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 | { |
2dc1bbeb | 40 | unsigned int min_bs = td->o.rw_min_bs; |
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 | ||
e0a04ac1 JA |
160 | static int __get_next_rand_offset_zoned(struct thread_data *td, |
161 | struct fio_file *f, enum fio_ddir ddir, | |
162 | uint64_t *b) | |
163 | { | |
164 | unsigned int v, send, stotal; | |
165 | uint64_t offset, lastb; | |
166 | static int warned; | |
167 | struct zone_split_index *zsi; | |
168 | ||
169 | lastb = last_block(td, f, ddir); | |
170 | if (!lastb) | |
171 | return 1; | |
172 | ||
173 | if (!td->o.zone_split_nr[ddir]) { | |
174 | bail: | |
175 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
176 | } | |
177 | ||
178 | /* | |
179 | * Generate a value, v, between 1 and 100, both inclusive | |
180 | */ | |
36dd3379 | 181 | v = rand32_between(&td->zone_state, 1, 100); |
e0a04ac1 JA |
182 | |
183 | zsi = &td->zone_state_index[ddir][v - 1]; | |
184 | stotal = zsi->size_perc_prev; | |
185 | send = zsi->size_perc; | |
186 | ||
187 | /* | |
188 | * Should never happen | |
189 | */ | |
190 | if (send == -1U) { | |
191 | if (!warned) { | |
192 | log_err("fio: bug in zoned generation\n"); | |
193 | warned = 1; | |
194 | } | |
195 | goto bail; | |
196 | } | |
197 | ||
198 | /* | |
199 | * 'send' is some percentage below or equal to 100 that | |
200 | * marks the end of the current IO range. 'stotal' marks | |
201 | * the start, in percent. | |
202 | */ | |
203 | if (stotal) | |
204 | offset = stotal * lastb / 100ULL; | |
205 | else | |
206 | offset = 0; | |
207 | ||
208 | lastb = lastb * (send - stotal) / 100ULL; | |
209 | ||
210 | /* | |
211 | * Generate index from 0..send-of-lastb | |
212 | */ | |
213 | if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1) | |
214 | return 1; | |
215 | ||
216 | /* | |
217 | * Add our start offset, if any | |
218 | */ | |
219 | if (offset) | |
220 | *b += offset; | |
221 | ||
222 | return 0; | |
223 | } | |
56d9fa4b | 224 | |
1ae83d45 JA |
225 | static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b) |
226 | { | |
227 | struct rand_off *r1 = flist_entry(a, struct rand_off, list); | |
228 | struct rand_off *r2 = flist_entry(b, struct rand_off, list); | |
229 | ||
230 | return r1->off - r2->off; | |
231 | } | |
232 | ||
233 | static int get_off_from_method(struct thread_data *td, struct fio_file *f, | |
234 | enum fio_ddir ddir, uint64_t *b) | |
e25839d4 | 235 | { |
e0a04ac1 JA |
236 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) { |
237 | uint64_t lastb; | |
238 | ||
239 | lastb = last_block(td, f, ddir); | |
240 | if (!lastb) | |
241 | return 1; | |
242 | ||
243 | return __get_next_rand_offset(td, f, ddir, b, lastb); | |
244 | } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) | |
e25839d4 | 245 | return __get_next_rand_offset_zipf(td, f, ddir, b); |
925fee33 JA |
246 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
247 | return __get_next_rand_offset_pareto(td, f, ddir, b); | |
56d9fa4b JA |
248 | else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS) |
249 | return __get_next_rand_offset_gauss(td, f, ddir, b); | |
e0a04ac1 JA |
250 | else if (td->o.random_distribution == FIO_RAND_DIST_ZONED) |
251 | return __get_next_rand_offset_zoned(td, f, ddir, b); | |
e25839d4 JA |
252 | |
253 | log_err("fio: unknown random distribution: %d\n", td->o.random_distribution); | |
254 | return 1; | |
255 | } | |
256 | ||
bcd5abfa JA |
257 | /* |
258 | * Sort the reads for a verify phase in batches of verifysort_nr, if | |
259 | * specified. | |
260 | */ | |
e39c0676 | 261 | static inline bool should_sort_io(struct thread_data *td) |
bcd5abfa JA |
262 | { |
263 | if (!td->o.verifysort_nr || !td->o.do_verify) | |
e39c0676 | 264 | return false; |
bcd5abfa | 265 | if (!td_random(td)) |
e39c0676 | 266 | return false; |
bcd5abfa | 267 | if (td->runstate != TD_VERIFYING) |
e39c0676 | 268 | return false; |
c3546b53 JA |
269 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE || |
270 | td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) | |
e39c0676 | 271 | return false; |
bcd5abfa | 272 | |
e39c0676 | 273 | return true; |
bcd5abfa JA |
274 | } |
275 | ||
e39c0676 | 276 | static bool should_do_random(struct thread_data *td, enum fio_ddir ddir) |
211c9b89 JA |
277 | { |
278 | unsigned int v; | |
211c9b89 | 279 | |
d9472271 | 280 | if (td->o.perc_rand[ddir] == 100) |
e39c0676 | 281 | return true; |
211c9b89 | 282 | |
36dd3379 | 283 | v = rand32_between(&td->seq_rand_state[ddir], 1, 100); |
211c9b89 | 284 | |
d9472271 | 285 | return v <= td->o.perc_rand[ddir]; |
211c9b89 JA |
286 | } |
287 | ||
1ae83d45 JA |
288 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
289 | enum fio_ddir ddir, uint64_t *b) | |
290 | { | |
291 | struct rand_off *r; | |
292 | int i, ret = 1; | |
293 | ||
bcd5abfa | 294 | if (!should_sort_io(td)) |
1ae83d45 JA |
295 | return get_off_from_method(td, f, ddir, b); |
296 | ||
297 | if (!flist_empty(&td->next_rand_list)) { | |
1ae83d45 | 298 | fetch: |
9342d5f8 | 299 | r = flist_first_entry(&td->next_rand_list, struct rand_off, list); |
1ae83d45 JA |
300 | flist_del(&r->list); |
301 | *b = r->off; | |
302 | free(r); | |
303 | return 0; | |
304 | } | |
305 | ||
306 | for (i = 0; i < td->o.verifysort_nr; i++) { | |
307 | r = malloc(sizeof(*r)); | |
308 | ||
309 | ret = get_off_from_method(td, f, ddir, &r->off); | |
310 | if (ret) { | |
311 | free(r); | |
312 | break; | |
313 | } | |
314 | ||
315 | flist_add(&r->list, &td->next_rand_list); | |
316 | } | |
317 | ||
318 | if (ret && !i) | |
319 | return ret; | |
320 | ||
321 | assert(!flist_empty(&td->next_rand_list)); | |
322 | flist_sort(NULL, &td->next_rand_list, flist_cmp); | |
323 | goto fetch; | |
324 | } | |
325 | ||
38dad62d | 326 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 327 | enum fio_ddir ddir, uint64_t *b) |
38dad62d | 328 | { |
c04e4661 DE |
329 | if (!get_next_rand_offset(td, f, ddir, b)) |
330 | return 0; | |
331 | ||
8c07860d JA |
332 | if (td->o.time_based || |
333 | (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) { | |
33c48814 | 334 | fio_file_reset(td, f); |
c04e4661 DE |
335 | if (!get_next_rand_offset(td, f, ddir, b)) |
336 | return 0; | |
38dad62d JA |
337 | } |
338 | ||
c04e4661 | 339 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", |
f1dfb668 | 340 | f->file_name, (unsigned long long) f->last_pos[ddir], |
4b91ee8f | 341 | (unsigned long long) f->real_file_size); |
c04e4661 | 342 | return 1; |
38dad62d JA |
343 | } |
344 | ||
37cf9e3c | 345 | static int get_next_seq_offset(struct thread_data *td, struct fio_file *f, |
1ae83d45 | 346 | enum fio_ddir ddir, uint64_t *offset) |
38dad62d | 347 | { |
ac002339 JA |
348 | struct thread_options *o = &td->o; |
349 | ||
ff58fced JA |
350 | assert(ddir_rw(ddir)); |
351 | ||
f1dfb668 | 352 | if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) && |
19ddc35b JA |
353 | o->time_based) { |
354 | struct thread_options *o = &td->o; | |
355 | uint64_t io_size = f->io_size + (f->io_size % o->min_bs[ddir]); | |
356 | ||
357 | if (io_size > f->last_pos[ddir]) | |
358 | f->last_pos[ddir] = 0; | |
359 | else | |
360 | f->last_pos[ddir] = f->last_pos[ddir] - io_size; | |
361 | } | |
c04e4661 | 362 | |
f1dfb668 | 363 | if (f->last_pos[ddir] < f->real_file_size) { |
1ae83d45 | 364 | uint64_t pos; |
059b0802 | 365 | |
c22825bb JA |
366 | if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) { |
367 | if (f->real_file_size > f->io_size) | |
368 | f->last_pos[ddir] = f->io_size; | |
369 | else | |
370 | f->last_pos[ddir] = f->real_file_size; | |
371 | } | |
a66da7a2 | 372 | |
f1dfb668 | 373 | pos = f->last_pos[ddir] - f->file_offset; |
ac002339 JA |
374 | if (pos && o->ddir_seq_add) { |
375 | pos += o->ddir_seq_add; | |
376 | ||
377 | /* | |
378 | * If we reach beyond the end of the file | |
379 | * with holed IO, wrap around to the | |
b0a84f48 JA |
380 | * beginning again. If we're doing backwards IO, |
381 | * wrap to the end. | |
ac002339 | 382 | */ |
b0a84f48 JA |
383 | if (pos >= f->real_file_size) { |
384 | if (o->ddir_seq_add > 0) | |
385 | pos = f->file_offset; | |
c22825bb JA |
386 | else { |
387 | if (f->real_file_size > f->io_size) | |
388 | pos = f->io_size; | |
389 | else | |
390 | pos = f->real_file_size; | |
391 | ||
392 | pos += o->ddir_seq_add; | |
393 | } | |
b0a84f48 | 394 | } |
ac002339 | 395 | } |
059b0802 | 396 | |
37cf9e3c | 397 | *offset = pos; |
38dad62d JA |
398 | return 0; |
399 | } | |
400 | ||
401 | return 1; | |
402 | } | |
403 | ||
404 | static int get_next_block(struct thread_data *td, struct io_u *io_u, | |
6aca9b3d JA |
405 | enum fio_ddir ddir, int rw_seq, |
406 | unsigned int *is_random) | |
38dad62d JA |
407 | { |
408 | struct fio_file *f = io_u->file; | |
1ae83d45 | 409 | uint64_t b, offset; |
38dad62d JA |
410 | int ret; |
411 | ||
ff58fced JA |
412 | assert(ddir_rw(ddir)); |
413 | ||
37cf9e3c JA |
414 | b = offset = -1ULL; |
415 | ||
38dad62d | 416 | if (rw_seq) { |
211c9b89 | 417 | if (td_random(td)) { |
6aca9b3d | 418 | if (should_do_random(td, ddir)) { |
211c9b89 | 419 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
420 | *is_random = 1; |
421 | } else { | |
422 | *is_random = 0; | |
1651e431 | 423 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
211c9b89 JA |
424 | ret = get_next_seq_offset(td, f, ddir, &offset); |
425 | if (ret) | |
426 | ret = get_next_rand_block(td, f, ddir, &b); | |
427 | } | |
6aca9b3d JA |
428 | } else { |
429 | *is_random = 0; | |
37cf9e3c | 430 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 431 | } |
38dad62d | 432 | } else { |
1651e431 | 433 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
6aca9b3d | 434 | *is_random = 0; |
38dad62d JA |
435 | |
436 | if (td->o.rw_seq == RW_SEQ_SEQ) { | |
37cf9e3c | 437 | ret = get_next_seq_offset(td, f, ddir, &offset); |
6aca9b3d | 438 | if (ret) { |
37cf9e3c | 439 | ret = get_next_rand_block(td, f, ddir, &b); |
6aca9b3d JA |
440 | *is_random = 0; |
441 | } | |
38dad62d | 442 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
f1dfb668 JA |
443 | if (f->last_start[ddir] != -1ULL) |
444 | offset = f->last_start[ddir] - f->file_offset; | |
38dad62d | 445 | else |
37cf9e3c | 446 | offset = 0; |
38dad62d JA |
447 | ret = 0; |
448 | } else { | |
449 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); | |
450 | ret = 1; | |
451 | } | |
452 | } | |
6d68b997 | 453 | |
37cf9e3c JA |
454 | if (!ret) { |
455 | if (offset != -1ULL) | |
456 | io_u->offset = offset; | |
457 | else if (b != -1ULL) | |
458 | io_u->offset = b * td->o.ba[ddir]; | |
459 | else { | |
4e0a8fa2 | 460 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
37cf9e3c JA |
461 | ret = 1; |
462 | } | |
463 | } | |
464 | ||
38dad62d JA |
465 | return ret; |
466 | } | |
467 | ||
10ba535a JA |
468 | /* |
469 | * For random io, generate a random new block and see if it's used. Repeat | |
470 | * until we find a free one. For sequential io, just return the end of | |
471 | * the last io issued. | |
472 | */ | |
6aca9b3d JA |
473 | static int __get_next_offset(struct thread_data *td, struct io_u *io_u, |
474 | unsigned int *is_random) | |
10ba535a | 475 | { |
9bf2061e | 476 | struct fio_file *f = io_u->file; |
4ba66134 | 477 | enum fio_ddir ddir = io_u->ddir; |
38dad62d | 478 | int rw_seq_hit = 0; |
10ba535a | 479 | |
ff58fced JA |
480 | assert(ddir_rw(ddir)); |
481 | ||
38dad62d JA |
482 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
483 | rw_seq_hit = 1; | |
5736c10d | 484 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
38dad62d | 485 | } |
211097b2 | 486 | |
6aca9b3d | 487 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
38dad62d | 488 | return 1; |
10ba535a | 489 | |
009bd847 JA |
490 | if (io_u->offset >= f->io_size) { |
491 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", | |
4b91ee8f JA |
492 | (unsigned long long) io_u->offset, |
493 | (unsigned long long) f->io_size); | |
009bd847 JA |
494 | return 1; |
495 | } | |
496 | ||
497 | io_u->offset += f->file_offset; | |
2ba1c290 JA |
498 | if (io_u->offset >= f->real_file_size) { |
499 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", | |
4b91ee8f JA |
500 | (unsigned long long) io_u->offset, |
501 | (unsigned long long) f->real_file_size); | |
10ba535a | 502 | return 1; |
2ba1c290 | 503 | } |
10ba535a JA |
504 | |
505 | return 0; | |
506 | } | |
507 | ||
6aca9b3d JA |
508 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
509 | unsigned int *is_random) | |
15dc1934 | 510 | { |
d72be545 JA |
511 | if (td->flags & TD_F_PROFILE_OPS) { |
512 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 513 | |
d72be545 | 514 | if (ops->fill_io_u_off) |
6aca9b3d | 515 | return ops->fill_io_u_off(td, io_u, is_random); |
d72be545 | 516 | } |
15dc1934 | 517 | |
6aca9b3d | 518 | return __get_next_offset(td, io_u, is_random); |
15dc1934 JA |
519 | } |
520 | ||
e39c0676 JA |
521 | static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u, |
522 | unsigned int buflen) | |
79944128 JA |
523 | { |
524 | struct fio_file *f = io_u->file; | |
525 | ||
bedc9dc2 | 526 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
79944128 JA |
527 | } |
528 | ||
6aca9b3d JA |
529 | static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u, |
530 | unsigned int is_random) | |
10ba535a | 531 | { |
6aca9b3d | 532 | int ddir = io_u->ddir; |
24d23ca7 | 533 | unsigned int buflen = 0; |
f3059de1 | 534 | unsigned int minbs, maxbs; |
3dd29f7c | 535 | uint64_t frand_max, r; |
7c961359 | 536 | bool power_2; |
10ba535a | 537 | |
9ee1c647 | 538 | assert(ddir_rw(ddir)); |
6aca9b3d JA |
539 | |
540 | if (td->o.bs_is_seq_rand) | |
541 | ddir = is_random ? DDIR_WRITE: DDIR_READ; | |
ff58fced | 542 | |
f3059de1 JA |
543 | minbs = td->o.min_bs[ddir]; |
544 | maxbs = td->o.max_bs[ddir]; | |
545 | ||
79944128 JA |
546 | if (minbs == maxbs) |
547 | return minbs; | |
548 | ||
52c58027 JA |
549 | /* |
550 | * If we can't satisfy the min block size from here, then fail | |
551 | */ | |
552 | if (!io_u_fits(td, io_u, minbs)) | |
553 | return 0; | |
554 | ||
2f282cec | 555 | frand_max = rand_max(&td->bsrange_state[ddir]); |
79944128 | 556 | do { |
2f282cec | 557 | r = __rand(&td->bsrange_state[ddir]); |
4c07ad86 | 558 | |
720e84ad | 559 | if (!td->o.bssplit_nr[ddir]) { |
f3059de1 | 560 | buflen = 1 + (unsigned int) ((double) maxbs * |
c3546b53 | 561 | (r / (frand_max + 1.0))); |
f3059de1 JA |
562 | if (buflen < minbs) |
563 | buflen = minbs; | |
5ec10eaa | 564 | } else { |
3dd29f7c | 565 | long long perc = 0; |
564ca972 JA |
566 | unsigned int i; |
567 | ||
720e84ad JA |
568 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
569 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; | |
564ca972 JA |
570 | |
571 | buflen = bsp->bs; | |
572 | perc += bsp->perc; | |
3dd29f7c JA |
573 | if (!perc) |
574 | break; | |
575 | if ((r / perc <= frand_max / 100ULL) && | |
79944128 | 576 | io_u_fits(td, io_u, buflen)) |
564ca972 JA |
577 | break; |
578 | } | |
579 | } | |
79944128 | 580 | |
17a6b702 PL |
581 | power_2 = is_power_of_2(minbs); |
582 | if (!td->o.bs_unaligned && power_2) | |
7c306bb1 | 583 | buflen &= ~(minbs - 1); |
17a6b702 PL |
584 | else if (!td->o.bs_unaligned && !power_2) |
585 | buflen -= buflen % minbs; | |
79944128 | 586 | } while (!io_u_fits(td, io_u, buflen)); |
6a5e6884 | 587 | |
10ba535a JA |
588 | return buflen; |
589 | } | |
590 | ||
6aca9b3d JA |
591 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u, |
592 | unsigned int is_random) | |
15dc1934 | 593 | { |
d72be545 JA |
594 | if (td->flags & TD_F_PROFILE_OPS) { |
595 | struct prof_io_ops *ops = &td->prof_io_ops; | |
7eb36574 | 596 | |
d72be545 | 597 | if (ops->fill_io_u_size) |
6aca9b3d | 598 | return ops->fill_io_u_size(td, io_u, is_random); |
d72be545 | 599 | } |
15dc1934 | 600 | |
6aca9b3d | 601 | return __get_next_buflen(td, io_u, is_random); |
15dc1934 JA |
602 | } |
603 | ||
afe24a5a JA |
604 | static void set_rwmix_bytes(struct thread_data *td) |
605 | { | |
afe24a5a JA |
606 | unsigned int diff; |
607 | ||
608 | /* | |
609 | * we do time or byte based switch. this is needed because | |
610 | * buffered writes may issue a lot quicker than they complete, | |
611 | * whereas reads do not. | |
612 | */ | |
e47f799f | 613 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
04c540d9 | 614 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
e47f799f JA |
615 | } |
616 | ||
617 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) | |
618 | { | |
619 | unsigned int v; | |
e47f799f | 620 | |
36dd3379 | 621 | v = rand32_between(&td->rwmix_state, 1, 100); |
4c07ad86 | 622 | |
04c540d9 | 623 | if (v <= td->o.rwmix[DDIR_READ]) |
e47f799f JA |
624 | return DDIR_READ; |
625 | ||
626 | return DDIR_WRITE; | |
afe24a5a JA |
627 | } |
628 | ||
0d593542 | 629 | int io_u_quiesce(struct thread_data *td) |
002e7183 | 630 | { |
0d593542 JA |
631 | int completed = 0; |
632 | ||
002e7183 JA |
633 | /* |
634 | * We are going to sleep, ensure that we flush anything pending as | |
635 | * not to skew our latency numbers. | |
636 | * | |
637 | * Changed to only monitor 'in flight' requests here instead of the | |
638 | * td->cur_depth, b/c td->cur_depth does not accurately represent | |
639 | * io's that have been actually submitted to an async engine, | |
640 | * and cur_depth is meaningless for sync engines. | |
641 | */ | |
9cc80b6d JA |
642 | if (td->io_u_queued || td->cur_depth) { |
643 | int fio_unused ret; | |
644 | ||
645 | ret = td_io_commit(td); | |
646 | } | |
647 | ||
002e7183 | 648 | while (td->io_u_in_flight) { |
89fd0573 | 649 | int ret; |
002e7183 | 650 | |
55312f9f | 651 | ret = io_u_queued_complete(td, 1); |
0d593542 JA |
652 | if (ret > 0) |
653 | completed += ret; | |
002e7183 | 654 | } |
0d593542 | 655 | |
6be06c46 JA |
656 | if (td->flags & TD_F_REGROW_LOGS) |
657 | regrow_logs(td); | |
658 | ||
0d593542 | 659 | return completed; |
002e7183 JA |
660 | } |
661 | ||
581e7141 JA |
662 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
663 | { | |
664 | enum fio_ddir odir = ddir ^ 1; | |
90eff1c9 SW |
665 | long usec; |
666 | uint64_t now; | |
581e7141 | 667 | |
ff58fced | 668 | assert(ddir_rw(ddir)); |
50a8ce86 | 669 | now = utime_since_now(&td->start); |
ff58fced | 670 | |
50a8ce86 D |
671 | /* |
672 | * if rate_next_io_time is in the past, need to catch up to rate | |
673 | */ | |
674 | if (td->rate_next_io_time[ddir] <= now) | |
581e7141 JA |
675 | return ddir; |
676 | ||
677 | /* | |
50a8ce86 | 678 | * We are ahead of rate in this direction. See if we |
581e7141 JA |
679 | * should switch. |
680 | */ | |
315fcfec | 681 | if (td_rw(td) && td->o.rwmix[odir]) { |
581e7141 | 682 | /* |
50a8ce86 | 683 | * Other direction is behind rate, switch |
581e7141 | 684 | */ |
50a8ce86 | 685 | if (td->rate_next_io_time[odir] <= now) |
581e7141 JA |
686 | return odir; |
687 | ||
688 | /* | |
50a8ce86 D |
689 | * Both directions are ahead of rate. sleep the min |
690 | * switch if necissary | |
581e7141 | 691 | */ |
50a8ce86 D |
692 | if (td->rate_next_io_time[ddir] <= |
693 | td->rate_next_io_time[odir]) { | |
694 | usec = td->rate_next_io_time[ddir] - now; | |
581e7141 | 695 | } else { |
50a8ce86 | 696 | usec = td->rate_next_io_time[odir] - now; |
581e7141 JA |
697 | ddir = odir; |
698 | } | |
699 | } else | |
50a8ce86 | 700 | usec = td->rate_next_io_time[ddir] - now; |
581e7141 | 701 | |
a9da8ab2 JA |
702 | if (td->o.io_submit_mode == IO_MODE_INLINE) |
703 | io_u_quiesce(td); | |
78c1eda5 | 704 | |
9cc80b6d | 705 | usec = usec_sleep(td, usec); |
581e7141 | 706 | |
581e7141 JA |
707 | return ddir; |
708 | } | |
709 | ||
10ba535a JA |
710 | /* |
711 | * Return the data direction for the next io_u. If the job is a | |
712 | * mixed read/write workload, check the rwmix cycle and switch if | |
713 | * necessary. | |
714 | */ | |
1e97cce9 | 715 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
10ba535a | 716 | { |
581e7141 JA |
717 | enum fio_ddir ddir; |
718 | ||
5f9099ea | 719 | /* |
f6860972 TK |
720 | * See if it's time to fsync/fdatasync/sync_file_range first, |
721 | * and if not then move on to check regular I/Os. | |
5f9099ea | 722 | */ |
f6860972 TK |
723 | if (should_fsync(td)) { |
724 | if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] && | |
725 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks)) | |
726 | return DDIR_SYNC; | |
727 | ||
728 | if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] && | |
729 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks)) | |
730 | return DDIR_DATASYNC; | |
731 | ||
732 | if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] && | |
733 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr)) | |
734 | return DDIR_SYNC_FILE_RANGE; | |
735 | } | |
44f29692 | 736 | |
10ba535a | 737 | if (td_rw(td)) { |
10ba535a JA |
738 | /* |
739 | * Check if it's time to seed a new data direction. | |
740 | */ | |
e4928662 | 741 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
e47f799f JA |
742 | /* |
743 | * Put a top limit on how many bytes we do for | |
744 | * one data direction, to avoid overflowing the | |
745 | * ranges too much | |
746 | */ | |
747 | ddir = get_rand_ddir(td); | |
e47f799f JA |
748 | |
749 | if (ddir != td->rwmix_ddir) | |
750 | set_rwmix_bytes(td); | |
751 | ||
752 | td->rwmix_ddir = ddir; | |
10ba535a | 753 | } |
581e7141 | 754 | ddir = td->rwmix_ddir; |
10ba535a | 755 | } else if (td_read(td)) |
581e7141 | 756 | ddir = DDIR_READ; |
6eaf09d6 | 757 | else if (td_write(td)) |
581e7141 | 758 | ddir = DDIR_WRITE; |
5c8e84ca | 759 | else if (td_trim(td)) |
6eaf09d6 | 760 | ddir = DDIR_TRIM; |
5c8e84ca TK |
761 | else |
762 | ddir = DDIR_INVAL; | |
581e7141 JA |
763 | |
764 | td->rwmix_ddir = rate_ddir(td, ddir); | |
765 | return td->rwmix_ddir; | |
10ba535a JA |
766 | } |
767 | ||
1ef2b6be JA |
768 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
769 | { | |
0e4dd95c DE |
770 | enum fio_ddir ddir = get_rw_ddir(td); |
771 | ||
82a90686 | 772 | if (td_trimwrite(td)) { |
0e4dd95c DE |
773 | struct fio_file *f = io_u->file; |
774 | if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM]) | |
775 | ddir = DDIR_TRIM; | |
776 | else | |
777 | ddir = DDIR_WRITE; | |
778 | } | |
779 | ||
780 | io_u->ddir = io_u->acct_ddir = ddir; | |
1ef2b6be | 781 | |
9b87f09b | 782 | if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) && |
1ef2b6be JA |
783 | td->o.barrier_blocks && |
784 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && | |
785 | td->io_issues[DDIR_WRITE]) | |
1651e431 | 786 | io_u_set(td, io_u, IO_U_F_BARRIER); |
1ef2b6be JA |
787 | } |
788 | ||
e8462bd8 | 789 | void put_file_log(struct thread_data *td, struct fio_file *f) |
60f2c658 | 790 | { |
71b84caa | 791 | unsigned int ret = put_file(td, f); |
60f2c658 JA |
792 | |
793 | if (ret) | |
794 | td_verror(td, ret, "file close"); | |
795 | } | |
796 | ||
10ba535a JA |
797 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
798 | { | |
a9da8ab2 JA |
799 | if (td->parent) |
800 | td = td->parent; | |
801 | ||
e8462bd8 JA |
802 | td_io_u_lock(td); |
803 | ||
f8b0bd10 | 804 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
60f2c658 | 805 | put_file_log(td, io_u->file); |
f8b0bd10 | 806 | |
10ba535a | 807 | io_u->file = NULL; |
1651e431 | 808 | io_u_set(td, io_u, IO_U_F_FREE); |
d7ee2a7d | 809 | |
a9da8ab2 | 810 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 811 | td->cur_depth--; |
a9da8ab2 JA |
812 | assert(!(td->flags & TD_F_CHILD)); |
813 | } | |
2ae0b204 | 814 | io_u_qpush(&td->io_u_freelist, io_u); |
e8462bd8 JA |
815 | td_io_u_unlock(td); |
816 | td_io_u_free_notify(td); | |
10ba535a JA |
817 | } |
818 | ||
f2bba182 RR |
819 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
820 | { | |
1651e431 | 821 | io_u_clear(td, io_u, IO_U_F_FLIGHT); |
f2bba182 RR |
822 | put_io_u(td, io_u); |
823 | } | |
824 | ||
755200a3 JA |
825 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
826 | { | |
827 | struct io_u *__io_u = *io_u; | |
bcd5abfa | 828 | enum fio_ddir ddir = acct_ddir(__io_u); |
755200a3 | 829 | |
465221b0 JA |
830 | dprint(FD_IO, "requeue %p\n", __io_u); |
831 | ||
a9da8ab2 JA |
832 | if (td->parent) |
833 | td = td->parent; | |
834 | ||
e8462bd8 JA |
835 | td_io_u_lock(td); |
836 | ||
1651e431 | 837 | io_u_set(td, __io_u, IO_U_F_FREE); |
bcd5abfa JA |
838 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
839 | td->io_issues[ddir]--; | |
5ec10eaa | 840 | |
1651e431 | 841 | io_u_clear(td, __io_u, IO_U_F_FLIGHT); |
a9da8ab2 | 842 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
0c41214f | 843 | td->cur_depth--; |
a9da8ab2 JA |
844 | assert(!(td->flags & TD_F_CHILD)); |
845 | } | |
2ae0b204 JA |
846 | |
847 | io_u_rpush(&td->io_u_requeues, __io_u); | |
e8462bd8 | 848 | td_io_u_unlock(td); |
a9da8ab2 | 849 | td_io_u_free_notify(td); |
755200a3 JA |
850 | *io_u = NULL; |
851 | } | |
852 | ||
224b3093 | 853 | static void __fill_io_u_zone(struct thread_data *td, struct io_u *io_u) |
854 | { | |
855 | struct fio_file *f = io_u->file; | |
856 | ||
857 | /* | |
858 | * See if it's time to switch to a new zone | |
859 | */ | |
860 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { | |
861 | td->zone_bytes = 0; | |
862 | f->file_offset += td->o.zone_range + td->o.zone_skip; | |
863 | ||
864 | /* | |
865 | * Wrap from the beginning, if we exceed the file size | |
866 | */ | |
867 | if (f->file_offset >= f->real_file_size) | |
868 | f->file_offset = f->real_file_size - f->file_offset; | |
869 | f->last_pos[io_u->ddir] = f->file_offset; | |
870 | td->io_skip_bytes += td->o.zone_skip; | |
871 | } | |
872 | ||
873 | /* | |
874 | * If zone_size > zone_range, then maintain the same zone until | |
875 | * zone_bytes >= zone_size. | |
876 | */ | |
877 | if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) { | |
878 | dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n", | |
879 | f->file_offset, f->last_pos[io_u->ddir]); | |
880 | f->last_pos[io_u->ddir] = f->file_offset; | |
881 | } | |
882 | ||
883 | /* | |
884 | * For random: if 'norandommap' is not set and zone_size > zone_range, | |
885 | * map needs to be reset as it's done with zone_range everytime. | |
886 | */ | |
887 | if ((td->zone_bytes % td->o.zone_range) == 0) { | |
888 | fio_file_reset(td, f); | |
889 | } | |
890 | } | |
891 | ||
9bf2061e | 892 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
10ba535a | 893 | { |
6aca9b3d JA |
894 | unsigned int is_random; |
895 | ||
9b87f09b | 896 | if (td_ioengine_flagged(td, FIO_NOIO)) |
b4c5e1ac JA |
897 | goto out; |
898 | ||
1ef2b6be | 899 | set_rw_ddir(td, io_u); |
5f9099ea | 900 | |
87dc1ab1 | 901 | /* |
ff58fced | 902 | * fsync() or fdatasync() or trim etc, we are done |
87dc1ab1 | 903 | */ |
ff58fced | 904 | if (!ddir_rw(io_u->ddir)) |
c38e9468 | 905 | goto out; |
a00735e6 | 906 | |
48f5abd3 | 907 | /* |
224b3093 | 908 | * When file is zoned zone_range is always positive |
48f5abd3 | 909 | */ |
224b3093 | 910 | if (td->o.zone_range) { |
911 | __fill_io_u_zone(td, io_u); | |
48f5abd3 JA |
912 | } |
913 | ||
10ba535a | 914 | /* |
c685b5b2 JA |
915 | * No log, let the seq/rand engine retrieve the next buflen and |
916 | * position. | |
10ba535a | 917 | */ |
6aca9b3d | 918 | if (get_next_offset(td, io_u, &is_random)) { |
2ba1c290 | 919 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
bca4ed4d | 920 | return 1; |
2ba1c290 | 921 | } |
10ba535a | 922 | |
6aca9b3d | 923 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
2ba1c290 JA |
924 | if (!io_u->buflen) { |
925 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); | |
bca4ed4d | 926 | return 1; |
2ba1c290 | 927 | } |
bca4ed4d | 928 | |
2ba1c290 | 929 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
48366f3a TK |
930 | dprint(FD_IO, "io_u %p, offset + buflen exceeds file size\n", |
931 | io_u); | |
932 | dprint(FD_IO, " offset=%llu/buflen=%lu > %llu\n", | |
4b91ee8f JA |
933 | (unsigned long long) io_u->offset, io_u->buflen, |
934 | (unsigned long long) io_u->file->real_file_size); | |
6a5e6884 | 935 | return 1; |
2ba1c290 | 936 | } |
6a5e6884 | 937 | |
bca4ed4d JA |
938 | /* |
939 | * mark entry before potentially trimming io_u | |
940 | */ | |
303032ae | 941 | if (td_random(td) && file_randommap(td, io_u->file)) |
9bf2061e | 942 | mark_random_map(td, io_u); |
bca4ed4d | 943 | |
c38e9468 | 944 | out: |
2ba1c290 | 945 | dprint_io_u(io_u, "fill_io_u"); |
d9d91e39 | 946 | td->zone_bytes += io_u->buflen; |
bca4ed4d | 947 | return 0; |
10ba535a JA |
948 | } |
949 | ||
838bc709 JA |
950 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
951 | { | |
2b13e716 | 952 | int idx = 0; |
838bc709 JA |
953 | |
954 | switch (nr) { | |
955 | default: | |
2b13e716 | 956 | idx = 6; |
838bc709 JA |
957 | break; |
958 | case 33 ... 64: | |
2b13e716 | 959 | idx = 5; |
838bc709 JA |
960 | break; |
961 | case 17 ... 32: | |
2b13e716 | 962 | idx = 4; |
838bc709 JA |
963 | break; |
964 | case 9 ... 16: | |
2b13e716 | 965 | idx = 3; |
838bc709 JA |
966 | break; |
967 | case 5 ... 8: | |
2b13e716 | 968 | idx = 2; |
838bc709 JA |
969 | break; |
970 | case 1 ... 4: | |
2b13e716 | 971 | idx = 1; |
838bc709 JA |
972 | case 0: |
973 | break; | |
974 | } | |
975 | ||
2b13e716 | 976 | map[idx]++; |
838bc709 JA |
977 | } |
978 | ||
979 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) | |
980 | { | |
981 | __io_u_mark_map(td->ts.io_u_submit, nr); | |
982 | td->ts.total_submit++; | |
983 | } | |
984 | ||
985 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) | |
986 | { | |
987 | __io_u_mark_map(td->ts.io_u_complete, nr); | |
988 | td->ts.total_complete++; | |
989 | } | |
990 | ||
d8005759 | 991 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
71619dc2 | 992 | { |
2b13e716 | 993 | int idx = 0; |
71619dc2 JA |
994 | |
995 | switch (td->cur_depth) { | |
996 | default: | |
2b13e716 | 997 | idx = 6; |
a783e61a | 998 | break; |
71619dc2 | 999 | case 32 ... 63: |
2b13e716 | 1000 | idx = 5; |
a783e61a | 1001 | break; |
71619dc2 | 1002 | case 16 ... 31: |
2b13e716 | 1003 | idx = 4; |
a783e61a | 1004 | break; |
71619dc2 | 1005 | case 8 ... 15: |
2b13e716 | 1006 | idx = 3; |
a783e61a | 1007 | break; |
71619dc2 | 1008 | case 4 ... 7: |
2b13e716 | 1009 | idx = 2; |
a783e61a | 1010 | break; |
71619dc2 | 1011 | case 2 ... 3: |
2b13e716 | 1012 | idx = 1; |
71619dc2 JA |
1013 | case 1: |
1014 | break; | |
1015 | } | |
1016 | ||
2b13e716 | 1017 | td->ts.io_u_map[idx] += nr; |
71619dc2 JA |
1018 | } |
1019 | ||
d6bb626e | 1020 | static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec) |
04a0feae | 1021 | { |
2b13e716 | 1022 | int idx = 0; |
04a0feae | 1023 | |
d6bb626e VF |
1024 | assert(nsec < 1000); |
1025 | ||
1026 | switch (nsec) { | |
1027 | case 750 ... 999: | |
1028 | idx = 9; | |
1029 | break; | |
1030 | case 500 ... 749: | |
1031 | idx = 8; | |
1032 | break; | |
1033 | case 250 ... 499: | |
1034 | idx = 7; | |
1035 | break; | |
1036 | case 100 ... 249: | |
1037 | idx = 6; | |
1038 | break; | |
1039 | case 50 ... 99: | |
1040 | idx = 5; | |
1041 | break; | |
1042 | case 20 ... 49: | |
1043 | idx = 4; | |
1044 | break; | |
1045 | case 10 ... 19: | |
1046 | idx = 3; | |
1047 | break; | |
1048 | case 4 ... 9: | |
1049 | idx = 2; | |
1050 | break; | |
1051 | case 2 ... 3: | |
1052 | idx = 1; | |
1053 | case 0 ... 1: | |
1054 | break; | |
1055 | } | |
1056 | ||
1057 | assert(idx < FIO_IO_U_LAT_N_NR); | |
1058 | td->ts.io_u_lat_n[idx]++; | |
1059 | } | |
1060 | ||
1061 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec) | |
1062 | { | |
1063 | int idx = 0; | |
1064 | ||
1065 | assert(usec < 1000 && usec >= 1); | |
04a0feae JA |
1066 | |
1067 | switch (usec) { | |
1068 | case 750 ... 999: | |
2b13e716 | 1069 | idx = 9; |
04a0feae JA |
1070 | break; |
1071 | case 500 ... 749: | |
2b13e716 | 1072 | idx = 8; |
04a0feae JA |
1073 | break; |
1074 | case 250 ... 499: | |
2b13e716 | 1075 | idx = 7; |
04a0feae JA |
1076 | break; |
1077 | case 100 ... 249: | |
2b13e716 | 1078 | idx = 6; |
04a0feae JA |
1079 | break; |
1080 | case 50 ... 99: | |
2b13e716 | 1081 | idx = 5; |
04a0feae JA |
1082 | break; |
1083 | case 20 ... 49: | |
2b13e716 | 1084 | idx = 4; |
04a0feae JA |
1085 | break; |
1086 | case 10 ... 19: | |
2b13e716 | 1087 | idx = 3; |
04a0feae JA |
1088 | break; |
1089 | case 4 ... 9: | |
2b13e716 | 1090 | idx = 2; |
04a0feae JA |
1091 | break; |
1092 | case 2 ... 3: | |
2b13e716 | 1093 | idx = 1; |
04a0feae JA |
1094 | case 0 ... 1: |
1095 | break; | |
1096 | } | |
1097 | ||
2b13e716 JA |
1098 | assert(idx < FIO_IO_U_LAT_U_NR); |
1099 | td->ts.io_u_lat_u[idx]++; | |
04a0feae JA |
1100 | } |
1101 | ||
d6bb626e | 1102 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec) |
ec118304 | 1103 | { |
2b13e716 | 1104 | int idx = 0; |
ec118304 | 1105 | |
d6bb626e VF |
1106 | assert(msec >= 1); |
1107 | ||
ec118304 JA |
1108 | switch (msec) { |
1109 | default: | |
2b13e716 | 1110 | idx = 11; |
04a0feae | 1111 | break; |
8abdce66 | 1112 | case 1000 ... 1999: |
2b13e716 | 1113 | idx = 10; |
04a0feae | 1114 | break; |
8abdce66 | 1115 | case 750 ... 999: |
2b13e716 | 1116 | idx = 9; |
04a0feae | 1117 | break; |
8abdce66 | 1118 | case 500 ... 749: |
2b13e716 | 1119 | idx = 8; |
04a0feae | 1120 | break; |
8abdce66 | 1121 | case 250 ... 499: |
2b13e716 | 1122 | idx = 7; |
04a0feae | 1123 | break; |
8abdce66 | 1124 | case 100 ... 249: |
2b13e716 | 1125 | idx = 6; |
04a0feae | 1126 | break; |
8abdce66 | 1127 | case 50 ... 99: |
2b13e716 | 1128 | idx = 5; |
04a0feae | 1129 | break; |
8abdce66 | 1130 | case 20 ... 49: |
2b13e716 | 1131 | idx = 4; |
04a0feae | 1132 | break; |
8abdce66 | 1133 | case 10 ... 19: |
2b13e716 | 1134 | idx = 3; |
04a0feae | 1135 | break; |
8abdce66 | 1136 | case 4 ... 9: |
2b13e716 | 1137 | idx = 2; |
04a0feae | 1138 | break; |
ec118304 | 1139 | case 2 ... 3: |
2b13e716 | 1140 | idx = 1; |
ec118304 JA |
1141 | case 0 ... 1: |
1142 | break; | |
1143 | } | |
1144 | ||
2b13e716 JA |
1145 | assert(idx < FIO_IO_U_LAT_M_NR); |
1146 | td->ts.io_u_lat_m[idx]++; | |
04a0feae JA |
1147 | } |
1148 | ||
d6bb626e | 1149 | static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec) |
04a0feae | 1150 | { |
d6bb626e VF |
1151 | if (nsec < 1000) |
1152 | io_u_mark_lat_nsec(td, nsec); | |
1153 | else if (nsec < 1000000) | |
1154 | io_u_mark_lat_usec(td, nsec / 1000); | |
04a0feae | 1155 | else |
d6bb626e | 1156 | io_u_mark_lat_msec(td, nsec / 1000000); |
ec118304 JA |
1157 | } |
1158 | ||
8c07860d JA |
1159 | static unsigned int __get_next_fileno_rand(struct thread_data *td) |
1160 | { | |
1161 | unsigned long fileno; | |
1162 | ||
1163 | if (td->o.file_service_type == FIO_FSERVICE_RANDOM) { | |
1164 | uint64_t frand_max = rand_max(&td->next_file_state); | |
1165 | unsigned long r; | |
1166 | ||
1167 | r = __rand(&td->next_file_state); | |
1168 | return (unsigned int) ((double) td->o.nr_files | |
1169 | * (r / (frand_max + 1.0))); | |
1170 | } | |
1171 | ||
1172 | if (td->o.file_service_type == FIO_FSERVICE_ZIPF) | |
1173 | fileno = zipf_next(&td->next_file_zipf); | |
1174 | else if (td->o.file_service_type == FIO_FSERVICE_PARETO) | |
1175 | fileno = pareto_next(&td->next_file_zipf); | |
1176 | else if (td->o.file_service_type == FIO_FSERVICE_GAUSS) | |
1177 | fileno = gauss_next(&td->next_file_gauss); | |
1178 | else { | |
1179 | log_err("fio: bad file service type: %d\n", td->o.file_service_type); | |
1180 | assert(0); | |
1181 | return 0; | |
1182 | } | |
1183 | ||
1184 | return fileno >> FIO_FSERVICE_SHIFT; | |
1185 | } | |
1186 | ||
0aabe160 JA |
1187 | /* |
1188 | * Get next file to service by choosing one at random | |
1189 | */ | |
2cc52930 JA |
1190 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
1191 | enum fio_file_flags goodf, | |
d6aed795 | 1192 | enum fio_file_flags badf) |
0aabe160 | 1193 | { |
0aabe160 | 1194 | struct fio_file *f; |
1c178180 | 1195 | int fno; |
0aabe160 JA |
1196 | |
1197 | do { | |
87b10676 | 1198 | int opened = 0; |
4c07ad86 | 1199 | |
8c07860d | 1200 | fno = __get_next_fileno_rand(td); |
7c83c089 | 1201 | |
126d65c6 | 1202 | f = td->files[fno]; |
d6aed795 | 1203 | if (fio_file_done(f)) |
059e63c0 | 1204 | continue; |
1c178180 | 1205 | |
d6aed795 | 1206 | if (!fio_file_open(f)) { |
87b10676 JA |
1207 | int err; |
1208 | ||
002fe734 JA |
1209 | if (td->nr_open_files >= td->o.open_files) |
1210 | return ERR_PTR(-EBUSY); | |
1211 | ||
87b10676 JA |
1212 | err = td_io_open_file(td, f); |
1213 | if (err) | |
1214 | continue; | |
1215 | opened = 1; | |
1216 | } | |
1217 | ||
2ba1c290 JA |
1218 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
1219 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); | |
0aabe160 | 1220 | return f; |
2ba1c290 | 1221 | } |
87b10676 JA |
1222 | if (opened) |
1223 | td_io_close_file(td, f); | |
0aabe160 JA |
1224 | } while (1); |
1225 | } | |
1226 | ||
1227 | /* | |
1228 | * Get next file to service by doing round robin between all available ones | |
1229 | */ | |
1c178180 JA |
1230 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
1231 | int badf) | |
3d7c391d JA |
1232 | { |
1233 | unsigned int old_next_file = td->next_file; | |
1234 | struct fio_file *f; | |
1235 | ||
1236 | do { | |
87b10676 JA |
1237 | int opened = 0; |
1238 | ||
126d65c6 | 1239 | f = td->files[td->next_file]; |
3d7c391d JA |
1240 | |
1241 | td->next_file++; | |
2dc1bbeb | 1242 | if (td->next_file >= td->o.nr_files) |
3d7c391d JA |
1243 | td->next_file = 0; |
1244 | ||
87b10676 | 1245 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
d6aed795 | 1246 | if (fio_file_done(f)) { |
d5ed68ea | 1247 | f = NULL; |
059e63c0 | 1248 | continue; |
d5ed68ea | 1249 | } |
059e63c0 | 1250 | |
d6aed795 | 1251 | if (!fio_file_open(f)) { |
87b10676 JA |
1252 | int err; |
1253 | ||
002fe734 JA |
1254 | if (td->nr_open_files >= td->o.open_files) |
1255 | return ERR_PTR(-EBUSY); | |
1256 | ||
87b10676 | 1257 | err = td_io_open_file(td, f); |
b5696bfc JA |
1258 | if (err) { |
1259 | dprint(FD_FILE, "error %d on open of %s\n", | |
1260 | err, f->file_name); | |
87c27b45 | 1261 | f = NULL; |
87b10676 | 1262 | continue; |
b5696bfc | 1263 | } |
87b10676 JA |
1264 | opened = 1; |
1265 | } | |
1266 | ||
0b9d69ec JA |
1267 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
1268 | f->flags); | |
1c178180 | 1269 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
3d7c391d JA |
1270 | break; |
1271 | ||
87b10676 JA |
1272 | if (opened) |
1273 | td_io_close_file(td, f); | |
1274 | ||
3d7c391d JA |
1275 | f = NULL; |
1276 | } while (td->next_file != old_next_file); | |
1277 | ||
2ba1c290 | 1278 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
3d7c391d JA |
1279 | return f; |
1280 | } | |
1281 | ||
7eb36574 | 1282 | static struct fio_file *__get_next_file(struct thread_data *td) |
bdb4e2e9 | 1283 | { |
1907dbc6 JA |
1284 | struct fio_file *f; |
1285 | ||
2dc1bbeb | 1286 | assert(td->o.nr_files <= td->files_index); |
1c178180 | 1287 | |
b5696bfc | 1288 | if (td->nr_done_files >= td->o.nr_files) { |
5ec10eaa JA |
1289 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
1290 | " nr_files=%d\n", td->nr_open_files, | |
1291 | td->nr_done_files, | |
1292 | td->o.nr_files); | |
bdb4e2e9 | 1293 | return NULL; |
2ba1c290 | 1294 | } |
bdb4e2e9 | 1295 | |
1907dbc6 | 1296 | f = td->file_service_file; |
d6aed795 | 1297 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
a086c257 JA |
1298 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
1299 | goto out; | |
1300 | if (td->file_service_left--) | |
1301 | goto out; | |
1302 | } | |
1907dbc6 | 1303 | |
a086c257 JA |
1304 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
1305 | td->o.file_service_type == FIO_FSERVICE_SEQ) | |
d6aed795 | 1306 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
bdb4e2e9 | 1307 | else |
d6aed795 | 1308 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
1907dbc6 | 1309 | |
002fe734 JA |
1310 | if (IS_ERR(f)) |
1311 | return f; | |
1312 | ||
1907dbc6 JA |
1313 | td->file_service_file = f; |
1314 | td->file_service_left = td->file_service_nr - 1; | |
2ba1c290 | 1315 | out: |
0dac421f JA |
1316 | if (f) |
1317 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); | |
1318 | else | |
1319 | dprint(FD_FILE, "get_next_file: NULL\n"); | |
1907dbc6 | 1320 | return f; |
bdb4e2e9 JA |
1321 | } |
1322 | ||
7eb36574 JA |
1323 | static struct fio_file *get_next_file(struct thread_data *td) |
1324 | { | |
372d8962 | 1325 | if (td->flags & TD_F_PROFILE_OPS) { |
d72be545 | 1326 | struct prof_io_ops *ops = &td->prof_io_ops; |
7eb36574 | 1327 | |
d72be545 JA |
1328 | if (ops->get_next_file) |
1329 | return ops->get_next_file(td); | |
1330 | } | |
7eb36574 JA |
1331 | |
1332 | return __get_next_file(td); | |
1333 | } | |
1334 | ||
002fe734 | 1335 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
429f6675 JA |
1336 | { |
1337 | struct fio_file *f; | |
1338 | ||
1339 | do { | |
1340 | f = get_next_file(td); | |
002fe734 JA |
1341 | if (IS_ERR_OR_NULL(f)) |
1342 | return PTR_ERR(f); | |
429f6675 | 1343 | |
429f6675 JA |
1344 | io_u->file = f; |
1345 | get_file(f); | |
1346 | ||
1347 | if (!fill_io_u(td, io_u)) | |
1348 | break; | |
1349 | ||
b5696bfc | 1350 | put_file_log(td, f); |
429f6675 | 1351 | td_io_close_file(td, f); |
b5696bfc | 1352 | io_u->file = NULL; |
8c07860d JA |
1353 | if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM) |
1354 | fio_file_reset(td, f); | |
1355 | else { | |
1356 | fio_file_set_done(f); | |
1357 | td->nr_done_files++; | |
1358 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, | |
0b9d69ec | 1359 | td->nr_done_files, td->o.nr_files); |
8c07860d | 1360 | } |
429f6675 JA |
1361 | } while (1); |
1362 | ||
1363 | return 0; | |
1364 | } | |
1365 | ||
3e260a46 JA |
1366 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
1367 | unsigned long tusec, unsigned long max_usec) | |
1368 | { | |
1369 | if (!td->error) | |
1370 | log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec); | |
1371 | td_verror(td, ETIMEDOUT, "max latency exceeded"); | |
1372 | icd->error = ETIMEDOUT; | |
1373 | } | |
1374 | ||
1375 | static void lat_new_cycle(struct thread_data *td) | |
1376 | { | |
1377 | fio_gettime(&td->latency_ts, NULL); | |
1378 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1379 | td->latency_failed = 0; | |
1380 | } | |
1381 | ||
1382 | /* | |
1383 | * We had an IO outside the latency target. Reduce the queue depth. If we | |
1384 | * are at QD=1, then it's time to give up. | |
1385 | */ | |
e39c0676 | 1386 | static bool __lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1387 | { |
1388 | if (td->latency_qd == 1) | |
e39c0676 | 1389 | return true; |
3e260a46 JA |
1390 | |
1391 | td->latency_qd_high = td->latency_qd; | |
6bb58215 JA |
1392 | |
1393 | if (td->latency_qd == td->latency_qd_low) | |
1394 | td->latency_qd_low--; | |
1395 | ||
3e260a46 JA |
1396 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
1397 | ||
1398 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
1399 | ||
1400 | /* | |
1401 | * When we ramp QD down, quiesce existing IO to prevent | |
1402 | * a storm of ramp downs due to pending higher depth. | |
1403 | */ | |
1404 | io_u_quiesce(td); | |
1405 | lat_new_cycle(td); | |
e39c0676 | 1406 | return false; |
3e260a46 JA |
1407 | } |
1408 | ||
e39c0676 | 1409 | static bool lat_target_failed(struct thread_data *td) |
3e260a46 JA |
1410 | { |
1411 | if (td->o.latency_percentile.u.f == 100.0) | |
1412 | return __lat_target_failed(td); | |
1413 | ||
1414 | td->latency_failed++; | |
e39c0676 | 1415 | return false; |
3e260a46 JA |
1416 | } |
1417 | ||
1418 | void lat_target_init(struct thread_data *td) | |
1419 | { | |
6bb58215 JA |
1420 | td->latency_end_run = 0; |
1421 | ||
3e260a46 JA |
1422 | if (td->o.latency_target) { |
1423 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); | |
1424 | fio_gettime(&td->latency_ts, NULL); | |
1425 | td->latency_qd = 1; | |
1426 | td->latency_qd_high = td->o.iodepth; | |
1427 | td->latency_qd_low = 1; | |
1428 | td->latency_ios = ddir_rw_sum(td->io_blocks); | |
1429 | } else | |
1430 | td->latency_qd = td->o.iodepth; | |
1431 | } | |
1432 | ||
6bb58215 JA |
1433 | void lat_target_reset(struct thread_data *td) |
1434 | { | |
1435 | if (!td->latency_end_run) | |
1436 | lat_target_init(td); | |
1437 | } | |
1438 | ||
3e260a46 JA |
1439 | static void lat_target_success(struct thread_data *td) |
1440 | { | |
1441 | const unsigned int qd = td->latency_qd; | |
6bb58215 | 1442 | struct thread_options *o = &td->o; |
3e260a46 JA |
1443 | |
1444 | td->latency_qd_low = td->latency_qd; | |
1445 | ||
1446 | /* | |
1447 | * If we haven't failed yet, we double up to a failing value instead | |
1448 | * of bisecting from highest possible queue depth. If we have set | |
1449 | * a limit other than td->o.iodepth, bisect between that. | |
1450 | */ | |
6bb58215 | 1451 | if (td->latency_qd_high != o->iodepth) |
3e260a46 JA |
1452 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
1453 | else | |
1454 | td->latency_qd *= 2; | |
1455 | ||
6bb58215 JA |
1456 | if (td->latency_qd > o->iodepth) |
1457 | td->latency_qd = o->iodepth; | |
3e260a46 JA |
1458 | |
1459 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); | |
6bb58215 | 1460 | |
3e260a46 | 1461 | /* |
6bb58215 JA |
1462 | * Same as last one, we are done. Let it run a latency cycle, so |
1463 | * we get only the results from the targeted depth. | |
3e260a46 | 1464 | */ |
6bb58215 JA |
1465 | if (td->latency_qd == qd) { |
1466 | if (td->latency_end_run) { | |
1467 | dprint(FD_RATE, "We are done\n"); | |
1468 | td->done = 1; | |
1469 | } else { | |
1470 | dprint(FD_RATE, "Quiesce and final run\n"); | |
1471 | io_u_quiesce(td); | |
1472 | td->latency_end_run = 1; | |
1473 | reset_all_stats(td); | |
1474 | reset_io_stats(td); | |
1475 | } | |
1476 | } | |
3e260a46 JA |
1477 | |
1478 | lat_new_cycle(td); | |
1479 | } | |
1480 | ||
1481 | /* | |
1482 | * Check if we can bump the queue depth | |
1483 | */ | |
1484 | void lat_target_check(struct thread_data *td) | |
1485 | { | |
1486 | uint64_t usec_window; | |
1487 | uint64_t ios; | |
1488 | double success_ios; | |
1489 | ||
1490 | usec_window = utime_since_now(&td->latency_ts); | |
1491 | if (usec_window < td->o.latency_window) | |
1492 | return; | |
1493 | ||
1494 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; | |
1495 | success_ios = (double) (ios - td->latency_failed) / (double) ios; | |
1496 | success_ios *= 100.0; | |
1497 | ||
1498 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); | |
1499 | ||
1500 | if (success_ios >= td->o.latency_percentile.u.f) | |
1501 | lat_target_success(td); | |
1502 | else | |
1503 | __lat_target_failed(td); | |
1504 | } | |
1505 | ||
1506 | /* | |
1507 | * If latency target is enabled, we might be ramping up or down and not | |
1508 | * using the full queue depth available. | |
1509 | */ | |
e39c0676 | 1510 | bool queue_full(const struct thread_data *td) |
3e260a46 JA |
1511 | { |
1512 | const int qempty = io_u_qempty(&td->io_u_freelist); | |
1513 | ||
1514 | if (qempty) | |
e39c0676 | 1515 | return true; |
3e260a46 | 1516 | if (!td->o.latency_target) |
e39c0676 | 1517 | return false; |
3e260a46 JA |
1518 | |
1519 | return td->cur_depth >= td->latency_qd; | |
1520 | } | |
429f6675 | 1521 | |
10ba535a JA |
1522 | struct io_u *__get_io_u(struct thread_data *td) |
1523 | { | |
0cae66f6 | 1524 | struct io_u *io_u = NULL; |
10ba535a | 1525 | |
ca09be4b JA |
1526 | if (td->stop_io) |
1527 | return NULL; | |
1528 | ||
e8462bd8 JA |
1529 | td_io_u_lock(td); |
1530 | ||
1531 | again: | |
2ae0b204 JA |
1532 | if (!io_u_rempty(&td->io_u_requeues)) |
1533 | io_u = io_u_rpop(&td->io_u_requeues); | |
3e260a46 | 1534 | else if (!queue_full(td)) { |
2ae0b204 | 1535 | io_u = io_u_qpop(&td->io_u_freelist); |
10ba535a | 1536 | |
225ba9e3 | 1537 | io_u->file = NULL; |
6040dabc | 1538 | io_u->buflen = 0; |
10ba535a | 1539 | io_u->resid = 0; |
d7762cf8 | 1540 | io_u->end_io = NULL; |
755200a3 JA |
1541 | } |
1542 | ||
1543 | if (io_u) { | |
0c6e7517 | 1544 | assert(io_u->flags & IO_U_F_FREE); |
1651e431 | 1545 | io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
f8b0bd10 JA |
1546 | IO_U_F_TRIMMED | IO_U_F_BARRIER | |
1547 | IO_U_F_VER_LIST); | |
0c6e7517 | 1548 | |
755200a3 | 1549 | io_u->error = 0; |
bcd5abfa | 1550 | io_u->acct_ddir = -1; |
10ba535a | 1551 | td->cur_depth++; |
a9da8ab2 | 1552 | assert(!(td->flags & TD_F_CHILD)); |
1651e431 | 1553 | io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH); |
f9401285 | 1554 | io_u->ipo = NULL; |
a9da8ab2 | 1555 | } else if (td_async_processing(td)) { |
1dec3e07 JA |
1556 | /* |
1557 | * We ran out, wait for async verify threads to finish and | |
1558 | * return one | |
1559 | */ | |
a9da8ab2 JA |
1560 | assert(!(td->flags & TD_F_CHILD)); |
1561 | assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock)); | |
1dec3e07 | 1562 | goto again; |
10ba535a JA |
1563 | } |
1564 | ||
e8462bd8 | 1565 | td_io_u_unlock(td); |
10ba535a JA |
1566 | return io_u; |
1567 | } | |
1568 | ||
e39c0676 | 1569 | static bool check_get_trim(struct thread_data *td, struct io_u *io_u) |
10ba535a | 1570 | { |
d72be545 | 1571 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
e39c0676 | 1572 | return false; |
d72be545 JA |
1573 | |
1574 | if (td->trim_entries) { | |
0d29de83 | 1575 | int get_trim = 0; |
10ba535a | 1576 | |
0d29de83 JA |
1577 | if (td->trim_batch) { |
1578 | td->trim_batch--; | |
1579 | get_trim = 1; | |
1580 | } else if (!(td->io_hist_len % td->o.trim_backlog) && | |
1581 | td->last_ddir != DDIR_READ) { | |
1582 | td->trim_batch = td->o.trim_batch; | |
1583 | if (!td->trim_batch) | |
1584 | td->trim_batch = td->o.trim_backlog; | |
1585 | get_trim = 1; | |
1586 | } | |
1587 | ||
917c64b3 | 1588 | if (get_trim && get_next_trim(td, io_u)) |
e39c0676 | 1589 | return true; |
2ba1c290 | 1590 | } |
10ba535a | 1591 | |
e39c0676 | 1592 | return false; |
0d29de83 JA |
1593 | } |
1594 | ||
e39c0676 | 1595 | static bool check_get_verify(struct thread_data *td, struct io_u *io_u) |
0d29de83 | 1596 | { |
d72be545 | 1597 | if (!(td->flags & TD_F_VER_BACKLOG)) |
e39c0676 | 1598 | return false; |
d72be545 JA |
1599 | |
1600 | if (td->io_hist_len) { | |
9e144189 JA |
1601 | int get_verify = 0; |
1602 | ||
d1ece0c7 | 1603 | if (td->verify_batch) |
9e144189 | 1604 | get_verify = 1; |
d1ece0c7 | 1605 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
9e144189 JA |
1606 | td->last_ddir != DDIR_READ) { |
1607 | td->verify_batch = td->o.verify_batch; | |
f8a75c99 JA |
1608 | if (!td->verify_batch) |
1609 | td->verify_batch = td->o.verify_backlog; | |
9e144189 JA |
1610 | get_verify = 1; |
1611 | } | |
1612 | ||
d1ece0c7 JA |
1613 | if (get_verify && !get_next_verify(td, io_u)) { |
1614 | td->verify_batch--; | |
e39c0676 | 1615 | return true; |
d1ece0c7 | 1616 | } |
9e144189 JA |
1617 | } |
1618 | ||
e39c0676 | 1619 | return false; |
0d29de83 JA |
1620 | } |
1621 | ||
de789769 JA |
1622 | /* |
1623 | * Fill offset and start time into the buffer content, to prevent too | |
23f394d5 JA |
1624 | * easy compressible data for simple de-dupe attempts. Do this for every |
1625 | * 512b block in the range, since that should be the smallest block size | |
1626 | * we can expect from a device. | |
de789769 JA |
1627 | */ |
1628 | static void small_content_scramble(struct io_u *io_u) | |
1629 | { | |
23f394d5 | 1630 | unsigned int i, nr_blocks = io_u->buflen / 512; |
1ae83d45 | 1631 | uint64_t boffset; |
23f394d5 | 1632 | unsigned int offset; |
fe8d7fee | 1633 | char *p, *end; |
de789769 | 1634 | |
23f394d5 JA |
1635 | if (!nr_blocks) |
1636 | return; | |
1637 | ||
1638 | p = io_u->xfer_buf; | |
fba76ee8 | 1639 | boffset = io_u->offset; |
81f0366c | 1640 | io_u->buf_filled_len = 0; |
fad82f76 | 1641 | |
23f394d5 JA |
1642 | for (i = 0; i < nr_blocks; i++) { |
1643 | /* | |
1644 | * Fill the byte offset into a "random" start offset of | |
1645 | * the buffer, given by the product of the usec time | |
1646 | * and the actual offset. | |
1647 | */ | |
8b6a404c | 1648 | offset = ((io_u->start_time.tv_nsec/1000) ^ boffset) & 511; |
1ae83d45 JA |
1649 | offset &= ~(sizeof(uint64_t) - 1); |
1650 | if (offset >= 512 - sizeof(uint64_t)) | |
1651 | offset -= sizeof(uint64_t); | |
fba76ee8 | 1652 | memcpy(p + offset, &boffset, sizeof(boffset)); |
23f394d5 JA |
1653 | |
1654 | end = p + 512 - sizeof(io_u->start_time); | |
1655 | memcpy(end, &io_u->start_time, sizeof(io_u->start_time)); | |
1656 | p += 512; | |
fad82f76 | 1657 | boffset += 512; |
23f394d5 | 1658 | } |
de789769 JA |
1659 | } |
1660 | ||
0d29de83 JA |
1661 | /* |
1662 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
1663 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
1664 | */ | |
1665 | struct io_u *get_io_u(struct thread_data *td) | |
1666 | { | |
1667 | struct fio_file *f; | |
1668 | struct io_u *io_u; | |
de789769 | 1669 | int do_scramble = 0; |
002fe734 | 1670 | long ret = 0; |
0d29de83 JA |
1671 | |
1672 | io_u = __get_io_u(td); | |
1673 | if (!io_u) { | |
1674 | dprint(FD_IO, "__get_io_u failed\n"); | |
1675 | return NULL; | |
1676 | } | |
1677 | ||
1678 | if (check_get_verify(td, io_u)) | |
1679 | goto out; | |
1680 | if (check_get_trim(td, io_u)) | |
1681 | goto out; | |
1682 | ||
755200a3 JA |
1683 | /* |
1684 | * from a requeue, io_u already setup | |
1685 | */ | |
1686 | if (io_u->file) | |
77f392bf | 1687 | goto out; |
755200a3 | 1688 | |
429f6675 JA |
1689 | /* |
1690 | * If using an iolog, grab next piece if any available. | |
1691 | */ | |
d72be545 | 1692 | if (td->flags & TD_F_READ_IOLOG) { |
429f6675 JA |
1693 | if (read_iolog_get(td, io_u)) |
1694 | goto err_put; | |
2ba1c290 | 1695 | } else if (set_io_u_file(td, io_u)) { |
002fe734 | 1696 | ret = -EBUSY; |
2ba1c290 | 1697 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
429f6675 | 1698 | goto err_put; |
2ba1c290 | 1699 | } |
5ec10eaa | 1700 | |
429f6675 | 1701 | f = io_u->file; |
002fe734 JA |
1702 | if (!f) { |
1703 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); | |
1704 | goto err_put; | |
1705 | } | |
1706 | ||
d6aed795 | 1707 | assert(fio_file_open(f)); |
97af62ce | 1708 | |
ff58fced | 1709 | if (ddir_rw(io_u->ddir)) { |
9b87f09b | 1710 | if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) { |
2ba1c290 | 1711 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
429f6675 | 1712 | goto err_put; |
2ba1c290 | 1713 | } |
10ba535a | 1714 | |
f1dfb668 JA |
1715 | f->last_start[io_u->ddir] = io_u->offset; |
1716 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; | |
10ba535a | 1717 | |
fd68418e | 1718 | if (io_u->ddir == DDIR_WRITE) { |
d72be545 | 1719 | if (td->flags & TD_F_REFILL_BUFFERS) { |
9c42684e | 1720 | io_u_fill_buffer(td, io_u, |
1066358a | 1721 | td->o.min_bs[DDIR_WRITE], |
9e129577 | 1722 | io_u->buflen); |
ff441ae8 | 1723 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
bedc9dc2 | 1724 | !(td->flags & TD_F_COMPRESS)) |
fd68418e | 1725 | do_scramble = 1; |
d72be545 | 1726 | if (td->flags & TD_F_VER_NONE) { |
629f1d71 JA |
1727 | populate_verify_io_u(td, io_u); |
1728 | do_scramble = 0; | |
1729 | } | |
fd68418e | 1730 | } else if (io_u->ddir == DDIR_READ) { |
cbe8d756 RR |
1731 | /* |
1732 | * Reset the buf_filled parameters so next time if the | |
1733 | * buffer is used for writes it is refilled. | |
1734 | */ | |
cbe8d756 RR |
1735 | io_u->buf_filled_len = 0; |
1736 | } | |
87dc1ab1 | 1737 | } |
10ba535a | 1738 | |
165faf16 JA |
1739 | /* |
1740 | * Set io data pointers. | |
1741 | */ | |
cec6b55d JA |
1742 | io_u->xfer_buf = io_u->buf; |
1743 | io_u->xfer_buflen = io_u->buflen; | |
5973cafb | 1744 | |
6ac7a331 | 1745 | out: |
0d29de83 | 1746 | assert(io_u->file); |
429f6675 | 1747 | if (!td_io_prep(td, io_u)) { |
3ecc8c67 | 1748 | if (!td->o.disable_lat) |
993bf48b | 1749 | fio_gettime(&io_u->start_time, NULL); |
03553853 | 1750 | |
de789769 JA |
1751 | if (do_scramble) |
1752 | small_content_scramble(io_u); | |
03553853 | 1753 | |
429f6675 | 1754 | return io_u; |
36167d82 | 1755 | } |
429f6675 | 1756 | err_put: |
2ba1c290 | 1757 | dprint(FD_IO, "get_io_u failed\n"); |
429f6675 | 1758 | put_io_u(td, io_u); |
002fe734 | 1759 | return ERR_PTR(ret); |
10ba535a JA |
1760 | } |
1761 | ||
a9da8ab2 | 1762 | static void __io_u_log_error(struct thread_data *td, struct io_u *io_u) |
5451792e | 1763 | { |
8b28bd41 | 1764 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
825f818e | 1765 | |
8b28bd41 DM |
1766 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
1767 | return; | |
5451792e | 1768 | |
709c8313 RE |
1769 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n", |
1770 | io_u->file ? " on file " : "", | |
1771 | io_u->file ? io_u->file->file_name : "", | |
1772 | strerror(io_u->error), | |
1773 | io_ddir_name(io_u->ddir), | |
1774 | io_u->offset, io_u->xfer_buflen); | |
5451792e | 1775 | |
5ad7be56 KD |
1776 | if (td->io_ops->errdetails) { |
1777 | char *err = td->io_ops->errdetails(io_u); | |
1778 | ||
1779 | log_err("fio: %s\n", err); | |
1780 | free(err); | |
1781 | } | |
1782 | ||
5451792e JA |
1783 | if (!td->error) |
1784 | td_verror(td, io_u->error, "io_u error"); | |
1785 | } | |
1786 | ||
a9da8ab2 JA |
1787 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
1788 | { | |
1789 | __io_u_log_error(td, io_u); | |
1790 | if (td->parent) | |
094e66cb | 1791 | __io_u_log_error(td->parent, io_u); |
a9da8ab2 JA |
1792 | } |
1793 | ||
e39c0676 | 1794 | static inline bool gtod_reduce(struct thread_data *td) |
aba6c951 | 1795 | { |
3ecc8c67 JA |
1796 | return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw) |
1797 | || td->o.gtod_reduce; | |
aba6c951 JA |
1798 | } |
1799 | ||
c8eeb9df JA |
1800 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
1801 | struct io_completion_data *icd, | |
1802 | const enum fio_ddir idx, unsigned int bytes) | |
1803 | { | |
a9da8ab2 | 1804 | const int no_reduce = !gtod_reduce(td); |
d6bb626e | 1805 | unsigned long long llnsec = 0; |
c8eeb9df | 1806 | |
75dc383e JA |
1807 | if (td->parent) |
1808 | td = td->parent; | |
1809 | ||
8243be59 JA |
1810 | if (!td->o.stats) |
1811 | return; | |
1812 | ||
a9da8ab2 | 1813 | if (no_reduce) |
d6bb626e | 1814 | llnsec = ntime_since(&io_u->issue_time, &icd->time); |
c8eeb9df JA |
1815 | |
1816 | if (!td->o.disable_lat) { | |
d6bb626e | 1817 | unsigned long long tnsec; |
c8eeb9df | 1818 | |
d6bb626e VF |
1819 | tnsec = ntime_since(&io_u->start_time, &icd->time); |
1820 | add_lat_sample(td, idx, tnsec, bytes, io_u->offset); | |
15501535 | 1821 | |
d4afedfd JA |
1822 | if (td->flags & TD_F_PROFILE_OPS) { |
1823 | struct prof_io_ops *ops = &td->prof_io_ops; | |
1824 | ||
1825 | if (ops->io_u_lat) | |
d6bb626e | 1826 | icd->error = ops->io_u_lat(td, tnsec/1000); |
d4afedfd JA |
1827 | } |
1828 | ||
d6bb626e VF |
1829 | if (td->o.max_latency && tnsec/1000 > td->o.max_latency) |
1830 | lat_fatal(td, icd, tnsec/1000, td->o.max_latency); | |
1831 | if (td->o.latency_target && tnsec/1000 > td->o.latency_target) { | |
3e260a46 | 1832 | if (lat_target_failed(td)) |
d6bb626e | 1833 | lat_fatal(td, icd, tnsec/1000, td->o.latency_target); |
15501535 | 1834 | } |
c8eeb9df JA |
1835 | } |
1836 | ||
a47591e4 JA |
1837 | if (ddir_rw(idx)) { |
1838 | if (!td->o.disable_clat) { | |
d6bb626e VF |
1839 | add_clat_sample(td, idx, llnsec, bytes, io_u->offset); |
1840 | io_u_mark_latency(td, llnsec); | |
a47591e4 | 1841 | } |
c8eeb9df | 1842 | |
a47591e4 | 1843 | if (!td->o.disable_bw && per_unit_log(td->bw_log)) |
d6bb626e | 1844 | add_bw_sample(td, io_u, bytes, llnsec); |
c8eeb9df | 1845 | |
a47591e4 JA |
1846 | if (no_reduce && per_unit_log(td->iops_log)) |
1847 | add_iops_sample(td, io_u, bytes); | |
1848 | } | |
66347cfa DE |
1849 | |
1850 | if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) { | |
1851 | uint32_t *info = io_u_block_info(td, io_u); | |
1852 | if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) { | |
1853 | if (io_u->ddir == DDIR_TRIM) { | |
1854 | *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, | |
1855 | BLOCK_INFO_TRIMS(*info) + 1); | |
1856 | } else if (io_u->ddir == DDIR_WRITE) { | |
1857 | *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN, | |
1858 | *info); | |
1859 | } | |
1860 | } | |
1861 | } | |
c8eeb9df JA |
1862 | } |
1863 | ||
94a6e1bb JA |
1864 | static void file_log_write_comp(const struct thread_data *td, struct fio_file *f, |
1865 | uint64_t offset, unsigned int bytes) | |
1866 | { | |
1867 | int idx; | |
1868 | ||
639ad1ea JA |
1869 | if (!f) |
1870 | return; | |
1871 | ||
94a6e1bb JA |
1872 | if (f->first_write == -1ULL || offset < f->first_write) |
1873 | f->first_write = offset; | |
1874 | if (f->last_write == -1ULL || ((offset + bytes) > f->last_write)) | |
1875 | f->last_write = offset + bytes; | |
1876 | ||
1877 | if (!f->last_write_comp) | |
1878 | return; | |
1879 | ||
1880 | idx = f->last_write_idx++; | |
1881 | f->last_write_comp[idx] = offset; | |
1882 | if (f->last_write_idx == td->o.iodepth) | |
1883 | f->last_write_idx = 0; | |
1884 | } | |
1885 | ||
f8b0bd10 | 1886 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
97601024 | 1887 | struct io_completion_data *icd) |
10ba535a | 1888 | { |
f8b0bd10 JA |
1889 | struct io_u *io_u = *io_u_ptr; |
1890 | enum fio_ddir ddir = io_u->ddir; | |
1891 | struct fio_file *f = io_u->file; | |
10ba535a | 1892 | |
2ba1c290 JA |
1893 | dprint_io_u(io_u, "io complete"); |
1894 | ||
0c6e7517 | 1895 | assert(io_u->flags & IO_U_F_FLIGHT); |
1651e431 | 1896 | io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
f9401285 JA |
1897 | |
1898 | /* | |
1899 | * Mark IO ok to verify | |
1900 | */ | |
1901 | if (io_u->ipo) { | |
890b6656 JA |
1902 | /* |
1903 | * Remove errored entry from the verification list | |
1904 | */ | |
1905 | if (io_u->error) | |
1906 | unlog_io_piece(td, io_u); | |
1907 | else { | |
1908 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; | |
1909 | write_barrier(); | |
1910 | } | |
f9401285 JA |
1911 | } |
1912 | ||
f8b0bd10 | 1913 | if (ddir_sync(ddir)) { |
87dc1ab1 | 1914 | td->last_was_sync = 1; |
44f29692 JA |
1915 | if (f) { |
1916 | f->first_write = -1ULL; | |
1917 | f->last_write = -1ULL; | |
1918 | } | |
87dc1ab1 JA |
1919 | return; |
1920 | } | |
1921 | ||
1922 | td->last_was_sync = 0; | |
f8b0bd10 | 1923 | td->last_ddir = ddir; |
87dc1ab1 | 1924 | |
f8b0bd10 | 1925 | if (!io_u->error && ddir_rw(ddir)) { |
10ba535a | 1926 | unsigned int bytes = io_u->buflen - io_u->resid; |
b29ee5b3 | 1927 | int ret; |
10ba535a | 1928 | |
f8b0bd10 JA |
1929 | td->io_blocks[ddir]++; |
1930 | td->this_io_blocks[ddir]++; | |
1931 | td->io_bytes[ddir] += bytes; | |
ae2fafc8 | 1932 | |
1933 | if (!(io_u->flags & IO_U_F_VER_LIST)) | |
f8b0bd10 JA |
1934 | td->this_io_bytes[ddir] += bytes; |
1935 | ||
639ad1ea | 1936 | if (ddir == DDIR_WRITE) |
94a6e1bb | 1937 | file_log_write_comp(td, f, io_u->offset, bytes); |
44f29692 | 1938 | |
6b1190fd | 1939 | if (ramp_time_over(td) && (td->runstate == TD_RUNNING || |
50a8ce86 | 1940 | td->runstate == TD_VERIFYING)) |
f8b0bd10 | 1941 | account_io_completion(td, io_u, icd, ddir, bytes); |
40e1a6f0 | 1942 | |
f8b0bd10 | 1943 | icd->bytes_done[ddir] += bytes; |
3af6ef39 | 1944 | |
d7762cf8 | 1945 | if (io_u->end_io) { |
f8b0bd10 JA |
1946 | ret = io_u->end_io(td, io_u_ptr); |
1947 | io_u = *io_u_ptr; | |
3af6ef39 JA |
1948 | if (ret && !icd->error) |
1949 | icd->error = ret; | |
1950 | } | |
ff58fced | 1951 | } else if (io_u->error) { |
10ba535a | 1952 | icd->error = io_u->error; |
5451792e JA |
1953 | io_u_log_error(td, io_u); |
1954 | } | |
8b28bd41 | 1955 | if (icd->error) { |
f8b0bd10 JA |
1956 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
1957 | ||
8b28bd41 DM |
1958 | if (!td_non_fatal_error(td, eb, icd->error)) |
1959 | return; | |
f8b0bd10 | 1960 | |
f2bba182 RR |
1961 | /* |
1962 | * If there is a non_fatal error, then add to the error count | |
1963 | * and clear all the errors. | |
1964 | */ | |
1965 | update_error_count(td, icd->error); | |
1966 | td_clear_error(td); | |
1967 | icd->error = 0; | |
f8b0bd10 JA |
1968 | if (io_u) |
1969 | io_u->error = 0; | |
f2bba182 | 1970 | } |
10ba535a JA |
1971 | } |
1972 | ||
9520ebb9 JA |
1973 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
1974 | int nr) | |
10ba535a | 1975 | { |
6eaf09d6 | 1976 | int ddir; |
aba6c951 JA |
1977 | |
1978 | if (!gtod_reduce(td)) | |
9520ebb9 | 1979 | fio_gettime(&icd->time, NULL); |
02bcaa8c | 1980 | |
3af6ef39 JA |
1981 | icd->nr = nr; |
1982 | ||
10ba535a | 1983 | icd->error = 0; |
c1f50f76 | 1984 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
6eaf09d6 | 1985 | icd->bytes_done[ddir] = 0; |
36167d82 JA |
1986 | } |
1987 | ||
97601024 JA |
1988 | static void ios_completed(struct thread_data *td, |
1989 | struct io_completion_data *icd) | |
36167d82 JA |
1990 | { |
1991 | struct io_u *io_u; | |
1992 | int i; | |
1993 | ||
10ba535a JA |
1994 | for (i = 0; i < icd->nr; i++) { |
1995 | io_u = td->io_ops->event(td, i); | |
1996 | ||
f8b0bd10 | 1997 | io_completed(td, &io_u, icd); |
e8462bd8 | 1998 | |
f8b0bd10 | 1999 | if (io_u) |
e8462bd8 | 2000 | put_io_u(td, io_u); |
10ba535a JA |
2001 | } |
2002 | } | |
97601024 | 2003 | |
e7e6cfb4 JA |
2004 | /* |
2005 | * Complete a single io_u for the sync engines. | |
2006 | */ | |
55312f9f | 2007 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
97601024 JA |
2008 | { |
2009 | struct io_completion_data icd; | |
55312f9f | 2010 | int ddir; |
97601024 | 2011 | |
9520ebb9 | 2012 | init_icd(td, &icd, 1); |
f8b0bd10 | 2013 | io_completed(td, &io_u, &icd); |
e8462bd8 | 2014 | |
f8b0bd10 | 2015 | if (io_u) |
e8462bd8 | 2016 | put_io_u(td, io_u); |
97601024 | 2017 | |
581e7141 JA |
2018 | if (icd.error) { |
2019 | td_verror(td, icd.error, "io_u_sync_complete"); | |
2020 | return -1; | |
2021 | } | |
97601024 | 2022 | |
c1f50f76 | 2023 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
55312f9f | 2024 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
581e7141 JA |
2025 | |
2026 | return 0; | |
97601024 JA |
2027 | } |
2028 | ||
e7e6cfb4 JA |
2029 | /* |
2030 | * Called to complete min_events number of io for the async engines. | |
2031 | */ | |
55312f9f | 2032 | int io_u_queued_complete(struct thread_data *td, int min_evts) |
97601024 | 2033 | { |
97601024 | 2034 | struct io_completion_data icd; |
00de55ef | 2035 | struct timespec *tvp = NULL; |
55312f9f | 2036 | int ret, ddir; |
4d06a338 | 2037 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
97601024 | 2038 | |
12bb8569 | 2039 | dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts); |
b271fe62 | 2040 | |
4950421a | 2041 | if (!min_evts) |
00de55ef | 2042 | tvp = &ts; |
5fb4b366 RE |
2043 | else if (min_evts > td->cur_depth) |
2044 | min_evts = td->cur_depth; | |
97601024 | 2045 | |
82407585 RP |
2046 | /* No worries, td_io_getevents fixes min and max if they are |
2047 | * set incorrectly */ | |
2048 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp); | |
97601024 | 2049 | if (ret < 0) { |
e1161c32 | 2050 | td_verror(td, -ret, "td_io_getevents"); |
97601024 JA |
2051 | return ret; |
2052 | } else if (!ret) | |
2053 | return ret; | |
2054 | ||
9520ebb9 | 2055 | init_icd(td, &icd, ret); |
97601024 | 2056 | ios_completed(td, &icd); |
581e7141 JA |
2057 | if (icd.error) { |
2058 | td_verror(td, icd.error, "io_u_queued_complete"); | |
2059 | return -1; | |
2060 | } | |
97601024 | 2061 | |
c1f50f76 | 2062 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
55312f9f | 2063 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
581e7141 | 2064 | |
0d593542 | 2065 | return ret; |
97601024 | 2066 | } |
7e77dd02 JA |
2067 | |
2068 | /* | |
2069 | * Call when io_u is really queued, to update the submission latency. | |
2070 | */ | |
2071 | void io_u_queued(struct thread_data *td, struct io_u *io_u) | |
2072 | { | |
8243be59 | 2073 | if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) { |
9520ebb9 | 2074 | unsigned long slat_time; |
7e77dd02 | 2075 | |
d6bb626e | 2076 | slat_time = ntime_since(&io_u->start_time, &io_u->issue_time); |
75dc383e JA |
2077 | |
2078 | if (td->parent) | |
2079 | td = td->parent; | |
2080 | ||
ae588852 JA |
2081 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
2082 | io_u->offset); | |
9520ebb9 | 2083 | } |
7e77dd02 | 2084 | } |
433afcb4 | 2085 | |
5c94b008 JA |
2086 | /* |
2087 | * See if we should reuse the last seed, if dedupe is enabled | |
2088 | */ | |
9451b93e | 2089 | static struct frand_state *get_buf_state(struct thread_data *td) |
5c94b008 JA |
2090 | { |
2091 | unsigned int v; | |
5c94b008 JA |
2092 | |
2093 | if (!td->o.dedupe_percentage) | |
2094 | return &td->buf_state; | |
732eedd0 | 2095 | else if (td->o.dedupe_percentage == 100) { |
9451b93e JA |
2096 | frand_copy(&td->buf_state_prev, &td->buf_state); |
2097 | return &td->buf_state; | |
732eedd0 | 2098 | } |
5c94b008 | 2099 | |
36dd3379 | 2100 | v = rand32_between(&td->dedupe_state, 1, 100); |
5c94b008 JA |
2101 | |
2102 | if (v <= td->o.dedupe_percentage) | |
2103 | return &td->buf_state_prev; | |
2104 | ||
2105 | return &td->buf_state; | |
2106 | } | |
2107 | ||
9451b93e | 2108 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) |
5c94b008 | 2109 | { |
9451b93e JA |
2110 | if (td->o.dedupe_percentage == 100) |
2111 | frand_copy(rs, &td->buf_state_prev); | |
2112 | else if (rs == &td->buf_state) | |
5c94b008 JA |
2113 | frand_copy(&td->buf_state_prev, rs); |
2114 | } | |
2115 | ||
cc86c395 JA |
2116 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write, |
2117 | unsigned int max_bs) | |
5973cafb | 2118 | { |
d1af2894 JA |
2119 | struct thread_options *o = &td->o; |
2120 | ||
15600335 JA |
2121 | if (o->mem_type == MEM_CUDA_MALLOC) |
2122 | return; | |
03553853 | 2123 | |
4eff3e57 | 2124 | if (o->compress_percentage || o->dedupe_percentage) { |
9c42684e | 2125 | unsigned int perc = td->o.compress_percentage; |
5c94b008 | 2126 | struct frand_state *rs; |
1066358a | 2127 | unsigned int left = max_bs; |
1e7f82e2 | 2128 | unsigned int this_write; |
5c94b008 | 2129 | |
1066358a | 2130 | do { |
9451b93e | 2131 | rs = get_buf_state(td); |
9c42684e | 2132 | |
1066358a | 2133 | min_write = min(min_write, left); |
f97a43a1 | 2134 | |
1066358a | 2135 | if (perc) { |
4b157ac6 JA |
2136 | this_write = min_not_zero(min_write, |
2137 | td->o.compress_chunk); | |
1e7f82e2 JA |
2138 | |
2139 | fill_random_buf_percentage(rs, buf, perc, | |
2140 | this_write, this_write, | |
2141 | o->buffer_pattern, | |
2142 | o->buffer_pattern_bytes); | |
2143 | } else { | |
1066358a | 2144 | fill_random_buf(rs, buf, min_write); |
1e7f82e2 JA |
2145 | this_write = min_write; |
2146 | } | |
1066358a | 2147 | |
1e7f82e2 JA |
2148 | buf += this_write; |
2149 | left -= this_write; | |
9451b93e | 2150 | save_buf_state(td, rs); |
1066358a | 2151 | } while (left); |
d1af2894 JA |
2152 | } else if (o->buffer_pattern_bytes) |
2153 | fill_buffer_pattern(td, buf, max_bs); | |
999d245e | 2154 | else if (o->zero_buffers) |
cc86c395 | 2155 | memset(buf, 0, max_bs); |
999d245e | 2156 | else |
9451b93e | 2157 | fill_random_buf(get_buf_state(td), buf, max_bs); |
cc86c395 JA |
2158 | } |
2159 | ||
2160 | /* | |
2161 | * "randomly" fill the buffer contents | |
2162 | */ | |
2163 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, | |
2164 | unsigned int min_write, unsigned int max_bs) | |
2165 | { | |
2166 | io_u->buf_filled_len = 0; | |
2167 | fill_io_buffer(td, io_u->buf, min_write, max_bs); | |
5973cafb | 2168 | } |
e2c75fc4 TK |
2169 | |
2170 | static int do_sync_file_range(const struct thread_data *td, | |
2171 | struct fio_file *f) | |
2172 | { | |
2173 | off64_t offset, nbytes; | |
2174 | ||
2175 | offset = f->first_write; | |
2176 | nbytes = f->last_write - f->first_write; | |
2177 | ||
2178 | if (!nbytes) | |
2179 | return 0; | |
2180 | ||
2181 | return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range); | |
2182 | } | |
2183 | ||
2184 | int do_io_u_sync(const struct thread_data *td, struct io_u *io_u) | |
2185 | { | |
2186 | int ret; | |
2187 | ||
2188 | if (io_u->ddir == DDIR_SYNC) { | |
2189 | ret = fsync(io_u->file->fd); | |
2190 | } else if (io_u->ddir == DDIR_DATASYNC) { | |
2191 | #ifdef CONFIG_FDATASYNC | |
2192 | ret = fdatasync(io_u->file->fd); | |
2193 | #else | |
2194 | ret = io_u->xfer_buflen; | |
2195 | io_u->error = EINVAL; | |
2196 | #endif | |
2197 | } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE) | |
2198 | ret = do_sync_file_range(td, io_u->file); | |
2199 | else { | |
2200 | ret = io_u->xfer_buflen; | |
2201 | io_u->error = EINVAL; | |
2202 | } | |
2203 | ||
2204 | if (ret < 0) | |
2205 | io_u->error = errno; | |
2206 | ||
2207 | return ret; | |
2208 | } | |
2209 | ||
2210 | int do_io_u_trim(const struct thread_data *td, struct io_u *io_u) | |
2211 | { | |
2212 | #ifndef FIO_HAVE_TRIM | |
2213 | io_u->error = EINVAL; | |
2214 | return 0; | |
2215 | #else | |
2216 | struct fio_file *f = io_u->file; | |
2217 | int ret; | |
2218 | ||
496b1f9e | 2219 | ret = os_trim(f, io_u->offset, io_u->xfer_buflen); |
e2c75fc4 TK |
2220 | if (!ret) |
2221 | return io_u->xfer_buflen; | |
2222 | ||
2223 | io_u->error = ret; | |
2224 | return 0; | |
2225 | #endif | |
2226 | } |