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> | |
0c6e7517 | 6 | #include <assert.h> |
10ba535a JA |
7 | |
8 | #include "fio.h" | |
5973cafb | 9 | #include "hash.h" |
4f5af7b2 | 10 | #include "verify.h" |
0d29de83 | 11 | #include "trim.h" |
1fbbf72e | 12 | #include "lib/rand.h" |
7ebd796f | 13 | #include "lib/axmap.h" |
002fe734 | 14 | #include "err.h" |
0f38bbef | 15 | #include "lib/pow2.h" |
4b157ac6 | 16 | #include "minmax.h" |
10ba535a | 17 | |
97601024 JA |
18 | struct io_completion_data { |
19 | int nr; /* input */ | |
97601024 JA |
20 | |
21 | int error; /* output */ | |
100f49f1 | 22 | uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */ |
8b6a404c | 23 | struct timespec time; /* output */ |
97601024 JA |
24 | }; |
25 | ||
10ba535a | 26 | /* |
7ebd796f | 27 | * The ->io_axmap contains a map of blocks we have or have not done io |
10ba535a JA |
28 | * to yet. Used to make sure we cover the entire range in a fair fashion. |
29 | */ | |
e39c0676 | 30 | static bool random_map_free(struct fio_file *f, const uint64_t block) |
10ba535a | 31 | { |
7ebd796f | 32 | return !axmap_isset(f->io_axmap, block); |
10ba535a JA |
33 | } |
34 | ||
df415585 JA |
35 | /* |
36 | * Mark a given offset as used in the map. | |
37 | */ | |
9bf2061e | 38 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
df415585 | 39 | { |
956e60ea | 40 | unsigned int min_bs = td->o.min_bs[io_u->ddir]; |
9bf2061e | 41 | struct fio_file *f = io_u->file; |
51ede0b1 | 42 | unsigned int nr_blocks; |
1ae83d45 | 43 | uint64_t block; |
df415585 | 44 | |
1ae83d45 | 45 | block = (io_u->offset - f->file_offset) / (uint64_t) min_bs; |
c685b5b2 | 46 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
df415585 | 47 | |
2ab9e98b | 48 | if (!(io_u->flags & IO_U_F_BUSY_OK)) |
7ebd796f | 49 | nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks); |
df415585 | 50 | |
51ede0b1 JA |
51 | if ((nr_blocks * min_bs) < io_u->buflen) |
52 | io_u->buflen = nr_blocks * min_bs; | |
df415585 JA |
53 | } |
54 | ||
74776733 JA |
55 | static uint64_t last_block(struct thread_data *td, struct fio_file *f, |
56 | enum fio_ddir ddir) | |
2ba1c290 | 57 | { |
74776733 JA |
58 | uint64_t max_blocks; |
59 | uint64_t max_size; | |
2ba1c290 | 60 | |
ff58fced JA |
61 | assert(ddir_rw(ddir)); |
62 | ||
d9dd70f7 JA |
63 | /* |
64 | * Hmm, should we make sure that ->io_size <= ->real_file_size? | |
79591fa9 | 65 | * -> not for now since there is code assuming it could go either. |
d9dd70f7 JA |
66 | */ |
67 | max_size = f->io_size; | |
68 | if (max_size > f->real_file_size) | |
69 | max_size = f->real_file_size; | |
70 | ||
ed335855 SN |
71 | if (td->o.zone_range) |
72 | max_size = td->o.zone_range; | |
73 | ||
0412d12e JE |
74 | if (td->o.min_bs[ddir] > td->o.ba[ddir]) |
75 | max_size -= td->o.min_bs[ddir] - td->o.ba[ddir]; | |
76 | ||
1ae83d45 | 77 | max_blocks = max_size / (uint64_t) td->o.ba[ddir]; |
2ba1c290 JA |
78 | if (!max_blocks) |
79 | return 0; | |
80 | ||
67778e88 | 81 | return max_blocks; |
2ba1c290 JA |
82 | } |
83 | ||
1ae83d45 JA |
84 | struct rand_off { |
85 | struct flist_head list; | |
86 | uint64_t off; | |
87 | }; | |
88 | ||
e25839d4 | 89 | static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
e0a04ac1 JA |
90 | enum fio_ddir ddir, uint64_t *b, |
91 | uint64_t lastb) | |
ec4015da | 92 | { |
6f49f8bc | 93 | uint64_t r; |
5e0baa7f | 94 | |
c3546b53 JA |
95 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE || |
96 | td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) { | |
6f49f8bc | 97 | |
d6b72507 | 98 | r = __rand(&td->random_state); |
8055e41d | 99 | |
4b91ee8f | 100 | dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r); |
8055e41d | 101 | |
e0a04ac1 | 102 | *b = lastb * (r / (rand_max(&td->random_state) + 1.0)); |
51ede0b1 | 103 | } else { |
8055e41d | 104 | uint64_t off = 0; |
43c63a78 | 105 | |
967d1b63 JA |
106 | assert(fio_file_lfsr(f)); |
107 | ||
6f49f8bc | 108 | if (lfsr_next(&f->lfsr, &off)) |
8055e41d | 109 | return 1; |
ec4015da | 110 | |
8055e41d JA |
111 | *b = off; |
112 | } | |
0ce8b119 | 113 | |
ec4015da | 114 | /* |
51ede0b1 | 115 | * if we are not maintaining a random map, we are done. |
ec4015da | 116 | */ |
51ede0b1 JA |
117 | if (!file_randommap(td, f)) |
118 | goto ret; | |
43c63a78 JA |
119 | |
120 | /* | |
51ede0b1 | 121 | * calculate map offset and check if it's free |
43c63a78 | 122 | */ |
51ede0b1 JA |
123 | if (random_map_free(f, *b)) |
124 | goto ret; | |
125 | ||
4b91ee8f JA |
126 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
127 | (unsigned long long) *b); | |
51ede0b1 | 128 | |
7ebd796f | 129 | *b = axmap_next_free(f->io_axmap, *b); |
51ede0b1 JA |
130 | if (*b == (uint64_t) -1ULL) |
131 | return 1; | |
0ce8b119 JA |
132 | ret: |
133 | return 0; | |
ec4015da JA |
134 | } |
135 | ||
925fee33 JA |
136 | static int __get_next_rand_offset_zipf(struct thread_data *td, |
137 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 138 | uint64_t *b) |
e25839d4 | 139 | { |
9c6f6316 | 140 | *b = zipf_next(&f->zipf); |
e25839d4 JA |
141 | return 0; |
142 | } | |
143 | ||
925fee33 JA |
144 | static int __get_next_rand_offset_pareto(struct thread_data *td, |
145 | struct fio_file *f, enum fio_ddir ddir, | |
1ae83d45 | 146 | uint64_t *b) |
925fee33 | 147 | { |
9c6f6316 | 148 | *b = pareto_next(&f->zipf); |
925fee33 JA |
149 | return 0; |
150 | } | |
151 | ||
56d9fa4b JA |
152 | static int __get_next_rand_offset_gauss(struct thread_data *td, |
153 | struct fio_file *f, enum fio_ddir ddir, | |
154 | uint64_t *b) | |
155 | { | |
156 | *b = gauss_next(&f->gauss); | |
157 | return 0; | |
158 | } | |
159 | ||
59466396 JA |
160 | static int __get_next_rand_offset_zoned_abs(struct thread_data *td, |
161 | struct fio_file *f, | |
162 | enum fio_ddir ddir, uint64_t *b) | |
163 | { | |
164 | struct zone_split_index *zsi; | |
165 | uint64_t offset, lastb; | |
166 | uint64_t send, stotal; | |
167 | static int warned; | |
168 | unsigned int v; | |
169 | ||
170 | lastb = last_block(td, f, ddir); | |
171 | if (!lastb) | |
172 | return 1; | |
173 | ||
174 | if (!td->o.zone_split_nr[ddir]) { | |
175 | bail: | |
176 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
177 | } | |
178 | ||
179 | /* | |
180 | * Generate a value, v, between 1 and 100, both inclusive | |
181 | */ | |
182 | v = rand32_between(&td->zone_state, 1, 100); | |
183 | ||
184 | zsi = &td->zone_state_index[ddir][v - 1]; | |
185 | stotal = zsi->size_prev / td->o.ba[ddir]; | |
186 | send = zsi->size / td->o.ba[ddir]; | |
187 | ||
188 | /* | |
189 | * Should never happen | |
190 | */ | |
191 | if (send == -1U) { | |
192 | if (!warned) { | |
193 | log_err("fio: bug in zoned generation\n"); | |
194 | warned = 1; | |
195 | } | |
196 | goto bail; | |
197 | } else if (send > lastb) { | |
198 | /* | |
199 | * This happens if the user specifies ranges that exceed | |
200 | * the file/device size. We can't handle that gracefully, | |
201 | * so error and exit. | |
202 | */ | |
203 | log_err("fio: zoned_abs sizes exceed file size\n"); | |
204 | return 1; | |
205 | } | |
206 | ||
207 | /* | |
208 | * 'send' is some percentage below or equal to 100 that | |
209 | * marks the end of the current IO range. 'stotal' marks | |
210 | * the start, in percent. | |
211 | */ | |
212 | if (stotal) | |
213 | offset = stotal; | |
214 | else | |
215 | offset = 0; | |
216 | ||
217 | lastb = send - stotal; | |
218 | ||
219 | /* | |
220 | * Generate index from 0..send-of-lastb | |
221 | */ | |
222 | if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1) | |
223 | return 1; | |
224 | ||
225 | /* | |
226 | * Add our start offset, if any | |
227 | */ | |
228 | if (offset) | |
229 | *b += offset; | |
230 | ||
231 | return 0; | |
232 | } | |
233 | ||
e0a04ac1 JA |
234 | static int __get_next_rand_offset_zoned(struct thread_data *td, |
235 | struct fio_file *f, enum fio_ddir ddir, | |
236 | uint64_t *b) | |
237 | { | |
238 | unsigned int v, send, stotal; | |
239 | uint64_t offset, lastb; | |
240 | static int warned; | |
241 | struct zone_split_index *zsi; | |
242 | ||
243 | lastb = last_block(td, f, ddir); | |
244 | if (!lastb) | |
245 | return 1; | |
246 | ||
247 | if (!td->o.zone_split_nr[ddir]) { | |
248 | bail: | |
249 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
250 | } | |
251 | ||
252 | /* | |
253 | * Generate a value, v, between 1 and 100, both inclusive | |
254 | */ | |
36dd3379 | 255 | v = rand32_between(&td->zone_state, 1, 100); |
e0a04ac1 JA |
256 | |
257 | zsi = &td->zone_state_index[ddir][v - 1]; | |
258 | stotal = zsi->size_perc_prev; | |
259 | send = zsi->size_perc; | |
260 | ||
261 | /* | |
262 | * Should never happen | |
263 | */ | |
264 | if (send == -1U) { | |
265 | if (!warned) { | |
266 | log_err("fio: bug in zoned generation\n"); | |
267 | warned = 1; | |
268 | } | |
269 | goto bail; | |
270 | } | |
271 | ||
272 | /* | |
273 | * 'send' is some percentage below or equal to 100 that | |
274 | * marks the end of the current IO range. 'stotal' marks | |
275 | * the start, in percent. | |
276 | */ | |
277 | if (stotal) | |
278 | offset = stotal * lastb / 100ULL; | |
279 | else | |
280 | offset = 0; | |
281 | ||
282 | lastb = lastb * (send - stotal) / 100ULL; | |
283 | ||
284 | /* | |
285 | * Generate index from 0..send-of-lastb | |
286 | */ | |
287 | if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1) | |
288 | return 1; | |
289 | ||
290 | /* | |
291 | * Add our start offset, if any | |
292 | */ | |
293 | if (offset) | |
294 | *b += offset; | |
295 | ||
296 | return 0; | |
297 | } | |
56d9fa4b | 298 | |
1ae83d45 JA |
299 | static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b) |
300 | { | |
301 | struct rand_off *r1 = flist_entry(a, struct rand_off, list); | |
302 | struct rand_off *r2 = flist_entry(b, struct rand_off, list); | |
303 | ||
304 | return r1->off - r2->off; | |
305 | } | |
306 | ||
307 | static int get_off_from_method(struct thread_data *td, struct fio_file *f, | |
308 | enum fio_ddir ddir, uint64_t *b) | |
e25839d4 | 309 | { |
e0a04ac1 JA |
310 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) { |
311 | uint64_t lastb; | |
312 | ||
313 | lastb = last_block(td, f, ddir); | |
314 | if (!lastb) | |
315 | return 1; | |
316 | ||
317 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
318 | } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) | |
e25839d4 | 319 | return __get_next_rand_offset_zipf(td, f, ddir, b); |
925fee33 JA |
320 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
321 | return __get_next_rand_offset_pareto(td, f, ddir, b); | |
56d9fa4b JA |
322 | else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS) |
323 | return __get_next_rand_offset_gauss(td, f, ddir, b); | |
e0a04ac1 JA |
324 | else if (td->o.random_distribution == FIO_RAND_DIST_ZONED) |
325 | return __get_next_rand_offset_zoned(td, f, ddir, b); | |
59466396 JA |
326 | else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS) |
327 | return __get_next_rand_offset_zoned_abs(td, f, ddir, b); | |
e25839d4 JA |
328 | |
329 | log_err("fio: unknown random distribution: %d\n", td->o.random_distribution); | |
330 | return 1; | |
331 | } | |
332 | ||
bcd5abfa JA |
333 | /* |
334 | * Sort the reads for a verify phase in batches of verifysort_nr, if | |
335 | * specified. | |
336 | */ | |
e39c0676 | 337 | static inline bool should_sort_io(struct thread_data *td) |
bcd5abfa JA |
338 | { |
339 | if (!td->o.verifysort_nr || !td->o.do_verify) | |
e39c0676 | 340 | return false; |
bcd5abfa | 341 | if (!td_random(td)) |
e39c0676 | 342 | return false; |
bcd5abfa | 343 | if (td->runstate != TD_VERIFYING) |
e39c0676 | 344 | return false; |
c3546b53 JA |
345 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE || |
346 | td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) | |
e39c0676 | 347 | return false; |
bcd5abfa | 348 | |
e39c0676 | 349 | return true; |
bcd5abfa JA |
350 | } |
351 | ||
e39c0676 | 352 | static bool should_do_random(struct thread_data *td, enum fio_ddir ddir) |
211c9b89 JA |
353 | { |
354 | unsigned int v; | |
211c9b89 | 355 | |
d9472271 | 356 | if (td->o.perc_rand[ddir] == 100) |
e39c0676 | 357 | return true; |
211c9b89 | 358 | |
36dd3379 | 359 | v = rand32_between(&td->seq_rand_state[ddir], 1, 100); |
211c9b89 | 360 | |
d9472271 | 361 | return v <= td->o.perc_rand[ddir]; |
211c9b89 JA |
362 | } |
363 | ||
1ae83d45 JA |
364 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
365 | enum fio_ddir ddir, uint64_t *b) | |
366 | { | |
367 | struct rand_off *r; | |
368 | int i, ret = 1; | |
369 | ||
bcd5abfa | 370 | if (!should_sort_io(td)) |
1ae83d45 JA |
371 | return get_off_from_method(td, f, ddir, b); |
372 | ||
373 | if (!flist_empty(&td->next_rand_list)) { | |
1ae83d45 | 374 | fetch: |
9342d5f8 | 375 | r = flist_first_entry(&td->next_rand_list, struct rand_off, list); |
1ae83d45 JA |
376 | flist_del(&r->list); |
377 | *b = r->off; | |
378 | free(r); | |
379 | return 0; | |
380 | } | |
381 | ||
382 | for (i = 0; i < td->o.verifysort_nr; i++) { | |
383 | r = malloc(sizeof(*r)); | |
384 | ||
385 | ret = get_off_from_method(td, f, ddir, &r->off); | |
386 | if (ret) { | |
387 | free(r); | |
388 | break; | |
389 | } | |
390 | ||
391 | flist_add(&r->list, &td->next_rand_list); | |
392 | } | |
393 | ||
394 | if (ret && !i) | |
395 | return ret; | |
396 | ||
397 | assert(!flist_empty(&td->next_rand_list)); | |
398 | flist_sort(NULL, &td->next_rand_list, flist_cmp); | |
399 | goto fetch; | |
400 | } | |
401 | ||
0bcf41cd JA |
402 | static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f) |
403 | { | |
404 | struct thread_options *o = &td->o; | |
405 | ||
406 | if (o->invalidate_cache && !o->odirect) { | |
407 | int fio_unused ret; | |
408 | ||
409 | ret = file_invalidate_cache(td, f); | |
410 | } | |
411 | } | |
412 | ||
38dad62d | 413 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 414 | enum fio_ddir ddir, uint64_t *b) |
38dad62d | 415 | { |
c04e4661 DE |
416 | if (!get_next_rand_offset(td, f, ddir, b)) |
417 | return 0; | |
418 | ||
8c07860d JA |
419 | if (td->o.time_based || |
420 | (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) { | |
33c48814 | 421 | fio_file_reset(td, f); |
c04e4661 DE |
422 | if (!get_next_rand_offset(td, f, ddir, b)) |
423 | return 0; | |
0bcf41cd | 424 | loop_cache_invalidate(td, f); |
38dad62d JA |
425 | } |
426 | ||
c04e4661 | 427 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", |
f1dfb668 | 428 | f->file_name, (unsigned long long) f->last_pos[ddir], |
4b91ee8f | 429 | (unsigned long long) f->real_file_size); |
c04e4661 | 430 | return 1; |
38dad62d JA |
431 | } |
432 | ||
37cf9e3c | 433 | static int get_next_seq_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 434 | enum fio_ddir ddir, uint64_t *offset) |
38dad62d | 435 | { |
ac002339 JA |
436 | struct thread_options *o = &td->o; |
437 | ||
ff58fced JA |
438 | assert(ddir_rw(ddir)); |
439 | ||
17373ce2 JA |
440 | /* |
441 | * If we reach the end for a time based run, reset us back to 0 | |
442 | * and invalidate the cache, if we need to. | |
443 | */ | |
f1dfb668 | 444 | if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) && |
19ddc35b | 445 | o->time_based) { |
c89daa4a | 446 | f->last_pos[ddir] = f->file_offset; |
0bcf41cd | 447 | loop_cache_invalidate(td, f); |
19ddc35b | 448 | } |
c04e4661 | 449 | |
f1dfb668 | 450 | if (f->last_pos[ddir] < f->real_file_size) { |
1ae83d45 | 451 | uint64_t pos; |
059b0802 | 452 | |
c22825bb JA |
453 | if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) { |
454 | if (f->real_file_size > f->io_size) | |
455 | f->last_pos[ddir] = f->io_size; | |
456 | else | |
457 | f->last_pos[ddir] = f->real_file_size; | |
458 | } | |
a66da7a2 | 459 | |
f1dfb668 | 460 | pos = f->last_pos[ddir] - f->file_offset; |
ac002339 JA |
461 | if (pos && o->ddir_seq_add) { |
462 | pos += o->ddir_seq_add; | |
463 | ||
464 | /* | |
465 | * If we reach beyond the end of the file | |
466 | * with holed IO, wrap around to the | |
b0a84f48 JA |
467 | * beginning again. If we're doing backwards IO, |
468 | * wrap to the end. | |
ac002339 | 469 | */ |
b0a84f48 JA |
470 | if (pos >= f->real_file_size) { |
471 | if (o->ddir_seq_add > 0) | |
472 | pos = f->file_offset; | |
c22825bb JA |
473 | else { |
474 | if (f->real_file_size > f->io_size) | |
475 | pos = f->io_size; | |
476 | else | |
477 | pos = f->real_file_size; | |
478 | ||
479 | pos += o->ddir_seq_add; | |
480 | } | |
b0a84f48 | 481 | } |
ac002339 | 482 | } |
059b0802 | 483 | |
37cf9e3c | 484 | *offset = pos; |
38dad62d JA |
485 | return 0; |
486 | } | |
487 | ||
488 | return 1; | |
489 | } | |
490 | ||
491 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
6aca9b3d JA |
492 | enum fio_ddir ddir, int rw_seq, |
493 | unsigned int *is_random) | |
38dad62d JA |
494 | { |
495 | struct fio_file *f = io_u->file; | |
1ae83d45 | 496 | uint64_t b, offset; |
38dad62d JA |
497 | int ret; |
498 | ||
ff58fced JA |
499 | assert(ddir_rw(ddir)); |
500 | ||
37cf9e3c JA |
501 | b = offset = -1ULL; |
502 | ||
38dad62d | 503 | if (rw_seq) { |
211c9b89 | 504 | if (td_random(td)) { |
6aca9b3d | 505 | if (should_do_random(td, ddir)) { |
211c9b89 | 506 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
507 | *is_random = 1; |
508 | } else { | |
509 | *is_random = 0; | |
1651e431 | 510 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
211c9b89 JA |
511 | ret = get_next_seq_offset(td, f, ddir, &offset); |
512 | if (ret) | |
513 | ret = get_next_rand_block(td, f, ddir, &b); | |
514 | } | |
6aca9b3d JA |
515 | } else { |
516 | *is_random = 0; | |
37cf9e3c | 517 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 518 | } |
38dad62d | 519 | } else { |
1651e431 | 520 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
6aca9b3d | 521 | *is_random = 0; |
38dad62d JA |
522 | |
523 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 524 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 525 | if (ret) { |
37cf9e3c | 526 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
527 | *is_random = 0; |
528 | } | |
38dad62d | 529 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
f1dfb668 JA |
530 | if (f->last_start[ddir] != -1ULL) |
531 | offset = f->last_start[ddir] - f->file_offset; | |
38dad62d | 532 | else |
37cf9e3c | 533 | offset = 0; |
38dad62d JA |
534 | ret = 0; |
535 | } else { | |
536 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
537 | ret = 1; | |
538 | } | |
539 | } | |
6d68b997 | 540 | |
37cf9e3c JA |
541 | if (!ret) { |
542 | if (offset != -1ULL) | |
543 | io_u->offset = offset; | |
544 | else if (b != -1ULL) | |
545 | io_u->offset = b * td->o.ba[ddir]; | |
546 | else { | |
4e0a8fa2 | 547 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
548 | ret = 1; |
549 | } | |
550 | } | |
551 | ||
38dad62d JA |
552 | return ret; |
553 | } | |
554 | ||
10ba535a JA |
555 | /* |
556 | * For random io, generate a random new block and see if it's used. Repeat | |
557 | * until we find a free one. For sequential io, just return the end of | |
558 | * the last io issued. | |
559 | */ | |
6aca9b3d JA |
560 | static int __get_next_offset(struct thread_data *td, struct io_u *io_u, |
561 | unsigned int *is_random) | |
10ba535a | 562 | { |
9bf2061e | 563 | struct fio_file *f = io_u->file; |
4ba66134 | 564 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 565 | int rw_seq_hit = 0; |
10ba535a | 566 | |
ff58fced JA |
567 | assert(ddir_rw(ddir)); |
568 | ||
38dad62d JA |
569 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
570 | rw_seq_hit = 1; | |
5736c10d | 571 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 572 | } |
211097b2 | 573 | |
6aca9b3d | 574 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 575 | return 1; |
10ba535a | 576 | |
009bd847 JA |
577 | if (io_u->offset >= f->io_size) { |
578 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
579 | (unsigned long long) io_u->offset, |
580 | (unsigned long long) f->io_size); | |
009bd847 JA |
581 | return 1; |
582 | } | |
583 | ||
584 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
585 | if (io_u->offset >= f->real_file_size) { |
586 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
587 | (unsigned long long) io_u->offset, |
588 | (unsigned long long) f->real_file_size); | |
10ba535a | 589 | return 1; |
2ba1c290 | 590 | } |
10ba535a JA |
591 | |
592 | return 0; | |
593 | } | |
594 | ||
6aca9b3d JA |
595 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
596 | unsigned int *is_random) | |
15dc1934 | 597 | { |
d72be545 JA |
598 | if (td->flags & TD_F_PROFILE_OPS) { |
599 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 600 | |
d72be545 | 601 | if (ops->fill_io_u_off) |
6aca9b3d | 602 | return ops->fill_io_u_off(td, io_u, is_random); |
d72be545 | 603 | } |
15dc1934 | 604 | |
6aca9b3d | 605 | return __get_next_offset(td, io_u, is_random); |
15dc1934 JA |
606 | } |
607 | ||
e39c0676 JA |
608 | static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u, |
609 | unsigned int buflen) | |
79944128 JA |
610 | { |
611 | struct fio_file *f = io_u->file; | |
612 | ||
bedc9dc2 | 613 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
614 | } |
615 | ||
6aca9b3d JA |
616 | static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u, |
617 | unsigned int is_random) | |
10ba535a | 618 | { |
6aca9b3d | 619 | int ddir = io_u->ddir; |
24d23ca7 | 620 | unsigned int buflen = 0; |
f3059de1 | 621 | unsigned int minbs, maxbs; |
3dd29f7c | 622 | uint64_t frand_max, r; |
7c961359 | 623 | bool power_2; |
10ba535a | 624 | |
9ee1c647 | 625 | assert(ddir_rw(ddir)); |
6aca9b3d JA |
626 | |
627 | if (td->o.bs_is_seq_rand) | |
628 | ddir = is_random ? DDIR_WRITE: DDIR_READ; | |
ff58fced | 629 | |
f3059de1 JA |
630 | minbs = td->o.min_bs[ddir]; |
631 | maxbs = td->o.max_bs[ddir]; | |
632 | ||
79944128 JA |
633 | if (minbs == maxbs) |
634 | return minbs; | |
635 | ||
52c58027 JA |
636 | /* |
637 | * If we can't satisfy the min block size from here, then fail | |
638 | */ | |
639 | if (!io_u_fits(td, io_u, minbs)) | |
640 | return 0; | |
641 | ||
2f282cec | 642 | frand_max = rand_max(&td->bsrange_state[ddir]); |
79944128 | 643 | do { |
2f282cec | 644 | r = __rand(&td->bsrange_state[ddir]); |
4c07ad86 | 645 | |
720e84ad | 646 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 | 647 | buflen = 1 + (unsigned int) ((double) maxbs * |
c3546b53 | 648 | (r / (frand_max + 1.0))); |
f3059de1 JA |
649 | if (buflen < minbs) |
650 | buflen = minbs; | |
5ec10eaa | 651 | } else { |
3dd29f7c | 652 | long long perc = 0; |
564ca972 JA |
653 | unsigned int i; |
654 | ||
720e84ad JA |
655 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
656 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
657 | |
658 | buflen = bsp->bs; | |
659 | perc += bsp->perc; | |
3dd29f7c JA |
660 | if (!perc) |
661 | break; | |
662 | if ((r / perc <= frand_max / 100ULL) && | |
79944128 | 663 | io_u_fits(td, io_u, buflen)) |
564ca972 JA |
664 | break; |
665 | } | |
666 | } | |
79944128 | 667 | |
17a6b702 PL |
668 | power_2 = is_power_of_2(minbs); |
669 | if (!td->o.bs_unaligned && power_2) | |
7c306bb1 | 670 | buflen &= ~(minbs - 1); |
17a6b702 PL |
671 | else if (!td->o.bs_unaligned && !power_2) |
672 | buflen -= buflen % minbs; | |
79944128 | 673 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 674 | |
10ba535a JA |
675 | return buflen; |
676 | } | |
677 | ||
6aca9b3d JA |
678 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u, |
679 | unsigned int is_random) | |
15dc1934 | 680 | { |
d72be545 JA |
681 | if (td->flags & TD_F_PROFILE_OPS) { |
682 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 683 | |
d72be545 | 684 | if (ops->fill_io_u_size) |
6aca9b3d | 685 | return ops->fill_io_u_size(td, io_u, is_random); |
d72be545 | 686 | } |
15dc1934 | 687 | |
6aca9b3d | 688 | return __get_next_buflen(td, io_u, is_random); |
15dc1934 JA |
689 | } |
690 | ||
afe24a5a JA |
691 | static void set_rwmix_bytes(struct thread_data *td) |
692 | { | |
afe24a5a JA |
693 | unsigned int diff; |
694 | ||
695 | /* | |
696 | * we do time or byte based switch. this is needed because | |
697 | * buffered writes may issue a lot quicker than they complete, | |
698 | * whereas reads do not. | |
699 | */ | |
e47f799f | 700 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 701 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
702 | } |
703 | ||
704 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
705 | { | |
706 | unsigned int v; | |
e47f799f | 707 | |
36dd3379 | 708 | v = rand32_between(&td->rwmix_state, 1, 100); |
4c07ad86 | 709 | |
04c540d9 | 710 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
711 | return DDIR_READ; |
712 | ||
713 | return DDIR_WRITE; | |
afe24a5a JA |
714 | } |
715 | ||
0d593542 | 716 | int io_u_quiesce(struct thread_data *td) |
002e7183 | 717 | { |
0d593542 JA |
718 | int completed = 0; |
719 | ||
002e7183 JA |
720 | /* |
721 | * We are going to sleep, ensure that we flush anything pending as | |
722 | * not to skew our latency numbers. | |
723 | * | |
724 | * Changed to only monitor 'in flight' requests here instead of the | |
725 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
726 | * io's that have been actually submitted to an async engine, | |
727 | * and cur_depth is meaningless for sync engines. | |
728 | */ | |
9cc80b6d JA |
729 | if (td->io_u_queued || td->cur_depth) { |
730 | int fio_unused ret; | |
731 | ||
732 | ret = td_io_commit(td); | |
733 | } | |
734 | ||
002e7183 | 735 | while (td->io_u_in_flight) { |
89fd0573 | 736 | int ret; |
002e7183 | 737 | |
55312f9f | 738 | ret = io_u_queued_complete(td, 1); |
0d593542 JA |
739 | if (ret > 0) |
740 | completed += ret; | |
002e7183 | 741 | } |
0d593542 | 742 | |
6be06c46 JA |
743 | if (td->flags & TD_F_REGROW_LOGS) |
744 | regrow_logs(td); | |
745 | ||
0d593542 | 746 | return completed; |
002e7183 JA |
747 | } |
748 | ||
581e7141 JA |
749 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
750 | { | |
751 | enum fio_ddir odir = ddir ^ 1; | |
b5407f8b | 752 | uint64_t usec; |
90eff1c9 | 753 | uint64_t now; |
581e7141 | 754 | |
ff58fced | 755 | assert(ddir_rw(ddir)); |
50a8ce86 | 756 | now = utime_since_now(&td->start); |
ff58fced | 757 | |
50a8ce86 D |
758 | /* |
759 | * if rate_next_io_time is in the past, need to catch up to rate | |
760 | */ | |
761 | if (td->rate_next_io_time[ddir] <= now) | |
581e7141 JA |
762 | return ddir; |
763 | ||
764 | /* | |
50a8ce86 | 765 | * We are ahead of rate in this direction. See if we |
581e7141 JA |
766 | * should switch. |
767 | */ | |
315fcfec | 768 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 | 769 | /* |
50a8ce86 | 770 | * Other direction is behind rate, switch |
581e7141 | 771 | */ |
50a8ce86 | 772 | if (td->rate_next_io_time[odir] <= now) |
581e7141 JA |
773 | return odir; |
774 | ||
775 | /* | |
50a8ce86 D |
776 | * Both directions are ahead of rate. sleep the min |
777 | * switch if necissary | |
581e7141 | 778 | */ |
50a8ce86 D |
779 | if (td->rate_next_io_time[ddir] <= |
780 | td->rate_next_io_time[odir]) { | |
781 | usec = td->rate_next_io_time[ddir] - now; | |
581e7141 | 782 | } else { |
50a8ce86 | 783 | usec = td->rate_next_io_time[odir] - now; |
581e7141 JA |
784 | ddir = odir; |
785 | } | |
786 | } else | |
50a8ce86 | 787 | usec = td->rate_next_io_time[ddir] - now; |
581e7141 | 788 | |
a9da8ab2 JA |
789 | if (td->o.io_submit_mode == IO_MODE_INLINE) |
790 | io_u_quiesce(td); | |
78c1eda5 | 791 | |
9cc80b6d | 792 | usec = usec_sleep(td, usec); |
581e7141 | 793 | |
581e7141 JA |
794 | return ddir; |
795 | } | |
796 | ||
10ba535a JA |
797 | /* |
798 | * Return the data direction for the next io_u. If the job is a | |
799 | * mixed read/write workload, check the rwmix cycle and switch if | |
800 | * necessary. | |
801 | */ | |
1e97cce9 | 802 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 803 | { |
581e7141 JA |
804 | enum fio_ddir ddir; |
805 | ||
5f9099ea | 806 | /* |
f6860972 TK |
807 | * See if it's time to fsync/fdatasync/sync_file_range first, |
808 | * and if not then move on to check regular I/Os. | |
5f9099ea | 809 | */ |
f6860972 TK |
810 | if (should_fsync(td)) { |
811 | if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] && | |
812 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks)) | |
813 | return DDIR_SYNC; | |
814 | ||
815 | if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] && | |
816 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks)) | |
817 | return DDIR_DATASYNC; | |
818 | ||
819 | if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] && | |
820 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr)) | |
821 | return DDIR_SYNC_FILE_RANGE; | |
822 | } | |
44f29692 | 823 | |
10ba535a | 824 | if (td_rw(td)) { |
10ba535a JA |
825 | /* |
826 | * Check if it's time to seed a new data direction. | |
827 | */ | |
e4928662 | 828 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
829 | /* |
830 | * Put a top limit on how many bytes we do for | |
831 | * one data direction, to avoid overflowing the | |
832 | * ranges too much | |
833 | */ | |
834 | ddir = get_rand_ddir(td); | |
e47f799f JA |
835 | |
836 | if (ddir != td->rwmix_ddir) | |
837 | set_rwmix_bytes(td); | |
838 | ||
839 | td->rwmix_ddir = ddir; | |
10ba535a | 840 | } |
581e7141 | 841 | ddir = td->rwmix_ddir; |
10ba535a | 842 | } else if (td_read(td)) |
581e7141 | 843 | ddir = DDIR_READ; |
6eaf09d6 | 844 | else if (td_write(td)) |
581e7141 | 845 | ddir = DDIR_WRITE; |
5c8e84ca | 846 | else if (td_trim(td)) |
6eaf09d6 | 847 | ddir = DDIR_TRIM; |
5c8e84ca TK |
848 | else |
849 | ddir = DDIR_INVAL; | |
581e7141 JA |
850 | |
851 | td->rwmix_ddir = rate_ddir(td, ddir); | |
852 | return td->rwmix_ddir; | |
10ba535a JA |
853 | } |
854 | ||
1ef2b6be JA |
855 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
856 | { | |
0e4dd95c DE |
857 | enum fio_ddir ddir = get_rw_ddir(td); |
858 | ||
82a90686 | 859 | if (td_trimwrite(td)) { |
0e4dd95c DE |
860 | struct fio_file *f = io_u->file; |
861 | if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM]) | |
862 | ddir = DDIR_TRIM; | |
863 | else | |
864 | ddir = DDIR_WRITE; | |
865 | } | |
866 | ||
867 | io_u->ddir = io_u->acct_ddir = ddir; | |
1ef2b6be | 868 | |
9b87f09b | 869 | if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) && |
1ef2b6be JA |
870 | td->o.barrier_blocks && |
871 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
872 | td->io_issues[DDIR_WRITE]) | |
1651e431 | 873 | io_u_set(td, io_u, IO_U_F_BARRIER); |
1ef2b6be JA |
874 | } |
875 | ||
e8462bd8 | 876 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 877 | { |
71b84caa | 878 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
879 | |
880 | if (ret) | |
881 | td_verror(td, ret, "file close"); | |
882 | } | |
883 | ||
10ba535a JA |
884 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
885 | { | |
a9da8ab2 JA |
886 | if (td->parent) |
887 | td = td->parent; | |
888 | ||
e8462bd8 JA |
889 | td_io_u_lock(td); |
890 | ||
f8b0bd10 | 891 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 892 | put_file_log(td, io_u->file); |
f8b0bd10 | 893 | |
10ba535a | 894 | io_u->file = NULL; |
1651e431 | 895 | io_u_set(td, io_u, IO_U_F_FREE); |
d7ee2a7d | 896 | |
a9da8ab2 | 897 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 898 | td->cur_depth--; |
a9da8ab2 JA |
899 | assert(!(td->flags & TD_F_CHILD)); |
900 | } | |
2ae0b204 | 901 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 JA |
902 | td_io_u_unlock(td); |
903 | td_io_u_free_notify(td); | |
10ba535a JA |
904 | } |
905 | ||
f2bba182 RR |
906 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
907 | { | |
1651e431 | 908 | io_u_clear(td, io_u, IO_U_F_FLIGHT); |
f2bba182 RR |
909 | put_io_u(td, io_u); |
910 | } | |
911 | ||
755200a3 JA |
912 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
913 | { | |
914 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 915 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 916 | |
465221b0 JA |
917 | dprint(FD_IO, "requeue %p\n", __io_u); |
918 | ||
a9da8ab2 JA |
919 | if (td->parent) |
920 | td = td->parent; | |
921 | ||
e8462bd8 JA |
922 | td_io_u_lock(td); |
923 | ||
1651e431 | 924 | io_u_set(td, __io_u, IO_U_F_FREE); |
bcd5abfa JA |
925 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
926 | td->io_issues[ddir]--; | |
5ec10eaa | 927 | |
1651e431 | 928 | io_u_clear(td, __io_u, IO_U_F_FLIGHT); |
a9da8ab2 | 929 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 930 | td->cur_depth--; |
a9da8ab2 JA |
931 | assert(!(td->flags & TD_F_CHILD)); |
932 | } | |
2ae0b204 JA |
933 | |
934 | io_u_rpush(&td->io_u_requeues, __io_u); | |
e8462bd8 | 935 | td_io_u_unlock(td); |
a9da8ab2 | 936 | td_io_u_free_notify(td); |
755200a3 JA |
937 | *io_u = NULL; |
938 | } | |
939 | ||
9bf2061e | 940 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 941 | { |
6aca9b3d JA |
942 | unsigned int is_random; |
943 | ||
9b87f09b | 944 | if (td_ioengine_flagged(td, FIO_NOIO)) |
b4c5e1ac JA |
945 | goto out; |
946 | ||
1ef2b6be | 947 | set_rw_ddir(td, io_u); |
5f9099ea | 948 | |
87dc1ab1 | 949 | /* |
ff58fced | 950 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 951 | */ |
ff58fced | 952 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 953 | goto out; |
a00735e6 | 954 | |
48f5abd3 JA |
955 | /* |
956 | * See if it's time to switch to a new zone | |
957 | */ | |
13af05ae | 958 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { |
4c8be5b1 JA |
959 | struct fio_file *f = io_u->file; |
960 | ||
48f5abd3 | 961 | td->zone_bytes = 0; |
4c8be5b1 JA |
962 | f->file_offset += td->o.zone_range + td->o.zone_skip; |
963 | ||
964 | /* | |
965 | * Wrap from the beginning, if we exceed the file size | |
966 | */ | |
967 | if (f->file_offset >= f->real_file_size) | |
968 | f->file_offset = f->real_file_size - f->file_offset; | |
f1dfb668 | 969 | f->last_pos[io_u->ddir] = f->file_offset; |
48f5abd3 JA |
970 | td->io_skip_bytes += td->o.zone_skip; |
971 | } | |
972 | ||
10ba535a | 973 | /* |
c685b5b2 JA |
974 | * No log, let the seq/rand engine retrieve the next buflen and |
975 | * position. | |
10ba535a | 976 | */ |
6aca9b3d | 977 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 978 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 979 | return 1; |
2ba1c290 | 980 | } |
10ba535a | 981 | |
6aca9b3d | 982 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
983 | if (!io_u->buflen) { |
984 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 985 | return 1; |
2ba1c290 | 986 | } |
bca4ed4d | 987 | |
2ba1c290 | 988 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
48366f3a TK |
989 | dprint(FD_IO, "io_u %p, offset + buflen exceeds file size\n", |
990 | io_u); | |
991 | dprint(FD_IO, " offset=%llu/buflen=%lu > %llu\n", | |
4b91ee8f JA |
992 | (unsigned long long) io_u->offset, io_u->buflen, |
993 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 994 | return 1; |
2ba1c290 | 995 | } |
6a5e6884 | 996 | |
bca4ed4d JA |
997 | /* |
998 | * mark entry before potentially trimming io_u | |
999 | */ | |
303032ae | 1000 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 1001 | mark_random_map(td, io_u); |
bca4ed4d | 1002 | |
c38e9468 | 1003 | out: |
2ba1c290 | 1004 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 1005 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 1006 | return 0; |
10ba535a JA |
1007 | } |
1008 | ||
838bc709 JA |
1009 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
1010 | { | |
2b13e716 | 1011 | int idx = 0; |
838bc709 JA |
1012 | |
1013 | switch (nr) { | |
1014 | default: | |
2b13e716 | 1015 | idx = 6; |
838bc709 JA |
1016 | break; |
1017 | case 33 ... 64: | |
2b13e716 | 1018 | idx = 5; |
838bc709 JA |
1019 | break; |
1020 | case 17 ... 32: | |
2b13e716 | 1021 | idx = 4; |
838bc709 JA |
1022 | break; |
1023 | case 9 ... 16: | |
2b13e716 | 1024 | idx = 3; |
838bc709 JA |
1025 | break; |
1026 | case 5 ... 8: | |
2b13e716 | 1027 | idx = 2; |
838bc709 JA |
1028 | break; |
1029 | case 1 ... 4: | |
2b13e716 | 1030 | idx = 1; |
838bc709 JA |
1031 | case 0: |
1032 | break; | |
1033 | } | |
1034 | ||
2b13e716 | 1035 | map[idx]++; |
838bc709 JA |
1036 | } |
1037 | ||
1038 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
1039 | { | |
1040 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
1041 | td->ts.total_submit++; | |
1042 | } | |
1043 | ||
1044 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
1045 | { | |
1046 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
1047 | td->ts.total_complete++; | |
1048 | } | |
1049 | ||
d8005759 | 1050 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 1051 | { |
2b13e716 | 1052 | int idx = 0; |
71619dc2 JA |
1053 | |
1054 | switch (td->cur_depth) { | |
1055 | default: | |
2b13e716 | 1056 | idx = 6; |
a783e61a | 1057 | break; |
71619dc2 | 1058 | case 32 ... 63: |
2b13e716 | 1059 | idx = 5; |
a783e61a | 1060 | break; |
71619dc2 | 1061 | case 16 ... 31: |
2b13e716 | 1062 | idx = 4; |
a783e61a | 1063 | break; |
71619dc2 | 1064 | case 8 ... 15: |
2b13e716 | 1065 | idx = 3; |
a783e61a | 1066 | break; |
71619dc2 | 1067 | case 4 ... 7: |
2b13e716 | 1068 | idx = 2; |
a783e61a | 1069 | break; |
71619dc2 | 1070 | case 2 ... 3: |
2b13e716 | 1071 | idx = 1; |
71619dc2 JA |
1072 | case 1: |
1073 | break; | |
1074 | } | |
1075 | ||
2b13e716 | 1076 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
1077 | } |
1078 | ||
d6bb626e | 1079 | static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec) |
04a0feae | 1080 | { |
2b13e716 | 1081 | int idx = 0; |
04a0feae | 1082 | |
d6bb626e VF |
1083 | assert(nsec < 1000); |
1084 | ||
1085 | switch (nsec) { | |
1086 | case 750 ... 999: | |
1087 | idx = 9; | |
1088 | break; | |
1089 | case 500 ... 749: | |
1090 | idx = 8; | |
1091 | break; | |
1092 | case 250 ... 499: | |
1093 | idx = 7; | |
1094 | break; | |
1095 | case 100 ... 249: | |
1096 | idx = 6; | |
1097 | break; | |
1098 | case 50 ... 99: | |
1099 | idx = 5; | |
1100 | break; | |
1101 | case 20 ... 49: | |
1102 | idx = 4; | |
1103 | break; | |
1104 | case 10 ... 19: | |
1105 | idx = 3; | |
1106 | break; | |
1107 | case 4 ... 9: | |
1108 | idx = 2; | |
1109 | break; | |
1110 | case 2 ... 3: | |
1111 | idx = 1; | |
1112 | case 0 ... 1: | |
1113 | break; | |
1114 | } | |
1115 | ||
1116 | assert(idx < FIO_IO_U_LAT_N_NR); | |
1117 | td->ts.io_u_lat_n[idx]++; | |
1118 | } | |
1119 | ||
1120 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec) | |
1121 | { | |
1122 | int idx = 0; | |
1123 | ||
1124 | assert(usec < 1000 && usec >= 1); | |
04a0feae JA |
1125 | |
1126 | switch (usec) { | |
1127 | case 750 ... 999: | |
2b13e716 | 1128 | idx = 9; |
04a0feae JA |
1129 | break; |
1130 | case 500 ... 749: | |
2b13e716 | 1131 | idx = 8; |
04a0feae JA |
1132 | break; |
1133 | case 250 ... 499: | |
2b13e716 | 1134 | idx = 7; |
04a0feae JA |
1135 | break; |
1136 | case 100 ... 249: | |
2b13e716 | 1137 | idx = 6; |
04a0feae JA |
1138 | break; |
1139 | case 50 ... 99: | |
2b13e716 | 1140 | idx = 5; |
04a0feae JA |
1141 | break; |
1142 | case 20 ... 49: | |
2b13e716 | 1143 | idx = 4; |
04a0feae JA |
1144 | break; |
1145 | case 10 ... 19: | |
2b13e716 | 1146 | idx = 3; |
04a0feae JA |
1147 | break; |
1148 | case 4 ... 9: | |
2b13e716 | 1149 | idx = 2; |
04a0feae JA |
1150 | break; |
1151 | case 2 ... 3: | |
2b13e716 | 1152 | idx = 1; |
04a0feae JA |
1153 | case 0 ... 1: |
1154 | break; | |
1155 | } | |
1156 | ||
2b13e716 JA |
1157 | assert(idx < FIO_IO_U_LAT_U_NR); |
1158 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
1159 | } |
1160 | ||
d6bb626e | 1161 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec) |
ec118304 | 1162 | { |
2b13e716 | 1163 | int idx = 0; |
ec118304 | 1164 | |
d6bb626e VF |
1165 | assert(msec >= 1); |
1166 | ||
ec118304 JA |
1167 | switch (msec) { |
1168 | default: | |
2b13e716 | 1169 | idx = 11; |
04a0feae | 1170 | break; |
8abdce66 | 1171 | case 1000 ... 1999: |
2b13e716 | 1172 | idx = 10; |
04a0feae | 1173 | break; |
8abdce66 | 1174 | case 750 ... 999: |
2b13e716 | 1175 | idx = 9; |
04a0feae | 1176 | break; |
8abdce66 | 1177 | case 500 ... 749: |
2b13e716 | 1178 | idx = 8; |
04a0feae | 1179 | break; |
8abdce66 | 1180 | case 250 ... 499: |
2b13e716 | 1181 | idx = 7; |
04a0feae | 1182 | break; |
8abdce66 | 1183 | case 100 ... 249: |
2b13e716 | 1184 | idx = 6; |
04a0feae | 1185 | break; |
8abdce66 | 1186 | case 50 ... 99: |
2b13e716 | 1187 | idx = 5; |
04a0feae | 1188 | break; |
8abdce66 | 1189 | case 20 ... 49: |
2b13e716 | 1190 | idx = 4; |
04a0feae | 1191 | break; |
8abdce66 | 1192 | case 10 ... 19: |
2b13e716 | 1193 | idx = 3; |
04a0feae | 1194 | break; |
8abdce66 | 1195 | case 4 ... 9: |
2b13e716 | 1196 | idx = 2; |
04a0feae | 1197 | break; |
ec118304 | 1198 | case 2 ... 3: |
2b13e716 | 1199 | idx = 1; |
ec118304 JA |
1200 | case 0 ... 1: |
1201 | break; | |
1202 | } | |
1203 | ||
2b13e716 JA |
1204 | assert(idx < FIO_IO_U_LAT_M_NR); |
1205 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
1206 | } |
1207 | ||
d6bb626e | 1208 | static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec) |
04a0feae | 1209 | { |
d6bb626e VF |
1210 | if (nsec < 1000) |
1211 | io_u_mark_lat_nsec(td, nsec); | |
1212 | else if (nsec < 1000000) | |
1213 | io_u_mark_lat_usec(td, nsec / 1000); | |
04a0feae | 1214 | else |
d6bb626e | 1215 | io_u_mark_lat_msec(td, nsec / 1000000); |
ec118304 JA |
1216 | } |
1217 | ||
8c07860d JA |
1218 | static unsigned int __get_next_fileno_rand(struct thread_data *td) |
1219 | { | |
1220 | unsigned long fileno; | |
1221 | ||
1222 | if (td->o.file_service_type == FIO_FSERVICE_RANDOM) { | |
1223 | uint64_t frand_max = rand_max(&td->next_file_state); | |
1224 | unsigned long r; | |
1225 | ||
1226 | r = __rand(&td->next_file_state); | |
1227 | return (unsigned int) ((double) td->o.nr_files | |
1228 | * (r / (frand_max + 1.0))); | |
1229 | } | |
1230 | ||
1231 | if (td->o.file_service_type == FIO_FSERVICE_ZIPF) | |
1232 | fileno = zipf_next(&td->next_file_zipf); | |
1233 | else if (td->o.file_service_type == FIO_FSERVICE_PARETO) | |
1234 | fileno = pareto_next(&td->next_file_zipf); | |
1235 | else if (td->o.file_service_type == FIO_FSERVICE_GAUSS) | |
1236 | fileno = gauss_next(&td->next_file_gauss); | |
1237 | else { | |
1238 | log_err("fio: bad file service type: %d\n", td->o.file_service_type); | |
1239 | assert(0); | |
1240 | return 0; | |
1241 | } | |
1242 | ||
1243 | return fileno >> FIO_FSERVICE_SHIFT; | |
1244 | } | |
1245 | ||
0aabe160 JA |
1246 | /* |
1247 | * Get next file to service by choosing one at random | |
1248 | */ | |
2cc52930 JA |
1249 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
1250 | enum fio_file_flags goodf, | |
d6aed795 | 1251 | enum fio_file_flags badf) |
0aabe160 | 1252 | { |
0aabe160 | 1253 | struct fio_file *f; |
1c178180 | 1254 | int fno; |
0aabe160 JA |
1255 | |
1256 | do { | |
87b10676 | 1257 | int opened = 0; |
4c07ad86 | 1258 | |
8c07860d | 1259 | fno = __get_next_fileno_rand(td); |
7c83c089 | 1260 | |
126d65c6 | 1261 | f = td->files[fno]; |
d6aed795 | 1262 | if (fio_file_done(f)) |
059e63c0 | 1263 | continue; |
1c178180 | 1264 | |
d6aed795 | 1265 | if (!fio_file_open(f)) { |
87b10676 JA |
1266 | int err; |
1267 | ||
002fe734 JA |
1268 | if (td->nr_open_files >= td->o.open_files) |
1269 | return ERR_PTR(-EBUSY); | |
1270 | ||
87b10676 JA |
1271 | err = td_io_open_file(td, f); |
1272 | if (err) | |
1273 | continue; | |
1274 | opened = 1; | |
1275 | } | |
1276 | ||
2ba1c290 JA |
1277 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
1278 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 1279 | return f; |
2ba1c290 | 1280 | } |
87b10676 JA |
1281 | if (opened) |
1282 | td_io_close_file(td, f); | |
0aabe160 JA |
1283 | } while (1); |
1284 | } | |
1285 | ||
1286 | /* | |
1287 | * Get next file to service by doing round robin between all available ones | |
1288 | */ | |
1c178180 JA |
1289 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1290 | int badf) | |
3d7c391d JA |
1291 | { |
1292 | unsigned int old_next_file = td->next_file; | |
1293 | struct fio_file *f; | |
1294 | ||
1295 | do { | |
87b10676 JA |
1296 | int opened = 0; |
1297 | ||
126d65c6 | 1298 | f = td->files[td->next_file]; |
3d7c391d JA |
1299 | |
1300 | td->next_file++; | |
2dc1bbeb | 1301 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1302 | td->next_file = 0; |
1303 | ||
87b10676 | 1304 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1305 | if (fio_file_done(f)) { |
d5ed68ea | 1306 | f = NULL; |
059e63c0 | 1307 | continue; |
d5ed68ea | 1308 | } |
059e63c0 | 1309 | |
d6aed795 | 1310 | if (!fio_file_open(f)) { |
87b10676 JA |
1311 | int err; |
1312 | ||
002fe734 JA |
1313 | if (td->nr_open_files >= td->o.open_files) |
1314 | return ERR_PTR(-EBUSY); | |
1315 | ||
87b10676 | 1316 | err = td_io_open_file(td, f); |
b5696bfc JA |
1317 | if (err) { |
1318 | dprint(FD_FILE, "error %d on open of %s\n", | |
1319 | err, f->file_name); | |
87c27b45 | 1320 | f = NULL; |
87b10676 | 1321 | continue; |
b5696bfc | 1322 | } |
87b10676 JA |
1323 | opened = 1; |
1324 | } | |
1325 | ||
0b9d69ec JA |
1326 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1327 | f->flags); | |
1c178180 | 1328 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1329 | break; |
1330 | ||
87b10676 JA |
1331 | if (opened) |
1332 | td_io_close_file(td, f); | |
1333 | ||
3d7c391d JA |
1334 | f = NULL; |
1335 | } while (td->next_file != old_next_file); | |
1336 | ||
2ba1c290 | 1337 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1338 | return f; |
1339 | } | |
1340 | ||
7eb36574 | 1341 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1342 | { |
1907dbc6 JA |
1343 | struct fio_file *f; |
1344 | ||
2dc1bbeb | 1345 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1346 | |
b5696bfc | 1347 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1348 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1349 | " nr_files=%d\n", td->nr_open_files, | |
1350 | td->nr_done_files, | |
1351 | td->o.nr_files); | |
bdb4e2e9 | 1352 | return NULL; |
2ba1c290 | 1353 | } |
bdb4e2e9 | 1354 | |
1907dbc6 | 1355 | f = td->file_service_file; |
d6aed795 | 1356 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1357 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1358 | goto out; | |
1359 | if (td->file_service_left--) | |
1360 | goto out; | |
1361 | } | |
1907dbc6 | 1362 | |
a086c257 JA |
1363 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1364 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1365 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1366 | else |
d6aed795 | 1367 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1368 | |
002fe734 JA |
1369 | if (IS_ERR(f)) |
1370 | return f; | |
1371 | ||
1907dbc6 JA |
1372 | td->file_service_file = f; |
1373 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1374 | out: |
0dac421f JA |
1375 | if (f) |
1376 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1377 | else | |
1378 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1379 | return f; |
bdb4e2e9 JA |
1380 | } |
1381 | ||
7eb36574 JA |
1382 | static struct fio_file *get_next_file(struct thread_data *td) |
1383 | { | |
372d8962 | 1384 | if (td->flags & TD_F_PROFILE_OPS) { |
d72be545 | 1385 | struct prof_io_ops *ops = &td->prof_io_ops; |
7eb36574 | 1386 | |
d72be545 JA |
1387 | if (ops->get_next_file) |
1388 | return ops->get_next_file(td); | |
1389 | } | |
7eb36574 JA |
1390 | |
1391 | return __get_next_file(td); | |
1392 | } | |
1393 | ||
002fe734 | 1394 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1395 | { |
1396 | struct fio_file *f; | |
1397 | ||
1398 | do { | |
1399 | f = get_next_file(td); | |
002fe734 JA |
1400 | if (IS_ERR_OR_NULL(f)) |
1401 | return PTR_ERR(f); | |
429f6675 | 1402 | |
429f6675 JA |
1403 | io_u->file = f; |
1404 | get_file(f); | |
1405 | ||
1406 | if (!fill_io_u(td, io_u)) | |
1407 | break; | |
1408 | ||
b5696bfc | 1409 | put_file_log(td, f); |
429f6675 | 1410 | td_io_close_file(td, f); |
b5696bfc | 1411 | io_u->file = NULL; |
8c07860d JA |
1412 | if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM) |
1413 | fio_file_reset(td, f); | |
1414 | else { | |
1415 | fio_file_set_done(f); | |
1416 | td->nr_done_files++; | |
1417 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, | |
0b9d69ec | 1418 | td->nr_done_files, td->o.nr_files); |
8c07860d | 1419 | } |
429f6675 JA |
1420 | } while (1); |
1421 | ||
1422 | return 0; | |
1423 | } | |
1424 | ||
3e260a46 JA |
1425 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
1426 | unsigned long tusec, unsigned long max_usec) | |
1427 | { | |
1428 | if (!td->error) | |
1429 | log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec); | |
1430 | td_verror(td, ETIMEDOUT, "max latency exceeded"); | |
1431 | icd->error = ETIMEDOUT; | |
1432 | } | |
1433 | ||
1434 | static void lat_new_cycle(struct thread_data *td) | |
1435 | { | |
1436 | fio_gettime(&td->latency_ts, NULL); | |
1437 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1438 | td->latency_failed = 0; | |
1439 | } | |
1440 | ||
1441 | /* | |
1442 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1443 | * are at QD=1, then it's time to give up. | |
1444 | */ | |
e39c0676 | 1445 | static bool __lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1446 | { |
1447 | if (td->latency_qd == 1) | |
e39c0676 | 1448 | return true; |
3e260a46 JA |
1449 | |
1450 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1451 | |
1452 | if (td->latency_qd == td->latency_qd_low) | |
1453 | td->latency_qd_low--; | |
1454 | ||
3e260a46 JA |
1455 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1456 | ||
1457 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1458 | ||
1459 | /* | |
1460 | * When we ramp QD down, quiesce existing IO to prevent | |
1461 | * a storm of ramp downs due to pending higher depth. | |
1462 | */ | |
1463 | io_u_quiesce(td); | |
1464 | lat_new_cycle(td); | |
e39c0676 | 1465 | return false; |
3e260a46 JA |
1466 | } |
1467 | ||
e39c0676 | 1468 | static bool lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1469 | { |
1470 | if (td->o.latency_percentile.u.f == 100.0) | |
1471 | return __lat_target_failed(td); | |
1472 | ||
1473 | td->latency_failed++; | |
e39c0676 | 1474 | return false; |
3e260a46 JA |
1475 | } |
1476 | ||
1477 | void lat_target_init(struct thread_data *td) | |
1478 | { | |
6bb58215 JA |
1479 | td->latency_end_run = 0; |
1480 | ||
3e260a46 JA |
1481 | if (td->o.latency_target) { |
1482 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1483 | fio_gettime(&td->latency_ts, NULL); | |
1484 | td->latency_qd = 1; | |
1485 | td->latency_qd_high = td->o.iodepth; | |
1486 | td->latency_qd_low = 1; | |
1487 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1488 | } else | |
1489 | td->latency_qd = td->o.iodepth; | |
1490 | } | |
1491 | ||
6bb58215 JA |
1492 | void lat_target_reset(struct thread_data *td) |
1493 | { | |
1494 | if (!td->latency_end_run) | |
1495 | lat_target_init(td); | |
1496 | } | |
1497 | ||
3e260a46 JA |
1498 | static void lat_target_success(struct thread_data *td) |
1499 | { | |
1500 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1501 | struct thread_options *o = &td->o; |
3e260a46 JA |
1502 | |
1503 | td->latency_qd_low = td->latency_qd; | |
1504 | ||
1505 | /* | |
1506 | * If we haven't failed yet, we double up to a failing value instead | |
1507 | * of bisecting from highest possible queue depth. If we have set | |
1508 | * a limit other than td->o.iodepth, bisect between that. | |
1509 | */ | |
6bb58215 | 1510 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1511 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1512 | else | |
1513 | td->latency_qd *= 2; | |
1514 | ||
6bb58215 JA |
1515 | if (td->latency_qd > o->iodepth) |
1516 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1517 | |
1518 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1519 | |
3e260a46 | 1520 | /* |
6bb58215 JA |
1521 | * Same as last one, we are done. Let it run a latency cycle, so |
1522 | * we get only the results from the targeted depth. | |
3e260a46 | 1523 | */ |
6bb58215 JA |
1524 | if (td->latency_qd == qd) { |
1525 | if (td->latency_end_run) { | |
1526 | dprint(FD_RATE, "We are done\n"); | |
1527 | td->done = 1; | |
1528 | } else { | |
1529 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1530 | io_u_quiesce(td); | |
1531 | td->latency_end_run = 1; | |
1532 | reset_all_stats(td); | |
1533 | reset_io_stats(td); | |
1534 | } | |
1535 | } | |
3e260a46 JA |
1536 | |
1537 | lat_new_cycle(td); | |
1538 | } | |
1539 | ||
1540 | /* | |
1541 | * Check if we can bump the queue depth | |
1542 | */ | |
1543 | void lat_target_check(struct thread_data *td) | |
1544 | { | |
1545 | uint64_t usec_window; | |
1546 | uint64_t ios; | |
1547 | double success_ios; | |
1548 | ||
1549 | usec_window = utime_since_now(&td->latency_ts); | |
1550 | if (usec_window < td->o.latency_window) | |
1551 | return; | |
1552 | ||
1553 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1554 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1555 | success_ios *= 100.0; | |
1556 | ||
1557 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1558 | ||
1559 | if (success_ios >= td->o.latency_percentile.u.f) | |
1560 | lat_target_success(td); | |
1561 | else | |
1562 | __lat_target_failed(td); | |
1563 | } | |
1564 | ||
1565 | /* | |
1566 | * If latency target is enabled, we might be ramping up or down and not | |
1567 | * using the full queue depth available. | |
1568 | */ | |
e39c0676 | 1569 | bool queue_full(const struct thread_data *td) |
3e260a46 JA |
1570 | { |
1571 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1572 | ||
1573 | if (qempty) | |
e39c0676 | 1574 | return true; |
3e260a46 | 1575 | if (!td->o.latency_target) |
e39c0676 | 1576 | return false; |
3e260a46 JA |
1577 | |
1578 | return td->cur_depth >= td->latency_qd; | |
1579 | } | |
429f6675 | 1580 | |
10ba535a JA |
1581 | struct io_u *__get_io_u(struct thread_data *td) |
1582 | { | |
0cae66f6 | 1583 | struct io_u *io_u = NULL; |
10ba535a | 1584 | |
ca09be4b JA |
1585 | if (td->stop_io) |
1586 | return NULL; | |
1587 | ||
e8462bd8 JA |
1588 | td_io_u_lock(td); |
1589 | ||
1590 | again: | |
2ae0b204 JA |
1591 | if (!io_u_rempty(&td->io_u_requeues)) |
1592 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1593 | else if (!queue_full(td)) { |
2ae0b204 | 1594 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1595 | |
225ba9e3 | 1596 | io_u->file = NULL; |
6040dabc | 1597 | io_u->buflen = 0; |
10ba535a | 1598 | io_u->resid = 0; |
d7762cf8 | 1599 | io_u->end_io = NULL; |
755200a3 JA |
1600 | } |
1601 | ||
1602 | if (io_u) { | |
0c6e7517 | 1603 | assert(io_u->flags & IO_U_F_FREE); |
1651e431 | 1604 | io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
f8b0bd10 JA |
1605 | IO_U_F_TRIMMED | IO_U_F_BARRIER | |
1606 | IO_U_F_VER_LIST); | |
0c6e7517 | 1607 | |
755200a3 | 1608 | io_u->error = 0; |
bcd5abfa | 1609 | io_u->acct_ddir = -1; |
10ba535a | 1610 | td->cur_depth++; |
a9da8ab2 | 1611 | assert(!(td->flags & TD_F_CHILD)); |
1651e431 | 1612 | io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH); |
f9401285 | 1613 | io_u->ipo = NULL; |
a9da8ab2 | 1614 | } else if (td_async_processing(td)) { |
1dec3e07 JA |
1615 | /* |
1616 | * We ran out, wait for async verify threads to finish and | |
1617 | * return one | |
1618 | */ | |
a9da8ab2 JA |
1619 | assert(!(td->flags & TD_F_CHILD)); |
1620 | assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock)); | |
1dec3e07 | 1621 | goto again; |
10ba535a JA |
1622 | } |
1623 | ||
e8462bd8 | 1624 | td_io_u_unlock(td); |
10ba535a JA |
1625 | return io_u; |
1626 | } | |
1627 | ||
e39c0676 | 1628 | static bool check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1629 | { |
d72be545 | 1630 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
e39c0676 | 1631 | return false; |
d72be545 JA |
1632 | |
1633 | if (td->trim_entries) { | |
0d29de83 | 1634 | int get_trim = 0; |
10ba535a | 1635 | |
0d29de83 JA |
1636 | if (td->trim_batch) { |
1637 | td->trim_batch--; | |
1638 | get_trim = 1; | |
1639 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1640 | td->last_ddir != DDIR_READ) { | |
1641 | td->trim_batch = td->o.trim_batch; | |
1642 | if (!td->trim_batch) | |
1643 | td->trim_batch = td->o.trim_backlog; | |
1644 | get_trim = 1; | |
1645 | } | |
1646 | ||
917c64b3 | 1647 | if (get_trim && get_next_trim(td, io_u)) |
e39c0676 | 1648 | return true; |
2ba1c290 | 1649 | } |
10ba535a | 1650 | |
e39c0676 | 1651 | return false; |
0d29de83 JA |
1652 | } |
1653 | ||
e39c0676 | 1654 | static bool check_get_verify(struct thread_data *td, struct io_u *io_u) |
0d29de83 | 1655 | { |
d72be545 | 1656 | if (!(td->flags & TD_F_VER_BACKLOG)) |
e39c0676 | 1657 | return false; |
d72be545 JA |
1658 | |
1659 | if (td->io_hist_len) { | |
9e144189 JA |
1660 | int get_verify = 0; |
1661 | ||
d1ece0c7 | 1662 | if (td->verify_batch) |
9e144189 | 1663 | get_verify = 1; |
d1ece0c7 | 1664 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1665 | td->last_ddir != DDIR_READ) { |
1666 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1667 | if (!td->verify_batch) |
1668 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1669 | get_verify = 1; |
1670 | } | |
1671 | ||
d1ece0c7 JA |
1672 | if (get_verify && !get_next_verify(td, io_u)) { |
1673 | td->verify_batch--; | |
e39c0676 | 1674 | return true; |
d1ece0c7 | 1675 | } |
9e144189 JA |
1676 | } |
1677 | ||
e39c0676 | 1678 | return false; |
0d29de83 JA |
1679 | } |
1680 | ||
de789769 JA |
1681 | /* |
1682 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1683 | * easy compressible data for simple de-dupe attempts. Do this for every |
1684 | * 512b block in the range, since that should be the smallest block size | |
1685 | * we can expect from a device. | |
de789769 JA |
1686 | */ |
1687 | static void small_content_scramble(struct io_u *io_u) | |
1688 | { | |
23f394d5 | 1689 | unsigned int i, nr_blocks = io_u->buflen / 512; |
1ae83d45 | 1690 | uint64_t boffset; |
23f394d5 | 1691 | unsigned int offset; |
fe8d7fee | 1692 | char *p, *end; |
de789769 | 1693 | |
23f394d5 JA |
1694 | if (!nr_blocks) |
1695 | return; | |
1696 | ||
1697 | p = io_u->xfer_buf; | |
fba76ee8 | 1698 | boffset = io_u->offset; |
81f0366c | 1699 | io_u->buf_filled_len = 0; |
fad82f76 | 1700 | |
23f394d5 JA |
1701 | for (i = 0; i < nr_blocks; i++) { |
1702 | /* | |
1703 | * Fill the byte offset into a "random" start offset of | |
1704 | * the buffer, given by the product of the usec time | |
1705 | * and the actual offset. | |
1706 | */ | |
8b6a404c | 1707 | offset = ((io_u->start_time.tv_nsec/1000) ^ boffset) & 511; |
1ae83d45 JA |
1708 | offset &= ~(sizeof(uint64_t) - 1); |
1709 | if (offset >= 512 - sizeof(uint64_t)) | |
1710 | offset -= sizeof(uint64_t); | |
fba76ee8 | 1711 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1712 | |
1713 | end = p + 512 - sizeof(io_u->start_time); | |
1714 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1715 | p += 512; | |
fad82f76 | 1716 | boffset += 512; |
23f394d5 | 1717 | } |
de789769 JA |
1718 | } |
1719 | ||
0d29de83 JA |
1720 | /* |
1721 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1722 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1723 | */ | |
1724 | struct io_u *get_io_u(struct thread_data *td) | |
1725 | { | |
1726 | struct fio_file *f; | |
1727 | struct io_u *io_u; | |
de789769 | 1728 | int do_scramble = 0; |
002fe734 | 1729 | long ret = 0; |
0d29de83 JA |
1730 | |
1731 | io_u = __get_io_u(td); | |
1732 | if (!io_u) { | |
1733 | dprint(FD_IO, "__get_io_u failed\n"); | |
1734 | return NULL; | |
1735 | } | |
1736 | ||
1737 | if (check_get_verify(td, io_u)) | |
1738 | goto out; | |
1739 | if (check_get_trim(td, io_u)) | |
1740 | goto out; | |
1741 | ||
755200a3 JA |
1742 | /* |
1743 | * from a requeue, io_u already setup | |
1744 | */ | |
1745 | if (io_u->file) | |
77f392bf | 1746 | goto out; |
755200a3 | 1747 | |
429f6675 JA |
1748 | /* |
1749 | * If using an iolog, grab next piece if any available. | |
1750 | */ | |
d72be545 | 1751 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1752 | if (read_iolog_get(td, io_u)) |
1753 | goto err_put; | |
2ba1c290 | 1754 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1755 | ret = -EBUSY; |
2ba1c290 | 1756 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1757 | goto err_put; |
2ba1c290 | 1758 | } |
5ec10eaa | 1759 | |
429f6675 | 1760 | f = io_u->file; |
002fe734 JA |
1761 | if (!f) { |
1762 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1763 | goto err_put; | |
1764 | } | |
1765 | ||
d6aed795 | 1766 | assert(fio_file_open(f)); |
97af62ce | 1767 | |
ff58fced | 1768 | if (ddir_rw(io_u->ddir)) { |
9b87f09b | 1769 | if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) { |
2ba1c290 | 1770 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1771 | goto err_put; |
2ba1c290 | 1772 | } |
10ba535a | 1773 | |
f1dfb668 JA |
1774 | f->last_start[io_u->ddir] = io_u->offset; |
1775 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; | |
10ba535a | 1776 | |
fd68418e | 1777 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1778 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1779 | io_u_fill_buffer(td, io_u, |
1066358a | 1780 | td->o.min_bs[DDIR_WRITE], |
9e129577 | 1781 | io_u->buflen); |
ff441ae8 | 1782 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
bedc9dc2 | 1783 | !(td->flags & TD_F_COMPRESS)) |
fd68418e | 1784 | do_scramble = 1; |
d72be545 | 1785 | if (td->flags & TD_F_VER_NONE) { |
629f1d71 JA |
1786 | populate_verify_io_u(td, io_u); |
1787 | do_scramble = 0; | |
1788 | } | |
fd68418e | 1789 | } else if (io_u->ddir == DDIR_READ) { |
cbe8d756 RR |
1790 | /* |
1791 | * Reset the buf_filled parameters so next time if the | |
1792 | * buffer is used for writes it is refilled. | |
1793 | */ | |
cbe8d756 RR |
1794 | io_u->buf_filled_len = 0; |
1795 | } | |
87dc1ab1 | 1796 | } |
10ba535a | 1797 | |
165faf16 JA |
1798 | /* |
1799 | * Set io data pointers. | |
1800 | */ | |
cec6b55d JA |
1801 | io_u->xfer_buf = io_u->buf; |
1802 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1803 | |
6ac7a331 | 1804 | out: |
0d29de83 | 1805 | assert(io_u->file); |
429f6675 | 1806 | if (!td_io_prep(td, io_u)) { |
3ecc8c67 | 1807 | if (!td->o.disable_lat) |
993bf48b | 1808 | fio_gettime(&io_u->start_time, NULL); |
03553853 | 1809 | |
de789769 JA |
1810 | if (do_scramble) |
1811 | small_content_scramble(io_u); | |
03553853 | 1812 | |
429f6675 | 1813 | return io_u; |
36167d82 | 1814 | } |
429f6675 | 1815 | err_put: |
2ba1c290 | 1816 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1817 | put_io_u(td, io_u); |
002fe734 | 1818 | return ERR_PTR(ret); |
10ba535a JA |
1819 | } |
1820 | ||
a9da8ab2 | 1821 | static void __io_u_log_error(struct thread_data *td, struct io_u *io_u) |
5451792e | 1822 | { |
8b28bd41 | 1823 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1824 | |
8b28bd41 DM |
1825 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1826 | return; | |
5451792e | 1827 | |
709c8313 RE |
1828 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n", |
1829 | io_u->file ? " on file " : "", | |
1830 | io_u->file ? io_u->file->file_name : "", | |
1831 | strerror(io_u->error), | |
1832 | io_ddir_name(io_u->ddir), | |
1833 | io_u->offset, io_u->xfer_buflen); | |
5451792e | 1834 | |
5ad7be56 KD |
1835 | if (td->io_ops->errdetails) { |
1836 | char *err = td->io_ops->errdetails(io_u); | |
1837 | ||
1838 | log_err("fio: %s\n", err); | |
1839 | free(err); | |
1840 | } | |
1841 | ||
5451792e JA |
1842 | if (!td->error) |
1843 | td_verror(td, io_u->error, "io_u error"); | |
1844 | } | |
1845 | ||
a9da8ab2 JA |
1846 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1847 | { | |
1848 | __io_u_log_error(td, io_u); | |
1849 | if (td->parent) | |
094e66cb | 1850 | __io_u_log_error(td->parent, io_u); |
a9da8ab2 JA |
1851 | } |
1852 | ||
e39c0676 | 1853 | static inline bool gtod_reduce(struct thread_data *td) |
aba6c951 | 1854 | { |
3ecc8c67 JA |
1855 | return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw) |
1856 | || td->o.gtod_reduce; | |
aba6c951 JA |
1857 | } |
1858 | ||
c8eeb9df JA |
1859 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1860 | struct io_completion_data *icd, | |
1861 | const enum fio_ddir idx, unsigned int bytes) | |
1862 | { | |
a9da8ab2 | 1863 | const int no_reduce = !gtod_reduce(td); |
d6bb626e | 1864 | unsigned long long llnsec = 0; |
c8eeb9df | 1865 | |
75dc383e JA |
1866 | if (td->parent) |
1867 | td = td->parent; | |
1868 | ||
132b1ee4 | 1869 | if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS)) |
8243be59 JA |
1870 | return; |
1871 | ||
a9da8ab2 | 1872 | if (no_reduce) |
d6bb626e | 1873 | llnsec = ntime_since(&io_u->issue_time, &icd->time); |
c8eeb9df JA |
1874 | |
1875 | if (!td->o.disable_lat) { | |
d6bb626e | 1876 | unsigned long long tnsec; |
c8eeb9df | 1877 | |
d6bb626e VF |
1878 | tnsec = ntime_since(&io_u->start_time, &icd->time); |
1879 | add_lat_sample(td, idx, tnsec, bytes, io_u->offset); | |
15501535 | 1880 | |
d4afedfd JA |
1881 | if (td->flags & TD_F_PROFILE_OPS) { |
1882 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1883 | ||
1884 | if (ops->io_u_lat) | |
d6bb626e | 1885 | icd->error = ops->io_u_lat(td, tnsec/1000); |
d4afedfd JA |
1886 | } |
1887 | ||
d6bb626e VF |
1888 | if (td->o.max_latency && tnsec/1000 > td->o.max_latency) |
1889 | lat_fatal(td, icd, tnsec/1000, td->o.max_latency); | |
1890 | if (td->o.latency_target && tnsec/1000 > td->o.latency_target) { | |
3e260a46 | 1891 | if (lat_target_failed(td)) |
d6bb626e | 1892 | lat_fatal(td, icd, tnsec/1000, td->o.latency_target); |
15501535 | 1893 | } |
c8eeb9df JA |
1894 | } |
1895 | ||
a47591e4 JA |
1896 | if (ddir_rw(idx)) { |
1897 | if (!td->o.disable_clat) { | |
d6bb626e VF |
1898 | add_clat_sample(td, idx, llnsec, bytes, io_u->offset); |
1899 | io_u_mark_latency(td, llnsec); | |
a47591e4 | 1900 | } |
c8eeb9df | 1901 | |
a47591e4 | 1902 | if (!td->o.disable_bw && per_unit_log(td->bw_log)) |
d6bb626e | 1903 | add_bw_sample(td, io_u, bytes, llnsec); |
c8eeb9df | 1904 | |
a47591e4 JA |
1905 | if (no_reduce && per_unit_log(td->iops_log)) |
1906 | add_iops_sample(td, io_u, bytes); | |
1907 | } | |
66347cfa DE |
1908 | |
1909 | if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) { | |
1910 | uint32_t *info = io_u_block_info(td, io_u); | |
1911 | if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) { | |
1912 | if (io_u->ddir == DDIR_TRIM) { | |
1913 | *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, | |
1914 | BLOCK_INFO_TRIMS(*info) + 1); | |
1915 | } else if (io_u->ddir == DDIR_WRITE) { | |
1916 | *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN, | |
1917 | *info); | |
1918 | } | |
1919 | } | |
1920 | } | |
c8eeb9df JA |
1921 | } |
1922 | ||
94a6e1bb JA |
1923 | static void file_log_write_comp(const struct thread_data *td, struct fio_file *f, |
1924 | uint64_t offset, unsigned int bytes) | |
1925 | { | |
1926 | int idx; | |
1927 | ||
639ad1ea JA |
1928 | if (!f) |
1929 | return; | |
1930 | ||
94a6e1bb JA |
1931 | if (f->first_write == -1ULL || offset < f->first_write) |
1932 | f->first_write = offset; | |
1933 | if (f->last_write == -1ULL || ((offset + bytes) > f->last_write)) | |
1934 | f->last_write = offset + bytes; | |
1935 | ||
1936 | if (!f->last_write_comp) | |
1937 | return; | |
1938 | ||
1939 | idx = f->last_write_idx++; | |
1940 | f->last_write_comp[idx] = offset; | |
1941 | if (f->last_write_idx == td->o.iodepth) | |
1942 | f->last_write_idx = 0; | |
1943 | } | |
1944 | ||
f8b0bd10 | 1945 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1946 | struct io_completion_data *icd) |
10ba535a | 1947 | { |
f8b0bd10 JA |
1948 | struct io_u *io_u = *io_u_ptr; |
1949 | enum fio_ddir ddir = io_u->ddir; | |
1950 | struct fio_file *f = io_u->file; | |
10ba535a | 1951 | |
2ba1c290 JA |
1952 | dprint_io_u(io_u, "io complete"); |
1953 | ||
0c6e7517 | 1954 | assert(io_u->flags & IO_U_F_FLIGHT); |
1651e431 | 1955 | io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1956 | |
1957 | /* | |
1958 | * Mark IO ok to verify | |
1959 | */ | |
1960 | if (io_u->ipo) { | |
890b6656 JA |
1961 | /* |
1962 | * Remove errored entry from the verification list | |
1963 | */ | |
1964 | if (io_u->error) | |
1965 | unlog_io_piece(td, io_u); | |
1966 | else { | |
1967 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1968 | write_barrier(); | |
1969 | } | |
f9401285 JA |
1970 | } |
1971 | ||
f8b0bd10 | 1972 | if (ddir_sync(ddir)) { |
87dc1ab1 | 1973 | td->last_was_sync = 1; |
44f29692 JA |
1974 | if (f) { |
1975 | f->first_write = -1ULL; | |
1976 | f->last_write = -1ULL; | |
1977 | } | |
87dc1ab1 JA |
1978 | return; |
1979 | } | |
1980 | ||
1981 | td->last_was_sync = 0; | |
f8b0bd10 | 1982 | td->last_ddir = ddir; |
87dc1ab1 | 1983 | |
f8b0bd10 | 1984 | if (!io_u->error && ddir_rw(ddir)) { |
10ba535a | 1985 | unsigned int bytes = io_u->buflen - io_u->resid; |
b29ee5b3 | 1986 | int ret; |
10ba535a | 1987 | |
f8b0bd10 JA |
1988 | td->io_blocks[ddir]++; |
1989 | td->this_io_blocks[ddir]++; | |
1990 | td->io_bytes[ddir] += bytes; | |
ae2fafc8 | 1991 | |
1992 | if (!(io_u->flags & IO_U_F_VER_LIST)) | |
f8b0bd10 JA |
1993 | td->this_io_bytes[ddir] += bytes; |
1994 | ||
639ad1ea | 1995 | if (ddir == DDIR_WRITE) |
94a6e1bb | 1996 | file_log_write_comp(td, f, io_u->offset, bytes); |
44f29692 | 1997 | |
6b1190fd | 1998 | if (ramp_time_over(td) && (td->runstate == TD_RUNNING || |
50a8ce86 | 1999 | td->runstate == TD_VERIFYING)) |
f8b0bd10 | 2000 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 2001 | |
f8b0bd10 | 2002 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 2003 | |
d7762cf8 | 2004 | if (io_u->end_io) { |
f8b0bd10 JA |
2005 | ret = io_u->end_io(td, io_u_ptr); |
2006 | io_u = *io_u_ptr; | |
3af6ef39 JA |
2007 | if (ret && !icd->error) |
2008 | icd->error = ret; | |
2009 | } | |
ff58fced | 2010 | } else if (io_u->error) { |
10ba535a | 2011 | icd->error = io_u->error; |
5451792e JA |
2012 | io_u_log_error(td, io_u); |
2013 | } | |
8b28bd41 | 2014 | if (icd->error) { |
f8b0bd10 JA |
2015 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
2016 | ||
8b28bd41 DM |
2017 | if (!td_non_fatal_error(td, eb, icd->error)) |
2018 | return; | |
f8b0bd10 | 2019 | |
f2bba182 RR |
2020 | /* |
2021 | * If there is a non_fatal error, then add to the error count | |
2022 | * and clear all the errors. | |
2023 | */ | |
2024 | update_error_count(td, icd->error); | |
2025 | td_clear_error(td); | |
2026 | icd->error = 0; | |
f8b0bd10 JA |
2027 | if (io_u) |
2028 | io_u->error = 0; | |
f2bba182 | 2029 | } |
10ba535a JA |
2030 | } |
2031 | ||
9520ebb9 JA |
2032 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
2033 | int nr) | |
10ba535a | 2034 | { |
6eaf09d6 | 2035 | int ddir; |
aba6c951 JA |
2036 | |
2037 | if (!gtod_reduce(td)) | |
9520ebb9 | 2038 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 2039 | |
3af6ef39 JA |
2040 | icd->nr = nr; |
2041 | ||
10ba535a | 2042 | icd->error = 0; |
c1f50f76 | 2043 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
6eaf09d6 | 2044 | icd->bytes_done[ddir] = 0; |
36167d82 JA |
2045 | } |
2046 | ||
97601024 JA |
2047 | static void ios_completed(struct thread_data *td, |
2048 | struct io_completion_data *icd) | |
36167d82 JA |
2049 | { |
2050 | struct io_u *io_u; | |
2051 | int i; | |
2052 | ||
10ba535a JA |
2053 | for (i = 0; i < icd->nr; i++) { |
2054 | io_u = td->io_ops->event(td, i); | |
2055 | ||
f8b0bd10 | 2056 | io_completed(td, &io_u, icd); |
e8462bd8 | 2057 | |
f8b0bd10 | 2058 | if (io_u) |
e8462bd8 | 2059 | put_io_u(td, io_u); |
10ba535a JA |
2060 | } |
2061 | } | |
97601024 | 2062 | |
e7e6cfb4 JA |
2063 | /* |
2064 | * Complete a single io_u for the sync engines. | |
2065 | */ | |
55312f9f | 2066 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
2067 | { |
2068 | struct io_completion_data icd; | |
55312f9f | 2069 | int ddir; |
97601024 | 2070 | |
9520ebb9 | 2071 | init_icd(td, &icd, 1); |
f8b0bd10 | 2072 | io_completed(td, &io_u, &icd); |
e8462bd8 | 2073 | |
f8b0bd10 | 2074 | if (io_u) |
e8462bd8 | 2075 | put_io_u(td, io_u); |
97601024 | 2076 | |
581e7141 JA |
2077 | if (icd.error) { |
2078 | td_verror(td, icd.error, "io_u_sync_complete"); | |
2079 | return -1; | |
2080 | } | |
97601024 | 2081 | |
c1f50f76 | 2082 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
55312f9f | 2083 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
581e7141 JA |
2084 | |
2085 | return 0; | |
97601024 JA |
2086 | } |
2087 | ||
e7e6cfb4 JA |
2088 | /* |
2089 | * Called to complete min_events number of io for the async engines. | |
2090 | */ | |
55312f9f | 2091 | int io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 2092 | { |
97601024 | 2093 | struct io_completion_data icd; |
00de55ef | 2094 | struct timespec *tvp = NULL; |
55312f9f | 2095 | int ret, ddir; |
4d06a338 | 2096 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 2097 | |
12bb8569 | 2098 | dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts); |
b271fe62 | 2099 | |
4950421a | 2100 | if (!min_evts) |
00de55ef | 2101 | tvp = &ts; |
5fb4b366 RE |
2102 | else if (min_evts > td->cur_depth) |
2103 | min_evts = td->cur_depth; | |
97601024 | 2104 | |
82407585 RP |
2105 | /* No worries, td_io_getevents fixes min and max if they are |
2106 | * set incorrectly */ | |
2107 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp); | |
97601024 | 2108 | if (ret < 0) { |
e1161c32 | 2109 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
2110 | return ret; |
2111 | } else if (!ret) | |
2112 | return ret; | |
2113 | ||
9520ebb9 | 2114 | init_icd(td, &icd, ret); |
97601024 | 2115 | ios_completed(td, &icd); |
581e7141 JA |
2116 | if (icd.error) { |
2117 | td_verror(td, icd.error, "io_u_queued_complete"); | |
2118 | return -1; | |
2119 | } | |
97601024 | 2120 | |
c1f50f76 | 2121 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
55312f9f | 2122 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
581e7141 | 2123 | |
0d593542 | 2124 | return ret; |
97601024 | 2125 | } |
7e77dd02 JA |
2126 | |
2127 | /* | |
2128 | * Call when io_u is really queued, to update the submission latency. | |
2129 | */ | |
2130 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
2131 | { | |
8243be59 | 2132 | if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) { |
9520ebb9 | 2133 | unsigned long slat_time; |
7e77dd02 | 2134 | |
d6bb626e | 2135 | slat_time = ntime_since(&io_u->start_time, &io_u->issue_time); |
75dc383e JA |
2136 | |
2137 | if (td->parent) | |
2138 | td = td->parent; | |
2139 | ||
ae588852 JA |
2140 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
2141 | io_u->offset); | |
9520ebb9 | 2142 | } |
7e77dd02 | 2143 | } |
433afcb4 | 2144 | |
5c94b008 JA |
2145 | /* |
2146 | * See if we should reuse the last seed, if dedupe is enabled | |
2147 | */ | |
9451b93e | 2148 | static struct frand_state *get_buf_state(struct thread_data *td) |
5c94b008 JA |
2149 | { |
2150 | unsigned int v; | |
5c94b008 JA |
2151 | |
2152 | if (!td->o.dedupe_percentage) | |
2153 | return &td->buf_state; | |
732eedd0 | 2154 | else if (td->o.dedupe_percentage == 100) { |
9451b93e JA |
2155 | frand_copy(&td->buf_state_prev, &td->buf_state); |
2156 | return &td->buf_state; | |
732eedd0 | 2157 | } |
5c94b008 | 2158 | |
36dd3379 | 2159 | v = rand32_between(&td->dedupe_state, 1, 100); |
5c94b008 JA |
2160 | |
2161 | if (v <= td->o.dedupe_percentage) | |
2162 | return &td->buf_state_prev; | |
2163 | ||
2164 | return &td->buf_state; | |
2165 | } | |
2166 | ||
9451b93e | 2167 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) |
5c94b008 | 2168 | { |
9451b93e JA |
2169 | if (td->o.dedupe_percentage == 100) |
2170 | frand_copy(rs, &td->buf_state_prev); | |
2171 | else if (rs == &td->buf_state) | |
5c94b008 JA |
2172 | frand_copy(&td->buf_state_prev, rs); |
2173 | } | |
2174 | ||
cc86c395 JA |
2175 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write, |
2176 | unsigned int max_bs) | |
5973cafb | 2177 | { |
d1af2894 JA |
2178 | struct thread_options *o = &td->o; |
2179 | ||
15600335 JA |
2180 | if (o->mem_type == MEM_CUDA_MALLOC) |
2181 | return; | |
03553853 | 2182 | |
4eff3e57 | 2183 | if (o->compress_percentage || o->dedupe_percentage) { |
9c42684e | 2184 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 2185 | struct frand_state *rs; |
1066358a | 2186 | unsigned int left = max_bs; |
1e7f82e2 | 2187 | unsigned int this_write; |
5c94b008 | 2188 | |
1066358a | 2189 | do { |
9451b93e | 2190 | rs = get_buf_state(td); |
9c42684e | 2191 | |
1066358a | 2192 | min_write = min(min_write, left); |
f97a43a1 | 2193 | |
1066358a | 2194 | if (perc) { |
4b157ac6 JA |
2195 | this_write = min_not_zero(min_write, |
2196 | td->o.compress_chunk); | |
1e7f82e2 JA |
2197 | |
2198 | fill_random_buf_percentage(rs, buf, perc, | |
2199 | this_write, this_write, | |
2200 | o->buffer_pattern, | |
2201 | o->buffer_pattern_bytes); | |
2202 | } else { | |
1066358a | 2203 | fill_random_buf(rs, buf, min_write); |
1e7f82e2 JA |
2204 | this_write = min_write; |
2205 | } | |
1066358a | 2206 | |
1e7f82e2 JA |
2207 | buf += this_write; |
2208 | left -= this_write; | |
9451b93e | 2209 | save_buf_state(td, rs); |
1066358a | 2210 | } while (left); |
d1af2894 JA |
2211 | } else if (o->buffer_pattern_bytes) |
2212 | fill_buffer_pattern(td, buf, max_bs); | |
999d245e | 2213 | else if (o->zero_buffers) |
cc86c395 | 2214 | memset(buf, 0, max_bs); |
999d245e | 2215 | else |
9451b93e | 2216 | fill_random_buf(get_buf_state(td), buf, max_bs); |
cc86c395 JA |
2217 | } |
2218 | ||
2219 | /* | |
2220 | * "randomly" fill the buffer contents | |
2221 | */ | |
2222 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
2223 | unsigned int min_write, unsigned int max_bs) | |
2224 | { | |
2225 | io_u->buf_filled_len = 0; | |
2226 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 2227 | } |
e2c75fc4 TK |
2228 | |
2229 | static int do_sync_file_range(const struct thread_data *td, | |
2230 | struct fio_file *f) | |
2231 | { | |
2232 | off64_t offset, nbytes; | |
2233 | ||
2234 | offset = f->first_write; | |
2235 | nbytes = f->last_write - f->first_write; | |
2236 | ||
2237 | if (!nbytes) | |
2238 | return 0; | |
2239 | ||
2240 | return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range); | |
2241 | } | |
2242 | ||
2243 | int do_io_u_sync(const struct thread_data *td, struct io_u *io_u) | |
2244 | { | |
2245 | int ret; | |
2246 | ||
2247 | if (io_u->ddir == DDIR_SYNC) { | |
2248 | ret = fsync(io_u->file->fd); | |
2249 | } else if (io_u->ddir == DDIR_DATASYNC) { | |
2250 | #ifdef CONFIG_FDATASYNC | |
2251 | ret = fdatasync(io_u->file->fd); | |
2252 | #else | |
2253 | ret = io_u->xfer_buflen; | |
2254 | io_u->error = EINVAL; | |
2255 | #endif | |
2256 | } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE) | |
2257 | ret = do_sync_file_range(td, io_u->file); | |
2258 | else { | |
2259 | ret = io_u->xfer_buflen; | |
2260 | io_u->error = EINVAL; | |
2261 | } | |
2262 | ||
2263 | if (ret < 0) | |
2264 | io_u->error = errno; | |
2265 | ||
2266 | return ret; | |
2267 | } | |
2268 | ||
2269 | int do_io_u_trim(const struct thread_data *td, struct io_u *io_u) | |
2270 | { | |
2271 | #ifndef FIO_HAVE_TRIM | |
2272 | io_u->error = EINVAL; | |
2273 | return 0; | |
2274 | #else | |
2275 | struct fio_file *f = io_u->file; | |
2276 | int ret; | |
2277 | ||
496b1f9e | 2278 | ret = os_trim(f, io_u->offset, io_u->xfer_buflen); |
e2c75fc4 TK |
2279 | if (!ret) |
2280 | return io_u->xfer_buflen; | |
2281 | ||
2282 | io_u->error = ret; | |
2283 | return 0; | |
2284 | #endif | |
2285 | } |