Commit | Line | Data |
---|---|---|
10ba535a JA |
1 | #include <unistd.h> |
2 | #include <fcntl.h> | |
3 | #include <string.h> | |
4 | #include <signal.h> | |
5 | #include <time.h> | |
10ba535a JA |
6 | |
7 | #include "fio.h" | |
8 | #include "os.h" | |
9 | ||
97601024 JA |
10 | struct io_completion_data { |
11 | int nr; /* input */ | |
12 | endio_handler *handler; /* input */ | |
13 | ||
14 | int error; /* output */ | |
15 | unsigned long bytes_done[2]; /* output */ | |
16 | struct timeval time; /* output */ | |
17 | }; | |
18 | ||
10ba535a JA |
19 | /* |
20 | * The ->file_map[] contains a map of blocks we have or have not done io | |
21 | * to yet. Used to make sure we cover the entire range in a fair fashion. | |
22 | */ | |
23 | static int random_map_free(struct thread_data *td, struct fio_file *f, | |
24 | unsigned long long block) | |
25 | { | |
26 | unsigned int idx = RAND_MAP_IDX(td, f, block); | |
27 | unsigned int bit = RAND_MAP_BIT(td, f, block); | |
28 | ||
29 | return (f->file_map[idx] & (1UL << bit)) == 0; | |
30 | } | |
31 | ||
df415585 JA |
32 | /* |
33 | * Mark a given offset as used in the map. | |
34 | */ | |
35 | static void mark_random_map(struct thread_data *td, struct fio_file *f, | |
36 | struct io_u *io_u) | |
37 | { | |
afdbe580 | 38 | unsigned int min_bs = td->rw_min_bs; |
a00735e6 JA |
39 | unsigned long long block; |
40 | unsigned int blocks; | |
c685b5b2 | 41 | unsigned int nr_blocks; |
df415585 | 42 | |
a00735e6 JA |
43 | block = io_u->offset / (unsigned long long) min_bs; |
44 | blocks = 0; | |
c685b5b2 JA |
45 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
46 | ||
47 | while (blocks < nr_blocks) { | |
df415585 JA |
48 | unsigned int idx, bit; |
49 | ||
50 | if (!random_map_free(td, f, block)) | |
51 | break; | |
52 | ||
53 | idx = RAND_MAP_IDX(td, f, block); | |
54 | bit = RAND_MAP_BIT(td, f, block); | |
55 | ||
0032bf9f | 56 | fio_assert(td, idx < f->num_maps); |
df415585 JA |
57 | |
58 | f->file_map[idx] |= (1UL << bit); | |
59 | block++; | |
60 | blocks++; | |
61 | } | |
62 | ||
a00735e6 JA |
63 | if ((blocks * min_bs) < io_u->buflen) |
64 | io_u->buflen = blocks * min_bs; | |
df415585 JA |
65 | } |
66 | ||
10ba535a JA |
67 | /* |
68 | * Return the next free block in the map. | |
69 | */ | |
70 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, | |
71 | unsigned long long *b) | |
72 | { | |
73 | int i; | |
74 | ||
c685b5b2 JA |
75 | i = f->last_free_lookup; |
76 | *b = (i * BLOCKS_PER_MAP); | |
ee88470c | 77 | while ((*b) * td->rw_min_bs < f->real_file_size) { |
10ba535a JA |
78 | if (f->file_map[i] != -1UL) { |
79 | *b += ffz(f->file_map[i]); | |
c685b5b2 | 80 | f->last_free_lookup = i; |
10ba535a JA |
81 | return 0; |
82 | } | |
83 | ||
84 | *b += BLOCKS_PER_MAP; | |
85 | i++; | |
86 | } | |
87 | ||
88 | return 1; | |
89 | } | |
90 | ||
91 | /* | |
92 | * For random io, generate a random new block and see if it's used. Repeat | |
93 | * until we find a free one. For sequential io, just return the end of | |
94 | * the last io issued. | |
95 | */ | |
96 | static int get_next_offset(struct thread_data *td, struct fio_file *f, | |
c685b5b2 | 97 | struct io_u *io_u) |
10ba535a | 98 | { |
c685b5b2 | 99 | const int ddir = io_u->ddir; |
10ba535a JA |
100 | unsigned long long b, rb; |
101 | long r; | |
102 | ||
103 | if (!td->sequential) { | |
f697125a | 104 | unsigned long long max_blocks = f->file_size / td->min_bs[ddir]; |
c685b5b2 | 105 | int loops = 5; |
10ba535a JA |
106 | |
107 | do { | |
108 | r = os_random_long(&td->random_state); | |
109 | b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0)); | |
bb8895e0 JA |
110 | if (td->norandommap) |
111 | break; | |
a00735e6 | 112 | rb = b + (f->file_offset / td->min_bs[ddir]); |
10ba535a JA |
113 | loops--; |
114 | } while (!random_map_free(td, f, rb) && loops); | |
115 | ||
bca4ed4d JA |
116 | /* |
117 | * if we failed to retrieve a truly random offset within | |
118 | * the loops assigned, see if there are free ones left at all | |
119 | */ | |
120 | if (!loops && get_next_free_block(td, f, &b)) | |
121 | return 1; | |
10ba535a | 122 | } else |
a00735e6 | 123 | b = f->last_pos / td->min_bs[ddir]; |
10ba535a | 124 | |
c685b5b2 | 125 | io_u->offset = (b * td->min_bs[ddir]) + f->file_offset; |
bca4ed4d | 126 | if (io_u->offset >= f->real_file_size) |
10ba535a JA |
127 | return 1; |
128 | ||
129 | return 0; | |
130 | } | |
131 | ||
bca4ed4d JA |
132 | static unsigned int get_next_buflen(struct thread_data *td, struct fio_file *f, |
133 | struct io_u *io_u) | |
10ba535a | 134 | { |
bca4ed4d | 135 | const int ddir = io_u->ddir; |
10ba535a JA |
136 | unsigned int buflen; |
137 | long r; | |
138 | ||
a00735e6 JA |
139 | if (td->min_bs[ddir] == td->max_bs[ddir]) |
140 | buflen = td->min_bs[ddir]; | |
10ba535a JA |
141 | else { |
142 | r = os_random_long(&td->bsrange_state); | |
1e97cce9 | 143 | buflen = (unsigned int) (1 + (double) (td->max_bs[ddir] - 1) * r / (RAND_MAX + 1.0)); |
690adba3 | 144 | if (!td->bs_unaligned) |
a00735e6 | 145 | buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1); |
10ba535a JA |
146 | } |
147 | ||
bca4ed4d | 148 | while (buflen + io_u->offset > f->real_file_size) { |
bca4ed4d JA |
149 | if (buflen == td->min_bs[ddir]) |
150 | return 0; | |
151 | ||
152 | buflen = td->min_bs[ddir]; | |
10ba535a JA |
153 | } |
154 | ||
155 | return buflen; | |
156 | } | |
157 | ||
158 | /* | |
159 | * Return the data direction for the next io_u. If the job is a | |
160 | * mixed read/write workload, check the rwmix cycle and switch if | |
161 | * necessary. | |
162 | */ | |
1e97cce9 | 163 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a JA |
164 | { |
165 | if (td_rw(td)) { | |
166 | struct timeval now; | |
167 | unsigned long elapsed; | |
168 | ||
02bcaa8c | 169 | fio_gettime(&now, NULL); |
10ba535a JA |
170 | elapsed = mtime_since_now(&td->rwmix_switch); |
171 | ||
172 | /* | |
173 | * Check if it's time to seed a new data direction. | |
174 | */ | |
175 | if (elapsed >= td->rwmixcycle) { | |
176 | unsigned int v; | |
177 | long r; | |
178 | ||
179 | r = os_random_long(&td->rwmix_state); | |
180 | v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); | |
181 | if (v < td->rwmixread) | |
182 | td->rwmix_ddir = DDIR_READ; | |
183 | else | |
184 | td->rwmix_ddir = DDIR_WRITE; | |
185 | memcpy(&td->rwmix_switch, &now, sizeof(now)); | |
186 | } | |
187 | return td->rwmix_ddir; | |
188 | } else if (td_read(td)) | |
189 | return DDIR_READ; | |
190 | else | |
191 | return DDIR_WRITE; | |
192 | } | |
193 | ||
10ba535a JA |
194 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
195 | { | |
196 | io_u->file = NULL; | |
197 | list_del(&io_u->list); | |
198 | list_add(&io_u->list, &td->io_u_freelist); | |
199 | td->cur_depth--; | |
200 | } | |
201 | ||
755200a3 JA |
202 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
203 | { | |
204 | struct io_u *__io_u = *io_u; | |
205 | ||
206 | list_del(&__io_u->list); | |
207 | list_add_tail(&__io_u->list, &td->io_u_requeues); | |
208 | td->cur_depth--; | |
209 | *io_u = NULL; | |
210 | } | |
211 | ||
10ba535a JA |
212 | static int fill_io_u(struct thread_data *td, struct fio_file *f, |
213 | struct io_u *io_u) | |
214 | { | |
215 | /* | |
216 | * If using an iolog, grab next piece if any available. | |
217 | */ | |
218 | if (td->read_iolog) | |
219 | return read_iolog_get(td, io_u); | |
220 | ||
87dc1ab1 JA |
221 | /* |
222 | * see if it's time to sync | |
223 | */ | |
755200a3 JA |
224 | if (td->fsync_blocks && !(td->io_issues[DDIR_WRITE] % td->fsync_blocks) |
225 | && td->io_issues[DDIR_WRITE] && should_fsync(td)) { | |
87dc1ab1 JA |
226 | io_u->ddir = DDIR_SYNC; |
227 | io_u->file = f; | |
228 | return 0; | |
229 | } | |
230 | ||
a00735e6 JA |
231 | io_u->ddir = get_rw_ddir(td); |
232 | ||
10ba535a | 233 | /* |
c685b5b2 JA |
234 | * No log, let the seq/rand engine retrieve the next buflen and |
235 | * position. | |
10ba535a | 236 | */ |
bca4ed4d JA |
237 | if (get_next_offset(td, f, io_u)) |
238 | return 1; | |
10ba535a | 239 | |
bca4ed4d JA |
240 | io_u->buflen = get_next_buflen(td, f, io_u); |
241 | if (!io_u->buflen) | |
242 | return 1; | |
243 | ||
244 | /* | |
245 | * mark entry before potentially trimming io_u | |
246 | */ | |
247 | if (!td->read_iolog && !td->sequential && !td->norandommap) | |
248 | mark_random_map(td, f, io_u); | |
249 | ||
250 | /* | |
251 | * If using a write iolog, store this entry. | |
252 | */ | |
253 | if (td->write_iolog_file) | |
254 | write_iolog_put(td, io_u); | |
255 | ||
256 | io_u->file = f; | |
257 | return 0; | |
10ba535a JA |
258 | } |
259 | ||
71619dc2 JA |
260 | static void io_u_mark_depth(struct thread_data *td) |
261 | { | |
262 | int index = 0; | |
263 | ||
264 | switch (td->cur_depth) { | |
265 | default: | |
266 | index++; | |
267 | case 32 ... 63: | |
268 | index++; | |
269 | case 16 ... 31: | |
270 | index++; | |
271 | case 8 ... 15: | |
272 | index++; | |
273 | case 4 ... 7: | |
274 | index++; | |
275 | case 2 ... 3: | |
276 | index++; | |
277 | case 1: | |
278 | break; | |
279 | } | |
280 | ||
281 | td->io_u_map[index]++; | |
282 | td->total_io_u++; | |
283 | } | |
284 | ||
ec118304 JA |
285 | static void io_u_mark_latency(struct thread_data *td, unsigned long msec) |
286 | { | |
287 | int index = 0; | |
288 | ||
289 | switch (msec) { | |
290 | default: | |
291 | index++; | |
292 | case 1024 ... 2047: | |
293 | index++; | |
294 | case 512 ... 1023: | |
295 | index++; | |
296 | case 256 ... 511: | |
297 | index++; | |
298 | case 128 ... 255: | |
299 | index++; | |
300 | case 64 ... 127: | |
301 | index++; | |
302 | case 32 ... 63: | |
303 | index++; | |
304 | case 16 ... 31: | |
305 | index++; | |
306 | case 8 ... 15: | |
307 | index++; | |
308 | case 4 ... 7: | |
309 | index++; | |
310 | case 2 ... 3: | |
311 | index++; | |
312 | case 0 ... 1: | |
313 | break; | |
314 | } | |
315 | ||
316 | td->io_u_lat[index]++; | |
317 | } | |
318 | ||
3d7c391d JA |
319 | static struct fio_file *get_next_file(struct thread_data *td) |
320 | { | |
321 | unsigned int old_next_file = td->next_file; | |
322 | struct fio_file *f; | |
323 | ||
324 | do { | |
325 | f = &td->files[td->next_file]; | |
326 | ||
327 | td->next_file++; | |
328 | if (td->next_file >= td->nr_files) | |
329 | td->next_file = 0; | |
330 | ||
331 | if (f->fd != -1) | |
332 | break; | |
333 | ||
334 | f = NULL; | |
335 | } while (td->next_file != old_next_file); | |
336 | ||
337 | return f; | |
338 | } | |
339 | ||
10ba535a JA |
340 | struct io_u *__get_io_u(struct thread_data *td) |
341 | { | |
342 | struct io_u *io_u = NULL; | |
343 | ||
755200a3 JA |
344 | if (!list_empty(&td->io_u_requeues)) |
345 | io_u = list_entry(td->io_u_requeues.next, struct io_u, list); | |
346 | else if (!queue_full(td)) { | |
10ba535a JA |
347 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); |
348 | ||
6040dabc | 349 | io_u->buflen = 0; |
10ba535a | 350 | io_u->resid = 0; |
755200a3 JA |
351 | io_u->file = NULL; |
352 | } | |
353 | ||
354 | if (io_u) { | |
355 | io_u->error = 0; | |
10ba535a JA |
356 | list_del(&io_u->list); |
357 | list_add(&io_u->list, &td->io_u_busylist); | |
358 | td->cur_depth++; | |
71619dc2 | 359 | io_u_mark_depth(td); |
10ba535a JA |
360 | } |
361 | ||
362 | return io_u; | |
363 | } | |
364 | ||
365 | /* | |
366 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
367 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
368 | */ | |
3d7c391d | 369 | struct io_u *get_io_u(struct thread_data *td) |
10ba535a | 370 | { |
3d7c391d | 371 | struct fio_file *f; |
10ba535a JA |
372 | struct io_u *io_u; |
373 | ||
374 | io_u = __get_io_u(td); | |
375 | if (!io_u) | |
376 | return NULL; | |
377 | ||
755200a3 JA |
378 | /* |
379 | * from a requeue, io_u already setup | |
380 | */ | |
381 | if (io_u->file) | |
77f392bf | 382 | goto out; |
755200a3 | 383 | |
3d7c391d JA |
384 | f = get_next_file(td); |
385 | if (!f) { | |
386 | put_io_u(td, io_u); | |
387 | return NULL; | |
388 | } | |
389 | ||
390 | io_u->file = f; | |
391 | ||
10ba535a JA |
392 | if (td->zone_bytes >= td->zone_size) { |
393 | td->zone_bytes = 0; | |
394 | f->last_pos += td->zone_skip; | |
395 | } | |
396 | ||
397 | if (fill_io_u(td, f, io_u)) { | |
398 | put_io_u(td, io_u); | |
399 | return NULL; | |
400 | } | |
401 | ||
ee88470c | 402 | if (io_u->buflen + io_u->offset > f->real_file_size) { |
10ba535a JA |
403 | if (td->io_ops->flags & FIO_RAWIO) { |
404 | put_io_u(td, io_u); | |
405 | return NULL; | |
406 | } | |
407 | ||
ee88470c | 408 | io_u->buflen = f->real_file_size - io_u->offset; |
10ba535a JA |
409 | } |
410 | ||
87dc1ab1 JA |
411 | if (io_u->ddir != DDIR_SYNC) { |
412 | if (!io_u->buflen) { | |
413 | put_io_u(td, io_u); | |
414 | return NULL; | |
415 | } | |
10ba535a | 416 | |
36167d82 | 417 | f->last_pos = io_u->offset + io_u->buflen; |
10ba535a | 418 | |
87dc1ab1 JA |
419 | if (td->verify != VERIFY_NONE) |
420 | populate_verify_io_u(td, io_u); | |
421 | } | |
10ba535a | 422 | |
165faf16 JA |
423 | /* |
424 | * Set io data pointers. | |
425 | */ | |
77f392bf | 426 | out: |
cec6b55d JA |
427 | io_u->xfer_buf = io_u->buf; |
428 | io_u->xfer_buflen = io_u->buflen; | |
165faf16 | 429 | |
36167d82 JA |
430 | if (td_io_prep(td, io_u)) { |
431 | put_io_u(td, io_u); | |
432 | return NULL; | |
433 | } | |
434 | ||
02bcaa8c | 435 | fio_gettime(&io_u->start_time, NULL); |
10ba535a JA |
436 | return io_u; |
437 | } | |
438 | ||
97601024 JA |
439 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
440 | struct io_completion_data *icd) | |
10ba535a | 441 | { |
10ba535a JA |
442 | unsigned long msec; |
443 | ||
87dc1ab1 JA |
444 | if (io_u->ddir == DDIR_SYNC) { |
445 | td->last_was_sync = 1; | |
446 | return; | |
447 | } | |
448 | ||
449 | td->last_was_sync = 0; | |
450 | ||
10ba535a JA |
451 | if (!io_u->error) { |
452 | unsigned int bytes = io_u->buflen - io_u->resid; | |
1e97cce9 | 453 | const enum fio_ddir idx = io_u->ddir; |
3af6ef39 | 454 | int ret; |
10ba535a JA |
455 | |
456 | td->io_blocks[idx]++; | |
457 | td->io_bytes[idx] += bytes; | |
458 | td->zone_bytes += bytes; | |
459 | td->this_io_bytes[idx] += bytes; | |
460 | ||
02bcaa8c JA |
461 | io_u->file->last_completed_pos = io_u->offset + io_u->buflen; |
462 | ||
463 | msec = mtime_since(&io_u->issue_time, &icd->time); | |
10ba535a JA |
464 | |
465 | add_clat_sample(td, idx, msec); | |
02bcaa8c | 466 | add_bw_sample(td, idx, &icd->time); |
ec118304 | 467 | io_u_mark_latency(td, msec); |
10ba535a JA |
468 | |
469 | if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE) | |
470 | log_io_piece(td, io_u); | |
471 | ||
472 | icd->bytes_done[idx] += bytes; | |
3af6ef39 JA |
473 | |
474 | if (icd->handler) { | |
475 | ret = icd->handler(io_u); | |
476 | if (ret && !icd->error) | |
477 | icd->error = ret; | |
478 | } | |
10ba535a JA |
479 | } else |
480 | icd->error = io_u->error; | |
481 | } | |
482 | ||
97601024 JA |
483 | static void init_icd(struct io_completion_data *icd, endio_handler *handler, |
484 | int nr) | |
10ba535a | 485 | { |
02bcaa8c JA |
486 | fio_gettime(&icd->time, NULL); |
487 | ||
3af6ef39 JA |
488 | icd->handler = handler; |
489 | icd->nr = nr; | |
490 | ||
10ba535a JA |
491 | icd->error = 0; |
492 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
36167d82 JA |
493 | } |
494 | ||
97601024 JA |
495 | static void ios_completed(struct thread_data *td, |
496 | struct io_completion_data *icd) | |
36167d82 JA |
497 | { |
498 | struct io_u *io_u; | |
499 | int i; | |
500 | ||
10ba535a JA |
501 | for (i = 0; i < icd->nr; i++) { |
502 | io_u = td->io_ops->event(td, i); | |
503 | ||
504 | io_completed(td, io_u, icd); | |
505 | put_io_u(td, io_u); | |
506 | } | |
507 | } | |
97601024 | 508 | |
e7e6cfb4 JA |
509 | /* |
510 | * Complete a single io_u for the sync engines. | |
511 | */ | |
97601024 JA |
512 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u, |
513 | endio_handler *handler) | |
514 | { | |
515 | struct io_completion_data icd; | |
516 | ||
517 | init_icd(&icd, handler, 1); | |
518 | io_completed(td, io_u, &icd); | |
519 | put_io_u(td, io_u); | |
520 | ||
521 | if (!icd.error) | |
522 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
523 | ||
97601024 JA |
524 | return -1; |
525 | } | |
526 | ||
e7e6cfb4 JA |
527 | /* |
528 | * Called to complete min_events number of io for the async engines. | |
529 | */ | |
97601024 JA |
530 | long io_u_queued_complete(struct thread_data *td, int min_events, |
531 | endio_handler *handler) | |
532 | ||
533 | { | |
97601024 | 534 | struct io_completion_data icd; |
00de55ef | 535 | struct timespec *tvp = NULL; |
97601024 JA |
536 | int ret; |
537 | ||
755200a3 | 538 | if (min_events > 0) { |
755200a3 JA |
539 | ret = td_io_commit(td); |
540 | if (ret < 0) { | |
541 | td_verror(td, -ret); | |
542 | return ret; | |
543 | } | |
00de55ef JA |
544 | } else { |
545 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; | |
546 | ||
547 | tvp = &ts; | |
755200a3 | 548 | } |
97601024 | 549 | |
00de55ef | 550 | ret = td_io_getevents(td, min_events, td->cur_depth, tvp); |
97601024 JA |
551 | if (ret < 0) { |
552 | td_verror(td, -ret); | |
553 | return ret; | |
554 | } else if (!ret) | |
555 | return ret; | |
556 | ||
557 | init_icd(&icd, handler, ret); | |
558 | ios_completed(td, &icd); | |
559 | if (!icd.error) | |
560 | return icd.bytes_done[0] + icd.bytes_done[1]; | |
561 | ||
97601024 JA |
562 | return -1; |
563 | } | |
7e77dd02 JA |
564 | |
565 | /* | |
566 | * Call when io_u is really queued, to update the submission latency. | |
567 | */ | |
568 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
569 | { | |
570 | unsigned long slat_time; | |
571 | ||
572 | slat_time = mtime_since(&io_u->start_time, &io_u->issue_time); | |
573 | add_slat_sample(td, io_u->ddir, slat_time); | |
574 | } |