Fix race on io_u->flags
[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"
5973cafb 9#include "hash.h"
4f5af7b2 10#include "verify.h"
10ba535a 11
97601024
JA
12struct io_completion_data {
13 int nr; /* input */
97601024
JA
14
15 int error; /* output */
16 unsigned long bytes_done[2]; /* output */
17 struct timeval time; /* output */
18};
19
10ba535a
JA
20/*
21 * The ->file_map[] contains a map of blocks we have or have not done io
22 * to yet. Used to make sure we cover the entire range in a fair fashion.
23 */
aec2de20 24static int random_map_free(struct fio_file *f, const unsigned long long block)
10ba535a 25{
aec2de20
JA
26 unsigned int idx = RAND_MAP_IDX(f, block);
27 unsigned int bit = RAND_MAP_BIT(f, block);
10ba535a 28
84422acd
JA
29 dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
30
de605666 31 return (f->file_map[idx] & (1 << bit)) == 0;
10ba535a
JA
32}
33
df415585
JA
34/*
35 * Mark a given offset as used in the map.
36 */
9bf2061e 37static void mark_random_map(struct thread_data *td, struct io_u *io_u)
df415585 38{
2dc1bbeb 39 unsigned int min_bs = td->o.rw_min_bs;
9bf2061e 40 struct fio_file *f = io_u->file;
a00735e6 41 unsigned long long block;
3e3357b1 42 unsigned int blocks, nr_blocks;
df415585 43
b9c5b644 44 block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
c685b5b2 45 nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
3e3357b1 46 blocks = 0;
c685b5b2 47
3e3357b1
JA
48 while (nr_blocks) {
49 unsigned int this_blocks, mask;
df415585
JA
50 unsigned int idx, bit;
51
1e3d53ac
JA
52 /*
53 * If we have a mixed random workload, we may
54 * encounter blocks we already did IO to.
55 */
8b113b4f
JA
56 if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) {
57 if (!blocks)
58 blocks = 1;
df415585 59 break;
8b113b4f 60 }
df415585 61
aec2de20
JA
62 idx = RAND_MAP_IDX(f, block);
63 bit = RAND_MAP_BIT(f, block);
df415585 64
0032bf9f 65 fio_assert(td, idx < f->num_maps);
df415585 66
3e3357b1
JA
67 this_blocks = nr_blocks;
68 if (this_blocks + bit > BLOCKS_PER_MAP)
69 this_blocks = BLOCKS_PER_MAP - bit;
70
71 if (this_blocks == BLOCKS_PER_MAP)
72 mask = -1U;
73 else
74 mask = ((1U << this_blocks) - 1) << bit;
75
3e3357b1
JA
76 f->file_map[idx] |= mask;
77 nr_blocks -= this_blocks;
78 blocks += this_blocks;
c9dd34b2 79 block += this_blocks;
df415585
JA
80 }
81
a00735e6
JA
82 if ((blocks * min_bs) < io_u->buflen)
83 io_u->buflen = blocks * min_bs;
df415585
JA
84}
85
3e3357b1
JA
86static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
87 enum fio_ddir ddir)
2ba1c290
JA
88{
89 unsigned long long max_blocks;
d9dd70f7 90 unsigned long long max_size;
2ba1c290 91
d9dd70f7
JA
92 /*
93 * Hmm, should we make sure that ->io_size <= ->real_file_size?
94 */
95 max_size = f->io_size;
96 if (max_size > f->real_file_size)
97 max_size = f->real_file_size;
98
2b7a01d0 99 max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
2ba1c290
JA
100 if (!max_blocks)
101 return 0;
102
67778e88 103 return max_blocks;
2ba1c290
JA
104}
105
10ba535a
JA
106/*
107 * Return the next free block in the map.
108 */
109static int get_next_free_block(struct thread_data *td, struct fio_file *f,
4ba66134 110 enum fio_ddir ddir, unsigned long long *b)
10ba535a 111{
84422acd 112 unsigned long long min_bs = td->o.rw_min_bs;
10ba535a
JA
113 int i;
114
c685b5b2
JA
115 i = f->last_free_lookup;
116 *b = (i * BLOCKS_PER_MAP);
899c29f2
RR
117 while ((*b) * min_bs < f->real_file_size &&
118 (*b) * min_bs < f->io_size) {
de605666 119 if (f->file_map[i] != (unsigned int) -1) {
697a606c 120 *b += ffz(f->file_map[i]);
4ba66134 121 if (*b > last_block(td, f, ddir))
2ba1c290 122 break;
c685b5b2 123 f->last_free_lookup = i;
10ba535a
JA
124 return 0;
125 }
126
127 *b += BLOCKS_PER_MAP;
128 i++;
129 }
130
2ba1c290 131 dprint(FD_IO, "failed finding a free block\n");
10ba535a
JA
132 return 1;
133}
134
ec4015da 135static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
4ba66134 136 enum fio_ddir ddir, unsigned long long *b)
ec4015da 137{
84422acd 138 unsigned long long r;
ec4015da
JA
139 int loops = 5;
140
141 do {
142 r = os_random_long(&td->random_state);
84422acd 143 dprint(FD_RANDOM, "off rand %llu\n", r);
5ec10eaa 144 *b = (last_block(td, f, ddir) - 1)
dc873b6f 145 * (r / ((unsigned long long) OS_RAND_MAX + 1.0));
2ba1c290 146
43c63a78
JA
147 /*
148 * if we are not maintaining a random map, we are done.
149 */
303032ae 150 if (!file_randommap(td, f))
43c63a78
JA
151 return 0;
152
153 /*
84422acd 154 * calculate map offset and check if it's free
43c63a78 155 */
aec2de20 156 if (random_map_free(f, *b))
43c63a78
JA
157 return 0;
158
84422acd
JA
159 dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
160 *b);
43c63a78 161 } while (--loops);
ec4015da
JA
162
163 /*
43c63a78
JA
164 * we get here, if we didn't suceed in looking up a block. generate
165 * a random start offset into the filemap, and find the first free
166 * block from there.
ec4015da 167 */
43c63a78
JA
168 loops = 10;
169 do {
dc873b6f
JA
170 f->last_free_lookup = (f->num_maps - 1) *
171 (r / (OS_RAND_MAX + 1.0));
4ba66134 172 if (!get_next_free_block(td, f, ddir, b))
43c63a78 173 return 0;
ec4015da 174
43c63a78
JA
175 r = os_random_long(&td->random_state);
176 } while (--loops);
177
178 /*
179 * that didn't work either, try exhaustive search from the start
180 */
181 f->last_free_lookup = 0;
4ba66134 182 return get_next_free_block(td, f, ddir, b);
ec4015da
JA
183}
184
10ba535a
JA
185/*
186 * For random io, generate a random new block and see if it's used. Repeat
187 * until we find a free one. For sequential io, just return the end of
188 * the last io issued.
189 */
9bf2061e 190static int get_next_offset(struct thread_data *td, struct io_u *io_u)
10ba535a 191{
9bf2061e 192 struct fio_file *f = io_u->file;
ec4015da 193 unsigned long long b;
4ba66134 194 enum fio_ddir ddir = io_u->ddir;
10ba535a 195
ec4015da
JA
196 if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
197 td->ddir_nr = td->o.ddir_nr;
211097b2 198
683023e8
JA
199 if (get_next_rand_offset(td, f, ddir, &b)) {
200 dprint(FD_IO, "%s: getting rand offset failed\n",
201 f->file_name);
bca4ed4d 202 return 1;
683023e8 203 }
43063a1c 204 } else {
0c3d768a 205 if (f->last_pos >= f->real_file_size) {
4ba66134 206 if (!td_random(td) ||
683023e8
JA
207 get_next_rand_offset(td, f, ddir, &b)) {
208 dprint(FD_IO, "%s: pos %llu > size %llu\n",
209 f->file_name, f->last_pos,
210 f->real_file_size);
0c3d768a 211 return 1;
683023e8 212 }
bcdedd0a 213 } else
4ba66134 214 b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
43063a1c 215 }
10ba535a 216
2b7a01d0 217 io_u->offset = b * td->o.ba[ddir];
009bd847
JA
218 if (io_u->offset >= f->io_size) {
219 dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
220 io_u->offset, f->io_size);
221 return 1;
222 }
223
224 io_u->offset += f->file_offset;
2ba1c290
JA
225 if (io_u->offset >= f->real_file_size) {
226 dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
227 io_u->offset, f->real_file_size);
10ba535a 228 return 1;
2ba1c290 229 }
10ba535a
JA
230
231 return 0;
232}
233
9bf2061e 234static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
10ba535a 235{
bca4ed4d 236 const int ddir = io_u->ddir;
f3f552b9 237 unsigned int uninitialized_var(buflen);
f3059de1 238 unsigned int minbs, maxbs;
10ba535a
JA
239 long r;
240
f3059de1
JA
241 minbs = td->o.min_bs[ddir];
242 maxbs = td->o.max_bs[ddir];
243
244 if (minbs == maxbs)
245 buflen = minbs;
10ba535a
JA
246 else {
247 r = os_random_long(&td->bsrange_state);
720e84ad 248 if (!td->o.bssplit_nr[ddir]) {
f3059de1
JA
249 buflen = 1 + (unsigned int) ((double) maxbs *
250 (r / (OS_RAND_MAX + 1.0)));
251 if (buflen < minbs)
252 buflen = minbs;
5ec10eaa 253 } else {
564ca972
JA
254 long perc = 0;
255 unsigned int i;
256
720e84ad
JA
257 for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
258 struct bssplit *bsp = &td->o.bssplit[ddir][i];
564ca972
JA
259
260 buflen = bsp->bs;
261 perc += bsp->perc;
f3059de1 262 if (r <= ((OS_RAND_MAX / 100L) * perc))
564ca972
JA
263 break;
264 }
265 }
f3059de1
JA
266 if (!td->o.bs_unaligned && is_power_of_2(minbs))
267 buflen = (buflen + minbs - 1) & ~(minbs - 1);
10ba535a
JA
268 }
269
4ba66134
JA
270 if (io_u->offset + buflen > io_u->file->real_file_size) {
271 dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
f3059de1
JA
272 minbs, ddir);
273 buflen = minbs;
4ba66134 274 }
6a5e6884 275
10ba535a
JA
276 return buflen;
277}
278
afe24a5a
JA
279static void set_rwmix_bytes(struct thread_data *td)
280{
afe24a5a
JA
281 unsigned int diff;
282
283 /*
284 * we do time or byte based switch. this is needed because
285 * buffered writes may issue a lot quicker than they complete,
286 * whereas reads do not.
287 */
e47f799f 288 diff = td->o.rwmix[td->rwmix_ddir ^ 1];
04c540d9 289 td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
e47f799f
JA
290}
291
292static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
293{
294 unsigned int v;
295 long r;
296
297 r = os_random_long(&td->rwmix_state);
dc873b6f 298 v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
04c540d9 299 if (v <= td->o.rwmix[DDIR_READ])
e47f799f
JA
300 return DDIR_READ;
301
302 return DDIR_WRITE;
afe24a5a
JA
303}
304
581e7141
JA
305static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
306{
307 enum fio_ddir odir = ddir ^ 1;
308 struct timeval t;
309 long usec;
310
311 if (td->rate_pending_usleep[ddir] <= 0)
312 return ddir;
313
314 /*
315 * We have too much pending sleep in this direction. See if we
316 * should switch.
317 */
318 if (td_rw(td)) {
319 /*
320 * Other direction does not have too much pending, switch
321 */
322 if (td->rate_pending_usleep[odir] < 100000)
323 return odir;
324
325 /*
326 * Both directions have pending sleep. Sleep the minimum time
327 * and deduct from both.
328 */
329 if (td->rate_pending_usleep[ddir] <=
330 td->rate_pending_usleep[odir]) {
331 usec = td->rate_pending_usleep[ddir];
332 } else {
333 usec = td->rate_pending_usleep[odir];
334 ddir = odir;
335 }
336 } else
337 usec = td->rate_pending_usleep[ddir];
338
339 fio_gettime(&t, NULL);
340 usec_sleep(td, usec);
341 usec = utime_since_now(&t);
342
343 td->rate_pending_usleep[ddir] -= usec;
344
345 odir = ddir ^ 1;
346 if (td_rw(td) && __should_check_rate(td, odir))
347 td->rate_pending_usleep[odir] -= usec;
0b9d69ec 348
581e7141
JA
349 return ddir;
350}
351
10ba535a
JA
352/*
353 * Return the data direction for the next io_u. If the job is a
354 * mixed read/write workload, check the rwmix cycle and switch if
355 * necessary.
356 */
1e97cce9 357static enum fio_ddir get_rw_ddir(struct thread_data *td)
10ba535a 358{
581e7141
JA
359 enum fio_ddir ddir;
360
5f9099ea
JA
361 /*
362 * see if it's time to fsync
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 return DDIR_SYNC;
368
369 /*
370 * see if it's time to fdatasync
371 */
372 if (td->o.fdatasync_blocks &&
373 !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
374 td->io_issues[DDIR_WRITE] && should_fsync(td))
375 return DDIR_DATASYNC;
376
10ba535a 377 if (td_rw(td)) {
10ba535a
JA
378 /*
379 * Check if it's time to seed a new data direction.
380 */
e4928662 381 if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
e47f799f
JA
382 /*
383 * Put a top limit on how many bytes we do for
384 * one data direction, to avoid overflowing the
385 * ranges too much
386 */
387 ddir = get_rand_ddir(td);
e47f799f
JA
388
389 if (ddir != td->rwmix_ddir)
390 set_rwmix_bytes(td);
391
392 td->rwmix_ddir = ddir;
10ba535a 393 }
581e7141 394 ddir = td->rwmix_ddir;
10ba535a 395 } else if (td_read(td))
581e7141 396 ddir = DDIR_READ;
10ba535a 397 else
581e7141
JA
398 ddir = DDIR_WRITE;
399
400 td->rwmix_ddir = rate_ddir(td, ddir);
401 return td->rwmix_ddir;
10ba535a
JA
402}
403
e8462bd8 404void put_file_log(struct thread_data *td, struct fio_file *f)
60f2c658
JA
405{
406 int ret = put_file(td, f);
407
408 if (ret)
409 td_verror(td, ret, "file close");
410}
411
10ba535a
JA
412void put_io_u(struct thread_data *td, struct io_u *io_u)
413{
e8462bd8
JA
414 td_io_u_lock(td);
415
0c6e7517 416 io_u->flags |= IO_U_F_FREE;
e8462bd8 417 io_u->flags &= ~IO_U_F_FREE_DEF;
0c6e7517 418
60f2c658
JA
419 if (io_u->file)
420 put_file_log(td, io_u->file);
2dbdab7e 421
10ba535a 422 io_u->file = NULL;
0c41214f
RR
423 if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
424 td->cur_depth--;
e8462bd8 425 flist_del_init(&io_u->list);
01743ee1 426 flist_add(&io_u->list, &td->io_u_freelist);
e8462bd8
JA
427 td_io_u_unlock(td);
428 td_io_u_free_notify(td);
10ba535a
JA
429}
430
f2bba182
RR
431void clear_io_u(struct thread_data *td, struct io_u *io_u)
432{
433 io_u->flags &= ~IO_U_F_FLIGHT;
434 put_io_u(td, io_u);
435}
436
755200a3
JA
437void requeue_io_u(struct thread_data *td, struct io_u **io_u)
438{
439 struct io_u *__io_u = *io_u;
440
465221b0
JA
441 dprint(FD_IO, "requeue %p\n", __io_u);
442
e8462bd8
JA
443 td_io_u_lock(td);
444
4d2e0f49 445 __io_u->flags |= IO_U_F_FREE;
5f9099ea 446 if ((__io_u->flags & IO_U_F_FLIGHT) && !ddir_sync(__io_u->ddir))
e4f54adb 447 td->io_issues[__io_u->ddir]--;
5ec10eaa 448
4d2e0f49 449 __io_u->flags &= ~IO_U_F_FLIGHT;
0c41214f
RR
450 if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
451 td->cur_depth--;
01743ee1
JA
452 flist_del(&__io_u->list);
453 flist_add_tail(&__io_u->list, &td->io_u_requeues);
e8462bd8 454 td_io_u_unlock(td);
755200a3
JA
455 *io_u = NULL;
456}
457
9bf2061e 458static int fill_io_u(struct thread_data *td, struct io_u *io_u)
10ba535a 459{
b4c5e1ac
JA
460 if (td->io_ops->flags & FIO_NOIO)
461 goto out;
462
5f9099ea
JA
463 io_u->ddir = get_rw_ddir(td);
464
87dc1ab1 465 /*
5f9099ea 466 * fsync() or fdatasync(), we are done
87dc1ab1 467 */
5f9099ea 468 if (ddir_sync(io_u->ddir))
c38e9468 469 goto out;
a00735e6 470
48f5abd3
JA
471 /*
472 * See if it's time to switch to a new zone
473 */
474 if (td->zone_bytes >= td->o.zone_size) {
475 td->zone_bytes = 0;
476 io_u->file->last_pos += td->o.zone_skip;
477 td->io_skip_bytes += td->o.zone_skip;
478 }
479
10ba535a 480 /*
c685b5b2
JA
481 * No log, let the seq/rand engine retrieve the next buflen and
482 * position.
10ba535a 483 */
2ba1c290
JA
484 if (get_next_offset(td, io_u)) {
485 dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
bca4ed4d 486 return 1;
2ba1c290 487 }
10ba535a 488
9bf2061e 489 io_u->buflen = get_next_buflen(td, io_u);
2ba1c290
JA
490 if (!io_u->buflen) {
491 dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
bca4ed4d 492 return 1;
2ba1c290 493 }
bca4ed4d 494
2ba1c290
JA
495 if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
496 dprint(FD_IO, "io_u %p, offset too large\n", io_u);
4ba66134
JA
497 dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset,
498 io_u->buflen, io_u->file->real_file_size);
6a5e6884 499 return 1;
2ba1c290 500 }
6a5e6884 501
bca4ed4d
JA
502 /*
503 * mark entry before potentially trimming io_u
504 */
303032ae 505 if (td_random(td) && file_randommap(td, io_u->file))
9bf2061e 506 mark_random_map(td, io_u);
bca4ed4d
JA
507
508 /*
509 * If using a write iolog, store this entry.
510 */
c38e9468 511out:
2ba1c290 512 dprint_io_u(io_u, "fill_io_u");
d9d91e39 513 td->zone_bytes += io_u->buflen;
f29b25a3 514 log_io_u(td, io_u);
bca4ed4d 515 return 0;
10ba535a
JA
516}
517
838bc709
JA
518static void __io_u_mark_map(unsigned int *map, unsigned int nr)
519{
520 int index = 0;
521
522 switch (nr) {
523 default:
524 index = 6;
525 break;
526 case 33 ... 64:
527 index = 5;
528 break;
529 case 17 ... 32:
530 index = 4;
531 break;
532 case 9 ... 16:
533 index = 3;
534 break;
535 case 5 ... 8:
536 index = 2;
537 break;
538 case 1 ... 4:
539 index = 1;
540 case 0:
541 break;
542 }
543
544 map[index]++;
545}
546
547void io_u_mark_submit(struct thread_data *td, unsigned int nr)
548{
549 __io_u_mark_map(td->ts.io_u_submit, nr);
550 td->ts.total_submit++;
551}
552
553void io_u_mark_complete(struct thread_data *td, unsigned int nr)
554{
555 __io_u_mark_map(td->ts.io_u_complete, nr);
556 td->ts.total_complete++;
557}
558
d8005759 559void io_u_mark_depth(struct thread_data *td, unsigned int nr)
71619dc2
JA
560{
561 int index = 0;
562
563 switch (td->cur_depth) {
564 default:
a783e61a
JA
565 index = 6;
566 break;
71619dc2 567 case 32 ... 63:
a783e61a
JA
568 index = 5;
569 break;
71619dc2 570 case 16 ... 31:
a783e61a
JA
571 index = 4;
572 break;
71619dc2 573 case 8 ... 15:
a783e61a
JA
574 index = 3;
575 break;
71619dc2 576 case 4 ... 7:
a783e61a
JA
577 index = 2;
578 break;
71619dc2 579 case 2 ... 3:
a783e61a 580 index = 1;
71619dc2
JA
581 case 1:
582 break;
583 }
584
3bec7ae0 585 td->ts.io_u_map[index] += nr;
71619dc2
JA
586}
587
04a0feae
JA
588static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
589{
590 int index = 0;
591
592 assert(usec < 1000);
593
594 switch (usec) {
595 case 750 ... 999:
596 index = 9;
597 break;
598 case 500 ... 749:
599 index = 8;
600 break;
601 case 250 ... 499:
602 index = 7;
603 break;
604 case 100 ... 249:
605 index = 6;
606 break;
607 case 50 ... 99:
608 index = 5;
609 break;
610 case 20 ... 49:
611 index = 4;
612 break;
613 case 10 ... 19:
614 index = 3;
615 break;
616 case 4 ... 9:
617 index = 2;
618 break;
619 case 2 ... 3:
620 index = 1;
621 case 0 ... 1:
622 break;
623 }
624
625 assert(index < FIO_IO_U_LAT_U_NR);
626 td->ts.io_u_lat_u[index]++;
627}
628
629static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
ec118304
JA
630{
631 int index = 0;
632
633 switch (msec) {
634 default:
04a0feae
JA
635 index = 11;
636 break;
8abdce66 637 case 1000 ... 1999:
04a0feae
JA
638 index = 10;
639 break;
8abdce66 640 case 750 ... 999:
04a0feae
JA
641 index = 9;
642 break;
8abdce66 643 case 500 ... 749:
04a0feae
JA
644 index = 8;
645 break;
8abdce66 646 case 250 ... 499:
04a0feae
JA
647 index = 7;
648 break;
8abdce66 649 case 100 ... 249:
04a0feae
JA
650 index = 6;
651 break;
8abdce66 652 case 50 ... 99:
04a0feae
JA
653 index = 5;
654 break;
8abdce66 655 case 20 ... 49:
04a0feae
JA
656 index = 4;
657 break;
8abdce66 658 case 10 ... 19:
04a0feae
JA
659 index = 3;
660 break;
8abdce66 661 case 4 ... 9:
04a0feae
JA
662 index = 2;
663 break;
ec118304 664 case 2 ... 3:
04a0feae 665 index = 1;
ec118304
JA
666 case 0 ... 1:
667 break;
668 }
669
04a0feae
JA
670 assert(index < FIO_IO_U_LAT_M_NR);
671 td->ts.io_u_lat_m[index]++;
672}
673
674static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
675{
676 if (usec < 1000)
677 io_u_mark_lat_usec(td, usec);
678 else
679 io_u_mark_lat_msec(td, usec / 1000);
ec118304
JA
680}
681
0aabe160
JA
682/*
683 * Get next file to service by choosing one at random
684 */
2cc52930
JA
685static struct fio_file *get_next_file_rand(struct thread_data *td,
686 enum fio_file_flags goodf,
d6aed795 687 enum fio_file_flags badf)
0aabe160 688{
0aabe160 689 struct fio_file *f;
1c178180 690 int fno;
0aabe160
JA
691
692 do {
7c83c089 693 long r = os_random_long(&td->next_file_state);
87b10676 694 int opened = 0;
7c83c089 695
5ec10eaa 696 fno = (unsigned int) ((double) td->o.nr_files
dc873b6f 697 * (r / (OS_RAND_MAX + 1.0)));
126d65c6 698 f = td->files[fno];
d6aed795 699 if (fio_file_done(f))
059e63c0 700 continue;
1c178180 701
d6aed795 702 if (!fio_file_open(f)) {
87b10676
JA
703 int err;
704
705 err = td_io_open_file(td, f);
706 if (err)
707 continue;
708 opened = 1;
709 }
710
2ba1c290
JA
711 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
712 dprint(FD_FILE, "get_next_file_rand: %p\n", f);
0aabe160 713 return f;
2ba1c290 714 }
87b10676
JA
715 if (opened)
716 td_io_close_file(td, f);
0aabe160
JA
717 } while (1);
718}
719
720/*
721 * Get next file to service by doing round robin between all available ones
722 */
1c178180
JA
723static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
724 int badf)
3d7c391d
JA
725{
726 unsigned int old_next_file = td->next_file;
727 struct fio_file *f;
728
729 do {
87b10676
JA
730 int opened = 0;
731
126d65c6 732 f = td->files[td->next_file];
3d7c391d
JA
733
734 td->next_file++;
2dc1bbeb 735 if (td->next_file >= td->o.nr_files)
3d7c391d
JA
736 td->next_file = 0;
737
87b10676 738 dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
d6aed795 739 if (fio_file_done(f)) {
d5ed68ea 740 f = NULL;
059e63c0 741 continue;
d5ed68ea 742 }
059e63c0 743
d6aed795 744 if (!fio_file_open(f)) {
87b10676
JA
745 int err;
746
747 err = td_io_open_file(td, f);
b5696bfc
JA
748 if (err) {
749 dprint(FD_FILE, "error %d on open of %s\n",
750 err, f->file_name);
87c27b45 751 f = NULL;
87b10676 752 continue;
b5696bfc 753 }
87b10676
JA
754 opened = 1;
755 }
756
0b9d69ec
JA
757 dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
758 f->flags);
1c178180 759 if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
3d7c391d
JA
760 break;
761
87b10676
JA
762 if (opened)
763 td_io_close_file(td, f);
764
3d7c391d
JA
765 f = NULL;
766 } while (td->next_file != old_next_file);
767
2ba1c290 768 dprint(FD_FILE, "get_next_file_rr: %p\n", f);
3d7c391d
JA
769 return f;
770}
771
bdb4e2e9
JA
772static struct fio_file *get_next_file(struct thread_data *td)
773{
1907dbc6
JA
774 struct fio_file *f;
775
2dc1bbeb 776 assert(td->o.nr_files <= td->files_index);
1c178180 777
b5696bfc 778 if (td->nr_done_files >= td->o.nr_files) {
5ec10eaa
JA
779 dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
780 " nr_files=%d\n", td->nr_open_files,
781 td->nr_done_files,
782 td->o.nr_files);
bdb4e2e9 783 return NULL;
2ba1c290 784 }
bdb4e2e9 785
1907dbc6 786 f = td->file_service_file;
d6aed795 787 if (f && fio_file_open(f) && !fio_file_closing(f)) {
a086c257
JA
788 if (td->o.file_service_type == FIO_FSERVICE_SEQ)
789 goto out;
790 if (td->file_service_left--)
791 goto out;
792 }
1907dbc6 793
a086c257
JA
794 if (td->o.file_service_type == FIO_FSERVICE_RR ||
795 td->o.file_service_type == FIO_FSERVICE_SEQ)
d6aed795 796 f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
bdb4e2e9 797 else
d6aed795 798 f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
1907dbc6
JA
799
800 td->file_service_file = f;
801 td->file_service_left = td->file_service_nr - 1;
2ba1c290 802out:
683023e8 803 dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
1907dbc6 804 return f;
bdb4e2e9
JA
805}
806
429f6675
JA
807static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
808{
809 struct fio_file *f;
810
811 do {
812 f = get_next_file(td);
813 if (!f)
814 return 1;
815
429f6675
JA
816 io_u->file = f;
817 get_file(f);
818
819 if (!fill_io_u(td, io_u))
820 break;
821
b5696bfc 822 put_file_log(td, f);
429f6675 823 td_io_close_file(td, f);
b5696bfc 824 io_u->file = NULL;
d6aed795 825 fio_file_set_done(f);
429f6675 826 td->nr_done_files++;
0b9d69ec
JA
827 dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
828 td->nr_done_files, td->o.nr_files);
429f6675
JA
829 } while (1);
830
831 return 0;
832}
833
834
10ba535a
JA
835struct io_u *__get_io_u(struct thread_data *td)
836{
837 struct io_u *io_u = NULL;
838
e8462bd8
JA
839 td_io_u_lock(td);
840
841again:
01743ee1
JA
842 if (!flist_empty(&td->io_u_requeues))
843 io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
755200a3 844 else if (!queue_full(td)) {
01743ee1 845 io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
10ba535a 846
6040dabc 847 io_u->buflen = 0;
10ba535a 848 io_u->resid = 0;
755200a3 849 io_u->file = NULL;
d7762cf8 850 io_u->end_io = NULL;
755200a3
JA
851 }
852
e8462bd8
JA
853 /*
854 * We ran out, wait for async verify threads to finish and return one
855 */
856 if (!io_u && td->o.verify_async) {
857 pthread_cond_wait(&td->free_cond, &td->io_u_lock);
858 goto again;
859 }
860
755200a3 861 if (io_u) {
0c6e7517 862 assert(io_u->flags & IO_U_F_FREE);
2ecc1b57 863 io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
0c6e7517 864
755200a3 865 io_u->error = 0;
01743ee1
JA
866 flist_del(&io_u->list);
867 flist_add(&io_u->list, &td->io_u_busylist);
10ba535a 868 td->cur_depth++;
0c41214f 869 io_u->flags |= IO_U_F_IN_CUR_DEPTH;
10ba535a
JA
870 }
871
e8462bd8 872 td_io_u_unlock(td);
10ba535a
JA
873 return io_u;
874}
875
876/*
877 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
878 * etc. The returned io_u is fully ready to be prepped and submitted.
879 */
3d7c391d 880struct io_u *get_io_u(struct thread_data *td)
10ba535a 881{
3d7c391d 882 struct fio_file *f;
10ba535a
JA
883 struct io_u *io_u;
884
885 io_u = __get_io_u(td);
2ba1c290
JA
886 if (!io_u) {
887 dprint(FD_IO, "__get_io_u failed\n");
10ba535a 888 return NULL;
2ba1c290 889 }
10ba535a 890
755200a3
JA
891 /*
892 * from a requeue, io_u already setup
893 */
894 if (io_u->file)
77f392bf 895 goto out;
755200a3 896
429f6675
JA
897 /*
898 * If using an iolog, grab next piece if any available.
899 */
900 if (td->o.read_iolog_file) {
901 if (read_iolog_get(td, io_u))
902 goto err_put;
2ba1c290
JA
903 } else if (set_io_u_file(td, io_u)) {
904 dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
429f6675 905 goto err_put;
2ba1c290 906 }
5ec10eaa 907
429f6675 908 f = io_u->file;
d6aed795 909 assert(fio_file_open(f));
97af62ce 910
5f9099ea 911 if (!ddir_sync(io_u->ddir)) {
d0656a93 912 if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
2ba1c290 913 dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
429f6675 914 goto err_put;
2ba1c290 915 }
10ba535a 916
36167d82 917 f->last_pos = io_u->offset + io_u->buflen;
10ba535a 918
c311cd2a 919 if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
87dc1ab1 920 populate_verify_io_u(td, io_u);
e4dad9c6
JA
921 else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
922 io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
87dc1ab1 923 }
10ba535a 924
165faf16
JA
925 /*
926 * Set io data pointers.
927 */
cec6b55d
JA
928 io_u->xfer_buf = io_u->buf;
929 io_u->xfer_buflen = io_u->buflen;
5973cafb 930
6ac7a331 931out:
429f6675 932 if (!td_io_prep(td, io_u)) {
993bf48b
JA
933 if (!td->o.disable_slat)
934 fio_gettime(&io_u->start_time, NULL);
429f6675 935 return io_u;
36167d82 936 }
429f6675 937err_put:
2ba1c290 938 dprint(FD_IO, "get_io_u failed\n");
429f6675
JA
939 put_io_u(td, io_u);
940 return NULL;
10ba535a
JA
941}
942
5451792e
JA
943void io_u_log_error(struct thread_data *td, struct io_u *io_u)
944{
945 const char *msg[] = { "read", "write", "sync" };
946
947 log_err("fio: io_u error");
948
949 if (io_u->file)
950 log_err(" on file %s", io_u->file->file_name);
951
952 log_err(": %s\n", strerror(io_u->error));
953
5ec10eaa
JA
954 log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
955 io_u->offset, io_u->xfer_buflen);
5451792e
JA
956
957 if (!td->error)
958 td_verror(td, io_u->error, "io_u error");
959}
960
97601024
JA
961static void io_completed(struct thread_data *td, struct io_u *io_u,
962 struct io_completion_data *icd)
10ba535a 963{
dbad30b0
JA
964 /*
965 * Older gcc's are too dumb to realize that usec is always used
966 * initialized, silence that warning.
967 */
968 unsigned long uninitialized_var(usec);
10ba535a 969
2ba1c290
JA
970 dprint_io_u(io_u, "io complete");
971
2ecc1b57 972 td_io_u_lock(td);
0c6e7517
JA
973 assert(io_u->flags & IO_U_F_FLIGHT);
974 io_u->flags &= ~IO_U_F_FLIGHT;
2ecc1b57 975 td_io_u_unlock(td);
0c6e7517 976
5f9099ea 977 if (ddir_sync(io_u->ddir)) {
87dc1ab1
JA
978 td->last_was_sync = 1;
979 return;
980 }
981
982 td->last_was_sync = 0;
983
10ba535a
JA
984 if (!io_u->error) {
985 unsigned int bytes = io_u->buflen - io_u->resid;
1e97cce9 986 const enum fio_ddir idx = io_u->ddir;
b29ee5b3 987 int ret;
10ba535a 988
b29ee5b3
JA
989 td->io_blocks[idx]++;
990 td->io_bytes[idx] += bytes;
991 td->this_io_bytes[idx] += bytes;
10ba535a 992
b29ee5b3 993 if (ramp_time_over(td)) {
40e1a6f0
JA
994 unsigned long uninitialized_var(lusec);
995 unsigned long uninitialized_var(rusec);
996
997 if (!td->o.disable_clat || !td->o.disable_bw)
998 lusec = utime_since(&io_u->issue_time,
999 &icd->time);
1000 if (__should_check_rate(td, idx) ||
1001 __should_check_rate(td, idx ^ 1))
1002 rusec = utime_since(&io_u->start_time,
9520ebb9 1003 &icd->time);
721938ae 1004
9520ebb9 1005 if (!td->o.disable_clat) {
01a1caa5 1006 add_clat_sample(td, idx, lusec, bytes);
40e1a6f0 1007 io_u_mark_latency(td, lusec);
9520ebb9
JA
1008 }
1009 if (!td->o.disable_bw)
306ddc97 1010 add_bw_sample(td, idx, bytes, &icd->time);
b23b6a2f
JA
1011 if (__should_check_rate(td, idx)) {
1012 td->rate_pending_usleep[idx] +=
1013 (long) td->rate_usec_cycle[idx] - rusec;
1014 }
581e7141 1015 if (__should_check_rate(td, idx ^ 1))
b23b6a2f 1016 td->rate_pending_usleep[idx ^ 1] -= rusec;
721938ae 1017 }
10ba535a 1018
660a1cb5 1019 if (td_write(td) && idx == DDIR_WRITE &&
8670579e 1020 td->o.do_verify &&
41128405 1021 td->o.verify != VERIFY_NONE)
10ba535a
JA
1022 log_io_piece(td, io_u);
1023
b29ee5b3 1024 icd->bytes_done[idx] += bytes;
3af6ef39 1025
d7762cf8 1026 if (io_u->end_io) {
36690c9b 1027 ret = io_u->end_io(td, io_u);
3af6ef39
JA
1028 if (ret && !icd->error)
1029 icd->error = ret;
1030 }
5451792e 1031 } else {
10ba535a 1032 icd->error = io_u->error;
5451792e
JA
1033 io_u_log_error(td, io_u);
1034 }
f2bba182
RR
1035 if (td->o.continue_on_error && icd->error &&
1036 td_non_fatal_error(icd->error)) {
1037 /*
1038 * If there is a non_fatal error, then add to the error count
1039 * and clear all the errors.
1040 */
1041 update_error_count(td, icd->error);
1042 td_clear_error(td);
1043 icd->error = 0;
1044 io_u->error = 0;
1045 }
10ba535a
JA
1046}
1047
9520ebb9
JA
1048static void init_icd(struct thread_data *td, struct io_completion_data *icd,
1049 int nr)
10ba535a 1050{
9520ebb9
JA
1051 if (!td->o.disable_clat || !td->o.disable_bw)
1052 fio_gettime(&icd->time, NULL);
02bcaa8c 1053
3af6ef39
JA
1054 icd->nr = nr;
1055
10ba535a
JA
1056 icd->error = 0;
1057 icd->bytes_done[0] = icd->bytes_done[1] = 0;
36167d82
JA
1058}
1059
97601024
JA
1060static void ios_completed(struct thread_data *td,
1061 struct io_completion_data *icd)
36167d82
JA
1062{
1063 struct io_u *io_u;
1064 int i;
1065
10ba535a
JA
1066 for (i = 0; i < icd->nr; i++) {
1067 io_u = td->io_ops->event(td, i);
1068
1069 io_completed(td, io_u, icd);
e8462bd8
JA
1070
1071 if (!(io_u->flags & IO_U_F_FREE_DEF))
1072 put_io_u(td, io_u);
10ba535a
JA
1073 }
1074}
97601024 1075
e7e6cfb4
JA
1076/*
1077 * Complete a single io_u for the sync engines.
1078 */
581e7141
JA
1079int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
1080 unsigned long *bytes)
97601024
JA
1081{
1082 struct io_completion_data icd;
1083
9520ebb9 1084 init_icd(td, &icd, 1);
97601024 1085 io_completed(td, io_u, &icd);
e8462bd8
JA
1086
1087 if (!(io_u->flags & IO_U_F_FREE_DEF))
1088 put_io_u(td, io_u);
97601024 1089
581e7141
JA
1090 if (icd.error) {
1091 td_verror(td, icd.error, "io_u_sync_complete");
1092 return -1;
1093 }
97601024 1094
581e7141
JA
1095 if (bytes) {
1096 bytes[0] += icd.bytes_done[0];
1097 bytes[1] += icd.bytes_done[1];
1098 }
1099
1100 return 0;
97601024
JA
1101}
1102
e7e6cfb4
JA
1103/*
1104 * Called to complete min_events number of io for the async engines.
1105 */
581e7141
JA
1106int io_u_queued_complete(struct thread_data *td, int min_evts,
1107 unsigned long *bytes)
97601024 1108{
97601024 1109 struct io_completion_data icd;
00de55ef 1110 struct timespec *tvp = NULL;
97601024 1111 int ret;
4d06a338 1112 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
97601024 1113
4950421a 1114 dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
b271fe62 1115
4950421a 1116 if (!min_evts)
00de55ef 1117 tvp = &ts;
97601024 1118
4950421a 1119 ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
97601024 1120 if (ret < 0) {
e1161c32 1121 td_verror(td, -ret, "td_io_getevents");
97601024
JA
1122 return ret;
1123 } else if (!ret)
1124 return ret;
1125
9520ebb9 1126 init_icd(td, &icd, ret);
97601024 1127 ios_completed(td, &icd);
581e7141
JA
1128 if (icd.error) {
1129 td_verror(td, icd.error, "io_u_queued_complete");
1130 return -1;
1131 }
97601024 1132
581e7141
JA
1133 if (bytes) {
1134 bytes[0] += icd.bytes_done[0];
1135 bytes[1] += icd.bytes_done[1];
1136 }
1137
1138 return 0;
97601024 1139}
7e77dd02
JA
1140
1141/*
1142 * Call when io_u is really queued, to update the submission latency.
1143 */
1144void io_u_queued(struct thread_data *td, struct io_u *io_u)
1145{
9520ebb9
JA
1146 if (!td->o.disable_slat) {
1147 unsigned long slat_time;
7e77dd02 1148
9520ebb9 1149 slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
29a90ddb 1150 add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
9520ebb9 1151 }
7e77dd02 1152}
433afcb4 1153
5973cafb
JA
1154/*
1155 * "randomly" fill the buffer contents
1156 */
1157void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
1158 unsigned int max_bs)
1159{
1160 long *ptr = io_u->buf;
1161
1162 if (!td->o.zero_buffers) {
1163 while ((void *) ptr - io_u->buf < max_bs) {
1164 *ptr = rand() * GOLDEN_RATIO_PRIME;
1165 ptr++;
1166 }
1167 } else
1168 memset(ptr, 0, max_bs);
1169}