| 1 | #include <unistd.h> |
| 2 | #include <string.h> |
| 3 | #include <assert.h> |
| 4 | |
| 5 | #include "fio.h" |
| 6 | #include "verify.h" |
| 7 | #include "trim.h" |
| 8 | #include "lib/rand.h" |
| 9 | #include "lib/axmap.h" |
| 10 | #include "err.h" |
| 11 | #include "lib/pow2.h" |
| 12 | #include "minmax.h" |
| 13 | |
| 14 | struct io_completion_data { |
| 15 | int nr; /* input */ |
| 16 | |
| 17 | int error; /* output */ |
| 18 | uint64_t bytes_done[DDIR_RWDIR_CNT]; /* output */ |
| 19 | struct timespec time; /* output */ |
| 20 | }; |
| 21 | |
| 22 | /* |
| 23 | * The ->io_axmap contains a map of blocks we have or have not done io |
| 24 | * to yet. Used to make sure we cover the entire range in a fair fashion. |
| 25 | */ |
| 26 | static bool random_map_free(struct fio_file *f, const uint64_t block) |
| 27 | { |
| 28 | return !axmap_isset(f->io_axmap, block); |
| 29 | } |
| 30 | |
| 31 | /* |
| 32 | * Mark a given offset as used in the map. |
| 33 | */ |
| 34 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
| 35 | { |
| 36 | unsigned int min_bs = td->o.min_bs[io_u->ddir]; |
| 37 | struct fio_file *f = io_u->file; |
| 38 | unsigned int nr_blocks; |
| 39 | uint64_t block; |
| 40 | |
| 41 | block = (io_u->offset - f->file_offset) / (uint64_t) min_bs; |
| 42 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
| 43 | |
| 44 | if (!(io_u->flags & IO_U_F_BUSY_OK)) |
| 45 | nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks); |
| 46 | |
| 47 | if ((nr_blocks * min_bs) < io_u->buflen) |
| 48 | io_u->buflen = nr_blocks * min_bs; |
| 49 | } |
| 50 | |
| 51 | static uint64_t last_block(struct thread_data *td, struct fio_file *f, |
| 52 | enum fio_ddir ddir) |
| 53 | { |
| 54 | uint64_t max_blocks; |
| 55 | uint64_t max_size; |
| 56 | |
| 57 | assert(ddir_rw(ddir)); |
| 58 | |
| 59 | /* |
| 60 | * Hmm, should we make sure that ->io_size <= ->real_file_size? |
| 61 | * -> not for now since there is code assuming it could go either. |
| 62 | */ |
| 63 | max_size = f->io_size; |
| 64 | if (max_size > f->real_file_size) |
| 65 | max_size = f->real_file_size; |
| 66 | |
| 67 | if (td->o.zone_range) |
| 68 | max_size = td->o.zone_range; |
| 69 | |
| 70 | if (td->o.min_bs[ddir] > td->o.ba[ddir]) |
| 71 | max_size -= td->o.min_bs[ddir] - td->o.ba[ddir]; |
| 72 | |
| 73 | max_blocks = max_size / (uint64_t) td->o.ba[ddir]; |
| 74 | if (!max_blocks) |
| 75 | return 0; |
| 76 | |
| 77 | return max_blocks; |
| 78 | } |
| 79 | |
| 80 | struct rand_off { |
| 81 | struct flist_head list; |
| 82 | uint64_t off; |
| 83 | }; |
| 84 | |
| 85 | static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
| 86 | enum fio_ddir ddir, uint64_t *b, |
| 87 | uint64_t lastb) |
| 88 | { |
| 89 | uint64_t r; |
| 90 | |
| 91 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE || |
| 92 | td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) { |
| 93 | |
| 94 | r = __rand(&td->random_state); |
| 95 | |
| 96 | dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r); |
| 97 | |
| 98 | *b = lastb * (r / (rand_max(&td->random_state) + 1.0)); |
| 99 | } else { |
| 100 | uint64_t off = 0; |
| 101 | |
| 102 | assert(fio_file_lfsr(f)); |
| 103 | |
| 104 | if (lfsr_next(&f->lfsr, &off)) |
| 105 | return 1; |
| 106 | |
| 107 | *b = off; |
| 108 | } |
| 109 | |
| 110 | /* |
| 111 | * if we are not maintaining a random map, we are done. |
| 112 | */ |
| 113 | if (!file_randommap(td, f)) |
| 114 | goto ret; |
| 115 | |
| 116 | /* |
| 117 | * calculate map offset and check if it's free |
| 118 | */ |
| 119 | if (random_map_free(f, *b)) |
| 120 | goto ret; |
| 121 | |
| 122 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
| 123 | (unsigned long long) *b); |
| 124 | |
| 125 | *b = axmap_next_free(f->io_axmap, *b); |
| 126 | if (*b == (uint64_t) -1ULL) |
| 127 | return 1; |
| 128 | ret: |
| 129 | return 0; |
| 130 | } |
| 131 | |
| 132 | static int __get_next_rand_offset_zipf(struct thread_data *td, |
| 133 | struct fio_file *f, enum fio_ddir ddir, |
| 134 | uint64_t *b) |
| 135 | { |
| 136 | *b = zipf_next(&f->zipf); |
| 137 | return 0; |
| 138 | } |
| 139 | |
| 140 | static int __get_next_rand_offset_pareto(struct thread_data *td, |
| 141 | struct fio_file *f, enum fio_ddir ddir, |
| 142 | uint64_t *b) |
| 143 | { |
| 144 | *b = pareto_next(&f->zipf); |
| 145 | return 0; |
| 146 | } |
| 147 | |
| 148 | static int __get_next_rand_offset_gauss(struct thread_data *td, |
| 149 | struct fio_file *f, enum fio_ddir ddir, |
| 150 | uint64_t *b) |
| 151 | { |
| 152 | *b = gauss_next(&f->gauss); |
| 153 | return 0; |
| 154 | } |
| 155 | |
| 156 | static int __get_next_rand_offset_zoned_abs(struct thread_data *td, |
| 157 | struct fio_file *f, |
| 158 | enum fio_ddir ddir, uint64_t *b) |
| 159 | { |
| 160 | struct zone_split_index *zsi; |
| 161 | uint64_t lastb, send, stotal; |
| 162 | unsigned int v; |
| 163 | |
| 164 | lastb = last_block(td, f, ddir); |
| 165 | if (!lastb) |
| 166 | return 1; |
| 167 | |
| 168 | if (!td->o.zone_split_nr[ddir]) { |
| 169 | bail: |
| 170 | return __get_next_rand_offset(td, f, ddir, b, lastb); |
| 171 | } |
| 172 | |
| 173 | /* |
| 174 | * Generate a value, v, between 1 and 100, both inclusive |
| 175 | */ |
| 176 | v = rand32_between(&td->zone_state, 1, 100); |
| 177 | |
| 178 | /* |
| 179 | * Find our generated table. 'send' is the end block of this zone, |
| 180 | * 'stotal' is our start offset. |
| 181 | */ |
| 182 | zsi = &td->zone_state_index[ddir][v - 1]; |
| 183 | stotal = zsi->size_prev / td->o.ba[ddir]; |
| 184 | send = zsi->size / td->o.ba[ddir]; |
| 185 | |
| 186 | /* |
| 187 | * Should never happen |
| 188 | */ |
| 189 | if (send == -1U) { |
| 190 | if (!fio_did_warn(FIO_WARN_ZONED_BUG)) |
| 191 | log_err("fio: bug in zoned generation\n"); |
| 192 | goto bail; |
| 193 | } else if (send > lastb) { |
| 194 | /* |
| 195 | * This happens if the user specifies ranges that exceed |
| 196 | * the file/device size. We can't handle that gracefully, |
| 197 | * so error and exit. |
| 198 | */ |
| 199 | log_err("fio: zoned_abs sizes exceed file size\n"); |
| 200 | return 1; |
| 201 | } |
| 202 | |
| 203 | /* |
| 204 | * Generate index from 0..send-stotal |
| 205 | */ |
| 206 | if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1) |
| 207 | return 1; |
| 208 | |
| 209 | *b += stotal; |
| 210 | return 0; |
| 211 | } |
| 212 | |
| 213 | static int __get_next_rand_offset_zoned(struct thread_data *td, |
| 214 | struct fio_file *f, enum fio_ddir ddir, |
| 215 | uint64_t *b) |
| 216 | { |
| 217 | unsigned int v, send, stotal; |
| 218 | uint64_t offset, lastb; |
| 219 | struct zone_split_index *zsi; |
| 220 | |
| 221 | lastb = last_block(td, f, ddir); |
| 222 | if (!lastb) |
| 223 | return 1; |
| 224 | |
| 225 | if (!td->o.zone_split_nr[ddir]) { |
| 226 | bail: |
| 227 | return __get_next_rand_offset(td, f, ddir, b, lastb); |
| 228 | } |
| 229 | |
| 230 | /* |
| 231 | * Generate a value, v, between 1 and 100, both inclusive |
| 232 | */ |
| 233 | v = rand32_between(&td->zone_state, 1, 100); |
| 234 | |
| 235 | zsi = &td->zone_state_index[ddir][v - 1]; |
| 236 | stotal = zsi->size_perc_prev; |
| 237 | send = zsi->size_perc; |
| 238 | |
| 239 | /* |
| 240 | * Should never happen |
| 241 | */ |
| 242 | if (send == -1U) { |
| 243 | if (!fio_did_warn(FIO_WARN_ZONED_BUG)) |
| 244 | log_err("fio: bug in zoned generation\n"); |
| 245 | goto bail; |
| 246 | } |
| 247 | |
| 248 | /* |
| 249 | * 'send' is some percentage below or equal to 100 that |
| 250 | * marks the end of the current IO range. 'stotal' marks |
| 251 | * the start, in percent. |
| 252 | */ |
| 253 | if (stotal) |
| 254 | offset = stotal * lastb / 100ULL; |
| 255 | else |
| 256 | offset = 0; |
| 257 | |
| 258 | lastb = lastb * (send - stotal) / 100ULL; |
| 259 | |
| 260 | /* |
| 261 | * Generate index from 0..send-of-lastb |
| 262 | */ |
| 263 | if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1) |
| 264 | return 1; |
| 265 | |
| 266 | /* |
| 267 | * Add our start offset, if any |
| 268 | */ |
| 269 | if (offset) |
| 270 | *b += offset; |
| 271 | |
| 272 | return 0; |
| 273 | } |
| 274 | |
| 275 | static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b) |
| 276 | { |
| 277 | struct rand_off *r1 = flist_entry(a, struct rand_off, list); |
| 278 | struct rand_off *r2 = flist_entry(b, struct rand_off, list); |
| 279 | |
| 280 | return r1->off - r2->off; |
| 281 | } |
| 282 | |
| 283 | static int get_off_from_method(struct thread_data *td, struct fio_file *f, |
| 284 | enum fio_ddir ddir, uint64_t *b) |
| 285 | { |
| 286 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) { |
| 287 | uint64_t lastb; |
| 288 | |
| 289 | lastb = last_block(td, f, ddir); |
| 290 | if (!lastb) |
| 291 | return 1; |
| 292 | |
| 293 | return __get_next_rand_offset(td, f, ddir, b, lastb); |
| 294 | } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) |
| 295 | return __get_next_rand_offset_zipf(td, f, ddir, b); |
| 296 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
| 297 | return __get_next_rand_offset_pareto(td, f, ddir, b); |
| 298 | else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS) |
| 299 | return __get_next_rand_offset_gauss(td, f, ddir, b); |
| 300 | else if (td->o.random_distribution == FIO_RAND_DIST_ZONED) |
| 301 | return __get_next_rand_offset_zoned(td, f, ddir, b); |
| 302 | else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS) |
| 303 | return __get_next_rand_offset_zoned_abs(td, f, ddir, b); |
| 304 | |
| 305 | log_err("fio: unknown random distribution: %d\n", td->o.random_distribution); |
| 306 | return 1; |
| 307 | } |
| 308 | |
| 309 | /* |
| 310 | * Sort the reads for a verify phase in batches of verifysort_nr, if |
| 311 | * specified. |
| 312 | */ |
| 313 | static inline bool should_sort_io(struct thread_data *td) |
| 314 | { |
| 315 | if (!td->o.verifysort_nr || !td->o.do_verify) |
| 316 | return false; |
| 317 | if (!td_random(td)) |
| 318 | return false; |
| 319 | if (td->runstate != TD_VERIFYING) |
| 320 | return false; |
| 321 | if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE || |
| 322 | td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) |
| 323 | return false; |
| 324 | |
| 325 | return true; |
| 326 | } |
| 327 | |
| 328 | static bool should_do_random(struct thread_data *td, enum fio_ddir ddir) |
| 329 | { |
| 330 | unsigned int v; |
| 331 | |
| 332 | if (td->o.perc_rand[ddir] == 100) |
| 333 | return true; |
| 334 | |
| 335 | v = rand32_between(&td->seq_rand_state[ddir], 1, 100); |
| 336 | |
| 337 | return v <= td->o.perc_rand[ddir]; |
| 338 | } |
| 339 | |
| 340 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
| 341 | enum fio_ddir ddir, uint64_t *b) |
| 342 | { |
| 343 | struct rand_off *r; |
| 344 | int i, ret = 1; |
| 345 | |
| 346 | if (!should_sort_io(td)) |
| 347 | return get_off_from_method(td, f, ddir, b); |
| 348 | |
| 349 | if (!flist_empty(&td->next_rand_list)) { |
| 350 | fetch: |
| 351 | r = flist_first_entry(&td->next_rand_list, struct rand_off, list); |
| 352 | flist_del(&r->list); |
| 353 | *b = r->off; |
| 354 | free(r); |
| 355 | return 0; |
| 356 | } |
| 357 | |
| 358 | for (i = 0; i < td->o.verifysort_nr; i++) { |
| 359 | r = malloc(sizeof(*r)); |
| 360 | |
| 361 | ret = get_off_from_method(td, f, ddir, &r->off); |
| 362 | if (ret) { |
| 363 | free(r); |
| 364 | break; |
| 365 | } |
| 366 | |
| 367 | flist_add(&r->list, &td->next_rand_list); |
| 368 | } |
| 369 | |
| 370 | if (ret && !i) |
| 371 | return ret; |
| 372 | |
| 373 | assert(!flist_empty(&td->next_rand_list)); |
| 374 | flist_sort(NULL, &td->next_rand_list, flist_cmp); |
| 375 | goto fetch; |
| 376 | } |
| 377 | |
| 378 | static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f) |
| 379 | { |
| 380 | struct thread_options *o = &td->o; |
| 381 | |
| 382 | if (o->invalidate_cache && !o->odirect) { |
| 383 | int fio_unused ret; |
| 384 | |
| 385 | ret = file_invalidate_cache(td, f); |
| 386 | } |
| 387 | } |
| 388 | |
| 389 | static int get_next_rand_block(struct thread_data *td, struct fio_file *f, |
| 390 | enum fio_ddir ddir, uint64_t *b) |
| 391 | { |
| 392 | if (!get_next_rand_offset(td, f, ddir, b)) |
| 393 | return 0; |
| 394 | |
| 395 | if (td->o.time_based || |
| 396 | (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) { |
| 397 | fio_file_reset(td, f); |
| 398 | if (!get_next_rand_offset(td, f, ddir, b)) |
| 399 | return 0; |
| 400 | loop_cache_invalidate(td, f); |
| 401 | } |
| 402 | |
| 403 | dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n", |
| 404 | f->file_name, (unsigned long long) f->last_pos[ddir], |
| 405 | (unsigned long long) f->real_file_size); |
| 406 | return 1; |
| 407 | } |
| 408 | |
| 409 | static int get_next_seq_offset(struct thread_data *td, struct fio_file *f, |
| 410 | enum fio_ddir ddir, uint64_t *offset) |
| 411 | { |
| 412 | struct thread_options *o = &td->o; |
| 413 | |
| 414 | assert(ddir_rw(ddir)); |
| 415 | |
| 416 | /* |
| 417 | * If we reach the end for a time based run, reset us back to 0 |
| 418 | * and invalidate the cache, if we need to. |
| 419 | */ |
| 420 | if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) && |
| 421 | o->time_based) { |
| 422 | f->last_pos[ddir] = f->file_offset; |
| 423 | loop_cache_invalidate(td, f); |
| 424 | } |
| 425 | |
| 426 | if (f->last_pos[ddir] < f->real_file_size) { |
| 427 | uint64_t pos; |
| 428 | |
| 429 | /* |
| 430 | * Only rewind if we already hit the end |
| 431 | */ |
| 432 | if (f->last_pos[ddir] == f->file_offset && |
| 433 | f->file_offset && o->ddir_seq_add < 0) { |
| 434 | if (f->real_file_size > f->io_size) |
| 435 | f->last_pos[ddir] = f->io_size; |
| 436 | else |
| 437 | f->last_pos[ddir] = f->real_file_size; |
| 438 | } |
| 439 | |
| 440 | pos = f->last_pos[ddir] - f->file_offset; |
| 441 | if (pos && o->ddir_seq_add) { |
| 442 | pos += o->ddir_seq_add; |
| 443 | |
| 444 | /* |
| 445 | * If we reach beyond the end of the file |
| 446 | * with holed IO, wrap around to the |
| 447 | * beginning again. If we're doing backwards IO, |
| 448 | * wrap to the end. |
| 449 | */ |
| 450 | if (pos >= f->real_file_size) { |
| 451 | if (o->ddir_seq_add > 0) |
| 452 | pos = f->file_offset; |
| 453 | else { |
| 454 | if (f->real_file_size > f->io_size) |
| 455 | pos = f->io_size; |
| 456 | else |
| 457 | pos = f->real_file_size; |
| 458 | |
| 459 | pos += o->ddir_seq_add; |
| 460 | } |
| 461 | } |
| 462 | } |
| 463 | |
| 464 | *offset = pos; |
| 465 | return 0; |
| 466 | } |
| 467 | |
| 468 | return 1; |
| 469 | } |
| 470 | |
| 471 | static int get_next_block(struct thread_data *td, struct io_u *io_u, |
| 472 | enum fio_ddir ddir, int rw_seq, |
| 473 | bool *is_random) |
| 474 | { |
| 475 | struct fio_file *f = io_u->file; |
| 476 | uint64_t b, offset; |
| 477 | int ret; |
| 478 | |
| 479 | assert(ddir_rw(ddir)); |
| 480 | |
| 481 | b = offset = -1ULL; |
| 482 | |
| 483 | if (rw_seq) { |
| 484 | if (td_random(td)) { |
| 485 | if (should_do_random(td, ddir)) { |
| 486 | ret = get_next_rand_block(td, f, ddir, &b); |
| 487 | *is_random = true; |
| 488 | } else { |
| 489 | *is_random = false; |
| 490 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
| 491 | ret = get_next_seq_offset(td, f, ddir, &offset); |
| 492 | if (ret) |
| 493 | ret = get_next_rand_block(td, f, ddir, &b); |
| 494 | } |
| 495 | } else { |
| 496 | *is_random = false; |
| 497 | ret = get_next_seq_offset(td, f, ddir, &offset); |
| 498 | } |
| 499 | } else { |
| 500 | io_u_set(td, io_u, IO_U_F_BUSY_OK); |
| 501 | *is_random = false; |
| 502 | |
| 503 | if (td->o.rw_seq == RW_SEQ_SEQ) { |
| 504 | ret = get_next_seq_offset(td, f, ddir, &offset); |
| 505 | if (ret) { |
| 506 | ret = get_next_rand_block(td, f, ddir, &b); |
| 507 | *is_random = false; |
| 508 | } |
| 509 | } else if (td->o.rw_seq == RW_SEQ_IDENT) { |
| 510 | if (f->last_start[ddir] != -1ULL) |
| 511 | offset = f->last_start[ddir] - f->file_offset; |
| 512 | else |
| 513 | offset = 0; |
| 514 | ret = 0; |
| 515 | } else { |
| 516 | log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq); |
| 517 | ret = 1; |
| 518 | } |
| 519 | } |
| 520 | |
| 521 | if (!ret) { |
| 522 | if (offset != -1ULL) |
| 523 | io_u->offset = offset; |
| 524 | else if (b != -1ULL) |
| 525 | io_u->offset = b * td->o.ba[ddir]; |
| 526 | else { |
| 527 | log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b); |
| 528 | ret = 1; |
| 529 | } |
| 530 | } |
| 531 | |
| 532 | return ret; |
| 533 | } |
| 534 | |
| 535 | /* |
| 536 | * For random io, generate a random new block and see if it's used. Repeat |
| 537 | * until we find a free one. For sequential io, just return the end of |
| 538 | * the last io issued. |
| 539 | */ |
| 540 | static int get_next_offset(struct thread_data *td, struct io_u *io_u, |
| 541 | bool *is_random) |
| 542 | { |
| 543 | struct fio_file *f = io_u->file; |
| 544 | enum fio_ddir ddir = io_u->ddir; |
| 545 | int rw_seq_hit = 0; |
| 546 | |
| 547 | assert(ddir_rw(ddir)); |
| 548 | |
| 549 | if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) { |
| 550 | rw_seq_hit = 1; |
| 551 | td->ddir_seq_nr = td->o.ddir_seq_nr; |
| 552 | } |
| 553 | |
| 554 | if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random)) |
| 555 | return 1; |
| 556 | |
| 557 | if (io_u->offset >= f->io_size) { |
| 558 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", |
| 559 | (unsigned long long) io_u->offset, |
| 560 | (unsigned long long) f->io_size); |
| 561 | return 1; |
| 562 | } |
| 563 | |
| 564 | io_u->offset += f->file_offset; |
| 565 | if (io_u->offset >= f->real_file_size) { |
| 566 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", |
| 567 | (unsigned long long) io_u->offset, |
| 568 | (unsigned long long) f->real_file_size); |
| 569 | return 1; |
| 570 | } |
| 571 | |
| 572 | return 0; |
| 573 | } |
| 574 | |
| 575 | static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u, |
| 576 | unsigned int buflen) |
| 577 | { |
| 578 | struct fio_file *f = io_u->file; |
| 579 | |
| 580 | return io_u->offset + buflen <= f->io_size + get_start_offset(td, f); |
| 581 | } |
| 582 | |
| 583 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u, |
| 584 | bool is_random) |
| 585 | { |
| 586 | int ddir = io_u->ddir; |
| 587 | unsigned int buflen = 0; |
| 588 | unsigned int minbs, maxbs; |
| 589 | uint64_t frand_max, r; |
| 590 | bool power_2; |
| 591 | |
| 592 | assert(ddir_rw(ddir)); |
| 593 | |
| 594 | if (td->o.bs_is_seq_rand) |
| 595 | ddir = is_random ? DDIR_WRITE : DDIR_READ; |
| 596 | |
| 597 | minbs = td->o.min_bs[ddir]; |
| 598 | maxbs = td->o.max_bs[ddir]; |
| 599 | |
| 600 | if (minbs == maxbs) |
| 601 | return minbs; |
| 602 | |
| 603 | /* |
| 604 | * If we can't satisfy the min block size from here, then fail |
| 605 | */ |
| 606 | if (!io_u_fits(td, io_u, minbs)) |
| 607 | return 0; |
| 608 | |
| 609 | frand_max = rand_max(&td->bsrange_state[ddir]); |
| 610 | do { |
| 611 | r = __rand(&td->bsrange_state[ddir]); |
| 612 | |
| 613 | if (!td->o.bssplit_nr[ddir]) { |
| 614 | buflen = 1 + (unsigned int) ((double) maxbs * |
| 615 | (r / (frand_max + 1.0))); |
| 616 | if (buflen < minbs) |
| 617 | buflen = minbs; |
| 618 | } else { |
| 619 | long long perc = 0; |
| 620 | unsigned int i; |
| 621 | |
| 622 | for (i = 0; i < td->o.bssplit_nr[ddir]; i++) { |
| 623 | struct bssplit *bsp = &td->o.bssplit[ddir][i]; |
| 624 | |
| 625 | buflen = bsp->bs; |
| 626 | perc += bsp->perc; |
| 627 | if (!perc) |
| 628 | break; |
| 629 | if ((r / perc <= frand_max / 100ULL) && |
| 630 | io_u_fits(td, io_u, buflen)) |
| 631 | break; |
| 632 | } |
| 633 | } |
| 634 | |
| 635 | power_2 = is_power_of_2(minbs); |
| 636 | if (!td->o.bs_unaligned && power_2) |
| 637 | buflen &= ~(minbs - 1); |
| 638 | else if (!td->o.bs_unaligned && !power_2) |
| 639 | buflen -= buflen % minbs; |
| 640 | } while (!io_u_fits(td, io_u, buflen)); |
| 641 | |
| 642 | return buflen; |
| 643 | } |
| 644 | |
| 645 | static void set_rwmix_bytes(struct thread_data *td) |
| 646 | { |
| 647 | unsigned int diff; |
| 648 | |
| 649 | /* |
| 650 | * we do time or byte based switch. this is needed because |
| 651 | * buffered writes may issue a lot quicker than they complete, |
| 652 | * whereas reads do not. |
| 653 | */ |
| 654 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
| 655 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
| 656 | } |
| 657 | |
| 658 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) |
| 659 | { |
| 660 | unsigned int v; |
| 661 | |
| 662 | v = rand32_between(&td->rwmix_state, 1, 100); |
| 663 | |
| 664 | if (v <= td->o.rwmix[DDIR_READ]) |
| 665 | return DDIR_READ; |
| 666 | |
| 667 | return DDIR_WRITE; |
| 668 | } |
| 669 | |
| 670 | int io_u_quiesce(struct thread_data *td) |
| 671 | { |
| 672 | int completed = 0; |
| 673 | |
| 674 | /* |
| 675 | * We are going to sleep, ensure that we flush anything pending as |
| 676 | * not to skew our latency numbers. |
| 677 | * |
| 678 | * Changed to only monitor 'in flight' requests here instead of the |
| 679 | * td->cur_depth, b/c td->cur_depth does not accurately represent |
| 680 | * io's that have been actually submitted to an async engine, |
| 681 | * and cur_depth is meaningless for sync engines. |
| 682 | */ |
| 683 | if (td->io_u_queued || td->cur_depth) { |
| 684 | int fio_unused ret; |
| 685 | |
| 686 | ret = td_io_commit(td); |
| 687 | } |
| 688 | |
| 689 | while (td->io_u_in_flight) { |
| 690 | int ret; |
| 691 | |
| 692 | ret = io_u_queued_complete(td, 1); |
| 693 | if (ret > 0) |
| 694 | completed += ret; |
| 695 | } |
| 696 | |
| 697 | if (td->flags & TD_F_REGROW_LOGS) |
| 698 | regrow_logs(td); |
| 699 | |
| 700 | return completed; |
| 701 | } |
| 702 | |
| 703 | static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir) |
| 704 | { |
| 705 | enum fio_ddir odir = ddir ^ 1; |
| 706 | uint64_t usec; |
| 707 | uint64_t now; |
| 708 | |
| 709 | assert(ddir_rw(ddir)); |
| 710 | now = utime_since_now(&td->start); |
| 711 | |
| 712 | /* |
| 713 | * if rate_next_io_time is in the past, need to catch up to rate |
| 714 | */ |
| 715 | if (td->rate_next_io_time[ddir] <= now) |
| 716 | return ddir; |
| 717 | |
| 718 | /* |
| 719 | * We are ahead of rate in this direction. See if we |
| 720 | * should switch. |
| 721 | */ |
| 722 | if (td_rw(td) && td->o.rwmix[odir]) { |
| 723 | /* |
| 724 | * Other direction is behind rate, switch |
| 725 | */ |
| 726 | if (td->rate_next_io_time[odir] <= now) |
| 727 | return odir; |
| 728 | |
| 729 | /* |
| 730 | * Both directions are ahead of rate. sleep the min, |
| 731 | * switch if necessary |
| 732 | */ |
| 733 | if (td->rate_next_io_time[ddir] <= |
| 734 | td->rate_next_io_time[odir]) { |
| 735 | usec = td->rate_next_io_time[ddir] - now; |
| 736 | } else { |
| 737 | usec = td->rate_next_io_time[odir] - now; |
| 738 | ddir = odir; |
| 739 | } |
| 740 | } else |
| 741 | usec = td->rate_next_io_time[ddir] - now; |
| 742 | |
| 743 | if (td->o.io_submit_mode == IO_MODE_INLINE) |
| 744 | io_u_quiesce(td); |
| 745 | |
| 746 | usec_sleep(td, usec); |
| 747 | return ddir; |
| 748 | } |
| 749 | |
| 750 | /* |
| 751 | * Return the data direction for the next io_u. If the job is a |
| 752 | * mixed read/write workload, check the rwmix cycle and switch if |
| 753 | * necessary. |
| 754 | */ |
| 755 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
| 756 | { |
| 757 | enum fio_ddir ddir; |
| 758 | |
| 759 | /* |
| 760 | * See if it's time to fsync/fdatasync/sync_file_range first, |
| 761 | * and if not then move on to check regular I/Os. |
| 762 | */ |
| 763 | if (should_fsync(td)) { |
| 764 | if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] && |
| 765 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks)) |
| 766 | return DDIR_SYNC; |
| 767 | |
| 768 | if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] && |
| 769 | !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks)) |
| 770 | return DDIR_DATASYNC; |
| 771 | |
| 772 | if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] && |
| 773 | !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr)) |
| 774 | return DDIR_SYNC_FILE_RANGE; |
| 775 | } |
| 776 | |
| 777 | if (td_rw(td)) { |
| 778 | /* |
| 779 | * Check if it's time to seed a new data direction. |
| 780 | */ |
| 781 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
| 782 | /* |
| 783 | * Put a top limit on how many bytes we do for |
| 784 | * one data direction, to avoid overflowing the |
| 785 | * ranges too much |
| 786 | */ |
| 787 | ddir = get_rand_ddir(td); |
| 788 | |
| 789 | if (ddir != td->rwmix_ddir) |
| 790 | set_rwmix_bytes(td); |
| 791 | |
| 792 | td->rwmix_ddir = ddir; |
| 793 | } |
| 794 | ddir = td->rwmix_ddir; |
| 795 | } else if (td_read(td)) |
| 796 | ddir = DDIR_READ; |
| 797 | else if (td_write(td)) |
| 798 | ddir = DDIR_WRITE; |
| 799 | else if (td_trim(td)) |
| 800 | ddir = DDIR_TRIM; |
| 801 | else |
| 802 | ddir = DDIR_INVAL; |
| 803 | |
| 804 | td->rwmix_ddir = rate_ddir(td, ddir); |
| 805 | return td->rwmix_ddir; |
| 806 | } |
| 807 | |
| 808 | static void set_rw_ddir(struct thread_data *td, struct io_u *io_u) |
| 809 | { |
| 810 | enum fio_ddir ddir = get_rw_ddir(td); |
| 811 | |
| 812 | if (td_trimwrite(td)) { |
| 813 | struct fio_file *f = io_u->file; |
| 814 | if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM]) |
| 815 | ddir = DDIR_TRIM; |
| 816 | else |
| 817 | ddir = DDIR_WRITE; |
| 818 | } |
| 819 | |
| 820 | io_u->ddir = io_u->acct_ddir = ddir; |
| 821 | |
| 822 | if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) && |
| 823 | td->o.barrier_blocks && |
| 824 | !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) && |
| 825 | td->io_issues[DDIR_WRITE]) |
| 826 | io_u_set(td, io_u, IO_U_F_BARRIER); |
| 827 | } |
| 828 | |
| 829 | void put_file_log(struct thread_data *td, struct fio_file *f) |
| 830 | { |
| 831 | unsigned int ret = put_file(td, f); |
| 832 | |
| 833 | if (ret) |
| 834 | td_verror(td, ret, "file close"); |
| 835 | } |
| 836 | |
| 837 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
| 838 | { |
| 839 | if (td->parent) |
| 840 | td = td->parent; |
| 841 | |
| 842 | td_io_u_lock(td); |
| 843 | |
| 844 | if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT)) |
| 845 | put_file_log(td, io_u->file); |
| 846 | |
| 847 | io_u->file = NULL; |
| 848 | io_u_set(td, io_u, IO_U_F_FREE); |
| 849 | |
| 850 | if (io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
| 851 | td->cur_depth--; |
| 852 | assert(!(td->flags & TD_F_CHILD)); |
| 853 | } |
| 854 | io_u_qpush(&td->io_u_freelist, io_u); |
| 855 | td_io_u_free_notify(td); |
| 856 | td_io_u_unlock(td); |
| 857 | } |
| 858 | |
| 859 | void clear_io_u(struct thread_data *td, struct io_u *io_u) |
| 860 | { |
| 861 | io_u_clear(td, io_u, IO_U_F_FLIGHT); |
| 862 | put_io_u(td, io_u); |
| 863 | } |
| 864 | |
| 865 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
| 866 | { |
| 867 | struct io_u *__io_u = *io_u; |
| 868 | enum fio_ddir ddir = acct_ddir(__io_u); |
| 869 | |
| 870 | dprint(FD_IO, "requeue %p\n", __io_u); |
| 871 | |
| 872 | if (td->parent) |
| 873 | td = td->parent; |
| 874 | |
| 875 | td_io_u_lock(td); |
| 876 | |
| 877 | io_u_set(td, __io_u, IO_U_F_FREE); |
| 878 | if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir)) |
| 879 | td->io_issues[ddir]--; |
| 880 | |
| 881 | io_u_clear(td, __io_u, IO_U_F_FLIGHT); |
| 882 | if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) { |
| 883 | td->cur_depth--; |
| 884 | assert(!(td->flags & TD_F_CHILD)); |
| 885 | } |
| 886 | |
| 887 | io_u_rpush(&td->io_u_requeues, __io_u); |
| 888 | td_io_u_free_notify(td); |
| 889 | td_io_u_unlock(td); |
| 890 | *io_u = NULL; |
| 891 | } |
| 892 | |
| 893 | static void __fill_io_u_zone(struct thread_data *td, struct io_u *io_u) |
| 894 | { |
| 895 | struct fio_file *f = io_u->file; |
| 896 | |
| 897 | /* |
| 898 | * See if it's time to switch to a new zone |
| 899 | */ |
| 900 | if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) { |
| 901 | td->zone_bytes = 0; |
| 902 | f->file_offset += td->o.zone_range + td->o.zone_skip; |
| 903 | |
| 904 | /* |
| 905 | * Wrap from the beginning, if we exceed the file size |
| 906 | */ |
| 907 | if (f->file_offset >= f->real_file_size) |
| 908 | f->file_offset = f->real_file_size - f->file_offset; |
| 909 | f->last_pos[io_u->ddir] = f->file_offset; |
| 910 | td->io_skip_bytes += td->o.zone_skip; |
| 911 | } |
| 912 | |
| 913 | /* |
| 914 | * If zone_size > zone_range, then maintain the same zone until |
| 915 | * zone_bytes >= zone_size. |
| 916 | */ |
| 917 | if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) { |
| 918 | dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n", |
| 919 | f->file_offset, f->last_pos[io_u->ddir]); |
| 920 | f->last_pos[io_u->ddir] = f->file_offset; |
| 921 | } |
| 922 | |
| 923 | /* |
| 924 | * For random: if 'norandommap' is not set and zone_size > zone_range, |
| 925 | * map needs to be reset as it's done with zone_range everytime. |
| 926 | */ |
| 927 | if ((td->zone_bytes % td->o.zone_range) == 0) { |
| 928 | fio_file_reset(td, f); |
| 929 | } |
| 930 | } |
| 931 | |
| 932 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
| 933 | { |
| 934 | bool is_random; |
| 935 | |
| 936 | if (td_ioengine_flagged(td, FIO_NOIO)) |
| 937 | goto out; |
| 938 | |
| 939 | set_rw_ddir(td, io_u); |
| 940 | |
| 941 | /* |
| 942 | * fsync() or fdatasync() or trim etc, we are done |
| 943 | */ |
| 944 | if (!ddir_rw(io_u->ddir)) |
| 945 | goto out; |
| 946 | |
| 947 | /* |
| 948 | * When file is zoned zone_range is always positive |
| 949 | */ |
| 950 | if (td->o.zone_range) { |
| 951 | __fill_io_u_zone(td, io_u); |
| 952 | } |
| 953 | |
| 954 | /* |
| 955 | * No log, let the seq/rand engine retrieve the next buflen and |
| 956 | * position. |
| 957 | */ |
| 958 | if (get_next_offset(td, io_u, &is_random)) { |
| 959 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
| 960 | return 1; |
| 961 | } |
| 962 | |
| 963 | io_u->buflen = get_next_buflen(td, io_u, is_random); |
| 964 | if (!io_u->buflen) { |
| 965 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); |
| 966 | return 1; |
| 967 | } |
| 968 | |
| 969 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
| 970 | dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%lx exceeds file size=0x%llx\n", |
| 971 | io_u, |
| 972 | (unsigned long long) io_u->offset, io_u->buflen, |
| 973 | (unsigned long long) io_u->file->real_file_size); |
| 974 | return 1; |
| 975 | } |
| 976 | |
| 977 | /* |
| 978 | * mark entry before potentially trimming io_u |
| 979 | */ |
| 980 | if (td_random(td) && file_randommap(td, io_u->file)) |
| 981 | mark_random_map(td, io_u); |
| 982 | |
| 983 | out: |
| 984 | dprint_io_u(io_u, "fill"); |
| 985 | td->zone_bytes += io_u->buflen; |
| 986 | return 0; |
| 987 | } |
| 988 | |
| 989 | static void __io_u_mark_map(uint64_t *map, unsigned int nr) |
| 990 | { |
| 991 | int idx = 0; |
| 992 | |
| 993 | switch (nr) { |
| 994 | default: |
| 995 | idx = 6; |
| 996 | break; |
| 997 | case 33 ... 64: |
| 998 | idx = 5; |
| 999 | break; |
| 1000 | case 17 ... 32: |
| 1001 | idx = 4; |
| 1002 | break; |
| 1003 | case 9 ... 16: |
| 1004 | idx = 3; |
| 1005 | break; |
| 1006 | case 5 ... 8: |
| 1007 | idx = 2; |
| 1008 | break; |
| 1009 | case 1 ... 4: |
| 1010 | idx = 1; |
| 1011 | case 0: |
| 1012 | break; |
| 1013 | } |
| 1014 | |
| 1015 | map[idx]++; |
| 1016 | } |
| 1017 | |
| 1018 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) |
| 1019 | { |
| 1020 | __io_u_mark_map(td->ts.io_u_submit, nr); |
| 1021 | td->ts.total_submit++; |
| 1022 | } |
| 1023 | |
| 1024 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) |
| 1025 | { |
| 1026 | __io_u_mark_map(td->ts.io_u_complete, nr); |
| 1027 | td->ts.total_complete++; |
| 1028 | } |
| 1029 | |
| 1030 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
| 1031 | { |
| 1032 | int idx = 0; |
| 1033 | |
| 1034 | switch (td->cur_depth) { |
| 1035 | default: |
| 1036 | idx = 6; |
| 1037 | break; |
| 1038 | case 32 ... 63: |
| 1039 | idx = 5; |
| 1040 | break; |
| 1041 | case 16 ... 31: |
| 1042 | idx = 4; |
| 1043 | break; |
| 1044 | case 8 ... 15: |
| 1045 | idx = 3; |
| 1046 | break; |
| 1047 | case 4 ... 7: |
| 1048 | idx = 2; |
| 1049 | break; |
| 1050 | case 2 ... 3: |
| 1051 | idx = 1; |
| 1052 | case 1: |
| 1053 | break; |
| 1054 | } |
| 1055 | |
| 1056 | td->ts.io_u_map[idx] += nr; |
| 1057 | } |
| 1058 | |
| 1059 | static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec) |
| 1060 | { |
| 1061 | int idx = 0; |
| 1062 | |
| 1063 | assert(nsec < 1000); |
| 1064 | |
| 1065 | switch (nsec) { |
| 1066 | case 750 ... 999: |
| 1067 | idx = 9; |
| 1068 | break; |
| 1069 | case 500 ... 749: |
| 1070 | idx = 8; |
| 1071 | break; |
| 1072 | case 250 ... 499: |
| 1073 | idx = 7; |
| 1074 | break; |
| 1075 | case 100 ... 249: |
| 1076 | idx = 6; |
| 1077 | break; |
| 1078 | case 50 ... 99: |
| 1079 | idx = 5; |
| 1080 | break; |
| 1081 | case 20 ... 49: |
| 1082 | idx = 4; |
| 1083 | break; |
| 1084 | case 10 ... 19: |
| 1085 | idx = 3; |
| 1086 | break; |
| 1087 | case 4 ... 9: |
| 1088 | idx = 2; |
| 1089 | break; |
| 1090 | case 2 ... 3: |
| 1091 | idx = 1; |
| 1092 | case 0 ... 1: |
| 1093 | break; |
| 1094 | } |
| 1095 | |
| 1096 | assert(idx < FIO_IO_U_LAT_N_NR); |
| 1097 | td->ts.io_u_lat_n[idx]++; |
| 1098 | } |
| 1099 | |
| 1100 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec) |
| 1101 | { |
| 1102 | int idx = 0; |
| 1103 | |
| 1104 | assert(usec < 1000 && usec >= 1); |
| 1105 | |
| 1106 | switch (usec) { |
| 1107 | case 750 ... 999: |
| 1108 | idx = 9; |
| 1109 | break; |
| 1110 | case 500 ... 749: |
| 1111 | idx = 8; |
| 1112 | break; |
| 1113 | case 250 ... 499: |
| 1114 | idx = 7; |
| 1115 | break; |
| 1116 | case 100 ... 249: |
| 1117 | idx = 6; |
| 1118 | break; |
| 1119 | case 50 ... 99: |
| 1120 | idx = 5; |
| 1121 | break; |
| 1122 | case 20 ... 49: |
| 1123 | idx = 4; |
| 1124 | break; |
| 1125 | case 10 ... 19: |
| 1126 | idx = 3; |
| 1127 | break; |
| 1128 | case 4 ... 9: |
| 1129 | idx = 2; |
| 1130 | break; |
| 1131 | case 2 ... 3: |
| 1132 | idx = 1; |
| 1133 | case 0 ... 1: |
| 1134 | break; |
| 1135 | } |
| 1136 | |
| 1137 | assert(idx < FIO_IO_U_LAT_U_NR); |
| 1138 | td->ts.io_u_lat_u[idx]++; |
| 1139 | } |
| 1140 | |
| 1141 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec) |
| 1142 | { |
| 1143 | int idx = 0; |
| 1144 | |
| 1145 | assert(msec >= 1); |
| 1146 | |
| 1147 | switch (msec) { |
| 1148 | default: |
| 1149 | idx = 11; |
| 1150 | break; |
| 1151 | case 1000 ... 1999: |
| 1152 | idx = 10; |
| 1153 | break; |
| 1154 | case 750 ... 999: |
| 1155 | idx = 9; |
| 1156 | break; |
| 1157 | case 500 ... 749: |
| 1158 | idx = 8; |
| 1159 | break; |
| 1160 | case 250 ... 499: |
| 1161 | idx = 7; |
| 1162 | break; |
| 1163 | case 100 ... 249: |
| 1164 | idx = 6; |
| 1165 | break; |
| 1166 | case 50 ... 99: |
| 1167 | idx = 5; |
| 1168 | break; |
| 1169 | case 20 ... 49: |
| 1170 | idx = 4; |
| 1171 | break; |
| 1172 | case 10 ... 19: |
| 1173 | idx = 3; |
| 1174 | break; |
| 1175 | case 4 ... 9: |
| 1176 | idx = 2; |
| 1177 | break; |
| 1178 | case 2 ... 3: |
| 1179 | idx = 1; |
| 1180 | case 0 ... 1: |
| 1181 | break; |
| 1182 | } |
| 1183 | |
| 1184 | assert(idx < FIO_IO_U_LAT_M_NR); |
| 1185 | td->ts.io_u_lat_m[idx]++; |
| 1186 | } |
| 1187 | |
| 1188 | static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec) |
| 1189 | { |
| 1190 | if (nsec < 1000) |
| 1191 | io_u_mark_lat_nsec(td, nsec); |
| 1192 | else if (nsec < 1000000) |
| 1193 | io_u_mark_lat_usec(td, nsec / 1000); |
| 1194 | else |
| 1195 | io_u_mark_lat_msec(td, nsec / 1000000); |
| 1196 | } |
| 1197 | |
| 1198 | static unsigned int __get_next_fileno_rand(struct thread_data *td) |
| 1199 | { |
| 1200 | unsigned long fileno; |
| 1201 | |
| 1202 | if (td->o.file_service_type == FIO_FSERVICE_RANDOM) { |
| 1203 | uint64_t frand_max = rand_max(&td->next_file_state); |
| 1204 | unsigned long r; |
| 1205 | |
| 1206 | r = __rand(&td->next_file_state); |
| 1207 | return (unsigned int) ((double) td->o.nr_files |
| 1208 | * (r / (frand_max + 1.0))); |
| 1209 | } |
| 1210 | |
| 1211 | if (td->o.file_service_type == FIO_FSERVICE_ZIPF) |
| 1212 | fileno = zipf_next(&td->next_file_zipf); |
| 1213 | else if (td->o.file_service_type == FIO_FSERVICE_PARETO) |
| 1214 | fileno = pareto_next(&td->next_file_zipf); |
| 1215 | else if (td->o.file_service_type == FIO_FSERVICE_GAUSS) |
| 1216 | fileno = gauss_next(&td->next_file_gauss); |
| 1217 | else { |
| 1218 | log_err("fio: bad file service type: %d\n", td->o.file_service_type); |
| 1219 | assert(0); |
| 1220 | return 0; |
| 1221 | } |
| 1222 | |
| 1223 | return fileno >> FIO_FSERVICE_SHIFT; |
| 1224 | } |
| 1225 | |
| 1226 | /* |
| 1227 | * Get next file to service by choosing one at random |
| 1228 | */ |
| 1229 | static struct fio_file *get_next_file_rand(struct thread_data *td, |
| 1230 | enum fio_file_flags goodf, |
| 1231 | enum fio_file_flags badf) |
| 1232 | { |
| 1233 | struct fio_file *f; |
| 1234 | int fno; |
| 1235 | |
| 1236 | do { |
| 1237 | int opened = 0; |
| 1238 | |
| 1239 | fno = __get_next_fileno_rand(td); |
| 1240 | |
| 1241 | f = td->files[fno]; |
| 1242 | if (fio_file_done(f)) |
| 1243 | continue; |
| 1244 | |
| 1245 | if (!fio_file_open(f)) { |
| 1246 | int err; |
| 1247 | |
| 1248 | if (td->nr_open_files >= td->o.open_files) |
| 1249 | return ERR_PTR(-EBUSY); |
| 1250 | |
| 1251 | err = td_io_open_file(td, f); |
| 1252 | if (err) |
| 1253 | continue; |
| 1254 | opened = 1; |
| 1255 | } |
| 1256 | |
| 1257 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
| 1258 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); |
| 1259 | return f; |
| 1260 | } |
| 1261 | if (opened) |
| 1262 | td_io_close_file(td, f); |
| 1263 | } while (1); |
| 1264 | } |
| 1265 | |
| 1266 | /* |
| 1267 | * Get next file to service by doing round robin between all available ones |
| 1268 | */ |
| 1269 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
| 1270 | int badf) |
| 1271 | { |
| 1272 | unsigned int old_next_file = td->next_file; |
| 1273 | struct fio_file *f; |
| 1274 | |
| 1275 | do { |
| 1276 | int opened = 0; |
| 1277 | |
| 1278 | f = td->files[td->next_file]; |
| 1279 | |
| 1280 | td->next_file++; |
| 1281 | if (td->next_file >= td->o.nr_files) |
| 1282 | td->next_file = 0; |
| 1283 | |
| 1284 | dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags); |
| 1285 | if (fio_file_done(f)) { |
| 1286 | f = NULL; |
| 1287 | continue; |
| 1288 | } |
| 1289 | |
| 1290 | if (!fio_file_open(f)) { |
| 1291 | int err; |
| 1292 | |
| 1293 | if (td->nr_open_files >= td->o.open_files) |
| 1294 | return ERR_PTR(-EBUSY); |
| 1295 | |
| 1296 | err = td_io_open_file(td, f); |
| 1297 | if (err) { |
| 1298 | dprint(FD_FILE, "error %d on open of %s\n", |
| 1299 | err, f->file_name); |
| 1300 | f = NULL; |
| 1301 | continue; |
| 1302 | } |
| 1303 | opened = 1; |
| 1304 | } |
| 1305 | |
| 1306 | dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, |
| 1307 | f->flags); |
| 1308 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
| 1309 | break; |
| 1310 | |
| 1311 | if (opened) |
| 1312 | td_io_close_file(td, f); |
| 1313 | |
| 1314 | f = NULL; |
| 1315 | } while (td->next_file != old_next_file); |
| 1316 | |
| 1317 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
| 1318 | return f; |
| 1319 | } |
| 1320 | |
| 1321 | static struct fio_file *__get_next_file(struct thread_data *td) |
| 1322 | { |
| 1323 | struct fio_file *f; |
| 1324 | |
| 1325 | assert(td->o.nr_files <= td->files_index); |
| 1326 | |
| 1327 | if (td->nr_done_files >= td->o.nr_files) { |
| 1328 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
| 1329 | " nr_files=%d\n", td->nr_open_files, |
| 1330 | td->nr_done_files, |
| 1331 | td->o.nr_files); |
| 1332 | return NULL; |
| 1333 | } |
| 1334 | |
| 1335 | f = td->file_service_file; |
| 1336 | if (f && fio_file_open(f) && !fio_file_closing(f)) { |
| 1337 | if (td->o.file_service_type == FIO_FSERVICE_SEQ) |
| 1338 | goto out; |
| 1339 | if (td->file_service_left--) |
| 1340 | goto out; |
| 1341 | } |
| 1342 | |
| 1343 | if (td->o.file_service_type == FIO_FSERVICE_RR || |
| 1344 | td->o.file_service_type == FIO_FSERVICE_SEQ) |
| 1345 | f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing); |
| 1346 | else |
| 1347 | f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing); |
| 1348 | |
| 1349 | if (IS_ERR(f)) |
| 1350 | return f; |
| 1351 | |
| 1352 | td->file_service_file = f; |
| 1353 | td->file_service_left = td->file_service_nr - 1; |
| 1354 | out: |
| 1355 | if (f) |
| 1356 | dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name); |
| 1357 | else |
| 1358 | dprint(FD_FILE, "get_next_file: NULL\n"); |
| 1359 | return f; |
| 1360 | } |
| 1361 | |
| 1362 | static struct fio_file *get_next_file(struct thread_data *td) |
| 1363 | { |
| 1364 | return __get_next_file(td); |
| 1365 | } |
| 1366 | |
| 1367 | static long set_io_u_file(struct thread_data *td, struct io_u *io_u) |
| 1368 | { |
| 1369 | struct fio_file *f; |
| 1370 | |
| 1371 | do { |
| 1372 | f = get_next_file(td); |
| 1373 | if (IS_ERR_OR_NULL(f)) |
| 1374 | return PTR_ERR(f); |
| 1375 | |
| 1376 | io_u->file = f; |
| 1377 | get_file(f); |
| 1378 | |
| 1379 | if (!fill_io_u(td, io_u)) |
| 1380 | break; |
| 1381 | |
| 1382 | put_file_log(td, f); |
| 1383 | td_io_close_file(td, f); |
| 1384 | io_u->file = NULL; |
| 1385 | if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM) |
| 1386 | fio_file_reset(td, f); |
| 1387 | else { |
| 1388 | fio_file_set_done(f); |
| 1389 | td->nr_done_files++; |
| 1390 | dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name, |
| 1391 | td->nr_done_files, td->o.nr_files); |
| 1392 | } |
| 1393 | } while (1); |
| 1394 | |
| 1395 | return 0; |
| 1396 | } |
| 1397 | |
| 1398 | static void lat_fatal(struct thread_data *td, struct io_completion_data *icd, |
| 1399 | unsigned long long tnsec, unsigned long long max_nsec) |
| 1400 | { |
| 1401 | if (!td->error) |
| 1402 | log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec); |
| 1403 | td_verror(td, ETIMEDOUT, "max latency exceeded"); |
| 1404 | icd->error = ETIMEDOUT; |
| 1405 | } |
| 1406 | |
| 1407 | static void lat_new_cycle(struct thread_data *td) |
| 1408 | { |
| 1409 | fio_gettime(&td->latency_ts, NULL); |
| 1410 | td->latency_ios = ddir_rw_sum(td->io_blocks); |
| 1411 | td->latency_failed = 0; |
| 1412 | } |
| 1413 | |
| 1414 | /* |
| 1415 | * We had an IO outside the latency target. Reduce the queue depth. If we |
| 1416 | * are at QD=1, then it's time to give up. |
| 1417 | */ |
| 1418 | static bool __lat_target_failed(struct thread_data *td) |
| 1419 | { |
| 1420 | if (td->latency_qd == 1) |
| 1421 | return true; |
| 1422 | |
| 1423 | td->latency_qd_high = td->latency_qd; |
| 1424 | |
| 1425 | if (td->latency_qd == td->latency_qd_low) |
| 1426 | td->latency_qd_low--; |
| 1427 | |
| 1428 | td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2; |
| 1429 | |
| 1430 | dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); |
| 1431 | |
| 1432 | /* |
| 1433 | * When we ramp QD down, quiesce existing IO to prevent |
| 1434 | * a storm of ramp downs due to pending higher depth. |
| 1435 | */ |
| 1436 | io_u_quiesce(td); |
| 1437 | lat_new_cycle(td); |
| 1438 | return false; |
| 1439 | } |
| 1440 | |
| 1441 | static bool lat_target_failed(struct thread_data *td) |
| 1442 | { |
| 1443 | if (td->o.latency_percentile.u.f == 100.0) |
| 1444 | return __lat_target_failed(td); |
| 1445 | |
| 1446 | td->latency_failed++; |
| 1447 | return false; |
| 1448 | } |
| 1449 | |
| 1450 | void lat_target_init(struct thread_data *td) |
| 1451 | { |
| 1452 | td->latency_end_run = 0; |
| 1453 | |
| 1454 | if (td->o.latency_target) { |
| 1455 | dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target); |
| 1456 | fio_gettime(&td->latency_ts, NULL); |
| 1457 | td->latency_qd = 1; |
| 1458 | td->latency_qd_high = td->o.iodepth; |
| 1459 | td->latency_qd_low = 1; |
| 1460 | td->latency_ios = ddir_rw_sum(td->io_blocks); |
| 1461 | } else |
| 1462 | td->latency_qd = td->o.iodepth; |
| 1463 | } |
| 1464 | |
| 1465 | void lat_target_reset(struct thread_data *td) |
| 1466 | { |
| 1467 | if (!td->latency_end_run) |
| 1468 | lat_target_init(td); |
| 1469 | } |
| 1470 | |
| 1471 | static void lat_target_success(struct thread_data *td) |
| 1472 | { |
| 1473 | const unsigned int qd = td->latency_qd; |
| 1474 | struct thread_options *o = &td->o; |
| 1475 | |
| 1476 | td->latency_qd_low = td->latency_qd; |
| 1477 | |
| 1478 | /* |
| 1479 | * If we haven't failed yet, we double up to a failing value instead |
| 1480 | * of bisecting from highest possible queue depth. If we have set |
| 1481 | * a limit other than td->o.iodepth, bisect between that. |
| 1482 | */ |
| 1483 | if (td->latency_qd_high != o->iodepth) |
| 1484 | td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2; |
| 1485 | else |
| 1486 | td->latency_qd *= 2; |
| 1487 | |
| 1488 | if (td->latency_qd > o->iodepth) |
| 1489 | td->latency_qd = o->iodepth; |
| 1490 | |
| 1491 | dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high); |
| 1492 | |
| 1493 | /* |
| 1494 | * Same as last one, we are done. Let it run a latency cycle, so |
| 1495 | * we get only the results from the targeted depth. |
| 1496 | */ |
| 1497 | if (td->latency_qd == qd) { |
| 1498 | if (td->latency_end_run) { |
| 1499 | dprint(FD_RATE, "We are done\n"); |
| 1500 | td->done = 1; |
| 1501 | } else { |
| 1502 | dprint(FD_RATE, "Quiesce and final run\n"); |
| 1503 | io_u_quiesce(td); |
| 1504 | td->latency_end_run = 1; |
| 1505 | reset_all_stats(td); |
| 1506 | reset_io_stats(td); |
| 1507 | } |
| 1508 | } |
| 1509 | |
| 1510 | lat_new_cycle(td); |
| 1511 | } |
| 1512 | |
| 1513 | /* |
| 1514 | * Check if we can bump the queue depth |
| 1515 | */ |
| 1516 | void lat_target_check(struct thread_data *td) |
| 1517 | { |
| 1518 | uint64_t usec_window; |
| 1519 | uint64_t ios; |
| 1520 | double success_ios; |
| 1521 | |
| 1522 | usec_window = utime_since_now(&td->latency_ts); |
| 1523 | if (usec_window < td->o.latency_window) |
| 1524 | return; |
| 1525 | |
| 1526 | ios = ddir_rw_sum(td->io_blocks) - td->latency_ios; |
| 1527 | success_ios = (double) (ios - td->latency_failed) / (double) ios; |
| 1528 | success_ios *= 100.0; |
| 1529 | |
| 1530 | dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f); |
| 1531 | |
| 1532 | if (success_ios >= td->o.latency_percentile.u.f) |
| 1533 | lat_target_success(td); |
| 1534 | else |
| 1535 | __lat_target_failed(td); |
| 1536 | } |
| 1537 | |
| 1538 | /* |
| 1539 | * If latency target is enabled, we might be ramping up or down and not |
| 1540 | * using the full queue depth available. |
| 1541 | */ |
| 1542 | bool queue_full(const struct thread_data *td) |
| 1543 | { |
| 1544 | const int qempty = io_u_qempty(&td->io_u_freelist); |
| 1545 | |
| 1546 | if (qempty) |
| 1547 | return true; |
| 1548 | if (!td->o.latency_target) |
| 1549 | return false; |
| 1550 | |
| 1551 | return td->cur_depth >= td->latency_qd; |
| 1552 | } |
| 1553 | |
| 1554 | struct io_u *__get_io_u(struct thread_data *td) |
| 1555 | { |
| 1556 | struct io_u *io_u = NULL; |
| 1557 | int ret; |
| 1558 | |
| 1559 | if (td->stop_io) |
| 1560 | return NULL; |
| 1561 | |
| 1562 | td_io_u_lock(td); |
| 1563 | |
| 1564 | again: |
| 1565 | if (!io_u_rempty(&td->io_u_requeues)) |
| 1566 | io_u = io_u_rpop(&td->io_u_requeues); |
| 1567 | else if (!queue_full(td)) { |
| 1568 | io_u = io_u_qpop(&td->io_u_freelist); |
| 1569 | |
| 1570 | io_u->file = NULL; |
| 1571 | io_u->buflen = 0; |
| 1572 | io_u->resid = 0; |
| 1573 | io_u->end_io = NULL; |
| 1574 | } |
| 1575 | |
| 1576 | if (io_u) { |
| 1577 | assert(io_u->flags & IO_U_F_FREE); |
| 1578 | io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT | |
| 1579 | IO_U_F_TRIMMED | IO_U_F_BARRIER | |
| 1580 | IO_U_F_VER_LIST); |
| 1581 | |
| 1582 | io_u->error = 0; |
| 1583 | io_u->acct_ddir = -1; |
| 1584 | td->cur_depth++; |
| 1585 | assert(!(td->flags & TD_F_CHILD)); |
| 1586 | io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH); |
| 1587 | io_u->ipo = NULL; |
| 1588 | } else if (td_async_processing(td)) { |
| 1589 | /* |
| 1590 | * We ran out, wait for async verify threads to finish and |
| 1591 | * return one |
| 1592 | */ |
| 1593 | assert(!(td->flags & TD_F_CHILD)); |
| 1594 | ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock); |
| 1595 | assert(ret == 0); |
| 1596 | goto again; |
| 1597 | } |
| 1598 | |
| 1599 | td_io_u_unlock(td); |
| 1600 | return io_u; |
| 1601 | } |
| 1602 | |
| 1603 | static bool check_get_trim(struct thread_data *td, struct io_u *io_u) |
| 1604 | { |
| 1605 | if (!(td->flags & TD_F_TRIM_BACKLOG)) |
| 1606 | return false; |
| 1607 | if (!td->trim_entries) |
| 1608 | return false; |
| 1609 | |
| 1610 | if (td->trim_batch) { |
| 1611 | td->trim_batch--; |
| 1612 | if (get_next_trim(td, io_u)) |
| 1613 | return true; |
| 1614 | } else if (!(td->io_hist_len % td->o.trim_backlog) && |
| 1615 | td->last_ddir != DDIR_READ) { |
| 1616 | td->trim_batch = td->o.trim_batch; |
| 1617 | if (!td->trim_batch) |
| 1618 | td->trim_batch = td->o.trim_backlog; |
| 1619 | if (get_next_trim(td, io_u)) |
| 1620 | return true; |
| 1621 | } |
| 1622 | |
| 1623 | return false; |
| 1624 | } |
| 1625 | |
| 1626 | static bool check_get_verify(struct thread_data *td, struct io_u *io_u) |
| 1627 | { |
| 1628 | if (!(td->flags & TD_F_VER_BACKLOG)) |
| 1629 | return false; |
| 1630 | |
| 1631 | if (td->io_hist_len) { |
| 1632 | int get_verify = 0; |
| 1633 | |
| 1634 | if (td->verify_batch) |
| 1635 | get_verify = 1; |
| 1636 | else if (!(td->io_hist_len % td->o.verify_backlog) && |
| 1637 | td->last_ddir != DDIR_READ) { |
| 1638 | td->verify_batch = td->o.verify_batch; |
| 1639 | if (!td->verify_batch) |
| 1640 | td->verify_batch = td->o.verify_backlog; |
| 1641 | get_verify = 1; |
| 1642 | } |
| 1643 | |
| 1644 | if (get_verify && !get_next_verify(td, io_u)) { |
| 1645 | td->verify_batch--; |
| 1646 | return true; |
| 1647 | } |
| 1648 | } |
| 1649 | |
| 1650 | return false; |
| 1651 | } |
| 1652 | |
| 1653 | /* |
| 1654 | * Fill offset and start time into the buffer content, to prevent too |
| 1655 | * easy compressible data for simple de-dupe attempts. Do this for every |
| 1656 | * 512b block in the range, since that should be the smallest block size |
| 1657 | * we can expect from a device. |
| 1658 | */ |
| 1659 | static void small_content_scramble(struct io_u *io_u) |
| 1660 | { |
| 1661 | unsigned int i, nr_blocks = io_u->buflen >> 9; |
| 1662 | unsigned int offset; |
| 1663 | uint64_t boffset, *iptr; |
| 1664 | char *p; |
| 1665 | |
| 1666 | if (!nr_blocks) |
| 1667 | return; |
| 1668 | |
| 1669 | p = io_u->xfer_buf; |
| 1670 | boffset = io_u->offset; |
| 1671 | |
| 1672 | if (io_u->buf_filled_len) |
| 1673 | io_u->buf_filled_len = 0; |
| 1674 | |
| 1675 | /* |
| 1676 | * Generate random index between 0..7. We do chunks of 512b, if |
| 1677 | * we assume a cacheline is 64 bytes, then we have 8 of those. |
| 1678 | * Scramble content within the blocks in the same cacheline to |
| 1679 | * speed things up. |
| 1680 | */ |
| 1681 | offset = (io_u->start_time.tv_nsec ^ boffset) & 7; |
| 1682 | |
| 1683 | for (i = 0; i < nr_blocks; i++) { |
| 1684 | /* |
| 1685 | * Fill offset into start of cacheline, time into end |
| 1686 | * of cacheline |
| 1687 | */ |
| 1688 | iptr = (void *) p + (offset << 6); |
| 1689 | *iptr = boffset; |
| 1690 | |
| 1691 | iptr = (void *) p + 64 - 2 * sizeof(uint64_t); |
| 1692 | iptr[0] = io_u->start_time.tv_sec; |
| 1693 | iptr[1] = io_u->start_time.tv_nsec; |
| 1694 | |
| 1695 | p += 512; |
| 1696 | boffset += 512; |
| 1697 | } |
| 1698 | } |
| 1699 | |
| 1700 | /* |
| 1701 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, |
| 1702 | * etc. The returned io_u is fully ready to be prepped, populated and submitted. |
| 1703 | */ |
| 1704 | struct io_u *get_io_u(struct thread_data *td) |
| 1705 | { |
| 1706 | struct fio_file *f; |
| 1707 | struct io_u *io_u; |
| 1708 | int do_scramble = 0; |
| 1709 | long ret = 0; |
| 1710 | |
| 1711 | io_u = __get_io_u(td); |
| 1712 | if (!io_u) { |
| 1713 | dprint(FD_IO, "__get_io_u failed\n"); |
| 1714 | return NULL; |
| 1715 | } |
| 1716 | |
| 1717 | if (check_get_verify(td, io_u)) |
| 1718 | goto out; |
| 1719 | if (check_get_trim(td, io_u)) |
| 1720 | goto out; |
| 1721 | |
| 1722 | /* |
| 1723 | * from a requeue, io_u already setup |
| 1724 | */ |
| 1725 | if (io_u->file) |
| 1726 | goto out; |
| 1727 | |
| 1728 | /* |
| 1729 | * If using an iolog, grab next piece if any available. |
| 1730 | */ |
| 1731 | if (td->flags & TD_F_READ_IOLOG) { |
| 1732 | if (read_iolog_get(td, io_u)) |
| 1733 | goto err_put; |
| 1734 | } else if (set_io_u_file(td, io_u)) { |
| 1735 | ret = -EBUSY; |
| 1736 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
| 1737 | goto err_put; |
| 1738 | } |
| 1739 | |
| 1740 | f = io_u->file; |
| 1741 | if (!f) { |
| 1742 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
| 1743 | goto err_put; |
| 1744 | } |
| 1745 | |
| 1746 | assert(fio_file_open(f)); |
| 1747 | |
| 1748 | if (ddir_rw(io_u->ddir)) { |
| 1749 | if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) { |
| 1750 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
| 1751 | goto err_put; |
| 1752 | } |
| 1753 | |
| 1754 | f->last_start[io_u->ddir] = io_u->offset; |
| 1755 | f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen; |
| 1756 | |
| 1757 | if (io_u->ddir == DDIR_WRITE) { |
| 1758 | if (td->flags & TD_F_REFILL_BUFFERS) { |
| 1759 | io_u_fill_buffer(td, io_u, |
| 1760 | td->o.min_bs[DDIR_WRITE], |
| 1761 | io_u->buflen); |
| 1762 | } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) && |
| 1763 | !(td->flags & TD_F_COMPRESS) && |
| 1764 | !(td->flags & TD_F_DO_VERIFY)) |
| 1765 | do_scramble = 1; |
| 1766 | } else if (io_u->ddir == DDIR_READ) { |
| 1767 | /* |
| 1768 | * Reset the buf_filled parameters so next time if the |
| 1769 | * buffer is used for writes it is refilled. |
| 1770 | */ |
| 1771 | io_u->buf_filled_len = 0; |
| 1772 | } |
| 1773 | } |
| 1774 | |
| 1775 | /* |
| 1776 | * Set io data pointers. |
| 1777 | */ |
| 1778 | io_u->xfer_buf = io_u->buf; |
| 1779 | io_u->xfer_buflen = io_u->buflen; |
| 1780 | |
| 1781 | out: |
| 1782 | assert(io_u->file); |
| 1783 | if (!td_io_prep(td, io_u)) { |
| 1784 | if (!td->o.disable_lat) |
| 1785 | fio_gettime(&io_u->start_time, NULL); |
| 1786 | |
| 1787 | if (do_scramble) |
| 1788 | small_content_scramble(io_u); |
| 1789 | |
| 1790 | return io_u; |
| 1791 | } |
| 1792 | err_put: |
| 1793 | dprint(FD_IO, "get_io_u failed\n"); |
| 1794 | put_io_u(td, io_u); |
| 1795 | return ERR_PTR(ret); |
| 1796 | } |
| 1797 | |
| 1798 | static void __io_u_log_error(struct thread_data *td, struct io_u *io_u) |
| 1799 | { |
| 1800 | enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error); |
| 1801 | |
| 1802 | if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump) |
| 1803 | return; |
| 1804 | |
| 1805 | log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n", |
| 1806 | io_u->file ? " on file " : "", |
| 1807 | io_u->file ? io_u->file->file_name : "", |
| 1808 | strerror(io_u->error), |
| 1809 | io_ddir_name(io_u->ddir), |
| 1810 | io_u->offset, io_u->xfer_buflen); |
| 1811 | |
| 1812 | if (td->io_ops->errdetails) { |
| 1813 | char *err = td->io_ops->errdetails(io_u); |
| 1814 | |
| 1815 | log_err("fio: %s\n", err); |
| 1816 | free(err); |
| 1817 | } |
| 1818 | |
| 1819 | if (!td->error) |
| 1820 | td_verror(td, io_u->error, "io_u error"); |
| 1821 | } |
| 1822 | |
| 1823 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
| 1824 | { |
| 1825 | __io_u_log_error(td, io_u); |
| 1826 | if (td->parent) |
| 1827 | __io_u_log_error(td->parent, io_u); |
| 1828 | } |
| 1829 | |
| 1830 | static inline bool gtod_reduce(struct thread_data *td) |
| 1831 | { |
| 1832 | return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw) |
| 1833 | || td->o.gtod_reduce; |
| 1834 | } |
| 1835 | |
| 1836 | static void account_io_completion(struct thread_data *td, struct io_u *io_u, |
| 1837 | struct io_completion_data *icd, |
| 1838 | const enum fio_ddir idx, unsigned int bytes) |
| 1839 | { |
| 1840 | const int no_reduce = !gtod_reduce(td); |
| 1841 | unsigned long long llnsec = 0; |
| 1842 | |
| 1843 | if (td->parent) |
| 1844 | td = td->parent; |
| 1845 | |
| 1846 | if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS)) |
| 1847 | return; |
| 1848 | |
| 1849 | if (no_reduce) |
| 1850 | llnsec = ntime_since(&io_u->issue_time, &icd->time); |
| 1851 | |
| 1852 | if (!td->o.disable_lat) { |
| 1853 | unsigned long long tnsec; |
| 1854 | |
| 1855 | tnsec = ntime_since(&io_u->start_time, &icd->time); |
| 1856 | add_lat_sample(td, idx, tnsec, bytes, io_u->offset); |
| 1857 | |
| 1858 | if (td->flags & TD_F_PROFILE_OPS) { |
| 1859 | struct prof_io_ops *ops = &td->prof_io_ops; |
| 1860 | |
| 1861 | if (ops->io_u_lat) |
| 1862 | icd->error = ops->io_u_lat(td, tnsec); |
| 1863 | } |
| 1864 | |
| 1865 | if (td->o.max_latency && tnsec > td->o.max_latency) |
| 1866 | lat_fatal(td, icd, tnsec, td->o.max_latency); |
| 1867 | if (td->o.latency_target && tnsec > td->o.latency_target) { |
| 1868 | if (lat_target_failed(td)) |
| 1869 | lat_fatal(td, icd, tnsec, td->o.latency_target); |
| 1870 | } |
| 1871 | } |
| 1872 | |
| 1873 | if (ddir_rw(idx)) { |
| 1874 | if (!td->o.disable_clat) { |
| 1875 | add_clat_sample(td, idx, llnsec, bytes, io_u->offset); |
| 1876 | io_u_mark_latency(td, llnsec); |
| 1877 | } |
| 1878 | |
| 1879 | if (!td->o.disable_bw && per_unit_log(td->bw_log)) |
| 1880 | add_bw_sample(td, io_u, bytes, llnsec); |
| 1881 | |
| 1882 | if (no_reduce && per_unit_log(td->iops_log)) |
| 1883 | add_iops_sample(td, io_u, bytes); |
| 1884 | } else if (ddir_sync(idx) && !td->o.disable_clat) |
| 1885 | add_sync_clat_sample(&td->ts, llnsec); |
| 1886 | |
| 1887 | if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) { |
| 1888 | uint32_t *info = io_u_block_info(td, io_u); |
| 1889 | if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) { |
| 1890 | if (io_u->ddir == DDIR_TRIM) { |
| 1891 | *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, |
| 1892 | BLOCK_INFO_TRIMS(*info) + 1); |
| 1893 | } else if (io_u->ddir == DDIR_WRITE) { |
| 1894 | *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN, |
| 1895 | *info); |
| 1896 | } |
| 1897 | } |
| 1898 | } |
| 1899 | } |
| 1900 | |
| 1901 | static void file_log_write_comp(const struct thread_data *td, struct fio_file *f, |
| 1902 | uint64_t offset, unsigned int bytes) |
| 1903 | { |
| 1904 | int idx; |
| 1905 | |
| 1906 | if (!f) |
| 1907 | return; |
| 1908 | |
| 1909 | if (f->first_write == -1ULL || offset < f->first_write) |
| 1910 | f->first_write = offset; |
| 1911 | if (f->last_write == -1ULL || ((offset + bytes) > f->last_write)) |
| 1912 | f->last_write = offset + bytes; |
| 1913 | |
| 1914 | if (!f->last_write_comp) |
| 1915 | return; |
| 1916 | |
| 1917 | idx = f->last_write_idx++; |
| 1918 | f->last_write_comp[idx] = offset; |
| 1919 | if (f->last_write_idx == td->o.iodepth) |
| 1920 | f->last_write_idx = 0; |
| 1921 | } |
| 1922 | |
| 1923 | static bool should_account(struct thread_data *td) |
| 1924 | { |
| 1925 | return ramp_time_over(td) && (td->runstate == TD_RUNNING || |
| 1926 | td->runstate == TD_VERIFYING); |
| 1927 | } |
| 1928 | |
| 1929 | static void io_completed(struct thread_data *td, struct io_u **io_u_ptr, |
| 1930 | struct io_completion_data *icd) |
| 1931 | { |
| 1932 | struct io_u *io_u = *io_u_ptr; |
| 1933 | enum fio_ddir ddir = io_u->ddir; |
| 1934 | struct fio_file *f = io_u->file; |
| 1935 | |
| 1936 | dprint_io_u(io_u, "complete"); |
| 1937 | |
| 1938 | assert(io_u->flags & IO_U_F_FLIGHT); |
| 1939 | io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK); |
| 1940 | |
| 1941 | /* |
| 1942 | * Mark IO ok to verify |
| 1943 | */ |
| 1944 | if (io_u->ipo) { |
| 1945 | /* |
| 1946 | * Remove errored entry from the verification list |
| 1947 | */ |
| 1948 | if (io_u->error) |
| 1949 | unlog_io_piece(td, io_u); |
| 1950 | else { |
| 1951 | io_u->ipo->flags &= ~IP_F_IN_FLIGHT; |
| 1952 | write_barrier(); |
| 1953 | } |
| 1954 | } |
| 1955 | |
| 1956 | if (ddir_sync(ddir)) { |
| 1957 | td->last_was_sync = true; |
| 1958 | if (f) { |
| 1959 | f->first_write = -1ULL; |
| 1960 | f->last_write = -1ULL; |
| 1961 | } |
| 1962 | if (should_account(td)) |
| 1963 | account_io_completion(td, io_u, icd, ddir, io_u->buflen); |
| 1964 | return; |
| 1965 | } |
| 1966 | |
| 1967 | td->last_was_sync = false; |
| 1968 | td->last_ddir = ddir; |
| 1969 | |
| 1970 | if (!io_u->error && ddir_rw(ddir)) { |
| 1971 | unsigned int bytes = io_u->buflen - io_u->resid; |
| 1972 | int ret; |
| 1973 | |
| 1974 | td->io_blocks[ddir]++; |
| 1975 | td->io_bytes[ddir] += bytes; |
| 1976 | |
| 1977 | if (!(io_u->flags & IO_U_F_VER_LIST)) { |
| 1978 | td->this_io_blocks[ddir]++; |
| 1979 | td->this_io_bytes[ddir] += bytes; |
| 1980 | } |
| 1981 | |
| 1982 | if (ddir == DDIR_WRITE) |
| 1983 | file_log_write_comp(td, f, io_u->offset, bytes); |
| 1984 | |
| 1985 | if (should_account(td)) |
| 1986 | account_io_completion(td, io_u, icd, ddir, bytes); |
| 1987 | |
| 1988 | icd->bytes_done[ddir] += bytes; |
| 1989 | |
| 1990 | if (io_u->end_io) { |
| 1991 | ret = io_u->end_io(td, io_u_ptr); |
| 1992 | io_u = *io_u_ptr; |
| 1993 | if (ret && !icd->error) |
| 1994 | icd->error = ret; |
| 1995 | } |
| 1996 | } else if (io_u->error) { |
| 1997 | icd->error = io_u->error; |
| 1998 | io_u_log_error(td, io_u); |
| 1999 | } |
| 2000 | if (icd->error) { |
| 2001 | enum error_type_bit eb = td_error_type(ddir, icd->error); |
| 2002 | |
| 2003 | if (!td_non_fatal_error(td, eb, icd->error)) |
| 2004 | return; |
| 2005 | |
| 2006 | /* |
| 2007 | * If there is a non_fatal error, then add to the error count |
| 2008 | * and clear all the errors. |
| 2009 | */ |
| 2010 | update_error_count(td, icd->error); |
| 2011 | td_clear_error(td); |
| 2012 | icd->error = 0; |
| 2013 | if (io_u) |
| 2014 | io_u->error = 0; |
| 2015 | } |
| 2016 | } |
| 2017 | |
| 2018 | static void init_icd(struct thread_data *td, struct io_completion_data *icd, |
| 2019 | int nr) |
| 2020 | { |
| 2021 | int ddir; |
| 2022 | |
| 2023 | if (!gtod_reduce(td)) |
| 2024 | fio_gettime(&icd->time, NULL); |
| 2025 | |
| 2026 | icd->nr = nr; |
| 2027 | |
| 2028 | icd->error = 0; |
| 2029 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
| 2030 | icd->bytes_done[ddir] = 0; |
| 2031 | } |
| 2032 | |
| 2033 | static void ios_completed(struct thread_data *td, |
| 2034 | struct io_completion_data *icd) |
| 2035 | { |
| 2036 | struct io_u *io_u; |
| 2037 | int i; |
| 2038 | |
| 2039 | for (i = 0; i < icd->nr; i++) { |
| 2040 | io_u = td->io_ops->event(td, i); |
| 2041 | |
| 2042 | io_completed(td, &io_u, icd); |
| 2043 | |
| 2044 | if (io_u) |
| 2045 | put_io_u(td, io_u); |
| 2046 | } |
| 2047 | } |
| 2048 | |
| 2049 | /* |
| 2050 | * Complete a single io_u for the sync engines. |
| 2051 | */ |
| 2052 | int io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
| 2053 | { |
| 2054 | struct io_completion_data icd; |
| 2055 | int ddir; |
| 2056 | |
| 2057 | init_icd(td, &icd, 1); |
| 2058 | io_completed(td, &io_u, &icd); |
| 2059 | |
| 2060 | if (io_u) |
| 2061 | put_io_u(td, io_u); |
| 2062 | |
| 2063 | if (icd.error) { |
| 2064 | td_verror(td, icd.error, "io_u_sync_complete"); |
| 2065 | return -1; |
| 2066 | } |
| 2067 | |
| 2068 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
| 2069 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
| 2070 | |
| 2071 | return 0; |
| 2072 | } |
| 2073 | |
| 2074 | /* |
| 2075 | * Called to complete min_events number of io for the async engines. |
| 2076 | */ |
| 2077 | int io_u_queued_complete(struct thread_data *td, int min_evts) |
| 2078 | { |
| 2079 | struct io_completion_data icd; |
| 2080 | struct timespec *tvp = NULL; |
| 2081 | int ret, ddir; |
| 2082 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
| 2083 | |
| 2084 | dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts); |
| 2085 | |
| 2086 | if (!min_evts) |
| 2087 | tvp = &ts; |
| 2088 | else if (min_evts > td->cur_depth) |
| 2089 | min_evts = td->cur_depth; |
| 2090 | |
| 2091 | /* No worries, td_io_getevents fixes min and max if they are |
| 2092 | * set incorrectly */ |
| 2093 | ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp); |
| 2094 | if (ret < 0) { |
| 2095 | td_verror(td, -ret, "td_io_getevents"); |
| 2096 | return ret; |
| 2097 | } else if (!ret) |
| 2098 | return ret; |
| 2099 | |
| 2100 | init_icd(td, &icd, ret); |
| 2101 | ios_completed(td, &icd); |
| 2102 | if (icd.error) { |
| 2103 | td_verror(td, icd.error, "io_u_queued_complete"); |
| 2104 | return -1; |
| 2105 | } |
| 2106 | |
| 2107 | for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++) |
| 2108 | td->bytes_done[ddir] += icd.bytes_done[ddir]; |
| 2109 | |
| 2110 | return ret; |
| 2111 | } |
| 2112 | |
| 2113 | /* |
| 2114 | * Call when io_u is really queued, to update the submission latency. |
| 2115 | */ |
| 2116 | void io_u_queued(struct thread_data *td, struct io_u *io_u) |
| 2117 | { |
| 2118 | if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) { |
| 2119 | unsigned long slat_time; |
| 2120 | |
| 2121 | slat_time = ntime_since(&io_u->start_time, &io_u->issue_time); |
| 2122 | |
| 2123 | if (td->parent) |
| 2124 | td = td->parent; |
| 2125 | |
| 2126 | add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen, |
| 2127 | io_u->offset); |
| 2128 | } |
| 2129 | } |
| 2130 | |
| 2131 | /* |
| 2132 | * See if we should reuse the last seed, if dedupe is enabled |
| 2133 | */ |
| 2134 | static struct frand_state *get_buf_state(struct thread_data *td) |
| 2135 | { |
| 2136 | unsigned int v; |
| 2137 | |
| 2138 | if (!td->o.dedupe_percentage) |
| 2139 | return &td->buf_state; |
| 2140 | else if (td->o.dedupe_percentage == 100) { |
| 2141 | frand_copy(&td->buf_state_prev, &td->buf_state); |
| 2142 | return &td->buf_state; |
| 2143 | } |
| 2144 | |
| 2145 | v = rand32_between(&td->dedupe_state, 1, 100); |
| 2146 | |
| 2147 | if (v <= td->o.dedupe_percentage) |
| 2148 | return &td->buf_state_prev; |
| 2149 | |
| 2150 | return &td->buf_state; |
| 2151 | } |
| 2152 | |
| 2153 | static void save_buf_state(struct thread_data *td, struct frand_state *rs) |
| 2154 | { |
| 2155 | if (td->o.dedupe_percentage == 100) |
| 2156 | frand_copy(rs, &td->buf_state_prev); |
| 2157 | else if (rs == &td->buf_state) |
| 2158 | frand_copy(&td->buf_state_prev, rs); |
| 2159 | } |
| 2160 | |
| 2161 | void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write, |
| 2162 | unsigned int max_bs) |
| 2163 | { |
| 2164 | struct thread_options *o = &td->o; |
| 2165 | |
| 2166 | if (o->mem_type == MEM_CUDA_MALLOC) |
| 2167 | return; |
| 2168 | |
| 2169 | if (o->compress_percentage || o->dedupe_percentage) { |
| 2170 | unsigned int perc = td->o.compress_percentage; |
| 2171 | struct frand_state *rs; |
| 2172 | unsigned int left = max_bs; |
| 2173 | unsigned int this_write; |
| 2174 | |
| 2175 | do { |
| 2176 | rs = get_buf_state(td); |
| 2177 | |
| 2178 | min_write = min(min_write, left); |
| 2179 | |
| 2180 | if (perc) { |
| 2181 | this_write = min_not_zero(min_write, |
| 2182 | td->o.compress_chunk); |
| 2183 | |
| 2184 | fill_random_buf_percentage(rs, buf, perc, |
| 2185 | this_write, this_write, |
| 2186 | o->buffer_pattern, |
| 2187 | o->buffer_pattern_bytes); |
| 2188 | } else { |
| 2189 | fill_random_buf(rs, buf, min_write); |
| 2190 | this_write = min_write; |
| 2191 | } |
| 2192 | |
| 2193 | buf += this_write; |
| 2194 | left -= this_write; |
| 2195 | save_buf_state(td, rs); |
| 2196 | } while (left); |
| 2197 | } else if (o->buffer_pattern_bytes) |
| 2198 | fill_buffer_pattern(td, buf, max_bs); |
| 2199 | else if (o->zero_buffers) |
| 2200 | memset(buf, 0, max_bs); |
| 2201 | else |
| 2202 | fill_random_buf(get_buf_state(td), buf, max_bs); |
| 2203 | } |
| 2204 | |
| 2205 | /* |
| 2206 | * "randomly" fill the buffer contents |
| 2207 | */ |
| 2208 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, |
| 2209 | unsigned int min_write, unsigned int max_bs) |
| 2210 | { |
| 2211 | io_u->buf_filled_len = 0; |
| 2212 | fill_io_buffer(td, io_u->buf, min_write, max_bs); |
| 2213 | } |
| 2214 | |
| 2215 | static int do_sync_file_range(const struct thread_data *td, |
| 2216 | struct fio_file *f) |
| 2217 | { |
| 2218 | off64_t offset, nbytes; |
| 2219 | |
| 2220 | offset = f->first_write; |
| 2221 | nbytes = f->last_write - f->first_write; |
| 2222 | |
| 2223 | if (!nbytes) |
| 2224 | return 0; |
| 2225 | |
| 2226 | return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range); |
| 2227 | } |
| 2228 | |
| 2229 | int do_io_u_sync(const struct thread_data *td, struct io_u *io_u) |
| 2230 | { |
| 2231 | int ret; |
| 2232 | |
| 2233 | if (io_u->ddir == DDIR_SYNC) { |
| 2234 | ret = fsync(io_u->file->fd); |
| 2235 | } else if (io_u->ddir == DDIR_DATASYNC) { |
| 2236 | #ifdef CONFIG_FDATASYNC |
| 2237 | ret = fdatasync(io_u->file->fd); |
| 2238 | #else |
| 2239 | ret = io_u->xfer_buflen; |
| 2240 | io_u->error = EINVAL; |
| 2241 | #endif |
| 2242 | } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE) |
| 2243 | ret = do_sync_file_range(td, io_u->file); |
| 2244 | else { |
| 2245 | ret = io_u->xfer_buflen; |
| 2246 | io_u->error = EINVAL; |
| 2247 | } |
| 2248 | |
| 2249 | if (ret < 0) |
| 2250 | io_u->error = errno; |
| 2251 | |
| 2252 | return ret; |
| 2253 | } |
| 2254 | |
| 2255 | int do_io_u_trim(const struct thread_data *td, struct io_u *io_u) |
| 2256 | { |
| 2257 | #ifndef FIO_HAVE_TRIM |
| 2258 | io_u->error = EINVAL; |
| 2259 | return 0; |
| 2260 | #else |
| 2261 | struct fio_file *f = io_u->file; |
| 2262 | int ret; |
| 2263 | |
| 2264 | ret = os_trim(f, io_u->offset, io_u->xfer_buflen); |
| 2265 | if (!ret) |
| 2266 | return io_u->xfer_buflen; |
| 2267 | |
| 2268 | io_u->error = ret; |
| 2269 | return 0; |
| 2270 | #endif |
| 2271 | } |