->rwmix_ddir = DDIR_READ;
188 else
189 td->rwmix_ddir = DDIR_WRITE;
190 memcpy(&td->rwmix_switch, &now, sizeof(now));
191 }
192 return td->rwmix_ddir;
193 } else if (td_read(td))
194 return DDIR_READ;
195 else
196 return DDIR_WRITE;
197}
198
10ba535a
JA
199void put_io_u(struct thread_data *td, struct io_u *io_u)
200{
0c6e7517
JA
201 assert((io_u->flags & IO_U_F_FREE) == 0);
202 io_u->flags |= IO_U_F_FREE;
203
10ba535a
JA
204 io_u->file = NULL;
205 list_del(&io_u->list);
206 list_add(&io_u->list, &td->io_u_freelist);
207 td->cur_depth--;
208}
209
755200a3
JA
210void requeue_io_u(struct thread_data *td, struct io_u **io_u)
211{
212 struct io_u *__io_u = *io_u;
213
4d2e0f49
JA
214 __io_u->flags |= IO_U_F_FREE;
215 __io_u->flags &= ~IO_U_F_FLIGHT;
216
755200a3
JA
217 list_del(&__io_u->list);
218 list_add_tail(&__io_u->list, &td->io_u_requeues);
219 td->cur_depth--;
220 *io_u = NULL;
221}
222
9bf2061e 223static int fill_io_u(struct thread_data *td, struct io_u *io_u)
10ba535a
JA
224{
225 /*
226 * If using an iolog, grab next piece if any available.
227 */
228 if (td->read_iolog)
229 return read_iolog_get(td, io_u);
230
87dc1ab1
JA
231 /*
232 * see if it's time to sync
233 */
755200a3
JA
234 if (td->fsync_blocks && !(td->io_issues[DDIR_WRITE] % td->fsync_blocks)
235 && td->io_issues[DDIR_WRITE] && should_fsync(td)) {
87dc1ab1 236 io_u->ddir = DDIR_SYNC;
87dc1ab1
JA
237 return 0;
238 }
239
a00735e6
JA
240 io_u->ddir = get_rw_ddir(td);
241
10ba535a 242 /*
c685b5b2
JA
243 * No log, let the seq/rand engine retrieve the next buflen and
244 * position.
10ba535a 245 */
9bf2061e 246 if (get_next_offset(td, io_u))
bca4ed4d 247 return 1;
10ba535a 248
9bf2061e 249 io_u->buflen = get_next_buflen(td, io_u);
bca4ed4d
JA
250 if (!io_u->buflen)
251 return 1;
252
253 /*
254 * mark entry before potentially trimming io_u
255 */
413dd459 256 if (!td->read_iolog && td_random(td) && !td->norandommap)
9bf2061e 257 mark_random_map(td, io_u);
bca4ed4d
JA
258
259 /*
260 * If using a write iolog, store this entry.
261 */
262 if (td->write_iolog_file)
263 write_iolog_put(td, io_u);
264
bca4ed4d 265 return 0;
10ba535a
JA
266}
267
71619dc2
JA
268static void io_u_mark_depth(struct thread_data *td)
269{
270 int index = 0;
271
272 switch (td->cur_depth) {
273 default:
274 index++;
275 case 32 ... 63:
276 index++;
277 case 16 ... 31:
278 index++;
279 case 8 ... 15:
280 index++;
281 case 4 ... 7:
282 index++;
283 case 2 ... 3:
284 index++;
285 case 1:
286 break;
287 }
288
289 td->io_u_map[index]++;
290 td->total_io_u++;
291}
292
ec118304
JA
293static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
294{
295 int index = 0;
296
297 switch (msec) {
298 default:
299 index++;
8abdce66 300 case 1000 ... 1999:
ec118304 301 index++;
8abdce66 302 case 750 ... 999:
ec118304 303 index++;
8abdce66 304 case 500 ... 749:
ec118304 305 index++;
8abdce66 306 case 250 ... 499:
ec118304 307 index++;
8abdce66 308 case 100 ... 249:
ec118304 309 index++;
8abdce66 310 case 50 ... 99:
ec118304 311 index++;
8abdce66 312 case 20 ... 49:
ec118304 313 index++;
8abdce66 314 case 10 ... 19:
ec118304 315 index++;
8abdce66 316 case 4 ... 9:
ec118304
JA
317 index++;
318 case 2 ... 3:
319 index++;
320 case 0 ... 1:
321 break;
322 }
323
324 td->io_u_lat[index]++;
325}
326
0aabe160
JA
327/*
328 * Get next file to service by choosing one at random
329 */
330static struct fio_file *get_next_file_rand(struct thread_data *td)
331{
0aabe160
JA
332 unsigned int fileno;
333 struct fio_file *f;
334
335 do {
7c83c089
JA
336 long r = os_random_long(&td->next_file_state);
337
0aabe160
JA
338 fileno = (unsigned int) ((double) (td->nr_files - 1) * r / (RAND_MAX + 1.0));
339 f = &td->files[fileno];
340 if (f->fd != -1)
341 return f;
342 } while (1);
343}
344
345/*
346 * Get next file to service by doing round robin between all available ones
347 */
348static struct fio_file *get_next_file_rr(struct thread_data *td)
3d7c391d
JA
349{
350 unsigned int old_next_file = td->next_file;
351 struct fio_file *f;
352
353 do {
354 f = &td->files[td->next_file];
355
356 td->next_file++;
357 if (td->next_file >= td->nr_files)
358 td->next_file = 0;
359
360 if (f->fd != -1)
361 break;
362
363 f = NULL;
364 } while (td->next_file != old_next_file);
365
366 return f;
367}
368
bdb4e2e9
JA
369static struct fio_file *get_next_file(struct thread_data *td)
370{
371 if (!td->nr_open_files)
372 return NULL;
373
374 if (td->file_service_type == FIO_FSERVICE_RR)
375 return get_next_file_rr(td);
376 else
377 return get_next_file_rand(td);
378}
379
10ba535a
JA
380struct io_u *__get_io_u(struct thread_data *td)
381{
382 struct io_u *io_u = NULL;
383
755200a3
JA
384 if (!list_empty(&td->io_u_requeues))
385 io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
386 else if (!queue_full(td)) {
10ba535a
JA
387 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
388
6040dabc 389 io_u->buflen = 0;
10ba535a 390 io_u->resid = 0;
755200a3 391 io_u->file = NULL;
d7762cf8 392 io_u->end_io = NULL;
755200a3
JA
393 }
394
395 if (io_u) {
0c6e7517
JA
396 assert(io_u->flags & IO_U_F_FREE);
397 io_u->flags &= ~IO_U_F_FREE;
398
755200a3 399 io_u->error = 0;
10ba535a
JA
400 list_del(&io_u->list);
401 list_add(&io_u->list, &td->io_u_busylist);
402 td->cur_depth++;
71619dc2 403 io_u_mark_depth(td);
10ba535a
JA
404 }
405
406 return io_u;
407}
408
409/*
410 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
411 * etc. The returned io_u is fully ready to be prepped and submitted.
412 */
3d7c391d 413struct io_u *get_io_u(struct thread_data *td)
10ba535a 414{
3d7c391d 415 struct fio_file *f;
10ba535a
JA
416 struct io_u *io_u;
417
418 io_u = __get_io_u(td);
419 if (!io_u)
420 return NULL;
421
755200a3
JA
422 /*
423 * from a requeue, io_u already setup
424 */
425 if (io_u->file)
77f392bf 426 goto out;
755200a3 427
bdb4e2e9
JA
428 do {
429 f = get_next_file(td);
430 if (!f) {
431 put_io_u(td, io_u);
432 return NULL;
433 }
3d7c391d 434
bdb4e2e9 435 io_u->file = f;
3d7c391d 436
bdb4e2e9
JA
437 if (!fill_io_u(td, io_u))
438 break;
10ba535a 439
bdb4e2e9
JA
440 /*
441 * No more to do for this file, close it
442 */
443 io_u->file = NULL;
444 close_file(td, f);
445 } while (1);
10ba535a 446
9bf2061e
JA
447 if (td->zone_bytes >= td->zone_size) {
448 td->zone_bytes = 0;
449 f->last_pos += td->zone_skip;
450 }
451
ee88470c 452 if (io_u->buflen + io_u->offset > f->real_file_size) {
10ba535a
JA
453 if (td->io_ops->flags & FIO_RAWIO) {
454 put_io_u(td, io_u);
455 return NULL;
456 }
457
ee88470c 458 io_u->buflen = f->real_file_size - io_u->offset;
10ba535a
JA
459 }
460
87dc1ab1
JA
461 if (io_u->ddir != DDIR_SYNC) {
462 if (!io_u->buflen) {
463 put_io_u(td, io_u);
464 return NULL;
465 }
10ba535a 466
36167d82 467 f->last_pos = io_u->offset + io_u->buflen;
10ba535a 468
87dc1ab1
JA
469 if (td->verify != VERIFY_NONE)
470 populate_verify_io_u(td, io_u);
471 }
10ba535a 472
165faf16
JA
473 /*
474 * Set io data pointers.
475 */
77f392bf 476out:
cec6b55d
JA
477 io_u->xfer_buf = io_u->buf;
478 io_u->xfer_buflen = io_u->buflen;
165faf16 479
36167d82
JA
480 if (td_io_prep(td, io_u)) {
481 put_io_u(td, io_u);
482 return NULL;
483 }
484
02bcaa8c 485 fio_gettime(&io_u->start_time, NULL);
10ba535a
JA
486 return io_u;
487}
488
97601024
JA
489static void io_completed(struct thread_data *td, struct io_u *io_u,
490 struct io_completion_data *icd)
10ba535a 491{
10ba535a
JA
492 unsigned long msec;
493
0c6e7517
JA
494 assert(io_u->flags & IO_U_F_FLIGHT);
495 io_u->flags &= ~IO_U_F_FLIGHT;
496
87dc1ab1
JA
497 if (io_u->ddir == DDIR_SYNC) {
498 td->last_was_sync = 1;
499 return;
500 }
501
502 td->last_was_sync = 0;
503
10ba535a
JA
504 if (!io_u->error) {
505 unsigned int bytes = io_u->buflen - io_u->resid;
1e97cce9 506 const enum fio_ddir idx = io_u->ddir;
3af6ef39 507 int ret;
10ba535a
JA
508
509 td->io_blocks[idx]++;
510 td->io_bytes[idx] += bytes;
511 td->zone_bytes += bytes;
512 td->this_io_bytes[idx] += bytes;
513
02bcaa8c
JA
514 io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
515
516 msec = mtime_since(&io_u->issue_time, &icd->time);
10ba535a
JA
517
518 add_clat_sample(td, idx, msec);
02bcaa8c 519 add_bw_sample(td, idx, &icd->time);
ec118304 520 io_u_mark_latency(td, msec);
10ba535a
JA
521
522 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
523 log_io_piece(td, io_u);
524
525 icd->bytes_done[idx] += bytes;
3af6ef39 526
d7762cf8
JA
527 if (io_u->end_io) {
528 ret = io_u->end_io(io_u);
3af6ef39
JA
529 if (ret && !icd->error)
530 icd->error = ret;
531 }
10ba535a
JA
532 } else
533 icd->error = io_u->error;
534}
535
d7762cf8 536static void init_icd(struct io_completion_data *icd, int nr)
10ba535a 537{
02bcaa8c
JA
538 fio_gettime(&icd->time, NULL);
539
3af6ef39
JA
540 icd->nr = nr;
541
10ba535a
JA
542 icd->error = 0;
543 icd->bytes_done[0] = icd->bytes_done[1] = 0;
36167d82
JA
544}
545
97601024
JA
546static void ios_completed(struct thread_data *td,
547 struct io_completion_data *icd)
36167d82
JA
548{
549 struct io_u *io_u;
550 int i;
551
10ba535a
JA
552 for (i = 0; i < icd->nr; i++) {
553 io_u = td->io_ops->event(td, i);
554
555 io_completed(td, io_u, icd);
556 put_io_u(td, io_u);
557 }
558}
97601024 559
e7e6cfb4
JA
560/*
561 * Complete a single io_u for the sync engines.
562 */
d7762cf8 563long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
97601024
JA
564{
565 struct io_completion_data icd;
566
d7762cf8 567 init_icd(&icd, 1);
97601024
JA
568 io_completed(td, io_u, &icd);
569 put_io_u(td, io_u);
570
571 if (!icd.error)
572 return icd.bytes_done[0] + icd.bytes_done[1];
573
37e974a5 574 td_verror(td, icd.error, "io_u_sync_complete");
97601024
JA
575 return -1;
576}
577
e7e6cfb4
JA
578/*
579 * Called to complete min_events number of io for the async engines.
580 */
d7762cf8 581long io_u_queued_complete(struct thread_data *td, int min_events)
97601024 582{
97601024 583 struct io_completion_data icd;
00de55ef 584 struct timespec *tvp = NULL;
97601024
JA
585 int ret;
586
face81b2 587 if (!min_events) {
00de55ef
JA
588 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
589
590 tvp = &ts;
755200a3 591 }
97601024 592
00de55ef 593 ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
97601024 594 if (ret < 0) {
e1161c32 595 td_verror(td, -ret, "td_io_getevents");
97601024
JA
596 return ret;
597 } else if (!ret)
598 return ret;
599
d7762cf8 600 init_icd(&icd, ret);
97601024
JA
601 ios_completed(td, &icd);
602 if (!icd.error)
603 return icd.bytes_done[0] + icd.bytes_done[1];
604
37e974a5 605 td_verror(td, icd.error, "io_u_queued_complete");
97601024
JA
606 return -1;
607}
7e77dd02
JA
608
609/*
610 * Call when io_u is really queued, to update the submission latency.
611 */
612void io_u_queued(struct thread_data *td, struct io_u *io_u)
613{
614 unsigned long slat_time;
615
616 slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
617 add_slat_sample(td, io_u->ddir, slat_time);
618}
433afcb4 619
55bc9728 620#ifdef FIO_USE_TIMEOUT
433afcb4
JA
621void io_u_set_timeout(struct thread_data *td)
622{
623 assert(td->cur_depth);
624
625 td->timer.it_interval.tv_sec = 0;
626 td->timer.it_interval.tv_usec = 0;
627 td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
628 td->timer.it_value.tv_usec = 0;
629 setitimer(ITIMER_REAL, &td->timer, NULL);
630 fio_gettime(&td->timeout_end, NULL);
631}
5945b9b4
JA
632
633static void io_u_dump(struct io_u *io_u)
634{
635 unsigned long t_start = mtime_since_now(&io_u->start_time);
636 unsigned long t_issue = mtime_since_now(&io_u->issue_time);
637
638 log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
639 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);
640 log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
641}
55bc9728
JA
642#else
643void io_u_set_timeout(struct thread_data fio_unused *td)
644{
645}
646#endif
433afcb4 647
55bc9728 648#ifdef FIO_USE_TIMEOUT
433afcb4
JA
649static void io_u_timeout_handler(int fio_unused sig)
650{
651 struct thread_data *td, *__td;
652 pid_t pid = getpid();
5945b9b4
JA
653 struct list_head *entry;
654 struct io_u *io_u;
433afcb4
JA
655 int i;
656
657 log_err("fio: io_u timeout\n");
658
659 /*
660 * TLS would be nice...
661 */
662 td = NULL;
663 for_each_td(__td, i) {
664 if (__td->pid == pid) {
665 td = __td;
666 break;
667 }
668 }
669
670 if (!td) {
671 log_err("fio: io_u timeout, can't find job\n");
672 exit(1);
673 }
674
675 if (!td->cur_depth) {
676 log_err("fio: timeout without pending work?\n");
677 return;
678 }
679
680 log_err("fio: io_u timeout: job=%s, pid=%d\n", td->name, td->pid);
5945b9b4
JA
681
682 list_for_each(entry, &td->io_u_busylist) {
683 io_u = list_entry(entry, struct io_u, list);
684
685 io_u_dump(io_u);
686 }
687
688 td_verror(td, ETIMEDOUT, "io_u timeout");
433afcb4
JA
689 exit(1);
690}
55bc9728 691#endif
433afcb4
JA
692
693void io_u_init_timeout(void)
694{
55bc9728 695#ifdef FIO_USE_TIMEOUT
433afcb4 696 signal(SIGALRM, io_u_timeout_handler);
55bc9728 697#endif
433afcb4 698}